Issue #2101 - Part 3: Update libyuv

Updated to version 1861, git revision 88b050f337cc0ca2a51800fe7bf4737222c87344 from
https://chromium.googlesource.com/libyuv/libyuv/

223 files changed, 110015 insertions, 36086 deletions

diff --git a/media/libyuv/.clang-format b/media/libyuv/.clang-format
new file mode 100644
index 0000000000..59d487051f
--- /dev/null
+++ b/media/libyuv/.clang-format
@@ -0,0 +1,6 @@
+# Defines the Chromium style for automatic reformatting.
+# http://clang.llvm.org/docs/ClangFormatStyleOptions.html
+BasedOnStyle: Chromium
+---
+Language: Java
+BasedOnStyle: Google
diff --git a/media/libyuv/.gn b/media/libyuv/.gn
new file mode 100644
index 0000000000..f9a5ee6ce8
--- /dev/null
+++ b/media/libyuv/.gn
@@ -0,0 +1,38 @@
+# Copyright 2015 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+import("//build/dotfile_settings.gni")
+
+# The location of the build configuration file.
+buildconfig = "//build/config/BUILDCONFIG.gn"
+
+# The python interpreter to use by default. On Windows, this will look
+# for python3.exe and python3.bat.
+script_executable = "python3"
+
+# The secondary source root is a parallel directory tree where
+# GN build files are placed when they can not be placed directly
+# in the source tree, e.g. for third party source trees.
+secondary_source = "//build/secondary/"
+
+# These are the targets to check headers for by default. The files in targets
+# matching these patterns (see "gn help label_pattern" for format) will have
+# their includes checked for proper dependencies when you run either
+# "gn check" or "gn gen --check".
+check_targets = [ "//libyuv/*" ]
+
+# These are the list of GN files that run exec_script. This whitelist exists
+# to force additional review for new uses of exec_script, which is strongly
+# discouraged except for gypi_to_gn calls.
+exec_script_whitelist = build_dotfile_settings.exec_script_whitelist +
+                        [ "//build_overrides/build.gni" ]
+
+default_args = {
+  mac_sdk_min = "10.12"
+  ios_deployment_target = "12.0"
+}
diff --git a/media/libyuv/.vpython b/media/libyuv/.vpython
new file mode 100644
index 0000000000..4a64fd21db
--- /dev/null
+++ b/media/libyuv/.vpython
@@ -0,0 +1,52 @@
+# This is a vpython "spec" file.
+#
+# It describes patterns for python wheel dependencies of the python scripts in
+# the chromium repo, particularly for dependencies that have compiled components
+# (since pure-python dependencies can be easily vendored into third_party).
+#
+# When vpython is invoked, it finds this file and builds a python VirtualEnv,
+# containing all of the dependencies described in this file, fetching them from
+# CIPD (the "Chrome Infrastructure Package Deployer" service). Unlike `pip`,
+# this never requires the end-user machine to have a working python extension
+# compilation environment. All of these packages are built using:
+#   https://chromium.googlesource.com/infra/infra/+/master/infra/tools/dockerbuild/
+#
+# All python scripts in the repo share this same spec, to avoid dependency
+# fragmentation.
+#
+# If you have depot_tools installed in your $PATH, you can invoke python scripts
+# in this repo by running them as you normally would run them, except
+# substituting `vpython` instead of `python` on the command line, e.g.:
+#   vpython path/to/script.py some --arguments
+#
+# Read more about `vpython` and how to modify this file here:
+#   https://chromium.googlesource.com/infra/infra/+/master/doc/users/vpython.md
+
+python_version: "2.7"
+
+# Used by:
+#   third_party/catapult
+wheel: <
+  name: "infra/python/wheels/psutil/${platform}_${py_python}_${py_abi}"
+  version: "version:5.2.2"
+>
+
+# Used by:
+#   third_party/catapult
+wheel: <
+  name: "infra/python/wheels/pypiwin32/${vpython_platform}"
+  version: "version:219"
+  match_tag: <
+    platform: "win32"
+  >
+  match_tag: <
+    platform: "win_amd64"
+  >
+>
+
+# Used by:
+#   build/android
+wheel: <
+  name: "infra/python/wheels/requests-py2_py3"
+  version: "version:2.13.0"
+>
diff --git a/media/libyuv/.vpython3 b/media/libyuv/.vpython3
new file mode 100644
index 0000000000..0a9aa38b9a
--- /dev/null
+++ b/media/libyuv/.vpython3
@@ -0,0 +1,405 @@
+# This is a vpython "spec" file.
+#
+# It describes patterns for python wheel dependencies of the python scripts in
+# the chromium repo, particularly for dependencies that have compiled components
+# (since pure-python dependencies can be easily vendored into third_party).
+#
+# When vpython is invoked, it finds this file and builds a python VirtualEnv,
+# containing all of the dependencies described in this file, fetching them from
+# CIPD (the "Chrome Infrastructure Package Deployer" service). Unlike `pip`,
+# this never requires the end-user machine to have a working python extension
+# compilation environment. All of these packages are built using:
+#   https://chromium.googlesource.com/infra/infra/+/main/infra/tools/dockerbuild/
+#
+# All python scripts in the repo share this same spec, to avoid dependency
+# fragmentation.
+#
+# If you have depot_tools installed in your $PATH, you can invoke python scripts
+# in this repo by running them as you normally would run them, except
+# substituting `vpython` instead of `python` on the command line, e.g.:
+#   vpython path/to/script.py some --arguments
+#
+# Read more about `vpython` and how to modify this file here:
+#   https://chromium.googlesource.com/infra/infra/+/main/doc/users/vpython.md
+
+python_version: "3.8"
+
+# The default set of platforms vpython checks does not yet include mac-arm64.
+# Setting `verify_pep425_tag` to the list of platforms we explicitly must support
+# allows us to ensure that vpython specs stay mac-arm64-friendly
+verify_pep425_tag: [
+    {python: "cp38", abi: "cp38", platform: "manylinux1_x86_64"},
+    {python: "cp38", abi: "cp38", platform: "linux_arm64"},
+
+    {python: "cp38", abi: "cp38", platform: "macosx_10_10_intel"},
+    {python: "cp38", abi: "cp38", platform: "macosx_11_0_arm64"},
+
+    {python: "cp38", abi: "cp38", platform: "win32"},
+    {python: "cp38", abi: "cp38", platform: "win_amd64"}
+]
+
+# Used by:
+#   build/android/pylib/local/emulator/avd.py
+#   components/policy/test_support/policy_testserver.py
+wheel: <
+  name: "infra/python/wheels/protobuf-py2_py3"
+  version: "version:3.15.8"
+>
+
+# TODO(https://crbug.com/898348): Add in necessary wheels as Python3 versions
+# become available.
+wheel: <
+  name: "infra/python/wheels/six-py2_py3"
+  version: "version:1.15.0"
+>
+
+# Common utilities.
+# Use the same versions specified by //third_party/catapult/.vpython3 so that
+# Chromium tests using Telemetry function properly.
+wheel: <
+  name: "infra/python/wheels/numpy/${vpython_platform}"
+  version: "version:1.20.3"
+  # A newer version of numpy is required on ARM64, but it breaks older OS versions.
+  not_match_tag <
+    platform: "macosx_11_0_arm64"
+  >
+>
+wheel: <
+  name: "infra/python/wheels/numpy/mac-arm64_cp38_cp38"
+  version: "version:1.21.1"
+  match_tag <
+    platform: "macosx_11_0_arm64"
+  >
+>
+wheel: <
+  name: "infra/python/wheels/psutil/${vpython_platform}"
+  version: "version:5.8.0.chromium.2"
+>
+wheel: <
+  name: "infra/python/wheels/requests-py2_py3"
+  version: "version:2.26.0"
+>
+
+# Used by various python unit tests.
+wheel: <
+  name: "infra/python/wheels/mock-py2_py3"
+  version: "version:2.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/parameterized-py2_py3"
+  version: "version:0.7.1"
+>
+wheel: <
+  name: "infra/python/wheels/pbr-py2_py3"
+  version: "version:3.0.0"
+>
+
+wheel: <
+  name: "infra/python/wheels/pyfakefs-py2_py3"
+  version: "version:3.7.2"
+>
+
+# Used by:
+#   build/chromeos/test_runner.py
+wheel: <
+  name: "infra/python/wheels/jsonlines-py2_py3"
+  version: "version:1.2.0"
+>
+wheel: <
+  name: "infra/python/wheels/python-dateutil-py2_py3"
+  version: "version:2.7.3"
+>
+
+# Used by WPT importer
+wheel: <
+  name: "infra/python/wheels/charset_normalizer-py3"
+  version: "version:2.0.4"
+>
+wheel: <
+  name: "infra/python/wheels/pyasn1-py2_py3"
+  version: "version:0.4.5"
+>
+wheel: <
+  name: "infra/python/wheels/pyasn1_modules-py2_py3"
+  version: "version:0.2.4"
+>
+wheel: <
+  name: "infra/python/wheels/rsa-py2_py3"
+  version: "version:3.4.2"
+>
+wheel: <
+  name: "infra/python/wheels/cachetools-py2_py3"
+  version: "version:2.0.1"
+>
+wheel: <
+  name: "infra/python/wheels/uritemplate-py2_py3"
+  version: "version:3.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/google-auth-py2_py3"
+  version: "version:1.25.0"
+>
+wheel: <
+  name: "infra/python/wheels/googleapis-common-protos-py2_py3"
+  version: "version:1.52.0"
+>
+wheel: <
+  name: "infra/python/wheels/google-api-core-py2_py3"
+  version: "version:1.25.1"
+>
+wheel: <
+  name: "infra/python/wheels/google-auth-httplib2-py2_py3"
+  version: "version:0.1.0"
+>
+wheel: <
+  name: "infra/python/wheels/google-api-python-client-py3"
+  version: "version:2.2.0"
+>
+wheel: <
+  name: "infra/python/wheels/oauth2client-py2_py3"
+  version: "version:3.0.0"
+>
+
+# Used by Web Platform Tests (WPT) codebase in
+# //third_party/blink/web_tests/external/wpt/tools/
+wheel: <
+  name: "infra/python/wheels/html5lib-py2_py3"
+  version: "version:1.0.1"
+>
+wheel: <
+  name: "infra/python/wheels/mozdebug-py2_py3"
+  version: "version:0.2"
+>
+wheel: <
+  name: "infra/python/wheels/mozinfo-py2_py3"
+  version: "version:1.2.2"
+>
+wheel: <
+  name: "infra/python/wheels/mozlog-py2_py3"
+  version: "version:7.1.0"
+>
+wheel: <
+  name: "infra/python/wheels/mozprocess-py2_py3"
+  version: "version:1.2.1"
+>
+wheel: <
+  name: "infra/python/wheels/urllib3-py2_py3"
+  version: "version:1.24.3"
+>
+wheel: <
+  name: "infra/python/wheels/blessings-py2_py3"
+  version: "version:1.7"
+>
+wheel: <
+  name: "infra/python/wheels/mozfile-py2_py3"
+  version: "version:2.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/mozterm-py2_py3"
+  version: "version:1.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/webencodings-py2_py3"
+  version: "version:0.5.1"
+>
+wheel: <
+  name: "infra/python/wheels/certifi-py2_py3"
+  version: "version:2020.11.8"
+>
+wheel: <
+  name: "infra/python/wheels/chardet-py2_py3"
+  version: "version:3.0.4"
+>
+wheel: <
+  name: "infra/python/wheels/idna-py2_py3"
+  version: "version:2.8"
+>
+wheel: <
+  name: "infra/python/wheels/distro-py2_py3"
+  version: "version:1.4.0"
+>
+wheel: <
+  name: "infra/python/wheels/pillow/linux-amd64_cp38_cp38"
+  version: "version:8.1.2"
+>
+wheel: <
+  name: "infra/python/wheels/aioquic/${vpython_platform}"
+  version: "version:0.9.15"
+>
+wheel: <
+  name: "infra/python/wheels/pylsqpack/${vpython_platform}"
+  version: "version:0.3.12"
+>
+wheel: <
+  name: "infra/python/wheels/cryptography/${vpython_platform}"
+  version: "version:3.3.1.chromium.1"
+>
+wheel: <
+  name: "infra/python/wheels/cffi/${vpython_platform}"
+  version: "version:1.14.5"
+>
+wheel: <
+  name: "infra/python/wheels/pycparser-py2_py3"
+  version: "version:2.19"
+>
+
+# Used by:
+#   chrome/test/chromedriver/test/run_webdriver_tests.py
+wheel: <
+  name: "infra/python/wheels/iniconfig-py3"
+  version: "version:1.1.1"
+>
+
+wheel: <
+  name: "infra/python/wheels/packaging-py2_py3"
+  version: "version:16.8"
+>
+
+wheel: <
+  name: "infra/python/wheels/pyparsing-py2_py3"
+  version: "version:2.4.7"
+>
+
+wheel: <
+  name: "infra/python/wheels/toml-py3"
+  version: "version:0.10.1"
+>
+
+wheel <
+  name: "infra/python/wheels/pytest-py3"
+  version: "version:6.2.2"
+>
+
+wheel <
+  name: "infra/python/wheels/pytest-asyncio-py3"
+  version: "version:0.14.0"
+>
+
+wheel <
+  name: "infra/python/wheels/attrs-py2_py3"
+  version: "version:20.3.0"
+>
+
+wheel <
+  name: "infra/python/wheels/six-py2_py3"
+  version: "version:1.15.0"
+>
+
+wheel <
+  name: "infra/python/wheels/more-itertools-py2_py3"
+  version: "version:4.1.0"
+>
+
+wheel <
+  name: "infra/python/wheels/pluggy-py3"
+  version: "version:0.13.1"
+>
+
+wheel <
+  name: "infra/python/wheels/py-py2_py3"
+  version: "version:1.10.0"
+>
+
+wheel <
+  name: "infra/python/wheels/funcsigs-py2_py3"
+  version: "version:1.0.2"
+>
+
+wheel: <
+  name: "infra/python/wheels/atomicwrites-py2_py3"
+  version: "version:1.3.0"
+>
+
+wheel: <
+  name: "infra/python/wheels/colorama-py2_py3"
+  version: "version:0.4.1"
+>
+
+# Used by:
+#   testing/buildbot/generate_buildbot_json_coveragetest.py
+wheel: <
+  name: "infra/python/wheels/coverage/${vpython_platform}"
+  version: "version:5.5.chromium.2"
+>
+
+# Used by:
+#   //content/test/gpu
+wheel: <
+  name: "infra/python/wheels/pathos/${vpython_platform}"
+  version: "version:0.2.7.chromium.4"
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "manylinux1_i686"
+  >
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "linux_mips64"
+  >
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "linux_armv6l"
+  >
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "linux_armv7l"
+  >
+>
+
+# Used by:
+#   //tools/infra/find_bad_builds.py
+wheel: <
+  name: "infra/python/wheels/pytz-py2_py3"
+  version: "version:2018.4"
+>
+
+# Used by:
+#   //third_party/blink/tools/blinkpy/web_tests/port/server_process.py
+wheel: <
+  name: "infra/python/wheels/pywin32/${vpython_platform}"
+  version: "version:300"
+  match_tag: <
+    platform: "win32"
+  >
+  match_tag: <
+    platform: "win_amd64"
+  >
+>
+
+# Used by:
+#   //content/test/gpu/gpu_tests/color_profile_manager_mac.py
+wheel: <
+  name: "infra/python/wheels/pyobjc/${vpython_platform}"
+  version: "version:7.3.chromium.1"
+  match_tag: <
+    platform: "macosx_10_10_intel"
+  >
+>
+
+# Used by:
+#   tools/perf/core/results_dashboard.py
+wheel: <
+  name: "infra/python/wheels/httplib2-py3"
+  version: "version:0.19.1"
+>
+
+# Used by:
+#   tools/perf/flakiness_cli
+wheel: <
+  name: "infra/python/wheels/pandas/${vpython_platform}"
+  version: "version:1.3.2.chromium.1"
+  match_tag: <
+    platform: "win32"
+  >
+  match_tag: <
+    platform: "win_amd64"
+  >
+  match_tag: <
+    platform: "manylinux1_i686"
+  >
+  match_tag: <
+    platform: "manylinux1_x86_64"
+  >
+  match_tag: <
+    platform: "macosx_10_6_intel"
+  >
+>
diff --git a/media/libyuv/AUTHORS b/media/libyuv/AUTHORS
index 9686ac13eb..28c08956a8 100644
--- a/media/libyuv/AUTHORS
+++ b/media/libyuv/AUTHORS
@@ -2,3 +2,5 @@
 # Name or Organization <email address>
 
 Google Inc.
+
+Ivan Pavlotskiy <ivan.pavlotskiy@lgepartner.com>
diff --git a/media/libyuv/Android.bp b/media/libyuv/Android.bp
new file mode 100644
index 0000000000..361562870d
--- /dev/null
+++ b/media/libyuv/Android.bp
@@ -0,0 +1,196 @@
+package {
+    default_applicable_licenses: ["external_libyuv_files_license"],
+}
+
+// Added automatically by a large-scale-change
+//
+// large-scale-change included anything that looked like it might be a license
+// text as a license_text. e.g. LICENSE, NOTICE, COPYING etc.
+//
+// Please consider removing redundant or irrelevant files from 'license_text:'.
+// See: http://go/android-license-faq
+license {
+    name: "external_libyuv_files_license",
+    visibility: [":__subpackages__"],
+    license_kinds: [
+        "SPDX-license-identifier-BSD",
+    ],
+    license_text: [
+        "LICENSE",
+        "PATENTS",
+    ],
+}
+
+cc_library {
+    name: "libyuv",
+    vendor_available: true,
+    product_available: true,
+    host_supported: true,
+    vndk: {
+        enabled: true,
+    },
+
+    srcs: [
+        "source/compare.cc",
+        "source/compare_common.cc",
+        "source/compare_gcc.cc",
+        "source/compare_msa.cc",
+        "source/compare_neon.cc",
+        "source/compare_neon64.cc",
+        "source/convert.cc",
+        "source/convert_argb.cc",
+        "source/convert_from.cc",
+        "source/convert_from_argb.cc",
+        "source/convert_jpeg.cc",
+        "source/convert_to_argb.cc",
+        "source/convert_to_i420.cc",
+        "source/cpu_id.cc",
+        "source/mjpeg_decoder.cc",
+        "source/mjpeg_validate.cc",
+        "source/planar_functions.cc",
+        "source/rotate.cc",
+        "source/rotate_any.cc",
+        "source/rotate_argb.cc",
+        "source/rotate_common.cc",
+        "source/rotate_gcc.cc",
+        "source/rotate_msa.cc",
+        "source/rotate_neon.cc",
+        "source/rotate_neon64.cc",
+        "source/row_any.cc",
+        "source/row_common.cc",
+        "source/row_gcc.cc",
+        "source/row_msa.cc",
+        "source/row_neon.cc",
+        "source/row_neon64.cc",
+        "source/scale.cc",
+        "source/scale_any.cc",
+        "source/scale_argb.cc",
+        "source/scale_common.cc",
+        "source/scale_gcc.cc",
+        "source/scale_msa.cc",
+        "source/scale_neon.cc",
+        "source/scale_neon64.cc",
+        "source/scale_rgb.cc",
+        "source/scale_uv.cc",
+        "source/video_common.cc",
+    ],
+
+    cflags: [
+        "-Wall",
+        "-Werror",
+        "-Wno-unused-parameter",
+        "-fexceptions",
+        "-DHAVE_JPEG",
+    ],
+
+    arch: {
+        arm: {
+            cflags: ["-mfpu=neon"],
+        },
+    },
+
+    shared_libs: ["libjpeg"],
+
+    export_include_dirs: ["include"],
+
+    apex_available: [
+        "//apex_available:platform",
+        "com.android.media.swcodec",
+    ],
+    min_sdk_version: "29",
+}
+
+// compatibilty static library until all uses of libyuv_static are replaced
+// with libyuv (b/37646797)
+cc_library_static {
+    name: "libyuv_static",
+    vendor_available: true,
+    whole_static_libs: ["libyuv"],
+    apex_available: [
+        "//apex_available:platform",
+        "com.android.media.swcodec",
+    ],
+    min_sdk_version: "29",
+}
+
+cc_test {
+    name: "libyuv_unittest",
+    static_libs: ["libyuv"],
+    shared_libs: ["libjpeg"],
+    cflags: ["-Wall", "-Werror"],
+    srcs: [
+        "unit_test/basictypes_test.cc",
+        "unit_test/color_test.cc",
+        "unit_test/compare_test.cc",
+        "unit_test/convert_test.cc",
+        "unit_test/cpu_test.cc",
+        "unit_test/cpu_thread_test.cc",
+        "unit_test/math_test.cc",
+        "unit_test/planar_test.cc",
+        "unit_test/rotate_argb_test.cc",
+        "unit_test/rotate_test.cc",
+        "unit_test/scale_argb_test.cc",
+        "unit_test/scale_rgb_test.cc",
+        "unit_test/scale_test.cc",
+        "unit_test/scale_uv_test.cc",
+        "unit_test/unit_test.cc",
+        "unit_test/video_common_test.cc",
+    ],
+}
+
+cc_test {
+    name: "compare",
+    gtest: false,
+    srcs: [
+        "util/compare.cc",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
+    name: "cpuid",
+    gtest: false,
+    srcs: [
+        "util/cpuid.c",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
+    name: "i444tonv12_eg",
+    gtest: false,
+    srcs: [
+        "util/i444tonv12_eg.cc",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
+    name: "psnr",
+    gtest: false,
+    srcs: [
+        "util/psnr_main.cc",
+        "util/psnr.cc",
+        "util/ssim.cc",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
+    name: "yuvconstants",
+    gtest: false,
+    srcs: [
+        "util/yuvconstants.c",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
+    name: "yuvconvert",
+    gtest: false,
+    srcs: [
+        "util/yuvconvert.cc",
+    ],
+    static_libs: ["libyuv"],
+    shared_libs: ["libjpeg"],
+}
diff --git a/media/libyuv/Android.mk b/media/libyuv/Android.mk
index 4d2092acf5..45d9daa8e3 100644
--- a/media/libyuv/Android.mk
+++ b/media/libyuv/Android.mk
@@ -1,4 +1,4 @@
-# This is the Android makefile for libyuv for both platform and NDK.
+# This is the Android makefile for libyuv for NDK.
 LOCAL_PATH:= $(call my-dir)
 
 include $(CLEAR_VARS)
@@ -8,8 +8,11 @@ LOCAL_CPP_EXTENSION := .cc
 LOCAL_SRC_FILES := \
     source/compare.cc           \
     source/compare_common.cc    \
-    source/compare_neon64.cc    \
     source/compare_gcc.cc       \
+    source/compare_msa.cc       \
+    source/compare_neon.cc      \
+    source/compare_neon64.cc    \
+    source/compare_win.cc       \
     source/convert.cc           \
     source/convert_argb.cc      \
     source/convert_from.cc      \
@@ -22,42 +25,84 @@ LOCAL_SRC_FILES := \
     source/rotate_any.cc        \
     source/rotate_argb.cc       \
     source/rotate_common.cc     \
-    source/rotate_mips.cc       \
-    source/rotate_neon64.cc     \
     source/rotate_gcc.cc        \
+    source/rotate_msa.cc        \
+    source/rotate_neon.cc       \
+    source/rotate_neon64.cc     \
+    source/rotate_win.cc        \
     source/row_any.cc           \
     source/row_common.cc        \
-    source/row_mips.cc          \
+    source/row_gcc.cc           \
+    source/row_msa.cc           \
+    source/row_neon.cc          \
     source/row_neon64.cc        \
-    source/row_gcc.cc	        \
+    source/row_win.cc           \
     source/scale.cc             \
     source/scale_any.cc         \
     source/scale_argb.cc        \
     source/scale_common.cc      \
-    source/scale_mips.cc        \
-    source/scale_neon64.cc      \
     source/scale_gcc.cc         \
+    source/scale_msa.cc         \
+    source/scale_neon.cc        \
+    source/scale_neon64.cc      \
+    source/scale_rgb.cc         \
+    source/scale_uv.cc          \
+    source/scale_win.cc         \
     source/video_common.cc
 
-# TODO(fbarchard): Enable mjpeg encoder.
-#   source/mjpeg_decoder.cc
-#   source/convert_jpeg.cc
-#   source/mjpeg_validate.cc
-
-ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
-    LOCAL_CFLAGS += -DLIBYUV_NEON
-    LOCAL_SRC_FILES += \
-        source/compare_neon.cc.neon    \
-        source/rotate_neon.cc.neon     \
-        source/row_neon.cc.neon        \
-        source/scale_neon.cc.neon
+common_CFLAGS := -Wall -fexceptions
+ifneq ($(LIBYUV_DISABLE_JPEG), "yes")
+LOCAL_SRC_FILES += \
+    source/convert_jpeg.cc      \
+    source/mjpeg_decoder.cc     \
+    source/mjpeg_validate.cc
+common_CFLAGS += -DHAVE_JPEG
+LOCAL_SHARED_LIBRARIES := libjpeg
 endif
 
+LOCAL_CFLAGS += $(common_CFLAGS)
 LOCAL_EXPORT_C_INCLUDES := $(LOCAL_PATH)/include
 LOCAL_C_INCLUDES += $(LOCAL_PATH)/include
+LOCAL_EXPORT_C_INCLUDE_DIRS := $(LOCAL_PATH)/include
 
 LOCAL_MODULE := libyuv_static
 LOCAL_MODULE_TAGS := optional
 
 include $(BUILD_STATIC_LIBRARY)
 
+include $(CLEAR_VARS)
+
+LOCAL_WHOLE_STATIC_LIBRARIES := libyuv_static
+LOCAL_MODULE := libyuv
+ifneq ($(LIBYUV_DISABLE_JPEG), "yes")
+LOCAL_SHARED_LIBRARIES := libjpeg
+endif
+
+include $(BUILD_SHARED_LIBRARY)
+
+include $(CLEAR_VARS)
+LOCAL_STATIC_LIBRARIES := libyuv_static
+LOCAL_SHARED_LIBRARIES := libjpeg
+LOCAL_MODULE_TAGS := tests
+LOCAL_CPP_EXTENSION := .cc
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/include
+LOCAL_SRC_FILES := \
+    unit_test/basictypes_test.cc  \
+    unit_test/color_test.cc       \
+    unit_test/compare_test.cc     \
+    unit_test/convert_test.cc     \
+    unit_test/cpu_test.cc         \
+    unit_test/cpu_thread_test.cc  \
+    unit_test/math_test.cc        \
+    unit_test/planar_test.cc      \
+    unit_test/rotate_argb_test.cc \
+    unit_test/rotate_test.cc      \
+    unit_test/scale_argb_test.cc  \
+    unit_test/scale_rgb_test.cc   \
+    unit_test/scale_test.cc       \
+    unit_test/scale_uv_test.cc    \
+    unit_test/unit_test.cc        \
+    unit_test/video_common_test.cc
+
+LOCAL_MODULE := libyuv_unittest
+include $(BUILD_NATIVE_TEST)
diff --git a/media/libyuv/BUILD.gn b/media/libyuv/BUILD.gn
index b091cbc22d..a72ff06558 100644
--- a/media/libyuv/BUILD.gn
+++ b/media/libyuv/BUILD.gn
@@ -6,19 +6,80 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
-import("//build/config/arm.gni")
-import("//build/config/sanitizers/sanitizers.gni")
+import("//testing/test.gni")
+import("libyuv.gni")
+
+declare_args() {
+  # Set to false to disable building with absl flags.
+  libyuv_use_absl_flags = true
+
+  # When building a shared library using a target in WebRTC or
+  # Chromium projects that depends on libyuv, setting this flag
+  # to true makes libyuv symbols visible inside that library.
+  libyuv_symbols_visible = false
+}
 
 config("libyuv_config") {
-  include_dirs = [
-    ".",
-    "include",
-  ]
+  include_dirs = [ "include" ]
+  if (is_android && current_cpu == "arm64") {
+    ldflags = [ "-Wl,--dynamic-linker,/system/bin/linker64" ]
+  }
+  if (is_android && current_cpu != "arm64") {
+    ldflags = [ "-Wl,--dynamic-linker,/system/bin/linker" ]
+  }
+
+  if (!libyuv_use_neon) {
+    defines = [ "LIBYUV_DISABLE_NEON" ]
+  }
+}
+
+# This target is built when no specific target is specified on the command line.
+group("default") {
+  testonly = true
+  deps = [ ":libyuv" ]
+  if (libyuv_include_tests) {
+    deps += [
+      ":compare",
+      ":cpuid",
+      ":i444tonv12_eg",
+      ":libyuv_unittest",
+      ":psnr",
+      ":yuvconstants",
+      ":yuvconvert",
+    ]
+  }
+}
+
+group("libyuv") {
+  all_dependent_configs = [ ":libyuv_config" ]
+  deps = []
+
+  if (is_win && target_cpu == "x64") {
+    # Compile with clang in order to get inline assembly
+    public_deps = [ ":libyuv_internal(//build/toolchain/win:win_clang_x64)" ]
+  } else {
+    public_deps = [ ":libyuv_internal" ]
+  }
+
+  if (libyuv_use_neon) {
+    deps += [ ":libyuv_neon" ]
+  }
+
+  if (libyuv_use_msa) {
+    deps += [ ":libyuv_msa" ]
+  }
+
+  if (!is_ios && !libyuv_disable_jpeg) {
+    # Make sure that clients of libyuv link with libjpeg. This can't go in
+    # libyuv_internal because in Windows x64 builds that will generate a clang
+    # build of libjpeg, and we don't want two copies.
+    deps += [ "//third_party:jpeg" ]
+  }
 }
 
-use_neon = current_cpu == "arm64" || (current_cpu == "arm" && (arm_use_neon || arm_optionally_use_neon))
+static_library("libyuv_internal") {
+  visibility = [ ":*" ]
 
-source_set("libyuv") {
   sources = [
     # Headers
     "include/libyuv.h",
@@ -37,7 +98,9 @@ source_set("libyuv") {
     "include/libyuv/row.h",
     "include/libyuv/scale.h",
     "include/libyuv/scale_argb.h",
+    "include/libyuv/scale_rgb.h",
     "include/libyuv/scale_row.h",
+    "include/libyuv/scale_uv.h",
     "include/libyuv/version.h",
     "include/libyuv/video_common.h",
 
@@ -61,57 +124,61 @@ source_set("libyuv") {
     "source/rotate_any.cc",
     "source/rotate_argb.cc",
     "source/rotate_common.cc",
-    "source/rotate_mips.cc",
     "source/rotate_gcc.cc",
     "source/rotate_win.cc",
     "source/row_any.cc",
     "source/row_common.cc",
-    "source/row_mips.cc",
     "source/row_gcc.cc",
     "source/row_win.cc",
     "source/scale.cc",
     "source/scale_any.cc",
     "source/scale_argb.cc",
     "source/scale_common.cc",
-    "source/scale_mips.cc",
     "source/scale_gcc.cc",
+    "source/scale_rgb.cc",
+    "source/scale_uv.cc",
     "source/scale_win.cc",
     "source/video_common.cc",
   ]
 
-  configs -= [ "//build/config/compiler:chromium_code" ]
-  configs += [ "//build/config/compiler:no_chromium_code" ]
-
-  public_configs = [ ":libyuv_config" ]
-
+  configs += [ ":libyuv_config" ]
   defines = []
+  deps = []
 
-  if (!is_ios) {
-    defines += [ "HAVE_JPEG" ]
+  if (libyuv_symbols_visible) {
+    configs -= [ "//build/config/gcc:symbol_visibility_hidden" ]
+    configs += [ "//build/config/gcc:symbol_visibility_default" ]
   }
 
-  if (is_msan) {
-    # MemorySanitizer does not support assembly code yet.
-    # http://crbug.com/344505
-    defines += [ "LIBYUV_DISABLE_X86" ]
-  }
-
-  deps = [
-    "//third_party:jpeg",
-  ]
+  if (!is_ios && !libyuv_disable_jpeg) {
+    defines += [ "HAVE_JPEG" ]
 
-  if (use_neon) {
-    deps += [ ":libyuv_neon" ]
+    # Needed to pull in libjpeg headers. Can't add //third_party:jpeg to deps
+    # because in Windows x64 build it will get compiled with clang.
+    deps += [ "//third_party:jpeg_includes" ]
   }
 
-  if (is_nacl) {
-    # Always enable optimization under NaCl to workaround crbug.com/538243 .
+  # Always enable optimization for Release and NaCl builds (to workaround
+  # crbug.com/538243).
+  if (!is_debug || is_nacl) {
     configs -= [ "//build/config/compiler:default_optimization" ]
+
+    # Enable optimize for speed (-O2) over size (-Os).
     configs += [ "//build/config/compiler:optimize_max" ]
   }
+
+  # To enable AVX2 or other cpu optimization, pass flag here
+  if (!is_win) {
+    cflags = [
+      # "-mpopcnt",
+      # "-mavx2",
+      # "-mfma",
+      "-ffp-contract=fast",  # Enable fma vectorization for NEON.
+    ]
+  }
 }
 
-if (use_neon) {
+if (libyuv_use_neon) {
   static_library("libyuv_neon") {
     sources = [
       # ARM Source Files
@@ -125,11 +192,197 @@ if (use_neon) {
       "source/scale_neon64.cc",
     ]
 
+    deps = [ ":libyuv_internal" ]
+
     public_configs = [ ":libyuv_config" ]
 
+    # Always enable optimization for Release and NaCl builds (to workaround
+    # crbug.com/538243).
+    if (!is_debug) {
+      configs -= [ "//build/config/compiler:default_optimization" ]
+
+      # Enable optimize for speed (-O2) over size (-Os).
+      # TODO(fbarchard): Consider optimize_speed which is O3.
+      configs += [ "//build/config/compiler:optimize_max" ]
+    }
+
     if (current_cpu != "arm64") {
       configs -= [ "//build/config/compiler:compiler_arm_fpu" ]
       cflags = [ "-mfpu=neon" ]
     }
   }
 }
+
+if (libyuv_use_msa) {
+  static_library("libyuv_msa") {
+    sources = [
+      # MSA Source Files
+      "source/compare_msa.cc",
+      "source/rotate_msa.cc",
+      "source/row_msa.cc",
+      "source/scale_msa.cc",
+    ]
+
+    deps = [ ":libyuv_internal" ]
+
+    public_configs = [ ":libyuv_config" ]
+  }
+}
+
+if (libyuv_include_tests) {
+  config("libyuv_unittest_warnings_config") {
+    if (!is_win) {
+      cflags = [
+        # TODO(fbarchard): Fix sign and unused variable warnings.
+        "-Wno-sign-compare",
+        "-Wno-unused-variable",
+      ]
+    }
+    if (is_win) {
+      cflags = [
+        "/wd4245",  # signed/unsigned mismatch
+        "/wd4189",  # local variable is initialized but not referenced
+      ]
+    }
+  }
+  config("libyuv_unittest_config") {
+    defines = [ "GTEST_RELATIVE_PATH" ]
+  }
+
+  test("libyuv_unittest") {
+    testonly = true
+
+    sources = [
+      "unit_test/basictypes_test.cc",
+      "unit_test/color_test.cc",
+      "unit_test/compare_test.cc",
+      "unit_test/convert_test.cc",
+      "unit_test/cpu_test.cc",
+      "unit_test/cpu_thread_test.cc",
+      "unit_test/math_test.cc",
+      "unit_test/planar_test.cc",
+      "unit_test/rotate_argb_test.cc",
+      "unit_test/rotate_test.cc",
+      "unit_test/scale_argb_test.cc",
+      "unit_test/scale_rgb_test.cc",
+      "unit_test/scale_test.cc",
+      "unit_test/scale_uv_test.cc",
+      "unit_test/unit_test.cc",
+      "unit_test/unit_test.h",
+      "unit_test/video_common_test.cc",
+    ]
+
+    deps = [
+      ":libyuv",
+      "//testing/gtest",
+    ]
+
+    defines = []
+    if (libyuv_use_absl_flags) {
+      defines += [ "LIBYUV_USE_ABSL_FLAGS" ]
+      deps += [
+        "//third_party/abseil-cpp/absl/flags:flag",
+        "//third_party/abseil-cpp/absl/flags:parse",
+      ]
+    }
+
+    configs += [ ":libyuv_unittest_warnings_config" ]
+
+    public_deps = [ "//testing/gtest" ]
+    public_configs = [ ":libyuv_unittest_config" ]
+
+    if (is_linux || is_chromeos) {
+      cflags = [ "-fexceptions" ]
+    }
+    if (is_ios) {
+      configs -= [ "//build/config/compiler:default_symbols" ]
+      configs += [ "//build/config/compiler:symbols" ]
+      cflags = [ "-Wno-sometimes-uninitialized" ]
+    }
+    if (!is_ios && !libyuv_disable_jpeg) {
+      defines += [ "HAVE_JPEG" ]
+    }
+    if (is_android) {
+      deps += [ "//testing/android/native_test:native_test_native_code" ]
+    }
+
+    # TODO(YangZhang): These lines can be removed when high accuracy
+    # YUV to RGB to Neon is ported.
+    if ((target_cpu == "armv7" || target_cpu == "armv7s" ||
+         (target_cpu == "arm" && arm_version >= 7) || target_cpu == "arm64") &&
+        (arm_use_neon || arm_optionally_use_neon)) {
+      defines += [ "LIBYUV_NEON" ]
+    }
+
+    defines += [
+      # Enable the following 3 macros to turn off assembly for specified CPU.
+      # "LIBYUV_DISABLE_X86",
+      # "LIBYUV_DISABLE_NEON",
+      # Enable the following macro to build libyuv as a shared library (dll).
+      # "LIBYUV_USING_SHARED_LIBRARY"
+    ]
+  }
+
+  executable("compare") {
+    sources = [
+      # sources
+      "util/compare.cc",
+    ]
+    deps = [ ":libyuv" ]
+    if (is_linux || is_chromeos) {
+      cflags = [ "-fexceptions" ]
+    }
+  }
+
+  executable("yuvconvert") {
+    sources = [
+      # sources
+      "util/yuvconvert.cc",
+    ]
+    deps = [ ":libyuv" ]
+    if (is_linux || is_chromeos) {
+      cflags = [ "-fexceptions" ]
+    }
+  }
+
+  executable("yuvconstants") {
+    sources = [
+      # sources
+      "util/yuvconstants.c",
+    ]
+    deps = [ ":libyuv" ]
+    if (is_linux || is_chromeos) {
+      cflags = [ "-fexceptions" ]
+    }
+  }
+
+  executable("psnr") {
+    sources = [
+      # sources
+      "util/psnr.cc",
+      "util/psnr_main.cc",
+      "util/ssim.cc",
+    ]
+    deps = [ ":libyuv" ]
+
+    if (!is_ios && !libyuv_disable_jpeg) {
+      defines = [ "HAVE_JPEG" ]
+    }
+  }
+
+  executable("i444tonv12_eg") {
+    sources = [
+      # sources
+      "util/i444tonv12_eg.cc",
+    ]
+    deps = [ ":libyuv" ]
+  }
+
+  executable("cpuid") {
+    sources = [
+      # sources
+      "util/cpuid.c",
+    ]
+    deps = [ ":libyuv" ]
+  }
+}
diff --git a/media/libyuv/CM_linux_packages.cmake b/media/libyuv/CM_linux_packages.cmake
new file mode 100644
index 0000000000..a073edfaec
--- /dev/null
+++ b/media/libyuv/CM_linux_packages.cmake
@@ -0,0 +1,69 @@
+# determine the version number from the #define in libyuv/version.h
+EXECUTE_PROCESS (
+	COMMAND grep --perl-regex --only-matching "(?<=LIBYUV_VERSION )[0-9]+" include/libyuv/version.h
+	WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}
+	OUTPUT_VARIABLE YUV_VERSION_NUMBER
+	OUTPUT_STRIP_TRAILING_WHITESPACE )
+SET ( YUV_VER_MAJOR 0 )
+SET ( YUV_VER_MINOR 0 )
+SET ( YUV_VER_PATCH ${YUV_VERSION_NUMBER} )
+SET ( YUV_VERSION ${YUV_VER_MAJOR}.${YUV_VER_MINOR}.${YUV_VER_PATCH} )
+MESSAGE ( VERBOSE "Building ver.: ${YUV_VERSION}" )
+
+# is this a 32-bit or 64-bit build?
+IF ( CMAKE_SIZEOF_VOID_P EQUAL 8 )
+	SET ( YUV_BIT_SIZE 64 )
+ELSEIF ( CMAKE_SIZEOF_VOID_P EQUAL 4 )
+	SET ( YUV_BIT_SIZE 32 )
+ELSE ()
+	MESSAGE ( FATAL_ERROR "CMAKE_SIZEOF_VOID_P=${CMAKE_SIZEOF_VOID_P}" )
+ENDIF ()
+
+# detect if this is a ARM build
+STRING (FIND "${CMAKE_CXX_COMPILER}" "arm-linux-gnueabihf-g++" pos)
+IF ( ${pos} EQUAL -1 )
+	SET ( YUV_CROSS_COMPILE_FOR_ARM7 FALSE )
+ELSE ()
+	MESSAGE ( "Cross compiling for ARM7" )
+	SET ( YUV_CROSS_COMPILE_FOR_ARM7 TRUE )
+ENDIF ()
+STRING (FIND "${CMAKE_SYSTEM_PROCESSOR}" "arm" pos)
+IF ( ${pos} EQUAL -1 )
+	SET ( YUV_COMPILE_FOR_ARM7 FALSE )
+ELSE ()
+	MESSAGE ( "Compiling for ARM" )
+	SET ( YUV_COMPILE_FOR_ARM7 TRUE )
+ENDIF ()
+
+# setup the sytem name, such as "x86-32", "amd-64", and "arm-32
+IF ( ${YUV_CROSS_COMPILE_FOR_ARM7} OR ${YUV_COMPILE_FOR_ARM7} )
+	SET ( YUV_SYSTEM_NAME "armhf-${YUV_BIT_SIZE}" )
+ELSE ()
+	IF ( YUV_BIT_SIZE EQUAL 32 )
+		SET ( YUV_SYSTEM_NAME "x86-${YUV_BIT_SIZE}" )
+	ELSE ()
+		SET ( YUV_SYSTEM_NAME "amd-${YUV_BIT_SIZE}" )
+	ENDIF ()
+ENDIF ()
+MESSAGE ( VERBOSE "Packaging for: ${YUV_SYSTEM_NAME}" )
+
+# define all the variables needed by CPack to create .deb and .rpm packages
+SET ( CPACK_PACKAGE_VENDOR					"Frank Barchard" )
+SET ( CPACK_PACKAGE_CONTACT					"fbarchard@chromium.org" )
+SET ( CPACK_PACKAGE_VERSION					${YUV_VERSION} )
+SET ( CPACK_PACKAGE_VERSION_MAJOR			${YUV_VER_MAJOR} )
+SET ( CPACK_PACKAGE_VERSION_MINOR			${YUV_VER_MINOR} )
+SET ( CPACK_PACKAGE_VERSION_PATCH			${YUV_VER_PATCH} )
+SET ( CPACK_RESOURCE_FILE_LICENSE			${PROJECT_SOURCE_DIR}/LICENSE )
+SET ( CPACK_SYSTEM_NAME						"linux-${YUV_SYSTEM_NAME}" )
+SET ( CPACK_PACKAGE_NAME					"libyuv" )
+SET ( CPACK_PACKAGE_DESCRIPTION_SUMMARY		"YUV library" )
+SET ( CPACK_PACKAGE_DESCRIPTION				"YUV library and YUV conversion tool" )
+SET ( CPACK_DEBIAN_PACKAGE_SECTION			"other" )
+SET ( CPACK_DEBIAN_PACKAGE_PRIORITY			"optional" )
+SET ( CPACK_DEBIAN_PACKAGE_MAINTAINER		"Frank Barchard <fbarchard@chromium.org>" )
+SET ( CPACK_GENERATOR						"DEB;RPM" )
+
+# create the .deb and .rpm files (you'll need build-essential and rpm tools)
+INCLUDE( CPack )
+
diff --git a/media/libyuv/CMakeLists.txt b/media/libyuv/CMakeLists.txt
index 718b47ad32..bc64168550 100644
--- a/media/libyuv/CMakeLists.txt
+++ b/media/libyuv/CMakeLists.txt
@@ -1,110 +1,55 @@
-cmake_minimum_required(VERSION 2.8)
-
 # CMakeLists for libyuv
 # Originally created for "roxlu build system" to compile libyuv on windows
 # Run with -DTEST=ON to build unit tests
-option(TEST "Built unit tests" OFF)
-
-set(ly_base_dir ${CMAKE_CURRENT_LIST_DIR})
-set(ly_src_dir ${ly_base_dir}/source/)
-set(ly_inc_dir ${ly_base_dir}/include)
-set(ly_lib_name "yuv")
-
-set(ly_source_files
-  ${ly_src_dir}/compare.cc
-  ${ly_src_dir}/compare_common.cc
-  ${ly_src_dir}/compare_neon.cc
-  ${ly_src_dir}/compare_neon64.cc
-  ${ly_src_dir}/compare_gcc.cc
-  ${ly_src_dir}/compare_win.cc
-  ${ly_src_dir}/convert.cc
-  ${ly_src_dir}/convert_argb.cc
-  ${ly_src_dir}/convert_from.cc
-  ${ly_src_dir}/convert_from_argb.cc
-  ${ly_src_dir}/convert_jpeg.cc
-  ${ly_src_dir}/convert_to_argb.cc
-  ${ly_src_dir}/convert_to_i420.cc
-  ${ly_src_dir}/cpu_id.cc
-  ${ly_src_dir}/mjpeg_decoder.cc
-  ${ly_src_dir}/mjpeg_validate.cc
-  ${ly_src_dir}/planar_functions.cc
-  ${ly_src_dir}/rotate.cc
-  ${ly_src_dir}/rotate_any.cc
-  ${ly_src_dir}/rotate_argb.cc
-  ${ly_src_dir}/rotate_common.cc
-  ${ly_src_dir}/rotate_mips.cc
-  ${ly_src_dir}/rotate_neon.cc
-  ${ly_src_dir}/rotate_neon64.cc
-  ${ly_src_dir}/rotate_gcc.cc
-  ${ly_src_dir}/rotate_win.cc
-  ${ly_src_dir}/row_any.cc
-  ${ly_src_dir}/row_common.cc
-  ${ly_src_dir}/row_mips.cc
-  ${ly_src_dir}/row_neon.cc
-  ${ly_src_dir}/row_neon64.cc
-  ${ly_src_dir}/row_gcc.cc
-  ${ly_src_dir}/row_win.cc
-  ${ly_src_dir}/scale.cc
-  ${ly_src_dir}/scale_any.cc
-  ${ly_src_dir}/scale_argb.cc
-  ${ly_src_dir}/scale_common.cc
-  ${ly_src_dir}/scale_mips.cc
-  ${ly_src_dir}/scale_neon.cc
-  ${ly_src_dir}/scale_neon64.cc
-  ${ly_src_dir}/scale_gcc.cc
-  ${ly_src_dir}/scale_win.cc
-  ${ly_src_dir}/video_common.cc
-)
-
-set(ly_unittest_sources
-  ${ly_base_dir}/unit_test/basictypes_test.cc
-  ${ly_base_dir}/unit_test/color_test.cc
-  ${ly_base_dir}/unit_test/compare_test.cc
-  ${ly_base_dir}/unit_test/convert_test.cc
-  ${ly_base_dir}/unit_test/cpu_test.cc
-  ${ly_base_dir}/unit_test/math_test.cc
-  ${ly_base_dir}/unit_test/planar_test.cc
-  ${ly_base_dir}/unit_test/rotate_argb_test.cc
-  ${ly_base_dir}/unit_test/rotate_test.cc
-  ${ly_base_dir}/unit_test/scale_argb_test.cc
-  ${ly_base_dir}/unit_test/scale_test.cc
-  ${ly_base_dir}/unit_test/unit_test.cc
-  ${ly_base_dir}/unit_test/video_common_test.cc
-)
-
-set(ly_header_files
-  ${ly_inc_dir}/libyuv/basic_types.h
-  ${ly_inc_dir}/libyuv/compare.h
-  ${ly_inc_dir}/libyuv/convert.h
-  ${ly_inc_dir}/libyuv/convert_argb.h
-  ${ly_inc_dir}/libyuv/convert_from.h
-  ${ly_inc_dir}/libyuv/convert_from_argb.h
-  ${ly_inc_dir}/libyuv/cpu_id.h
-  ${ly_inc_dir}/libyuv/planar_functions.h
-  ${ly_inc_dir}/libyuv/rotate.h
-  ${ly_inc_dir}/libyuv/rotate_argb.h
-  ${ly_inc_dir}/libyuv/rotate_row.h
-  ${ly_inc_dir}/libyuv/row.h
-  ${ly_inc_dir}/libyuv/scale.h
-  ${ly_inc_dir}/libyuv/scale_argb.h
-  ${ly_inc_dir}/libyuv/scale_row.h
-  ${ly_inc_dir}/libyuv/version.h
-  ${ly_inc_dir}/libyuv/video_common.h
-  ${ly_inc_dir}/libyuv/mjpeg_decoder.h
-)
-
-include_directories(${ly_inc_dir})
-
-add_library(${ly_lib_name} STATIC ${ly_source_files})
-
-add_executable(convert ${ly_base_dir}/util/convert.cc)
-target_link_libraries(convert ${ly_lib_name})
-
-include(FindJPEG)
+
+PROJECT ( YUV C CXX )	# "C" is required even for C++ projects
+CMAKE_MINIMUM_REQUIRED( VERSION 2.8.12 )
+OPTION( TEST "Built unit tests" OFF )
+
+SET ( ly_base_dir	${PROJECT_SOURCE_DIR} )
+SET ( ly_src_dir	${ly_base_dir}/source )
+SET ( ly_inc_dir	${ly_base_dir}/include )
+SET ( ly_tst_dir	${ly_base_dir}/unit_test )
+SET ( ly_lib_name	yuv )
+SET ( ly_lib_static	${ly_lib_name} )
+SET ( ly_lib_shared	${ly_lib_name}_shared )
+
+FILE ( GLOB_RECURSE	ly_source_files ${ly_src_dir}/*.cc )
+LIST ( SORT			ly_source_files )
+
+FILE ( GLOB_RECURSE	ly_unittest_sources ${ly_tst_dir}/*.cc )
+LIST ( SORT			ly_unittest_sources )
+
+INCLUDE_DIRECTORIES( BEFORE ${ly_inc_dir} )
+
+if(MSVC)
+  ADD_DEFINITIONS ( -D_CRT_SECURE_NO_WARNINGS )
+endif()
+
+# this creates the static library (.a)
+ADD_LIBRARY				( ${ly_lib_static} STATIC ${ly_source_files} )
+
+# this creates the shared library (.so)
+ADD_LIBRARY				( ${ly_lib_shared} SHARED ${ly_source_files} )
+SET_TARGET_PROPERTIES	( ${ly_lib_shared} PROPERTIES OUTPUT_NAME "${ly_lib_name}" )
+SET_TARGET_PROPERTIES	( ${ly_lib_shared} PROPERTIES PREFIX "lib" )
+if(WIN32)
+  SET_TARGET_PROPERTIES	( ${ly_lib_shared} PROPERTIES IMPORT_PREFIX "lib" )
+endif()
+
+# this creates the conversion tool
+ADD_EXECUTABLE			( yuvconvert ${ly_base_dir}/util/yuvconvert.cc )
+TARGET_LINK_LIBRARIES	( yuvconvert ${ly_lib_static} )
+
+# this creates the yuvconstants tool
+ADD_EXECUTABLE      ( yuvconstants ${ly_base_dir}/util/yuvconstants.c )
+TARGET_LINK_LIBRARIES  ( yuvconstants ${ly_lib_static} )
+
+find_package ( JPEG )
 if (JPEG_FOUND)
-  include_directories(${JPEG_INCLUDE_DIR})
-  target_link_libraries(convert ${JPEG_LIBRARY})
-  add_definitions(-DHAVE_JPEG)
+  include_directories( ${JPEG_INCLUDE_DIR} )
+  target_link_libraries( ${ly_lib_shared} ${JPEG_LIBRARY} )
+  add_definitions( -DHAVE_JPEG )
 endif()
 
 if(TEST)
@@ -124,19 +69,33 @@ if(TEST)
   endif()
 
   add_executable(libyuv_unittest ${ly_unittest_sources})
-  target_link_libraries(libyuv_unittest ${ly_lib_name} ${GTEST_LIBRARY} pthread)
+  target_link_libraries(libyuv_unittest ${ly_lib_name} ${GTEST_LIBRARY})
+  find_library(PTHREAD_LIBRARY pthread)
+  if(NOT PTHREAD_LIBRARY STREQUAL "PTHREAD_LIBRARY-NOTFOUND")
+    target_link_libraries(libyuv_unittest pthread)
+  endif()
   if (JPEG_FOUND)
     target_link_libraries(libyuv_unittest ${JPEG_LIBRARY})
   endif()
-  
+
   if(NACL AND NACL_LIBC STREQUAL "newlib")
     target_link_libraries(libyuv_unittest glibc-compat)
   endif()
 
-  target_link_libraries(libyuv_unittest gflags)
-  
+  find_library(GFLAGS_LIBRARY gflags)
+  if(NOT GFLAGS_LIBRARY STREQUAL "GFLAGS_LIBRARY-NOTFOUND")
+    target_link_libraries(libyuv_unittest gflags)
+    add_definitions(-DLIBYUV_USE_GFLAGS)
+  endif()
 endif()
 
-install(TARGETS ${ly_lib_name} DESTINATION lib)
-install(FILES ${ly_header_files} DESTINATION include/libyuv)
-install(FILES ${ly_inc_dir}/libyuv.h DESTINATION include/)
+
+# install the conversion tool, .so, .a, and all the header files
+INSTALL ( PROGRAMS ${CMAKE_BINARY_DIR}/yuvconvert			DESTINATION bin )
+INSTALL ( TARGETS ${ly_lib_static}						DESTINATION lib )
+INSTALL ( TARGETS ${ly_lib_shared} LIBRARY				DESTINATION lib RUNTIME DESTINATION bin )
+INSTALL ( DIRECTORY ${PROJECT_SOURCE_DIR}/include/		DESTINATION include )
+
+# create the .deb and .rpm packages using cpack
+INCLUDE ( CM_linux_packages.cmake )
+
diff --git a/media/libyuv/DEPS b/media/libyuv/DEPS
index 0a450050b6..a56f9b0d48 100644
--- a/media/libyuv/DEPS
+++ b/media/libyuv/DEPS
@@ -1,42 +1,2438 @@
+gclient_gn_args_file = 'src/build/config/gclient_args.gni'
+gclient_gn_args = [
+  'generate_location_tags',
+]
+
 vars = {
-  # Override root_dir in your .gclient's custom_vars to specify a custom root
-  # folder name.
-  'root_dir': 'libyuv',
-  'extra_gyp_flag': '-Dextra_gyp_flag=0',
   'chromium_git': 'https://chromium.googlesource.com',
+  'chromium_revision': '504c0697552240028c5412dafd2a7306a7cd4be7',
+  'gn_version': 'git_revision:6f13aaac55a977e1948910942675c69f2b4f7a94',
+  # ninja CIPD package version.
+  # https://chrome-infra-packages.appspot.com/p/infra/3pp/tools/ninja
+  'ninja_version': 'version:2@1.8.2.chromium.3',
+  # reclient CIPD package version
+  'reclient_version': 're_client_version:0.81.1.0853992-gomaip',
 
-  # Roll the Chromium Git hash to pick up newer versions of all the
-  # dependencies and tools linked to in setup_links.py.
-  'chromium_revision': '2a818f54130d8c93f81490adce5a1e87307bf5f0',
+  # Keep the Chromium default of generating location tags.
+  'generate_location_tags': True,
+
+  # By default, download the fuchsia sdk from the public sdk directory.
+  'fuchsia_sdk_cipd_prefix': 'fuchsia/sdk/gn/',
+  'fuchsia_version': 'version:10.20221110.2.1',
+  # By default, download the fuchsia images from the fuchsia GCS bucket.
+  'fuchsia_images_bucket': 'fuchsia',
+  'checkout_fuchsia': False,
+  # Since the images are hundreds of MB, default to only downloading the image
+  # most commonly useful for developers. Bots and developers that need to use
+  # other images can override this with additional images.
+  'checkout_fuchsia_boot_images': "terminal.qemu-x64",
+  'checkout_fuchsia_product_bundles': '"{checkout_fuchsia_boot_images}" != ""',
 }
 
-# NOTE: Prefer revision numbers to tags for svn deps. Use http rather than
-# https; the latter can cause problems for users behind proxies.
 deps = {
-  Var('root_dir') + '/third_party/gflags/src':
-    Var('chromium_git') + '/external/github.com/gflags/gflags@03bebcb065c83beff83d50ae025a55a4bf94dfca',
+  'src/build':
+    Var('chromium_git') + '/chromium/src/build' + '@' + 'fe1231e1da1e95acb006f53d06caaad16756a376',
+  'src/buildtools':
+    Var('chromium_git') + '/chromium/src/buildtools' + '@' + '3c8fef071edb88facb7508060e978c5fb8608dd5',
+  'src/testing':
+    Var('chromium_git') + '/chromium/src/testing' + '@' + 'b4dc828e84ae95e1f5bf855f040c065287dac335',
+  'src/third_party':
+    Var('chromium_git') + '/chromium/src/third_party' + '@' + '73f7282fa28ca1fbe8401e391207fb6ccf34767f',
+
+  'src/buildtools/linux64': {
+    'packages': [
+      {
+        'package': 'gn/gn/linux-amd64',
+        'version': Var('gn_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+    'condition': 'checkout_linux',
+  },
+  'src/buildtools/mac': {
+    'packages': [
+      {
+        'package': 'gn/gn/mac-amd64',
+        'version': Var('gn_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+    'condition': 'checkout_mac',
+  },
+  'src/buildtools/win': {
+    'packages': [
+      {
+        'package': 'gn/gn/windows-amd64',
+        'version': Var('gn_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+    'condition': 'checkout_win',
+  },
+
+  'src/buildtools/reclient': {
+    'packages': [
+      {
+        'package': 'infra/rbe/client/${{platform}}',
+        'version': Var('reclient_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+  },
+
+  'src/buildtools/clang_format/script':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/clang/tools/clang-format.git' + '@' + '8b525d2747f2584fc35d8c7e612e66f377858df7',
+  'src/buildtools/third_party/libc++/trunk':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/libcxx.git' + '@' + 'cd0a05047451dfbdef5ba85f97ac4888e432a377',
+  'src/buildtools/third_party/libc++abi/trunk':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/libcxxabi.git' + '@' + '1a32724f721e1c3b6c590a07fe4a954344f15e48',
+  'src/buildtools/third_party/libunwind/trunk':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/libunwind.git' + '@' + '5870472fdd17f33d923b02e3e0acb9cbb18dbc9a',
+
+  'src/third_party/catapult':
+    Var('chromium_git') + '/catapult.git' + '@' + '4efb51be8574f2969273012958eaae85d01ede0b',
+  'src/third_party/colorama/src':
+    Var('chromium_git') + '/external/colorama.git' + '@' + '799604a1041e9b3bc5d2789ecbd7e8db2e18e6b8',
+  'src/third_party/depot_tools':
+    Var('chromium_git') + '/chromium/tools/depot_tools.git' + '@' + '2fc7e1ffd58b00601b47a5126201e5162911e244',
+  'src/third_party/freetype/src':
+    Var('chromium_git') + '/chromium/src/third_party/freetype2.git' + '@' + '1c44de209cb465d175279dc30cd95f9857f703dd',
+  'src/third_party/googletest/src':
+    Var('chromium_git') + '/external/github.com/google/googletest.git' + '@' + 'af29db7ec28d6df1c7f0f745186884091e602e07',
+  'src/third_party/harfbuzz-ng/src':
+    Var('chromium_git') + '/external/github.com/harfbuzz/harfbuzz.git' + '@' + '2822b589bc837fae6f66233e2cf2eef0f6ce8470',
+  'src/third_party/libjpeg_turbo':
+    Var('chromium_git') + '/chromium/deps/libjpeg_turbo.git' + '@' + 'ed683925e4897a84b3bffc5c1414c85b97a129a3',
+  'src/third_party/nasm':
+    Var('chromium_git') + '/chromium/deps/nasm.git' + '@' + '0873b2bae6a5388a1c55deac8456e3c60a47ca08',
+  'src/tools':
+    Var('chromium_git') + '/chromium/src/tools' + '@' + 'a20d904d021175f221bf58921a5a67fd48420ed9',
+
+  # libyuv-only dependencies (not present in Chromium).
+  'src/third_party/gtest-parallel':
+    Var('chromium_git') + '/external/webrtc/deps/third_party/gtest-parallel' + '@' + '1dad0e9f6d82ff994130b529d7d814b40eb32b0e',
+
+  'src/third_party/lss': {
+    'url': Var('chromium_git') + '/linux-syscall-support.git' + '@' + 'ce877209e11aa69dcfffbd53ef90ea1d07136521',
+    'condition': 'checkout_android or checkout_linux',
+  },
+
+  # Android deps:
+  'src/third_party/accessibility_test_framework': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/accessibility-test-framework',
+              'version': 'b5ec1e56e58e56bc1a0c77d43111c37f9b512c8a',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/boringssl/src':
+    'https://boringssl.googlesource.com/boringssl.git' + '@' + 'f0518d45119dd4dd322a884669daf8247bc3c992',
+  'src/base': {
+    'url': Var('chromium_git') + '/chromium/src/base' + '@' + 'f80120ff3265ba9bcb27416cc489343cfdc8bc61',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/bazel': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/bazel',
+              'version': 'VjMsf48QUWw8n7XtJP2AuSjIGmbQeYdWdwyxVvIRLmAC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/bouncycastle': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/bouncycastle',
+              'version': 'c078e87552ba26e776566fdaf0f22cd8712743d0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_ndk': {
+    'url': Var('chromium_git') + '/android_ndk.git' + '@' + '8388a2be5421311dc75c5f937aae13d821a27f3d',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/androidx': {
+    'packages': [
+      {
+          'package': 'chromium/third_party/androidx',
+          'version': 'DRqe-W5-XlO2ZySLCwsYKy7iqIaQ77O-Y91txXGY_hMC',
+      },
+    ],
+    'condition': 'checkout_android',
+    'dep_type': 'cipd',
+  },
+  'src/third_party/android_support_test_runner': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_support_test_runner',
+              'version': '96d4bf848cd210fdcbca6bcc8c1b4b39cbd93141',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_sdk/public': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_sdk/public/build-tools/33.0.0',
+              'version': '-VRKr36Uw8L_iFqqo9nevIBgNMggND5iWxjidyjnCgsC',
+          },
+          {
+              'package': 'chromium/third_party/android_sdk/public/emulator',
+              'version': '9lGp8nTUCRRWGMnI_96HcKfzjnxEJKUcfvfwmA3wXNkC',
+          },
+          {
+              'package': 'chromium/third_party/android_sdk/public/patcher',
+              'version': 'I6FNMhrXlpB-E1lOhMlvld7xt9lBVNOO83KIluXDyA0C',
+          },
+          {
+              'package': 'chromium/third_party/android_sdk/public/platform-tools',
+              'version': 'RSI3iwryh7URLGRgJHsCvUxj092woTPnKt4pwFcJ6L8C',
+          },
+          {
+              'package': 'chromium/third_party/android_sdk/public/platforms/android-33',
+              'version': 'eo5KvW6UVor92LwZai8Zulc624BQZoCu-yn7wa1z_YcC',
+          },
+          {
+              'package': 'chromium/third_party/android_sdk/public/sources/android-31',
+              'version': '_a_BcnANjPYw5mSKlNHa7GFY8yc1kdqj2rmQgac7yUcC',
+          },
+          {
+              'package': 'chromium/third_party/android_sdk/public/cmdline-tools',
+              'version': 'IPzAG-uU5zVMxohpg9-7-N0tQC1TCSW1VbrBFw7Ld04C',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/tools/clang/dsymutil': {
+    'packages': [
+      {
+        'package': 'chromium/llvm-build-tools/dsymutil',
+        'version': 'OWlhXkmj18li3yhJk59Kmjbc5KdgLh56TwCd1qBdzlIC',
+      }
+    ],
+    'condition': 'checkout_mac',
+    'dep_type': 'cipd',
+  },
+  'src/third_party/android_build_tools/aapt2': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_build_tools/aapt2',
+              'version': 'version:3.6.0-alpha03-5516695-cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/byte_buddy': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/byte_buddy',
+              'version': 'c9b53316603fc2d997c899c7ca1707f809b918cd',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/byte_buddy/android_sdk_build_tools_25_0_2': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_sdk/public/build-tools',
+              'version': 'kwIs2vdfTm93yEP8LG5aSnchN4BVEdVxbqQtF4XpPdkC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/ced/src': {
+    'url': Var('chromium_git') + '/external/github.com/google/compact_enc_det.git' + '@' + 'ba412eaaacd3186085babcd901679a48863c7dd5',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/errorprone/lib': {
+      'url': Var('chromium_git') + '/chromium/third_party/errorprone.git' + '@' + '980d49e839aa4984015efed34b0134d4b2c9b6d7',
+      'condition': 'checkout_android',
+  },
+  'src/third_party/findbugs': {
+    'url': Var('chromium_git') + '/chromium/deps/findbugs.git' + '@' + '4275d9ac8610db6b1bc9a5e887f97e41b33fac67',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/gson': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/gson',
+              'version': '681931c9778045903a0ed59856ce2dd8dd7bf7ca',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/guava': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/guava',
+              'version': 'a6fba501f3a0de88b9be1daa2052632de5b96a46',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/hamcrest': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/hamcrest',
+              'version': '37eccfc658fe79695d6abb6dd497463c4372032f',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/icu': {
+    'url': Var('chromium_git') + '/chromium/deps/icu.git' + '@' + '1b7d391f0528fb3a4976b7541b387ee04f915f83',
+  },
+  'src/third_party/icu4j': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/icu4j',
+              'version': 'e87e5bed2b4935913ee26a3ebd0b723ee2344354',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/intellij': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/intellij',
+              'version': '77c2721b024b36ee073402c08e6d8428c0295336',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/jdk': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/jdk',
+              'version': 'PfRSnxe8Od6WU4zBXomq-zsgcJgWmm3z4gMQNB-r2QcC',
+          },
+          {
+              'package': 'chromium/third_party/jdk/extras',
+              'version': 'fkhuOQ3r-zKtWEdKplpo6k0vKkjl-LY_rJTmtzFCQN4C',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/jsr-305/src': {
+    'url': Var('chromium_git') + '/external/jsr-305.git' + '@' + '642c508235471f7220af6d5df2d3210e3bfc0919',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/junit/src': {
+    'url': Var('chromium_git') + '/external/junit.git' + '@' + '05fe2a64f59127c02135be22f416e91260d6ede6',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/libunwindstack': {
+      'url': Var('chromium_git') + '/chromium/src/third_party/libunwindstack.git' + '@' + '4dbfa0e8c844c8e243b297bc185e54a99ff94f9e',
+      'condition': 'checkout_android',
+  },
+  'src/third_party/ninja': {
+    'packages': [
+      {
+        'package': 'infra/3pp/tools/ninja/${{platform}}',
+        'version': Var('ninja_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+  },
+  'src/third_party/mockito/src': {
+    'url': Var('chromium_git') + '/external/mockito/mockito.git' + '@' + '04a2a289a4222f80ad20717c25144981210d2eac',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/objenesis': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/objenesis',
+              'version': '9e367f55e5a65781ee77bfcbaa88fb82b30e75c0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/ow2_asm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/ow2_asm',
+              'version': 'NNAhdJzMdnutUVqfSJm5v0tVazA9l3Dd6CRwH6N4Q5kC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/r8': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/r8',
+              'version': 'szXK3tCGU7smsNs4r2mGqxme7d9KWLaOk0_ghbCJxUQC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  # This duplication is intentional, so we avoid updating the r8.jar used by
+  # dexing unless necessary, since each update invalidates all incremental
+  # dexing and unnecessarily slows down all bots.
+  'src/third_party/r8/d8': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/r8',
+              'version': 'Qn31g4m2cofkyvGgm46Uzkzds5DKdNYrdPePwRkVnv4C',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/proguard': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/proguard',
+              'version': 'Fd91BJFVlmiO6c46YMTsdy7n2f5Sk2hVVGlzPLvqZPsC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/requests/src': {
+    'url': Var('chromium_git') + '/external/github.com/kennethreitz/requests.git' + '@' + 'refs/tags/v2.23.0',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/robolectric': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/robolectric',
+              'version': 'iC6RDM5EH3GEAzR-1shW_Mg0FeeNE5shq1okkFfuuNQC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/sqlite4java': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/sqlite4java',
+              'version': '889660698187baa7c8b0d79f7bf58563125fbd66',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/turbine': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/turbine',
+              'version': 'Om6yIEXgJxuqghErK29h9RcMH6VaymMbxwScwXmcN6EC',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/ub-uiautomator/lib': {
+    'url': Var('chromium_git') + '/chromium/third_party/ub-uiautomator.git' + '@' + '00270549ce3161ae72ceb24712618ea28b4f9434',
+    'condition': 'checkout_android',
+  },
+
+  # iOS deps:
+  'src/ios': {
+    'url': Var('chromium_git') + '/chromium/src/ios' + '@' + '866ec86ecb27dad8a3ac7957590df7765a13834f',
+    'condition': 'checkout_ios'
+  },
+
+  # Everything coming after this is automatically updated by the auto-roller.
+  # === ANDROID_DEPS Generated Code Start ===
+  # Generated by //third_party/android_deps/fetch_all.py
+  'src/third_party/android_deps/libs/android_arch_core_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_core_common',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_core_runtime': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_core_runtime',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_lifecycle_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_common',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_lifecycle_common_java8': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_common_java8',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_lifecycle_livedata': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_livedata',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_lifecycle_livedata_core': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_livedata_core',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_lifecycle_runtime': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_runtime',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/android_arch_lifecycle_viewmodel': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_viewmodel',
+              'version': 'version:2@1.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_animated_vector_drawable': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_animated_vector_drawable',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_appcompat_v7': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_appcompat_v7',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_asynclayoutinflater': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_asynclayoutinflater',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_cardview_v7': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_cardview_v7',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_collections': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_collections',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_coordinatorlayout': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_coordinatorlayout',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_cursoradapter': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_cursoradapter',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_customview': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_customview',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_design': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_design',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_documentfile': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_documentfile',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_drawerlayout': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_drawerlayout',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_interpolator': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_interpolator',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_loader': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_loader',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_localbroadcastmanager': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_localbroadcastmanager',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_multidex': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_multidex',
+              'version': 'version:2@1.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_print': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_print',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_recyclerview_v7': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_recyclerview_v7',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_slidingpanelayout': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_slidingpanelayout',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_annotations',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_compat': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_compat',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_core_ui': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_core_ui',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_core_utils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_core_utils',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_fragment': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_fragment',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_media_compat': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_media_compat',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_v4': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_v4',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_support_vector_drawable': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_vector_drawable',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_swiperefreshlayout': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_swiperefreshlayout',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_transition': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_transition',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_versionedparcelable': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_versionedparcelable',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_support_viewpager': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_viewpager',
+              'version': 'version:2@28.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_tools_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_common',
+              'version': 'version:2@30.2.0-beta01.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs',
+              'version': 'version:2@1.1.5.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs_configuration': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs_configuration',
+              'version': 'version:2@1.1.5.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_tools_layoutlib_layoutlib_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_layoutlib_layoutlib_api',
+              'version': 'version:2@30.2.0-beta01.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_android_tools_sdk_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_sdk_common',
+              'version': 'version:2@30.2.0-beta01.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_github_ben_manes_caffeine_caffeine': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_github_ben_manes_caffeine_caffeine',
+              'version': 'version:2@2.8.8.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_github_kevinstern_software_and_algorithms': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_github_kevinstern_software_and_algorithms',
+              'version': 'version:2@1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_apps_common_testing_accessibility_framework_accessibility_test_framework': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_apps_common_testing_accessibility_framework_accessibility_test_framework',
+              'version': 'version:2@4.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_datatransport_transport_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_datatransport_transport_api',
+              'version': 'version:2@2.2.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_auth': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_auth',
+              'version': 'version:2@20.1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_auth_api_phone': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_auth_api_phone',
+              'version': 'version:2@18.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_auth_base': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_auth_base',
+              'version': 'version:2@18.0.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_base': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_base',
+              'version': 'version:2@18.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_basement': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_basement',
+              'version': 'version:2@18.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_cast': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_cast',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_cast_framework': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_cast_framework',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_clearcut': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_clearcut',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_cloud_messaging': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_cloud_messaging',
+              'version': 'version:2@16.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_fido': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_fido',
+              'version': 'version:2@19.0.0-beta.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_flags': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_flags',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_gcm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_gcm',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_iid': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_iid',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_instantapps': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_instantapps',
+              'version': 'version:2@18.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_location': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_location',
+              'version': 'version:2@19.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_phenotype': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_phenotype',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_places_placereport': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_places_placereport',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_stats': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_stats',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_tasks': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_tasks',
+              'version': 'version:2@18.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_vision': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_vision',
+              'version': 'version:2@20.1.3.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_vision_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_vision_common',
+              'version': 'version:2@19.1.3.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_material_material': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_material_material',
+              'version': 'version:2@1.7.0-alpha02.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_android_play_core': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_play_core',
+              'version': 'version:2@1.10.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_auto_auto_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_auto_common',
+              'version': 'version:2@1.2.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_auto_service_auto_service': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_service_auto_service',
+              'version': 'version:2@1.0-rc6.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_auto_service_auto_service_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_service_auto_service_annotations',
+              'version': 'version:2@1.0-rc6.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_auto_value_auto_value_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_value_auto_value_annotations',
+              'version': 'version:2@1.9.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_code_findbugs_jsr305': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_code_findbugs_jsr305',
+              'version': 'version:2@3.0.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_code_gson_gson': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_code_gson_gson',
+              'version': 'version:2@2.8.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_dagger_dagger': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger',
+              'version': 'version:2@2.30.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_dagger_dagger_compiler': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger_compiler',
+              'version': 'version:2@2.30.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_dagger_dagger_producers': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger_producers',
+              'version': 'version:2@2.30.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_dagger_dagger_spi': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger_spi',
+              'version': 'version:2@2.30.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_annotation': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_annotation',
+              'version': 'version:2@2.11.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_annotations',
+              'version': 'version:2@2.14.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_check_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_check_api',
+              'version': 'version:2@2.11.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_core': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_core',
+              'version': 'version:2@2.11.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_type_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_type_annotations',
+              'version': 'version:2@2.11.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_javac': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_javac',
+              'version': 'version:2@9+181-r4173-1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_errorprone_javac_shaded': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_javac_shaded',
+              'version': 'version:2@9-dev-r4023-3.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_annotations',
+              'version': 'version:2@16.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_common',
+              'version': 'version:2@19.5.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_components': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_components',
+              'version': 'version:2@16.1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_encoders': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_encoders',
+              'version': 'version:2@16.1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_encoders_json': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_encoders_json',
+              'version': 'version:2@17.1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_iid': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_iid',
+              'version': 'version:2@21.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_iid_interop': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_iid_interop',
+              'version': 'version:2@17.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_installations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_installations',
+              'version': 'version:2@16.3.5.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_installations_interop': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_installations_interop',
+              'version': 'version:2@16.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_measurement_connector': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_measurement_connector',
+              'version': 'version:2@18.0.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_messaging': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_messaging',
+              'version': 'version:2@21.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_flatbuffers_flatbuffers_java': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_flatbuffers_flatbuffers_java',
+              'version': 'version:2@2.0.3.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_googlejavaformat_google_java_format': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_googlejavaformat_google_java_format',
+              'version': 'version:2@1.5.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_guava_failureaccess': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_failureaccess',
+              'version': 'version:2@1.0.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_guava_guava': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_guava',
+              'version': 'version:2@31.0.1-jre.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_guava_guava_android': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_guava_android',
+              'version': 'version:2@31.0-android.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_guava_listenablefuture': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_listenablefuture',
+              'version': 'version:2@1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_j2objc_j2objc_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_j2objc_j2objc_annotations',
+              'version': 'version:2@1.3.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_protobuf_protobuf_java': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_protobuf_protobuf_java',
+              'version': 'version:2@3.19.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_google_protobuf_protobuf_javalite': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_protobuf_protobuf_javalite',
+              'version': 'version:2@3.19.3.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_googlecode_java_diff_utils_diffutils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_googlecode_java_diff_utils_diffutils',
+              'version': 'version:2@1.3.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_squareup_javapoet': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_squareup_javapoet',
+              'version': 'version:2@1.13.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/com_squareup_javawriter': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_squareup_javawriter',
+              'version': 'version:2@2.1.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/io_github_java_diff_utils_java_diff_utils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/io_github_java_diff_utils_java_diff_utils',
+              'version': 'version:2@4.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/javax_annotation_javax_annotation_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/javax_annotation_javax_annotation_api',
+              'version': 'version:2@1.3.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/javax_annotation_jsr250_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/javax_annotation_jsr250_api',
+              'version': 'version:2@1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/javax_inject_javax_inject': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/javax_inject_javax_inject',
+              'version': 'version:2@1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/net_bytebuddy_byte_buddy': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/net_bytebuddy_byte_buddy',
+              'version': 'version:2@1.12.13.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/net_bytebuddy_byte_buddy_agent': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/net_bytebuddy_byte_buddy_agent',
+              'version': 'version:2@1.12.13.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/net_ltgt_gradle_incap_incap': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/net_ltgt_gradle_incap_incap',
+              'version': 'version:2@0.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/net_sf_kxml_kxml2': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/net_sf_kxml_kxml2',
+              'version': 'version:2@2.3.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_bouncycastle_bcprov_jdk15on': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_bouncycastle_bcprov_jdk15on',
+              'version': 'version:2@1.68.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_ccil_cowan_tagsoup_tagsoup': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ccil_cowan_tagsoup_tagsoup',
+              'version': 'version:2@1.2.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_checkerframework_checker_compat_qual': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_checkerframework_checker_compat_qual',
+              'version': 'version:2@2.5.5.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_checkerframework_checker_qual': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_checkerframework_checker_qual',
+              'version': 'version:2@3.25.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_checkerframework_dataflow_errorprone': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_checkerframework_dataflow_errorprone',
+              'version': 'version:2@3.15.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_codehaus_mojo_animal_sniffer_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_codehaus_mojo_animal_sniffer_annotations',
+              'version': 'version:2@1.17.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_eclipse_jgit_org_eclipse_jgit': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_eclipse_jgit_org_eclipse_jgit',
+              'version': 'version:2@4.4.1.201607150455-r.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_hamcrest_hamcrest': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_hamcrest_hamcrest',
+              'version': 'version:2@2.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_annotations',
+              'version': 'version:2@13.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib',
+              'version': 'version:2@1.7.10.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_common',
+              'version': 'version:2@1.7.10.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk7': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk7',
+              'version': 'version:2@1.6.20.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk8': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk8',
+              'version': 'version:2@1.6.20.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_android': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_android',
+              'version': 'version:2@1.6.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_core_jvm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_core_jvm',
+              'version': 'version:2@1.6.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_metadata_jvm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_metadata_jvm',
+              'version': 'version:2@0.1.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_jsoup_jsoup': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jsoup_jsoup',
+              'version': 'version:2@1.15.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_mockito_mockito_core': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_mockito_mockito_core',
+              'version': 'version:2@4.7.0.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_objenesis_objenesis': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_objenesis_objenesis',
+              'version': 'version:2@3.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_ow2_asm_asm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm',
+              'version': 'version:2@9.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_analysis': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_analysis',
+              'version': 'version:2@9.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_commons': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_commons',
+              'version': 'version:2@9.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_tree': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_tree',
+              'version': 'version:2@9.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_util': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_util',
+              'version': 'version:2@9.2.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_pcollections_pcollections': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_pcollections_pcollections',
+              'version': 'version:2@3.1.4.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_annotations',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_junit': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_junit',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_nativeruntime': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_nativeruntime',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_pluginapi': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_pluginapi',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_plugins_maven_dependency_resolver': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_plugins_maven_dependency_resolver',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_resources': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_resources',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_robolectric': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_robolectric',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_sandbox': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_sandbox',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_shadowapi': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_shadowapi',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_shadows_framework': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_shadows_framework',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_shadows_playservices': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_shadows_playservices',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_utils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_utils',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  'src/third_party/android_deps/libs/org_robolectric_utils_reflector': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_utils_reflector',
+              'version': 'version:2@4.8.1.cr1',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+
+  # === ANDROID_DEPS Generated Code End ===
 }
 
-# Define rules for which include paths are allowed in our source.
-include_rules = [ '+gflags' ]
+pre_deps_hooks = [
+  {
+    # Remove any symlinks from before 177567c518b121731e507e9b9c4049c4dc96e4c8.
+    # TODO(kjellander): Remove this in March 2017.
+    'name': 'cleanup_links',
+    'pattern': '.',
+    'action': ['python3', 'src/cleanup_links.py'],
+  },
+]
 
 hooks = [
   {
-    # Clone chromium and its deps.
-    'name': 'sync chromium',
+    # This clobbers when necessary (based on get_landmines.py). It should be
+    # an early hook but it will need to be run after syncing Chromium and
+    # setting up the links, so the script actually exists.
+    'name': 'landmines',
     'pattern': '.',
-    'action': ['python', '-u', Var('root_dir') + '/sync_chromium.py',
-               '--target-revision', Var('chromium_revision')],
+    'action': [
+        'python3',
+        'src/build/landmines.py',
+        '--landmine-scripts',
+        'src/tools_libyuv/get_landmines.py',
+        '--src-dir',
+        'src',
+    ],
   },
+  # Downloads the current stable linux sysroot to build/linux/ if needed.
   {
-    # Create links to shared dependencies in Chromium.
-    'name': 'setup_links',
+    'name': 'sysroot_arm',
     'pattern': '.',
-    'action': ['python', Var('root_dir') + '/setup_links.py'],
+    'condition': 'checkout_linux and checkout_arm',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+               '--arch=arm'],
   },
   {
-    # A change to a .gyp, .gypi, or to GYP itself should run the generator.
+    'name': 'sysroot_arm64',
     'pattern': '.',
-    'action': ['python', Var('root_dir') + '/gyp_libyuv'],
+    'condition': 'checkout_linux and checkout_arm64',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+               '--arch=arm64'],
+  },
+  {
+    'name': 'sysroot_x86',
+    'pattern': '.',
+    'condition': 'checkout_linux and (checkout_x86 or checkout_x64)',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+               '--arch=x86'],
+  },
+  {
+    'name': 'sysroot_mips',
+    'pattern': '.',
+    'condition': 'checkout_linux and checkout_mips',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+               '--arch=mips'],
+  },
+  {
+    'name': 'sysroot_x64',
+    'pattern': '.',
+    'condition': 'checkout_linux and checkout_x64',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+               '--arch=x64'],
+  },
+  {
+    # Update the Windows toolchain if necessary.
+    'name': 'win_toolchain',
+    'pattern': '.',
+    'action': ['python3', 'src/build/vs_toolchain.py', 'update'],
+  },
+  {
+    # Update the Mac toolchain if necessary.
+    'name': 'mac_toolchain',
+    'pattern': '.',
+    'action': ['python3', 'src/build/mac_toolchain.py'],
+    'condition': 'checkout_mac',
+  },
+  {
+    'name': 'msan_chained_origins_focal',
+    'pattern': '.',
+    'condition': 'checkout_instrumented_libraries',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--no_auth',
+                '--bucket', 'chromium-instrumented-libraries',
+                '-s', 'src/third_party/instrumented_libraries/binaries/msan-chained-origins-focal.tgz.sha1',
+              ],
+  },
+  {
+    'name': 'msan_no_origins_focal',
+    'pattern': '.',
+    'condition': 'checkout_instrumented_libraries',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--no_auth',
+                '--bucket', 'chromium-instrumented-libraries',
+                '-s', 'src/third_party/instrumented_libraries/binaries/msan-no-origins-focal.tgz.sha1',
+              ],
+  },
+  {
+    'name': 'msan_chained_origins_xenial',
+    'pattern': '.',
+    'condition': 'checkout_instrumented_libraries',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--no_auth',
+                '--bucket', 'chromium-instrumented-libraries',
+                '-s', 'src/third_party/instrumented_libraries/binaries/msan-chained-origins-xenial.tgz.sha1',
+              ],
+  },
+  {
+    'name': 'msan_no_origins_xenial',
+    'pattern': '.',
+    'condition': 'checkout_instrumented_libraries',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--no_auth',
+                '--bucket', 'chromium-instrumented-libraries',
+                '-s', 'src/third_party/instrumented_libraries/binaries/msan-no-origins-xenial.tgz.sha1',
+              ],
+  },
+  {
+    'name': 'Download Fuchsia SDK from GCS',
+    'pattern': '.',
+    'condition': 'checkout_fuchsia',
+    'action': [
+      'python3',
+      'src/build/fuchsia/update_sdk.py',
+      '--cipd-prefix={fuchsia_sdk_cipd_prefix}',
+      '--version={fuchsia_version}',
+    ],
+  },
+  {
+    'name': 'Download Fuchsia system images',
+    'pattern': '.',
+    'condition': 'checkout_fuchsia and checkout_fuchsia_product_bundles',
+    'action': [
+      'python3',
+      'src/build/fuchsia/update_product_bundles.py',
+      '{checkout_fuchsia_boot_images}',
+    ],
+  },
+  {
+    # Pull clang if needed or requested via GYP_DEFINES.
+    # Note: On Win, this should run after win_toolchain, as it may use it.
+    'name': 'clang',
+    'pattern': '.',
+    'action': ['python3', 'src/tools/clang/scripts/update.py'],
+  },
+  {
+    # Update LASTCHANGE.
+    'name': 'lastchange',
+    'pattern': '.',
+    'action': ['python3', 'src/build/util/lastchange.py',
+               '-o', 'src/build/util/LASTCHANGE'],
+  },
+  # Pull clang-format binaries using checked-in hashes.
+  {
+    'name': 'clang_format_win',
+    'pattern': '.',
+    'condition': 'host_os == "win"',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--platform=win32',
+                '--no_auth',
+                '--bucket', 'chromium-clang-format',
+                '-s', 'src/buildtools/win/clang-format.exe.sha1',
+    ],
+  },
+  {
+    'name': 'clang_format_mac_x64',
+    'pattern': '.',
+    'condition': 'host_os == "mac" and host_cpu == "x64"',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--platform=darwin',
+                '--no_auth',
+                '--bucket', 'chromium-clang-format',
+                '-s', 'src/buildtools/mac/clang-format.x64.sha1',
+                '-o', 'src/buildtools/mac/clang-format',
+    ],
+  },
+  {
+    'name': 'clang_format_mac_arm64',
+    'pattern': '.',
+    'condition': 'host_os == "mac" and host_cpu == "arm64"',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--no_auth',
+                '--bucket', 'chromium-clang-format',
+                '-s', 'src/buildtools/mac/clang-format.arm64.sha1',
+                '-o', 'src/buildtools/mac/clang-format',
+     ],
+  },
+  {
+    'name': 'clang_format_linux',
+    'pattern': '.',
+    'condition': 'host_os == "linux"',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                '--no_resume',
+                '--platform=linux*',
+                '--no_auth',
+                '--bucket', 'chromium-clang-format',
+                '-s', 'src/buildtools/linux64/clang-format.sha1',
+    ],
+  },
+  # Pull luci-go binaries (isolate, swarming) using checked-in hashes.
+  {
+    'name': 'luci-go_win',
+    'pattern': '.',
+    'action': [ 'download_from_google_storage',
+                '--no_resume',
+                '--platform=win32',
+                '--no_auth',
+                '--bucket', 'chromium-luci',
+                '-d', 'src/tools/luci-go/win64',
+    ],
+  },
+  {
+    'name': 'luci-go_mac',
+    'pattern': '.',
+    'action': [ 'download_from_google_storage',
+                '--no_resume',
+                '--platform=darwin',
+                '--no_auth',
+                '--bucket', 'chromium-luci',
+                '-d', 'src/tools/luci-go/mac64',
+    ],
+  },
+  {
+    'name': 'luci-go_linux',
+    'pattern': '.',
+    'action': [ 'download_from_google_storage',
+                '--no_resume',
+                '--platform=linux*',
+                '--no_auth',
+                '--bucket', 'chromium-luci',
+                '-d', 'src/tools/luci-go/linux64',
+    ],
+  },
+  {
+    'name': 'Generate component metadata for tests',
+    'pattern': '.',
+    'action': [
+      'vpython3',
+      'src/testing/generate_location_tags.py',
+      '--out',
+      'src/testing/location_tags.json',
+    ],
   },
 ]
+
+recursedeps = []
diff --git a/media/libyuv/DIR_METADATA b/media/libyuv/DIR_METADATA
new file mode 100644
index 0000000000..8bc04f159c
--- /dev/null
+++ b/media/libyuv/DIR_METADATA
@@ -0,0 +1,3 @@
+monorail {
+  component: "Internals>Images>Codecs"
+}
diff --git a/media/libyuv/LICENSE_THIRD_PARTY b/media/libyuv/LICENSE_THIRD_PARTY
deleted file mode 100644
index a71591e771..0000000000
--- a/media/libyuv/LICENSE_THIRD_PARTY
+++ /dev/null
@@ -1,8 +0,0 @@
-This source tree contains third party source code which is governed by third
-party licenses. This file contains references to files which are under other
-licenses than the one provided in the LICENSE file in the root of the source
-tree.
-
-Files governed by third party licenses:
-source/x86inc.asm
-
diff --git a/media/libyuv/OWNERS b/media/libyuv/OWNERS
index 2db52d3079..f11a7bfdc5 100644
--- a/media/libyuv/OWNERS
+++ b/media/libyuv/OWNERS
@@ -1,13 +1,11 @@
+mbonadei@chromium.org
 fbarchard@chromium.org
 magjed@chromium.org
-torbjorng@chromium.org
+wtc@google.com
+jansson@google.com
 
-per-file *.gyp=kjellander@chromium.org
-per-file *.gn=kjellander@chromium.org
+per-file *.gn=mbonadei@chromium.org,jansson@google.com
 per-file .gitignore=*
 per-file AUTHORS=*
 per-file DEPS=*
-per-file PRESUBMIT.py=kjellander@chromium.org
-per-file gyp_libyuv.py=kjellander@chromium.org
-per-file setup_links.py=*
-per-file sync_chromium.py=kjellander@chromium.org
+per-file PRESUBMIT.py=mbonadei@chromium.org,jansson@google.com
diff --git a/media/libyuv/PRESUBMIT.py b/media/libyuv/PRESUBMIT.py
index e194fa35a6..d3901caf3a 100644
--- a/media/libyuv/PRESUBMIT.py
+++ b/media/libyuv/PRESUBMIT.py
@@ -1,4 +1,4 @@
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
+# Copyright 2017 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
 # that can be found in the LICENSE file in the root of the source
@@ -6,51 +6,47 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
-import re
-import sys
+# Runs PRESUBMIT.py in py3 mode by git cl presubmit.
+USE_PYTHON3 = True
 
+def _CommonChecks(input_api, output_api):
+  """Checks common to both upload and commit."""
+  results = []
+  results.extend(input_api.canned_checks.RunPylint(input_api, output_api,
+      files_to_skip=(r'^base[\\\/].*\.py$',
+                     r'^build[\\\/].*\.py$',
+                     r'^buildtools[\\\/].*\.py$',
+                     r'^ios[\\\/].*\.py$',
+                     r'^out.*[\\\/].*\.py$',
+                     r'^testing[\\\/].*\.py$',
+                     r'^third_party[\\\/].*\.py$',
+                     r'^tools[\\\/].*\.py$',
+                     # TODO(kjellander): should arguably be checked.
+                     r'^tools_libyuv[\\\/]valgrind[\\\/].*\.py$',
+                     r'^xcodebuild.*[\\\/].*\.py$',),
+      disabled_warnings=['F0401',  # Failed to import x
+                         'E0611',  # No package y in x
+                         'W0232',  # Class has no __init__ method
+                        ],
+      pylintrc='pylintrc',
+      version='2.7'))
+  return results
 
-def GetDefaultTryConfigs(bots=None):
-  """Returns a list of ('bot', set(['tests']), optionally filtered by [bots].
 
-  For WebRTC purposes, we always return an empty list of tests, since we want
-  to run all tests by default on all our trybots.
-  """
-  return { 'tryserver.libyuv': dict((bot, []) for bot in bots)}
+def CheckChangeOnUpload(input_api, output_api):
+  results = []
+  results.extend(_CommonChecks(input_api, output_api))
+  results.extend(
+      input_api.canned_checks.CheckGNFormatted(input_api, output_api))
+  return results
 
 
-# pylint: disable=W0613
-def GetPreferredTryMasters(project, change):
-  files = change.LocalPaths()
-  bots = [
-    'win',
-    'win_rel',
-    'win_x64_rel',
-    'win_x64_gn',
-    'win_x64_gn_rel',
-    'win_clang',
-    'win_clang_rel',
-    'win_x64_clang_rel',
-    'mac',
-    'mac_rel',
-    'mac_gn',
-    'mac_gn_rel',
-    'mac_asan',
-    'ios',
-    'ios_rel',
-    'ios_arm64',
-    'ios_arm64_rel',
-    'linux',
-    'linux_rel',
-    'linux_gn',
-    'linux_gn_rel',
-    'linux_memcheck',
-    'linux_tsan2',
-    'linux_asan',
-    'linux_msan',
-    'linux_ubsan',
-    'linux_ubsan_vptr',
-  ]
-  if not files or all(re.search(r'[\\/]OWNERS$', f) for f in files):
-    return {}
-  return GetDefaultTryConfigs(bots)
+def CheckChangeOnCommit(input_api, output_api):
+  results = []
+  results.extend(_CommonChecks(input_api, output_api))
+  results.extend(input_api.canned_checks.CheckOwners(input_api, output_api))
+  results.extend(input_api.canned_checks.CheckChangeWasUploaded(
+      input_api, output_api))
+  results.extend(input_api.canned_checks.CheckChangeHasDescription(
+      input_api, output_api))
+  return results
diff --git a/media/libyuv/README.chromium b/media/libyuv/README.chromium
index 251f8676fb..3d1777bed0 100644
--- a/media/libyuv/README.chromium
+++ b/media/libyuv/README.chromium
@@ -1,8 +1,9 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1602
+Version: 1861
 License: BSD
 License File: LICENSE
 
 Description:
 libyuv is an open source project that includes YUV conversion and scaling functionality.
+
diff --git a/media/libyuv/README.md b/media/libyuv/README.md
index b59b71c5d2..db70b7f08d 100644
--- a/media/libyuv/README.md
+++ b/media/libyuv/README.md
@@ -1,18 +1,18 @@
 **libyuv** is an open source project that includes YUV scaling and conversion functionality.
 
 * Scale YUV to prepare content for compression, with point, bilinear or box filter.
-* Convert to YUV from webcam formats.
-* Convert from YUV to formats for rendering/effects.
+* Convert to YUV from webcam formats for compression.
+* Convert to RGB formats for rendering/effects.
 * Rotate by 90/180/270 degrees to adjust for mobile devices in portrait mode.
-* Optimized for SSE2/SSSE3/AVX2 on x86/x64.
+* Optimized for SSSE3/AVX2 on x86/x64.
 * Optimized for Neon on Arm.
-* Optimized for DSP R2 on Mips.
+* Optimized for MSA on Mips.
 
 ### Development
 
-See [Getting started] [1] for instructions on how to get started developing.
+See [Getting started][1] for instructions on how to get started developing.
 
-You can also browse the [docs directory] [2] for more documentation.
+You can also browse the [docs directory][2] for more documentation.
 
-[1]: https://chromium.googlesource.com/libyuv/libyuv/+/master/docs/getting_started.md
-[2]: https://chromium.googlesource.com/libyuv/libyuv/+/master/docs/
+[1]: ./docs/getting_started.md
+[2]: ./docs/
diff --git a/media/libyuv/all.gyp b/media/libyuv/all.gyp
deleted file mode 100644
index 88a7484271..0000000000
--- a/media/libyuv/all.gyp
+++ /dev/null
@@ -1,21 +0,0 @@
-# Copyright 2013 The LibYuv Project Authors. All rights reserved.

-#

-# Use of this source code is governed by a BSD-style license

-# that can be found in the LICENSE file in the root of the source

-# tree. An additional intellectual property rights grant can be found

-# in the file PATENTS. All contributing project authors may

-# be found in the AUTHORS file in the root of the source tree.

-

-# all.gyp and All target are for benefit of android gyp build.

-{

-  'targets': [

-    {

-      'target_name': 'All',

-      'type': 'none',

-      'dependencies': [

-        'libyuv.gyp:*',

-        'libyuv_test.gyp:*',

-      ],

-    },

-  ],

-}

diff --git a/media/libyuv/build_overrides/build.gni b/media/libyuv/build_overrides/build.gni
index 6d3aa1ebc7..c849031338 100644
--- a/media/libyuv/build_overrides/build.gni
+++ b/media/libyuv/build_overrides/build.gni
@@ -6,10 +6,55 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
-# Using same overrides as WebRTC
-# See https://bugs.chromium.org/p/webrtc/issues/detail?id=5453.
-# Some WebRTC targets require the 10.7 deployment version of the Mac SDK and a
-# 10.11 min SDK but those targets are only used in non-Chromium builds. We can
-# remove this when Chromium drops 10.6 support and also requires 10.7.
-mac_sdk_min_build_override = "10.11"
-mac_deployment_target_build_override = "10.7"
+# Variable that can be used to support multiple build scenarios, like having
+# Chromium specific targets in a client project's GN file etc.
+build_with_chromium = false
+
+# Some non-Chromium builds don't support building java targets.
+enable_java_templates = true
+
+# Allow using custom suppressions files (currently not used by libyuv).
+asan_suppressions_file = "//build/sanitizers/asan_suppressions.cc"
+lsan_suppressions_file = "//build/sanitizers/lsan_suppressions.cc"
+tsan_suppressions_file = "//build/sanitizers/tsan_suppressions.cc"
+
+msan_blacklist_path =
+    rebase_path("//tools_libyuv/msan/blacklist.txt", root_build_dir)
+ubsan_blacklist_path =
+    rebase_path("//tools_libyuv/ubsan/blacklist.txt", root_build_dir)
+ubsan_vptr_blacklist_path =
+    rebase_path("//tools_libyuv/ubsan/vptr_blacklist.txt", root_build_dir)
+
+# For Chromium, Android 32-bit non-component, non-clang builds hit a 4GiB size
+# limit, making them requiring symbol_level=2. WebRTC doesn't hit that problem
+# so we just ignore that assert. See https://crbug.com/648948 for more info.
+ignore_elf32_limitations = true
+
+# Use bundled hermetic Xcode installation maintained by Chromium,
+# except for local iOS builds where it is unsupported.
+if (host_os == "mac") {
+  _result = exec_script("//build/mac/should_use_hermetic_xcode.py",
+                        [ target_os ],
+                        "value")
+  assert(_result != 2,
+         "Do not allow building targets with the default" +
+             "hermetic toolchain if the minimum OS version is not met.")
+  use_system_xcode = _result == 0
+}
+
+declare_args() {
+  # Tracing support requires //third_party/perfetto.
+  enable_base_tracing = false
+  use_perfetto_client_library = false
+
+  # Limits the defined //third_party/android_deps targets to only "buildCompile"
+  # and "buildCompileNoDeps" targets. This is useful for third-party
+  # repositories which do not use JUnit tests. For instance,
+  # limit_android_deps == true removes "gn gen" requirement for
+  # //third_party/robolectric .
+  limit_android_deps = false
+
+  # Allows googletest to pretty-print various absl types.
+  # Defined here rather than in gtest.gni to match chromium.
+  gtest_enable_absl_printers = true
+}
diff --git a/media/libyuv/build_overrides/gtest.gni b/media/libyuv/build_overrides/gtest.gni
new file mode 100644
index 0000000000..d3c3f68cef
--- /dev/null
+++ b/media/libyuv/build_overrides/gtest.gni
@@ -0,0 +1,19 @@
+# Copyright (c) 2016 The LibYuv project authors. All Rights Reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS.  All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+# Include support for registering main function in multi-process tests.
+gtest_include_multiprocess = true
+
+# Include support for platform-specific operations across unit tests.
+gtest_include_platform_test = true
+
+# Exclude support for testing Objective C code on OS X and iOS.
+gtest_include_objc_support = true
+
+# Exclude support for flushing coverage files on iOS.
+gtest_include_ios_coverage = true
diff --git a/media/libyuv/build_overrides/partition_alloc.gni b/media/libyuv/build_overrides/partition_alloc.gni
new file mode 100644
index 0000000000..dcf8ac2d25
--- /dev/null
+++ b/media/libyuv/build_overrides/partition_alloc.gni
@@ -0,0 +1,17 @@
+# Copyright 2022 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+# Use default values for PartitionAlloc as standalone library from
+# base/allocator/partition_allocator/build_overrides/partition_alloc.gni
+use_partition_alloc_as_malloc_default = false
+use_allocator_shim_default = false
+enable_backup_ref_ptr_support_default = false
+enable_mte_checked_ptr_support_default = false
+put_ref_count_in_previous_slot_default = false
+enable_backup_ref_ptr_slow_checks_default = false
+enable_dangling_raw_ptr_checks_default = false
diff --git a/media/libyuv/chromium/README b/media/libyuv/chromium/README
deleted file mode 100644
index 127f4b520f..0000000000
--- a/media/libyuv/chromium/README
+++ /dev/null
@@ -1,5 +0,0 @@
-This .gclient file is used to do download a copy of Chromium.
-Libyuv uses the Chromium build toolchain and a number of shared
-dependencies by creating symlinks to folders in this checkout,
-using the ../setup_links.py script.
-
diff --git a/media/libyuv/cleanup_links.py b/media/libyuv/cleanup_links.py
new file mode 100755
index 0000000000..7d1eba9b6b
--- /dev/null
+++ b/media/libyuv/cleanup_links.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env vpython3
+
+# Copyright 2017 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+# This is a copy of the file from WebRTC in:
+# https://chromium.googlesource.com/external/webrtc/+/master/cleanup_links.py
+
+"""Script to cleanup symlinks created from setup_links.py.
+
+Before 177567c518b121731e507e9b9c4049c4dc96e4c8 (#15754) we had a Chromium
+checkout which we created symlinks into. In order to do clean syncs after
+landing that change, this script cleans up any old symlinks, avoiding annoying
+manual cleanup needed in order to complete gclient sync.
+"""
+
+import argparse
+import logging
+import os
+import shelve
+import subprocess
+import sys
+
+
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+LINKS_DB = 'links'
+
+# Version management to make future upgrades/downgrades easier to support.
+SCHEMA_VERSION = 1
+
+class WebRTCLinkSetup():
+  def __init__(self, links_db, dry_run=False):
+    self._dry_run = dry_run
+    self._links_db = links_db
+
+  def CleanupLinks(self):
+    logging.debug('CleanupLinks')
+    for source, link_path  in self._links_db.tems():
+      if source == 'SCHEMA_VERSION':
+        continue
+      if os.path.islink(link_path) or sys.platform.startswith('win'):
+        # os.path.islink() always returns false on Windows
+        # See http://bugs.python.org/issue13143.
+        logging.debug('Removing link to %s at %s', source, link_path)
+        if not self._dry_run:
+          if os.path.exists(link_path):
+            if sys.platform.startswith('win') and os.path.isdir(link_path):
+              subprocess.check_call(['rmdir', '/q', '/s', link_path],
+                                    shell=True)
+            else:
+              os.remove(link_path)
+          del self._links_db[source]
+
+
+def _initialize_database(filename):
+  links_database = shelve.open(filename)
+  # Wipe the database if this version of the script ends up looking at a
+  # newer (future) version of the links db, just to be sure.
+  version = links_database.get('SCHEMA_VERSION')
+  if version and version != SCHEMA_VERSION:
+    logging.info('Found database with schema version %s while this script only '
+                 'supports %s. Wiping previous database contents.', version,
+                 SCHEMA_VERSION)
+    links_database.clear()
+  links_database['SCHEMA_VERSION'] = SCHEMA_VERSION
+  return links_database
+
+
+def main():
+  p = argparse.ArgumentParser()
+  p.add_argument('-d', '--dry-run', action='store_true', default=False,
+                 help='Print what would be done, but don\'t perform any '
+                      'operations. This will automatically set logging to '
+                      'verbose.')
+  p.add_argument('-v', '--verbose', action='store_const',
+                 const=logging.DEBUG, default=logging.INFO,
+                 help='Print verbose output for debugging.')
+  options = p.parse_args()
+
+  if options.dry_run:
+    options.verbose = logging.DEBUG
+  logging.basicConfig(format='%(message)s', level=options.verbose)
+
+  # Work from the root directory of the checkout.
+  script_dir = os.path.dirname(os.path.abspath(__file__))
+  os.chdir(script_dir)
+
+  # The database file gets .db appended on some platforms.
+  db_filenames = [LINKS_DB, LINKS_DB + '.db']
+  if any(os.path.isfile(f) for f in db_filenames):
+    links_database = _initialize_database(LINKS_DB)
+    try:
+      symlink_creator = WebRTCLinkSetup(links_database, options.dry_run)
+      symlink_creator.CleanupLinks()
+    finally:
+      for f in db_filenames:
+        if os.path.isfile(f):
+          os.remove(f)
+  return 0
+
+
+if __name__ == '__main__':
+  sys.exit(main())
diff --git a/media/libyuv/codereview.settings b/media/libyuv/codereview.settings
index 9b5380694e..b226fae57a 100644
--- a/media/libyuv/codereview.settings
+++ b/media/libyuv/codereview.settings
@@ -1,12 +1,5 @@
-# This file is used by gcl to get repository specific information.
+# This file is used by `git cl` to get repository specific information.
 CODE_REVIEW_SERVER: codereview.chromium.org
-#CC_LIST:
-VIEW_VC: https://chromium.googlesource.com/libyuv/libyuv/+/
-#STATUS:
-FORCE_HTTPS_COMMIT_URL: True
+GERRIT_HOST: True
 PROJECT: libyuv
-TRY_ON_UPLOAD: False
-TRYSERVER_ROOT: src
-TRYSERVER_SVN_URL: svn://svn.chromium.org/chrome-try/try-libyuv
-#GITCL_PREUPLOAD:
-#GITCL_PREDCOMMIT:
+VIEW_VC: https://chromium.googlesource.com/libyuv/libyuv/+/
diff --git a/media/libyuv/docs/deprecated_builds.md b/media/libyuv/docs/deprecated_builds.md
new file mode 100644
index 0000000000..ba42966cc2
--- /dev/null
+++ b/media/libyuv/docs/deprecated_builds.md
@@ -0,0 +1,441 @@
+# Deprecated Builds
+
+Older documentation on build configs which are no longer supported.
+
+## Pre-requisites
+
+You'll need to have depot tools installed: https://www.chromium.org/developers/how-tos/install-depot-tools
+Refer to chromium instructions for each platform for other prerequisites.
+
+## Getting the Code
+
+Create a working directory, enter it, and run:
+
+    gclient config https://chromium.googlesource.com/libyuv/libyuv
+    gclient sync
+
+
+Then you'll get a .gclient file like:
+
+    solutions = [
+      { "name"        : "libyuv",
+        "url"         : "https://chromium.googlesource.com/libyuv/libyuv",
+        "deps_file"   : "DEPS",
+        "managed"     : True,
+        "custom_deps" : {
+        },
+        "safesync_url": "",
+      },
+    ];
+
+
+For iOS add `;target_os=['ios'];` to your OSX .gclient and run `GYP_DEFINES="OS=ios" gclient sync.`
+
+Browse the Git reprository: https://chromium.googlesource.com/libyuv/libyuv/+/master
+
+### Android
+For Android add `;target_os=['android'];` to your Linux .gclient
+
+
+    solutions = [
+      { "name"        : "libyuv",
+        "url"         : "https://chromium.googlesource.com/libyuv/libyuv",
+        "deps_file"   : "DEPS",
+        "managed"     : True,
+        "custom_deps" : {
+        },
+        "safesync_url": "",
+      },
+    ];
+    target_os = ["android", "unix"];
+
+Then run:
+
+    export GYP_DEFINES="OS=android"
+    gclient sync
+
+Caveat: Theres an error with Google Play services updates.  If you get the error "Your version of the Google Play services library is not up to date", run the following:
+
+    cd chromium/src
+    ./build/android/play_services/update.py download
+    cd ../..
+
+For Windows the gclient sync must be done from an Administrator command prompt.
+
+The sync will generate native build files for your environment using gyp (Windows: Visual Studio, OSX: XCode, Linux: make). This generation can also be forced manually: `gclient runhooks`
+
+To get just the source (not buildable):
+
+    git clone https://chromium.googlesource.com/libyuv/libyuv
+
+
+## Building the Library and Unittests
+
+### Windows
+
+    set GYP_DEFINES=target_arch=ia32
+    call python gyp_libyuv -fninja -G msvs_version=2013
+    ninja -j7 -C out\Release
+    ninja -j7 -C out\Debug
+
+    set GYP_DEFINES=target_arch=x64
+    call python gyp_libyuv -fninja -G msvs_version=2013
+    ninja -C out\Debug_x64
+    ninja -C out\Release_x64
+
+#### Building with clangcl
+    set GYP_DEFINES=clang=1 target_arch=ia32
+    call python tools\clang\scripts\update.py
+    call python gyp_libyuv -fninja libyuv_test.gyp
+    ninja -C out\Debug
+    ninja -C out\Release
+
+### OSX
+
+Clang 64 bit shown. Remove `clang=1` for GCC and change x64 to ia32 for 32 bit.
+
+    GYP_DEFINES="clang=1 target_arch=x64" ./gyp_libyuv
+    ninja -j7 -C out/Debug
+    ninja -j7 -C out/Release
+
+    GYP_DEFINES="clang=1 target_arch=ia32" ./gyp_libyuv
+    ninja -j7 -C out/Debug
+    ninja -j7 -C out/Release
+
+### iOS
+http://www.chromium.org/developers/how-tos/build-instructions-ios
+
+Add to .gclient last line: `target_os=['ios'];`
+
+armv7
+
+    GYP_DEFINES="OS=ios target_arch=armv7 target_subarch=arm32" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_ios" ./gyp_libyuv
+    ninja -j7 -C out_ios/Debug-iphoneos libyuv_unittest
+    ninja -j7 -C out_ios/Release-iphoneos libyuv_unittest
+
+arm64
+
+    GYP_DEFINES="OS=ios target_arch=arm64 target_subarch=arm64" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_ios" ./gyp_libyuv
+    ninja -j7 -C out_ios/Debug-iphoneos libyuv_unittest
+    ninja -j7 -C out_ios/Release-iphoneos libyuv_unittest
+
+both armv7 and arm64 (fat)
+
+    GYP_DEFINES="OS=ios target_arch=armv7 target_subarch=both" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_ios" ./gyp_libyuv
+    ninja -j7 -C out_ios/Debug-iphoneos libyuv_unittest
+    ninja -j7 -C out_ios/Release-iphoneos libyuv_unittest
+
+simulator
+
+    GYP_DEFINES="OS=ios target_arch=ia32 target_subarch=arm32" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_sim" ./gyp_libyuv
+    ninja -j7 -C out_sim/Debug-iphonesimulator libyuv_unittest
+    ninja -j7 -C out_sim/Release-iphonesimulator libyuv_unittest
+
+### Android
+https://code.google.com/p/chromium/wiki/AndroidBuildInstructions
+
+Add to .gclient last line: `target_os=['android'];`
+
+armv7
+
+    GYP_DEFINES="OS=android" GYP_CROSSCOMPILE=1 ./gyp_libyuv
+    ninja -j7 -C out/Debug yuv_unittest_apk
+    ninja -j7 -C out/Release yuv_unittest_apk
+
+arm64
+
+    GYP_DEFINES="OS=android target_arch=arm64 target_subarch=arm64" GYP_CROSSCOMPILE=1 ./gyp_libyuv
+    ninja -j7 -C out/Debug yuv_unittest_apk
+    ninja -j7 -C out/Release yuv_unittest_apk
+
+ia32
+
+    GYP_DEFINES="OS=android target_arch=ia32" GYP_CROSSCOMPILE=1 ./gyp_libyuv
+    ninja -j7 -C out/Debug yuv_unittest_apk
+    ninja -j7 -C out/Release yuv_unittest_apk
+
+    GYP_DEFINES="OS=android target_arch=ia32 android_full_debug=1" GYP_CROSSCOMPILE=1 ./gyp_libyuv
+    ninja -j7 -C out/Debug yuv_unittest_apk
+
+mipsel
+
+    GYP_DEFINES="OS=android target_arch=mipsel" GYP_CROSSCOMPILE=1 ./gyp_libyuv
+    ninja -j7 -C out/Debug yuv_unittest_apk
+    ninja -j7 -C out/Release yuv_unittest_apk
+
+arm32 disassembly:
+
+    third_party/android_ndk/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/bin/arm-linux-androideabi-objdump -d out/Release/obj/source/libyuv.row_neon.o
+
+arm64 disassembly:
+
+    third_party/android_ndk/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/aarch64-linux-android-objdump -d out/Release/obj/source/libyuv.row_neon64.o
+
+Running tests:
+
+    build/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=*
+
+Running test as benchmark:
+
+    build/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=* -a "--libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=-1"
+
+Running test with C code:
+
+    build/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=* -a "--libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=1 --libyuv_cpu_info=1"
+
+#### Building with GN
+
+    gn gen out/Release "--args=is_debug=false target_cpu=\"x86\""
+    gn gen out/Debug "--args=is_debug=true target_cpu=\"x86\""
+    ninja -C out/Release
+    ninja -C out/Debug
+
+### Building Offical with GN
+
+    gn gen out/Official "--args=is_debug=false is_official_build=true is_chrome_branded=true"
+    ninja -C out/Official
+
+#### Building mips with GN
+
+mipsel
+    gn gen out/Default "--args=is_debug=false target_cpu=\"mipsel\" target_os = \"android\" mips_arch_variant = \"r6\" mips_use_msa = true is_component_build = true is_clang = false"
+    ninja -C out/Default
+
+mips64el
+    gn gen out/Default "--args=is_debug=false target_cpu=\"mips64el\" target_os = \"android\" mips_arch_variant = \"r6\" mips_use_msa = true is_component_build = true is_clang = false"
+    ninja -C out/Default
+
+### Linux
+
+    GYP_DEFINES="target_arch=x64" ./gyp_libyuv
+    ninja -j7 -C out/Debug
+    ninja -j7 -C out/Release
+
+    GYP_DEFINES="target_arch=ia32" ./gyp_libyuv
+    ninja -j7 -C out/Debug
+    ninja -j7 -C out/Release
+
+#### CentOS
+
+On CentOS 32 bit the following work around allows a sync:
+
+    export GYP_DEFINES="host_arch=ia32"
+    gclient sync
+
+### Windows Shared Library
+
+Modify libyuv.gyp from 'static_library' to 'shared_library', and add 'LIBYUV_BUILDING_SHARED_LIBRARY' to 'defines'.
+
+    gclient runhooks
+
+After this command follow the building the library instructions above.
+
+If you get a compile error for atlthunk.lib on Windows, read http://www.chromium.org/developers/how-tos/build-instructions-windows
+
+
+### Build targets
+
+    ninja -C out/Debug libyuv
+    ninja -C out/Debug libyuv_unittest
+    ninja -C out/Debug compare
+    ninja -C out/Debug yuvconvert
+    ninja -C out/Debug yuvconstants
+    ninja -C out/Debug psnr
+    ninja -C out/Debug cpuid
+
+
+## Building the Library with make
+
+### Linux
+
+    make -j7 V=1 -f linux.mk
+    make -j7 V=1 -f linux.mk clean
+    make -j7 V=1 -f linux.mk CXX=clang++
+
+## Building the Library with cmake
+
+Install cmake: http://www.cmake.org/
+
+Default debug build:
+
+    mkdir out
+    cd out
+    cmake ..
+    cmake --build .
+
+Release build/install
+
+    mkdir out
+    cd out
+    cmake -DCMAKE_INSTALL_PREFIX="/usr/lib" -DCMAKE_BUILD_TYPE="Release" ..
+    cmake --build . --config Release
+    sudo cmake --build . --target install --config Release
+
+### Windows 8 Phone
+
+Pre-requisite:
+
+* Install Visual Studio 2012 and Arm to your environment.<br>
+
+Then:
+
+    call "c:\Program Files (x86)\Microsoft Visual Studio 11.0\VC\bin\x86_arm\vcvarsx86_arm.bat"
+
+or with Visual Studio 2013:
+
+    call "c:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\x86_arm\vcvarsx86_arm.bat"
+    nmake /f winarm.mk clean
+    nmake /f winarm.mk
+
+### Windows Shared Library
+
+Modify libyuv.gyp from 'static_library' to 'shared_library', and add 'LIBYUV_BUILDING_SHARED_LIBRARY' to 'defines'. Then run this.
+
+    gclient runhooks
+
+After this command follow the building the library instructions above.
+
+If you get a compile error for atlthunk.lib on Windows, read http://www.chromium.org/developers/how-tos/build-instructions-windows
+
+### 64 bit Windows
+
+    set GYP_DEFINES=target_arch=x64
+    gclient runhooks V=1
+
+### ARM Linux
+
+    export GYP_DEFINES="target_arch=arm"
+    export CROSSTOOL=`<path>`/arm-none-linux-gnueabi
+    export CXX=$CROSSTOOL-g++
+    export CC=$CROSSTOOL-gcc
+    export AR=$CROSSTOOL-ar
+    export AS=$CROSSTOOL-as
+    export RANLIB=$CROSSTOOL-ranlib
+    gclient runhooks
+
+## Running Unittests
+
+### Windows
+
+    out\Release\libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter="*"
+
+### OSX
+
+    out/Release/libyuv_unittest --gtest_filter="*"
+
+### Linux
+
+    out/Release/libyuv_unittest --gtest_filter="*"
+
+Replace --gtest_filter="*" with specific unittest to run.  May include wildcards. e.g.
+
+    out/Release/libyuv_unittest --gtest_filter=libyuvTest.I420ToARGB_Opt
+
+## CPU Emulator tools
+
+### Intel SDE (Software Development Emulator)
+
+Pre-requisite: Install IntelSDE for Windows: http://software.intel.com/en-us/articles/intel-software-development-emulator
+
+Then run:
+
+    c:\intelsde\sde -hsw -- out\release\libyuv_unittest.exe --gtest_filter=*
+
+
+## Memory tools
+
+### Running Dr Memory memcheck for Windows
+
+Pre-requisite: Install Dr Memory for Windows and add it to your path: http://www.drmemory.org/docs/page_install_windows.html
+
+    set GYP_DEFINES=build_for_tool=drmemory target_arch=ia32
+    call python gyp_libyuv -fninja -G msvs_version=2013
+    ninja -C out\Debug
+    drmemory out\Debug\libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter=*
+
+### Running UBSan
+
+See Chromium instructions for sanitizers: https://www.chromium.org/developers/testing/undefinedbehaviorsanitizer
+
+Sanitizers available: TSan, MSan, ASan, UBSan, LSan
+
+    GYP_DEFINES='ubsan=1' gclient runhooks
+    ninja -C out/Release
+
+### Running Valgrind memcheck
+
+Memory errors and race conditions can be found by running tests under special memory tools. [Valgrind] [1] is an instrumentation framework for building dynamic analysis tools. Various tests and profilers are built upon it to find memory handling errors and memory leaks, for instance.
+
+[1]: http://valgrind.org
+
+    solutions = [
+      { "name"        : "libyuv",
+        "url"         : "https://chromium.googlesource.com/libyuv/libyuv",
+        "deps_file"   : "DEPS",
+        "managed"     : True,
+        "custom_deps" : {
+           "libyuv/chromium/src/third_party/valgrind": "https://chromium.googlesource.com/chromium/deps/valgrind/binaries",
+        },
+        "safesync_url": "",
+      },
+    ]
+
+Then run:
+
+    GYP_DEFINES="clang=0 target_arch=x64 build_for_tool=memcheck" python gyp_libyuv
+    ninja -C out/Debug
+    valgrind out/Debug/libyuv_unittest
+
+
+For more information, see http://www.chromium.org/developers/how-tos/using-valgrind
+
+### Running Thread Sanitizer (TSan)
+
+    GYP_DEFINES="clang=0 target_arch=x64 build_for_tool=tsan" python gyp_libyuv
+    ninja -C out/Debug
+    valgrind out/Debug/libyuv_unittest
+
+For more info, see http://www.chromium.org/developers/how-tos/using-valgrind/threadsanitizer
+
+### Running Address Sanitizer (ASan)
+
+    GYP_DEFINES="clang=0 target_arch=x64 build_for_tool=asan" python gyp_libyuv
+    ninja -C out/Debug
+    valgrind out/Debug/libyuv_unittest
+
+For more info, see http://dev.chromium.org/developers/testing/addresssanitizer
+
+## Benchmarking
+
+The unittests can be used to benchmark.
+
+### Windows
+
+    set LIBYUV_WIDTH=1280
+    set LIBYUV_HEIGHT=720
+    set LIBYUV_REPEAT=999
+    set LIBYUV_FLAGS=-1
+    out\Release\libyuv_unittest.exe --gtest_filter=*I420ToARGB_Opt
+
+
+### Linux and Mac
+
+    LIBYUV_WIDTH=1280 LIBYUV_HEIGHT=720 LIBYUV_REPEAT=1000 out/Release/libyuv_unittest --gtest_filter=*I420ToARGB_Opt
+
+    libyuvTest.I420ToARGB_Opt (547 ms)
+
+Indicates 0.547 ms/frame for 1280 x 720.
+
+## Making a change
+
+    gclient sync
+    git checkout -b mycl -t origin/master
+    git pull
+    <edit files>
+    git add -u
+    git commit -m "my change"
+    git cl lint
+    git cl try
+    git cl upload -r a-reviewer@chomium.org -s
+    <once approved..>
+    git cl land
diff --git a/media/libyuv/docs/environment_variables.md b/media/libyuv/docs/environment_variables.md
index 5802599e9d..dd5d59fbef 100644
--- a/media/libyuv/docs/environment_variables.md
+++ b/media/libyuv/docs/environment_variables.md
@@ -6,7 +6,10 @@ For test purposes, environment variables can be set to control libyuv behavior.
 
 By default the cpu is detected and the most advanced form of SIMD is used.  But you can disable instruction sets selectively, or completely, falling back on C code.  Set the variable to 1 to disable the specified instruction set.
 
+## All CPUs
     LIBYUV_DISABLE_ASM
+
+## Intel CPUs
     LIBYUV_DISABLE_X86
     LIBYUV_DISABLE_SSE2
     LIBYUV_DISABLE_SSSE3
@@ -14,12 +17,29 @@ By default the cpu is detected and the most advanced form of SIMD is used.  But
     LIBYUV_DISABLE_SSE42
     LIBYUV_DISABLE_AVX
     LIBYUV_DISABLE_AVX2
-    LIBYUV_DISABLE_AVX3
     LIBYUV_DISABLE_ERMS
     LIBYUV_DISABLE_FMA3
-    LIBYUV_DISABLE_DSPR2
+    LIBYUV_DISABLE_F16C
+    LIBYUV_DISABLE_AVX512BW
+    LIBYUV_DISABLE_AVX512VL
+    LIBYUV_DISABLE_AVX512VNNI
+    LIBYUV_DISABLE_AVX512VBMI
+    LIBYUV_DISABLE_AVX512VBMI2
+    LIBYUV_DISABLE_AVX512VBITALG
+    LIBYUV_DISABLE_AVX512VPOPCNTDQ
+    LIBYUV_DISABLE_GFNI
+
+## ARM CPUs
+
     LIBYUV_DISABLE_NEON
 
+## MIPS CPUs
+    LIBYUV_DISABLE_MSA
+
+## LOONGARCH CPUs
+    LIBYUV_DISABLE_LSX
+    LIBYUV_DISABLE_LASX
+
 # Test Width/Height/Repeat
 
 The unittests default to a small image (128x72) to run fast.  This can be set by environment variable to test a specific resolutions.
diff --git a/media/libyuv/docs/formats.md b/media/libyuv/docs/formats.md
index a7cfed8218..12ea9465e4 100644
--- a/media/libyuv/docs/formats.md
+++ b/media/libyuv/docs/formats.md
@@ -4,7 +4,9 @@ Formats (FOURCC) supported by libyuv are detailed here.
 
 # Core Formats
 
-There are 2 core formats supported by libyuv - I420 and ARGB.  All YUV formats can be converted to/from I420.  All RGB formats can be converted to/from ARGB.
+There are 2 core formats supported by libyuv - I420 and ARGB.
+  All YUV formats can be converted to/from I420.
+  All RGB formats can be converted to/from ARGB.
 
 Filtering functions such as scaling and planar functions work on I420 and/or ARGB.
 
@@ -35,50 +37,53 @@ This is how OSX formats map to libyuv
 # FOURCC (Four Charactacter Code) List
 
 The following is extracted from video_common.h as a complete list of formats supported by libyuv.
-
     enum FourCC {
-      // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+      // 10 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
       FOURCC_I420 = FOURCC('I', '4', '2', '0'),
       FOURCC_I422 = FOURCC('I', '4', '2', '2'),
       FOURCC_I444 = FOURCC('I', '4', '4', '4'),
-      FOURCC_I411 = FOURCC('I', '4', '1', '1'),
       FOURCC_I400 = FOURCC('I', '4', '0', '0'),
       FOURCC_NV21 = FOURCC('N', 'V', '2', '1'),
       FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
       FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
       FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
+      FOURCC_H010 = FOURCC('H', '0', '1', '0'),  // unofficial fourcc. 10 bit lsb
+      FOURCC_U010 = FOURCC('U', '0', '1', '0'),  // bt.2020, unofficial fourcc.
+                                                 // 10 bit lsb
 
-      // 2 Secondary YUV formats: row biplanar.
-      FOURCC_M420 = FOURCC('M', '4', '2', '0'),
-      FOURCC_Q420 = FOURCC('Q', '4', '2', '0'),
+      // 1 Secondary YUV format: row biplanar.
+      FOURCC_M420 = FOURCC('M', '4', '2', '0'),  // deprecated.
 
-      // 9 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp.
+      // 13 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp, 1 10 bpc, 2 64 bpp
       FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
       FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
       FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
+      FOURCC_AR30 = FOURCC('A', 'R', '3', '0'),  // 10 bit per channel. 2101010.
+      FOURCC_AB30 = FOURCC('A', 'B', '3', '0'),  // ABGR version of 10 bit
+      FOURCC_AR64 = FOURCC('A', 'R', '6', '4'),  // 16 bit per channel.
+      FOURCC_AB64 = FOURCC('A', 'B', '6', '4'),  // ABGR version of 16 bit
       FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
-      FOURCC_RAW  = FOURCC('r', 'a', 'w', ' '),
+      FOURCC_RAW = FOURCC('r', 'a', 'w', ' '),
       FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
       FOURCC_RGBP = FOURCC('R', 'G', 'B', 'P'),  // rgb565 LE.
       FOURCC_RGBO = FOURCC('R', 'G', 'B', 'O'),  // argb1555 LE.
       FOURCC_R444 = FOURCC('R', '4', '4', '4'),  // argb4444 LE.
 
-      // 4 Secondary RGB formats: 4 Bayer Patterns.
-      FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
-      FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
-      FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
-      FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
-
       // 1 Primary Compressed YUV format.
       FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
 
-      // 5 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+      // 11 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
       FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
       FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
       FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
       FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
       FOURCC_J420 = FOURCC('J', '4', '2', '0'),
-      FOURCC_J400 = FOURCC('J', '4', '0', '0'),
+      FOURCC_J400 = FOURCC('J', '4', '0', '0'),  // unofficial fourcc
+      FOURCC_H420 = FOURCC('H', '4', '2', '0'),  // unofficial fourcc
+      FOURCC_H422 = FOURCC('H', '4', '2', '2'),  // unofficial fourcc
+      FOURCC_U420 = FOURCC('U', '4', '2', '0'),  // bt.2020, unofficial fourcc
+      FOURCC_U422 = FOURCC('U', '4', '2', '2'),  // bt.2020, unofficial fourcc
+      FOURCC_U444 = FOURCC('U', '4', '4', '4'),  // bt.2020, unofficial fourcc
 
       // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
       FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
@@ -99,8 +104,35 @@ The following is extracted from video_common.h as a complete list of formats sup
       FOURCC_L565 = FOURCC('L', '5', '6', '5'),  // Alias for RGBP.
       FOURCC_5551 = FOURCC('5', '5', '5', '1'),  // Alias for RGBO.
 
-      // 1 Auxiliary compressed YUV format set aside for capturer.
-      FOURCC_H264 = FOURCC('H', '2', '6', '4'),
+# Planar YUV
+      The following formats contains a full size Y plane followed by 1 or 2
+        planes for UV: I420, I422, I444, I400, NV21, NV12, I400
+      The size (subsampling) of the UV varies.
+        I420, NV12 and NV21 are half width, half height
+        I422, NV16 and NV61 are half width, full height
+        I444, NV24 and NV42 are full width, full height
+        I400 and J400 have no chroma channel.
+
+# Color space
+      The YUV formats start with a letter to specify the color space. e.g. I420
+        I = BT.601 limited range
+        J = BT.601 full range     (J = JPEG that uses this)
+        H = BT.709 limited range  (H for HD)
+        F = BT.709 full range     (F for Full range)
+        U = BT.2020 limited range (U for UHD)
+        V = BT.2020 full range
+        For YUV to RGB conversions, a matrix can be passed.  See also convert_argh.h
+
+# HDR formats
+      Planar formats with 10 or 12 bits use the following fourcc:
+        I010, I012, P010, P012 are half width, half height
+        I210, I212, P210, P212 are half width, full height
+        I410, I412, P410, P412 are full width, full height
+      where
+        I is the color space (see above) and 3 planes: Y, U and V.
+        P is a biplanar format, similar to NV12 but 16 bits, with the valid bits in the high bits.  There is a Y plane and a UV plane.
+        0, 2 or 4 is the last digit of subsampling: 4:2:0, 4:2:2, or 4:4:4
+        10 or 12 is the bits per channel.  The bits are in the low bits of a 16 bit channel.
 
 # The ARGB FOURCC
 
@@ -131,3 +163,46 @@ Some are channel order agnostic (e.g. ARGBScale).
 Some functions are symmetric (e.g. ARGBToBGRA is the same as BGRAToARGB, so its a macro).
 
 ARGBBlend expects preattenuated ARGB. The R,G,B are premultiplied by alpha.  Other functions don't care.
+
+# RGB24 and RAW
+
+There are 2 RGB layouts - RGB24 (aka 24BG) and RAW
+
+RGB24 is B,G,R in memory
+RAW is R,G,B in memory
+
+# AR30 and XR30
+
+AR30 is 2 10 10 10 ARGB stored in little endian order.
+The 2 bit alpha has 4 values.  Here are the comparable 8 bit alpha values.
+0 - 0.    00000000b = 0x00 = 0
+1 - 33%.  01010101b = 0x55 = 85
+2 - 66%.  10101010b = 0xaa = 170
+3 - 100%. 11111111b = 0xff = 255
+The 10 bit RGB values range from 0 to 1023.
+XR30 is the same as AR30 but with no alpha channel.
+
+# AB64 and AR64
+
+AB64 is similar to ABGR, with 16 bit (2 bytes) per channel. Each channel stores an unsigned short.
+In memory R is the lowest and A is the highest.
+Each channel has value ranges from 0 to 65535.
+AR64 is similar to ARGB.
+
+# NV12 and NV21
+
+NV12 is a biplanar format with a full sized Y plane followed by a single
+chroma plane with weaved U and V values.
+NV21 is the same but with weaved V and U values.
+The 12 in NV12 refers to 12 bits per pixel.  NV12 has a half width and half
+height chroma channel, and therefore is a 420 subsampling.
+NV16 is 16 bits per pixel, with half width and full height.  aka 422.
+NV24 is 24 bits per pixel with full sized chroma channel. aka 444.
+Most NV12 functions allow the destination Y pointer to be NULL.
+
+# YUY2 and UYVY
+
+YUY2 is a packed YUV format with half width, full height.
+
+YUY2 is YUYV in memory
+UYVY is UYVY in memory
diff --git a/media/libyuv/docs/getting_started.md b/media/libyuv/docs/getting_started.md
index 7cd56167f1..15b19ab210 100644
--- a/media/libyuv/docs/getting_started.md
+++ b/media/libyuv/docs/getting_started.md
@@ -11,14 +11,13 @@ Refer to chromium instructions for each platform for other prerequisites.
 
 Create a working directory, enter it, and run:
 
-    gclient config https://chromium.googlesource.com/libyuv/libyuv
+    gclient config --name src https://chromium.googlesource.com/libyuv/libyuv
     gclient sync
 
-
 Then you'll get a .gclient file like:
 
     solutions = [
-      { "name"        : "libyuv",
+      { "name"        : "src",
         "url"         : "https://chromium.googlesource.com/libyuv/libyuv",
         "deps_file"   : "DEPS",
         "managed"     : True,
@@ -28,17 +27,15 @@ Then you'll get a .gclient file like:
       },
     ];
 
-
-For iOS add `;target_os=['ios'];` to your OSX .gclient and run `GYP_DEFINES="OS=ios" gclient sync.`
+For iOS add `;target_os=['ios'];` to your OSX .gclient and run `gclient sync.`
 
 Browse the Git reprository: https://chromium.googlesource.com/libyuv/libyuv/+/master
 
 ### Android
 For Android add `;target_os=['android'];` to your Linux .gclient
 
-
     solutions = [
-      { "name"        : "libyuv",
+      { "name"        : "src",
         "url"         : "https://chromium.googlesource.com/libyuv/libyuv",
         "deps_file"   : "DEPS",
         "managed"     : True,
@@ -47,23 +44,14 @@ For Android add `;target_os=['android'];` to your Linux .gclient
         "safesync_url": "",
       },
     ];
-    target_os = ["android", "unix"];
+    target_os = ["android", "linux"];
 
 Then run:
 
-    export GYP_DEFINES="OS=android"
     gclient sync
 
-Caveat: Theres an error with Google Play services updates.  If you get the error "Your version of the Google Play services library is not up to date", run the following:
-    cd chromium/src
-    ./build/android/play_services/update.py download
-    cd ../..
-
-For Windows the gclient sync must be done from an Administrator command prompt.
-
-The sync will generate native build files for your environment using gyp (Windows: Visual Studio, OSX: XCode, Linux: make). This generation can also be forced manually: `gclient runhooks`
-
 To get just the source (not buildable):
+
     git clone https://chromium.googlesource.com/libyuv/libyuv
 
 
@@ -71,187 +59,152 @@ To get just the source (not buildable):
 
 ### Windows
 
-    set GYP_DEFINES=target_arch=ia32
-    call python gyp_libyuv -fninja -G msvs_version=2013
-    ninja -j7 -C out\Release
-    ninja -j7 -C out\Debug
+    call gn gen out\Release "--args=is_debug=false target_cpu=\"x64\""
+    call gn gen out\Debug "--args=is_debug=true target_cpu=\"x64\""
+    ninja -v -C out\Release
+    ninja -v -C out\Debug
 
-    set GYP_DEFINES=target_arch=x64
-    call python gyp_libyuv -fninja -G msvs_version=2013
-    ninja -C out\Debug_x64
-    ninja -C out\Release_x64
+    call gn gen out\Release "--args=is_debug=false target_cpu=\"x86\""
+    call gn gen out\Debug "--args=is_debug=true target_cpu=\"x86\""
+    ninja -v -C out\Release
+    ninja -v -C out\Debug
 
-#### Building with clangcl
-    set GYP_DEFINES=clang=1 target_arch=ia32 libyuv_enable_svn=1
-    set LLVM_REPO_URL=svn://svn.chromium.org/llvm-project
-    call python tools\clang\scripts\update.py
-    call python gyp_libyuv -fninja libyuv_test.gyp
-    ninja -C out\Debug
-    ninja -C out\Release
+### macOS and Linux
 
-### OSX
+    gn gen out/Release "--args=is_debug=false"
+    gn gen out/Debug "--args=is_debug=true"
+    ninja -v -C out/Release
+    ninja -v -C out/Debug
 
-Clang 64 bit shown. Remove `clang=1` for GCC and change x64 to ia32 for 32 bit.
-
-    GYP_DEFINES="clang=1 target_arch=x64" ./gyp_libyuv
-    ninja -j7 -C out/Debug
-    ninja -j7 -C out/Release
+### Building Offical with GN
 
-    GYP_DEFINES="clang=1 target_arch=ia32" ./gyp_libyuv
-    ninja -j7 -C out/Debug
-    ninja -j7 -C out/Release
+    gn gen out/Official "--args=is_debug=false is_official_build=true is_chrome_branded=true"
+    ninja -C out/Official
 
 ### iOS
 http://www.chromium.org/developers/how-tos/build-instructions-ios
 
 Add to .gclient last line: `target_os=['ios'];`
 
-armv7
-
-    GYP_DEFINES="OS=ios target_arch=armv7 target_subarch=arm32" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_ios" ./gyp_libyuv
-    ninja -j7 -C out_ios/Debug-iphoneos libyuv_unittest
-    ninja -j7 -C out_ios/Release-iphoneos libyuv_unittest
-
 arm64
 
-    GYP_DEFINES="OS=ios target_arch=arm64 target_subarch=arm64" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_ios" ./gyp_libyuv
-    ninja -j7 -C out_ios/Debug-iphoneos libyuv_unittest
-    ninja -j7 -C out_ios/Release-iphoneos libyuv_unittest
+    gn gen out/Release "--args=is_debug=false target_os=\"ios\" ios_enable_code_signing=false target_cpu=\"arm64\""
+    gn gen out/Debug "--args=is_debug=true target_os=\"ios\" ios_enable_code_signing=false target_cpu=\"arm64\""
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
 
-both armv7 and arm64 (fat)
+ios simulator
 
-    GYP_DEFINES="OS=ios target_arch=armv7 target_subarch=both" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_ios" ./gyp_libyuv
-    ninja -j7 -C out_ios/Debug-iphoneos libyuv_unittest
-    ninja -j7 -C out_ios/Release-iphoneos libyuv_unittest
+    gn gen out/Release "--args=is_debug=false target_os=\"ios\" ios_enable_code_signing=false use_xcode_clang=true target_cpu=\"x86\""
+    gn gen out/Debug "--args=is_debug=true target_os=\"ios\" ios_enable_code_signing=false use_xcode_clang=true target_cpu=\"x86\""
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
 
-simulator
+ios disassembly
 
-    GYP_DEFINES="OS=ios target_arch=ia32 target_subarch=arm32" GYP_CROSSCOMPILE=1 GYP_GENERATOR_FLAGS="output_dir=out_sim" ./gyp_libyuv
-    ninja -j7 -C out_sim/Debug-iphonesimulator libyuv_unittest
-    ninja -j7 -C out_sim/Release-iphonesimulator libyuv_unittest
+    otool -tV ./out/Release/obj/libyuv_neon/row_neon64.o >row_neon64.txt
 
 ### Android
 https://code.google.com/p/chromium/wiki/AndroidBuildInstructions
 
 Add to .gclient last line: `target_os=['android'];`
 
-armv7
+arm64
 
-    GYP_DEFINES="OS=android" GYP_CROSSCOMPILE=1 ./gyp_libyuv
-    ninja -j7 -C out/Debug libyuv_unittest_apk
-    ninja -j7 -C out/Release libyuv_unittest_apk
+    gn gen out/Release "--args=is_debug=false target_os=\"android\" target_cpu=\"arm64\""
+    gn gen out/Debug "--args=is_debug=true target_os=\"android\" target_cpu=\"arm64\""
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
 
-arm64
+armv7
 
-    GYP_DEFINES="OS=android target_arch=arm64 target_subarch=arm64" GYP_CROSSCOMPILE=1 ./gyp_libyuv
-    ninja -j7 -C out/Debug libyuv_unittest_apk
-    ninja -j7 -C out/Release libyuv_unittest_apk
+    gn gen out/Release "--args=is_debug=false target_os=\"android\" target_cpu=\"arm\""
+    gn gen out/Debug "--args=is_debug=true target_os=\"android\" target_cpu=\"arm\""
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
 
 ia32
 
-    GYP_DEFINES="OS=android target_arch=ia32" GYP_CROSSCOMPILE=1 ./gyp_libyuv
-    ninja -j7 -C out/Debug libyuv_unittest_apk
-    ninja -j7 -C out/Release libyuv_unittest_apk
+    gn gen out/Release "--args=is_debug=false target_os=\"android\" target_cpu=\"x86\""
+    gn gen out/Debug "--args=is_debug=true target_os=\"android\" target_cpu=\"x86\""
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
 
-    GYP_DEFINES="OS=android target_arch=ia32 android_full_debug=1" GYP_CROSSCOMPILE=1 ./gyp_libyuv
-    ninja -j7 -C out/Debug libyuv_unittest_apk
+mips
 
-mipsel
+    gn gen out/Release "--args=is_debug=false target_os=\"android\" target_cpu=\"mips64el\" mips_arch_variant=\"r6\" mips_use_msa=true is_component_build=true"
+    gn gen out/Debug "--args=is_debug=true target_os=\"android\" target_cpu=\"mips64el\" mips_arch_variant=\"r6\" mips_use_msa=true is_component_build=true"
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
 
-    GYP_DEFINES="OS=android target_arch=mipsel" GYP_CROSSCOMPILE=1 ./gyp_libyuv
-    ninja -j7 -C out/Debug libyuv_unittest_apk
-    ninja -j7 -C out/Release libyuv_unittest_apk
+arm disassembly:
 
-arm32 disassembly:
+    third_party/android_ndk/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/aarch64-linux-android-objdump -d ./out/Release/obj/libyuv/row_common.o >row_common.txt
 
-    third_party/android_tools/ndk/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/bin/arm-linux-androideabi-objdump -d out/Release/obj/source/libyuv.row_neon.o
+    third_party/android_ndk/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/aarch64-linux-android-objdump -d ./out/Release/obj/libyuv_neon/row_neon.o >row_neon.txt
 
-arm64 disassembly:
+    third_party/android_ndk/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/aarch64-linux-android-objdump -d ./out/Release/obj/libyuv_neon/row_neon64.o >row_neon64.txt
 
-    third_party/android_tools/ndk/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/aarch64-linux-android-objdump -d out/Release/obj/source/libyuv.row_neon64.o
+    Caveat: Disassembly may require optimize_max be disabled in BUILD.gn
 
 Running tests:
 
-    util/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=*
+    out/Release/bin/run_libyuv_unittest -vv --gtest_filter=*
 
 Running test as benchmark:
 
-    util/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=* -a "--libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=-1"
+    out/Release/bin/run_libyuv_unittest -vv --gtest_filter=* --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=-1  --libyuv_cpu_info=-1
 
 Running test with C code:
 
-    util/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=* -a "--libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=1 --libyuv_cpu_info=1"
-
-#### Building with GN
-
-    gn gen out/Release "--args=is_debug=false target_cpu=\"x86\""
-    gn gen out/Debug "--args=is_debug=true target_cpu=\"x86\""
-    ninja -C out/Release
-    ninja -C out/Debug
-
-### Building Offical with GN
-
-    gn gen out/Official "--args=is_debug=false is_official_build=true is_chrome_branded=true"
-    ninja -C out/Official
-
-### Linux
-
-    GYP_DEFINES="target_arch=x64" ./gyp_libyuv
-    ninja -j7 -C out/Debug
-    ninja -j7 -C out/Release
-
-    GYP_DEFINES="target_arch=ia32" ./gyp_libyuv
-    ninja -j7 -C out/Debug
-    ninja -j7 -C out/Release
-
-#### CentOS
-
-On CentOS 32 bit the following work around allows a sync:
-
-    export GYP_DEFINES="host_arch=ia32"
-    gclient sync
-
-### Windows Shared Library
-
-Modify libyuv.gyp from 'static_library' to 'shared_library', and add 'LIBYUV_BUILDING_SHARED_LIBRARY' to 'defines'.
-
-    gclient runhooks
-
-After this command follow the building the library instructions above.
-
-If you get a compile error for atlthunk.lib on Windows, read http://www.chromium.org/developers/how-tos/build-instructions-windows
-
+    out/Release/bin/run_libyuv_unittest -vv --gtest_filter=* --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=1 --libyuv_cpu_info=1
 
 ### Build targets
 
     ninja -C out/Debug libyuv
     ninja -C out/Debug libyuv_unittest
     ninja -C out/Debug compare
-    ninja -C out/Debug convert
+    ninja -C out/Debug yuvconvert
+    ninja -C out/Debug yuvconstants
     ninja -C out/Debug psnr
     ninja -C out/Debug cpuid
 
+### ARM Linux
+
+    gn gen out/Release "--args=is_debug=false target_cpu=\"arm64\""
+    gn gen out/Debug "--args=is_debug=true target_cpu=\"arm64\""
+    ninja -v -C out/Debug libyuv_unittest
+    ninja -v -C out/Release libyuv_unittest
+
+### MIPS Linux
+
+mips
+
+   gn gen out/Release "--args=is_debug=false target_os=\"linux\" target_cpu=\"mips64el\" mips_arch_variant=\"loongson3\" is_component_build=false use_sysroot=false use_gold=false"
+   gn gen out/Debug "--args=is_debug=true target_os=\"linux\" target_cpu=\"mips64el\" mips_arch_variant=\"loongson3\" is_component_build=false use_sysroot=false use_gold=false"
+   ninja -v -C out/Debug libyuv_unittest
+   ninja -v -C out/Release libyuv_unittest
 
 ## Building the Library with make
 
 ### Linux
 
-    make -j7 V=1 -f linux.mk
-    make -j7 V=1 -f linux.mk clean
-    make -j7 V=1 -f linux.mk CXX=clang++
+    make V=1 -f linux.mk
+    make V=1 -f linux.mk clean
+    make V=1 -f linux.mk CXX=clang++ CC=clang
 
-## Building the Library with cmake
+## Building the library with cmake
 
 Install cmake: http://www.cmake.org/
 
-Default debug build:
+### Default debug build:
 
     mkdir out
     cd out
     cmake ..
     cmake --build .
 
-Release build/install
+### Release build/install
 
     mkdir out
     cd out
@@ -259,47 +212,31 @@ Release build/install
     cmake --build . --config Release
     sudo cmake --build . --target install --config Release
 
-### Windows 8 Phone
-
-Pre-requisite:
-
-* Install Visual Studio 2012 and Arm to your environment.<br>
-
-Then:
-
-    call "c:\Program Files (x86)\Microsoft Visual Studio 11.0\VC\bin\x86_arm\vcvarsx86_arm.bat"
-
-or with Visual Studio 2013:
-
-    call "c:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\x86_arm\vcvarsx86_arm.bat"
-    nmake /f winarm.mk clean
-    nmake /f winarm.mk
-
-### Windows Shared Library
-
-Modify libyuv.gyp from 'static_library' to 'shared_library', and add 'LIBYUV_BUILDING_SHARED_LIBRARY' to 'defines'. Then run this.
+### Build RPM/DEB packages
 
-    gclient runhooks
-
-After this command follow the building the library instructions above.
+    mkdir out
+    cd out
+    cmake -DCMAKE_BUILD_TYPE=Release ..
+    make -j4
+    make package
 
-If you get a compile error for atlthunk.lib on Windows, read http://www.chromium.org/developers/how-tos/build-instructions-windows
+## Setup for Arm Cross compile
 
-### 64 bit Windows
+See also https://www.ccoderun.ca/programming/2015-12-20_CrossCompiling/index.html
 
-    set GYP_DEFINES=target_arch=x64
-    gclient runhooks V=1
+    sudo apt-get install ssh dkms build-essential linux-headers-generic
+    sudo apt-get install kdevelop cmake git subversion
+    sudo apt-get install graphviz doxygen doxygen-gui
+    sudo apt-get install manpages manpages-dev manpages-posix manpages-posix-dev
+    sudo apt-get install libboost-all-dev libboost-dev libssl-dev
+    sudo apt-get install rpm terminator fish
+    sudo apt-get install g++-arm-linux-gnueabihf gcc-arm-linux-gnueabihf
 
-### ARM Linux
+### Build psnr tool
 
-    export GYP_DEFINES="target_arch=arm"
-    export CROSSTOOL=`<path>`/arm-none-linux-gnueabi
-    export CXX=$CROSSTOOL-g++
-    export CC=$CROSSTOOL-gcc
-    export AR=$CROSSTOOL-ar
-    export AS=$CROSSTOOL-as
-    export RANLIB=$CROSSTOOL-ranlib
-    gclient runhooks
+    cd util
+    arm-linux-gnueabihf-g++ psnr_main.cc psnr.cc ssim.cc -o psnr
+    arm-linux-gnueabihf-objdump -d psnr
 
 ## Running Unittests
 
@@ -307,123 +244,42 @@ If you get a compile error for atlthunk.lib on Windows, read http://www.chromium
 
     out\Release\libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter="*"
 
-### OSX
+### macOS and Linux
 
     out/Release/libyuv_unittest --gtest_filter="*"
 
-### Linux
-
-    out/Release/libyuv_unittest --gtest_filter="*"
-
-Replace --gtest_filter="*" with specific unittest to run.  May include wildcards. e.g.
-
-    out/Release/libyuv_unittest --gtest_filter=libyuvTest.I420ToARGB_Opt
+Replace --gtest_filter="*" with specific unittest to run.  May include wildcards.
+    out/Release/libyuv_unittest --gtest_filter=*I420ToARGB_Opt
 
 ## CPU Emulator tools
 
 ### Intel SDE (Software Development Emulator)
 
-Pre-requisite: Install IntelSDE for Windows: http://software.intel.com/en-us/articles/intel-software-development-emulator
+Pre-requisite: Install IntelSDE: http://software.intel.com/en-us/articles/intel-software-development-emulator
 
 Then run:
 
-    c:\intelsde\sde -hsw -- out\release\libyuv_unittest.exe --gtest_filter=*
-
-
-## Memory tools
-
-### Running Dr Memory memcheck for Windows
-
-Pre-requisite: Install Dr Memory for Windows and add it to your path: http://www.drmemory.org/docs/page_install_windows.html
-
-    set GYP_DEFINES=build_for_tool=drmemory target_arch=ia32
-    call python gyp_libyuv -fninja -G msvs_version=2013
-    ninja -C out\Debug
-    drmemory out\Debug\libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter=*
+    c:\intelsde\sde -hsw -- out\Release\libyuv_unittest.exe --gtest_filter=*
 
-### Running UBSan
+    ~/intelsde/sde -skx -- out/Release/libyuv_unittest --gtest_filter=**I420ToARGB_Opt
 
-See Chromium instructions for sanitizers: https://www.chromium.org/developers/testing/undefinedbehaviorsanitizer
+### Intel Architecture Code Analyzer
 
-Sanitizers available: TSan, MSan, ASan, UBSan, LSan
+Inset these 2 macros into assembly code to be analyzed:
+    IACA_ASM_START
+    IACA_ASM_END
+Build the code as usual, then run iaca on the object file.
+    ~/iaca-lin64/bin/iaca.sh -reduceout -arch HSW out/Release/obj/libyuv_internal/compare_gcc.o
 
-    GYP_DEFINES='ubsan=1' gclient runhooks
-    ninja -C out/Release
+## Sanitizers
 
-### Running Valgrind memcheck
+    gn gen out/Release "--args=is_debug=false is_msan=true"
+    ninja -v -C out/Release
 
-Memory errors and race conditions can be found by running tests under special memory tools. [Valgrind] [1] is an instrumentation framework for building dynamic analysis tools. Various tests and profilers are built upon it to find memory handling errors and memory leaks, for instance.
+Sanitizers available: asan, msan, tsan, ubsan, lsan, ubsan_vptr
 
-[1]: http://valgrind.org
-
-    solutions = [
-      { "name"        : "libyuv",
-        "url"         : "https://chromium.googlesource.com/libyuv/libyuv",
-        "deps_file"   : "DEPS",
-        "managed"     : True,
-        "custom_deps" : {
-           "libyuv/chromium/src/third_party/valgrind": "https://chromium.googlesource.com/chromium/deps/valgrind/binaries",
-        },
-        "safesync_url": "",
-      },
-    ]
-
-Then run:
-
-    GYP_DEFINES="clang=0 target_arch=x64 build_for_tool=memcheck" python gyp_libyuv
-    ninja -C out/Debug
-    valgrind out/Debug/libyuv_unittest
-
-
-For more information, see http://www.chromium.org/developers/how-tos/using-valgrind
-
-### Running Thread Sanitizer (TSan)
-
-    GYP_DEFINES="clang=0 target_arch=x64 build_for_tool=tsan" python gyp_libyuv
-    ninja -C out/Debug
-    valgrind out/Debug/libyuv_unittest
-
-For more info, see http://www.chromium.org/developers/how-tos/using-valgrind/threadsanitizer
-
-### Running Address Sanitizer (ASan)
-
-    GYP_DEFINES="clang=0 target_arch=x64 build_for_tool=asan" python gyp_libyuv
-    ninja -C out/Debug
-    valgrind out/Debug/libyuv_unittest
-
-For more info, see http://dev.chromium.org/developers/testing/addresssanitizer
-
-## Benchmarking
-
-The unittests can be used to benchmark.
-
-### Windows
-
-    set LIBYUV_WIDTH=1280
-    set LIBYUV_HEIGHT=720
-    set LIBYUV_REPEAT=999
-    set LIBYUV_FLAGS=-1
-    out\Release\libyuv_unittest.exe --gtest_filter=*I420ToARGB_Opt
-
-
-### Linux and Mac
-
-    LIBYUV_WIDTH=1280 LIBYUV_HEIGHT=720 LIBYUV_REPEAT=1000 out/Release/libyuv_unittest --gtest_filter=*I420ToARGB_Opt
-
-    libyuvTest.I420ToARGB_Opt (547 ms)
-
-Indicates 0.547 ms/frame for 1280 x 720.
+### Running Dr Memory memcheck for Windows
 
-## Making a change
+Pre-requisite: Install Dr Memory for Windows and add it to your path: http://www.drmemory.org/docs/page_install_windows.html
 
-    gclient sync
-    git checkout -b mycl -t origin/master
-    git pull
-    <edit files>
-    git add -u
-    git commit -m "my change"
-    git cl lint
-    git cl try
-    git cl upload -r a-reviewer@chomium.org -s
-    <once approved..>
-    git cl land
+    drmemory out\Debug\libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter=*
diff --git a/media/libyuv/docs/rotation.md b/media/libyuv/docs/rotation.md
index fb84fce5a9..a08430fded 100644
--- a/media/libyuv/docs/rotation.md
+++ b/media/libyuv/docs/rotation.md
@@ -100,4 +100,8 @@ Inverting can be achieved with almost any libyuv function by passing a negative
 
 I420Mirror and ARGBMirror can also be used to rotate by 180 degrees by passing a negative height.
 
+# Cropping - Vertical Flip
 
+When cropping from a subsampled format like NV21, the method of setting the start pointers wont work for odd crop start y on the UV plane.
+If the height after cropping will be odd, invert the source - point to the last row, negate the strides, and pass negative height, which
+will re-invert the image as the conversion outputs.
diff --git a/media/libyuv/download_vs_toolchain.py b/media/libyuv/download_vs_toolchain.py
index 4b3457899f..6bc086d68f 100644
--- a/media/libyuv/download_vs_toolchain.py
+++ b/media/libyuv/download_vs_toolchain.py
@@ -1,5 +1,5 @@
-#!/usr/bin/env python
-#
+#!/usr/bin/env vpython3
+
 # Copyright 2014 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
@@ -22,7 +22,7 @@ sys.path.insert(0, os.path.join(checkout_root, 'build'))
 sys.path.insert(0, os.path.join(checkout_root, 'tools', 'find_depot_tools'))
 
 
-import vs_toolchain
+import vs_toolchain  # pylint: disable=wrong-import-position
 
 
 if __name__ == '__main__':
diff --git a/media/libyuv/gyp_libyuv b/media/libyuv/gyp_libyuv
deleted file mode 100644
index bce1061a17..0000000000
--- a/media/libyuv/gyp_libyuv
+++ /dev/null
@@ -1,101 +0,0 @@
-#!/usr/bin/env python
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# This script is used to run GYP for libyuv. It contains selected parts of the
-# main function from the src/build/gyp_chromium file.
-
-import glob
-import os
-import shlex
-import sys
-
-checkout_root = os.path.dirname(os.path.realpath(__file__))
-
-sys.path.insert(0, os.path.join(checkout_root, 'build'))
-import gyp_chromium
-import gyp_helper
-import vs_toolchain
-
-sys.path.insert(0, os.path.join(checkout_root, 'tools', 'gyp', 'pylib'))
-import gyp
-
-def GetSupplementalFiles():
-  """Returns a list of the supplemental files that are included in all GYP
-  sources."""
-  # Can't use the one in gyp_chromium since the directory location of the root
-  # is different.
-  return glob.glob(os.path.join(checkout_root, '*', 'supplement.gypi'))
-
-
-if __name__ == '__main__':
-  args = sys.argv[1:]
-
-  if int(os.environ.get('GYP_CHROMIUM_NO_ACTION', 0)):
-    print 'Skipping gyp_libyuv due to GYP_CHROMIUM_NO_ACTION env var.'
-    sys.exit(0)
-
-  # This could give false positives since it doesn't actually do real option
-  # parsing.  Oh well.
-  gyp_file_specified = False
-  for arg in args:
-    if arg.endswith('.gyp'):
-      gyp_file_specified = True
-      break
-
-  # If we didn't get a file, assume 'all.gyp' in the root of the checkout.
-  if not gyp_file_specified:
-    # Because of a bug in gyp, simply adding the abspath to all.gyp doesn't
-    # work, but chdir'ing and adding the relative path does. Spooky :/
-    os.chdir(checkout_root)
-    args.append('all.gyp')
-
-  # There shouldn't be a circular dependency relationship between .gyp files,
-  args.append('--no-circular-check')
-
-  # Default to ninja unless GYP_GENERATORS is set.
-  if not os.environ.get('GYP_GENERATORS'):
-    os.environ['GYP_GENERATORS'] = 'ninja'
-
-  vs2013_runtime_dll_dirs = None
-  if int(os.environ.get('DEPOT_TOOLS_WIN_TOOLCHAIN', '1')):
-    vs2013_runtime_dll_dirs = vs_toolchain.SetEnvironmentAndGetRuntimeDllDirs()
-
-  # Enforce gyp syntax checking. This adds about 20% execution time.
-  args.append('--check')
-
-  supplemental_includes = gyp_chromium.GetSupplementalFiles()
-  gyp_vars_dict = gyp_chromium.GetGypVars(supplemental_includes)
-
-  # Automatically turn on crosscompile support for platforms that need it.
-  if all(('ninja' in os.environ.get('GYP_GENERATORS', ''),
-          gyp_vars_dict.get('OS') in ['ios'],
-          'GYP_CROSSCOMPILE' not in os.environ)):
-    os.environ['GYP_CROSSCOMPILE'] = '1'
-
-  args.extend(['-I' + i for i in
-               gyp_chromium.additional_include_files(supplemental_includes,
-                                                     args)])
-
-  # Set the gyp depth variable to the root of the checkout.
-  args.append('--depth=' + os.path.relpath(checkout_root))
-
-  print 'Updating projects from gyp files...'
-  sys.stdout.flush()
-
-  # Off we go...
-  gyp_rc = gyp.main(args)
-
-  if vs2013_runtime_dll_dirs:
-    x64_runtime, x86_runtime = vs2013_runtime_dll_dirs
-    vs_toolchain.CopyVsRuntimeDlls(
-        os.path.join(checkout_root, gyp_chromium.GetOutputDirectory()),
-        (x86_runtime, x64_runtime))
-
-  sys.exit(gyp_rc)
diff --git a/media/libyuv/gyp_libyuv.py b/media/libyuv/gyp_libyuv.py
deleted file mode 100644
index ac42038df3..0000000000
--- a/media/libyuv/gyp_libyuv.py
+++ /dev/null
@@ -1,28 +0,0 @@
-#!/usr/bin/env python
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-
-# This script is a modified copy of the src/build/gyp_chromium.py file. 
-# It is needed for parallel processing.
-
-# This file is (possibly, depending on python version) imported by
-# gyp_libyuv when GYP_PARALLEL=1 and it creates sub-processes
-# through the multiprocessing library.
-
-# Importing in Python 2.6 (fixed in 2.7) on Windows doesn't search for
-# imports that don't end in .py (and aren't directories with an
-# __init__.py). This wrapper makes "import gyp_libyuv" work with
-# those old versions and makes it possible to execute gyp_libyuv.py
-# directly on Windows where the extension is useful.
-
-import os
-
-path = os.path.abspath(os.path.split(__file__)[0])
-execfile(os.path.join(path, 'gyp_libyuv'))
diff --git a/media/libyuv/include/libyuv.h b/media/libyuv/include/libyuv.h
index de652836e0..a06e1233ab 100644
--- a/media/libyuv/include/libyuv.h
+++ b/media/libyuv/include/libyuv.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_H_
 #define INCLUDE_LIBYUV_H_
 
 #include "libyuv/basic_types.h"
@@ -26,7 +26,8 @@
 #include "libyuv/scale.h"
 #include "libyuv/scale_argb.h"
 #include "libyuv/scale_row.h"
+#include "libyuv/scale_uv.h"
 #include "libyuv/version.h"
 #include "libyuv/video_common.h"
 
-#endif  // INCLUDE_LIBYUV_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_H_
diff --git a/media/libyuv/include/libyuv/basic_types.h b/media/libyuv/include/libyuv/basic_types.h
index 4b5e4b619c..1bea67f2f2 100644
--- a/media/libyuv/include/libyuv/basic_types.h
+++ b/media/libyuv/include/libyuv/basic_types.h
@@ -8,96 +8,39 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_
 #define INCLUDE_LIBYUV_BASIC_TYPES_H_
 
-#include <stddef.h>  // for NULL, size_t
+#include <stddef.h>  // For size_t and NULL
 
-#if !(defined(_MSC_VER) && (_MSC_VER < 1600))
-#include <stdint.h>  // for uintptr_t
-#endif
+#if !defined(INT_TYPES_DEFINED) && !defined(GG_LONGLONG)
+#define INT_TYPES_DEFINED
 
-typedef uint64_t uint64;
-typedef int64_t  int64;
-#if defined(_MSC_VER)
-// nsprpub/pr/include/obsolete/protypes.h defines these weirdly
-typedef long int32;
-typedef unsigned long uint32;
+#if defined(_MSC_VER) && (_MSC_VER < 1600)
+#include <sys/types.h>  // for uintptr_t on x86
+typedef unsigned __int64 uint64_t;
+typedef __int64 int64_t;
+typedef unsigned int uint32_t;
+typedef int int32_t;
+typedef unsigned short uint16_t;
+typedef short int16_t;
+typedef unsigned char uint8_t;
+typedef signed char int8_t;
 #else
+#include <stdint.h>  // for uintptr_t and C99 types
+#endif               // defined(_MSC_VER) && (_MSC_VER < 1600)
+// Types are deprecated.  Enable this macro for legacy types.
+#ifdef LIBYUV_LEGACY_TYPES
+typedef uint64_t uint64;
+typedef int64_t int64;
 typedef uint32_t uint32;
-typedef int32_t  int32;
-#endif
+typedef int32_t int32;
 typedef uint16_t uint16;
-typedef int16_t  int16;
-typedef uint8_t  uint8;
-typedef int8_t   int8;
-#define INT_TYPES_DEFINED 1
-
-#ifndef GG_LONGLONG
-#ifndef INT_TYPES_DEFINED
-#define INT_TYPES_DEFINED
-#ifdef COMPILER_MSVC
-typedef unsigned __int64 uint64;
-typedef __int64 int64;
-#ifndef INT64_C
-#define INT64_C(x) x ## I64
-#endif
-#ifndef UINT64_C
-#define UINT64_C(x) x ## UI64
-#endif
-#define INT64_F "I64"
-#else  // COMPILER_MSVC
-#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
-typedef unsigned long uint64;  // NOLINT
-typedef long int64;  // NOLINT
-#ifndef INT64_C
-#define INT64_C(x) x ## L
-#endif
-#ifndef UINT64_C
-#define UINT64_C(x) x ## UL
-#endif
-#define INT64_F "l"
-#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
-typedef unsigned long long uint64;  // NOLINT
-typedef long long int64;  // NOLINT
-#ifndef INT64_C
-#define INT64_C(x) x ## LL
-#endif
-#ifndef UINT64_C
-#define UINT64_C(x) x ## ULL
-#endif
-#define INT64_F "ll"
-#endif  // __LP64__
-#endif  // COMPILER_MSVC
-typedef unsigned int uint32;
-typedef int int32;
-typedef unsigned short uint16;  // NOLINT
-typedef short int16;  // NOLINT
-typedef unsigned char uint8;
-typedef signed char int8;
+typedef int16_t int16;
+typedef uint8_t uint8;
+typedef int8_t int8;
+#endif  // LIBYUV_LEGACY_TYPES
 #endif  // INT_TYPES_DEFINED
-#endif  // GG_LONGLONG
-
-// Detect compiler is for x86 or x64.
-#if defined(__x86_64__) || defined(_M_X64) || \
-    defined(__i386__) || defined(_M_IX86)
-#define CPU_X86 1
-#endif
-// Detect compiler is for ARM.
-#if defined(__arm__) || defined(_M_ARM)
-#define CPU_ARM 1
-#endif
-
-#ifndef ALIGNP
-#ifdef __cplusplus
-#define ALIGNP(p, t) \
-    (reinterpret_cast<uint8*>(((reinterpret_cast<uintptr_t>(p) + \
-    ((t) - 1)) & ~((t) - 1))))
-#else
-#define ALIGNP(p, t) \
-    ((uint8*)((((uintptr_t)(p) + ((t) - 1)) & ~((t) - 1))))  /* NOLINT */
-#endif
-#endif
 
 #if !defined(LIBYUV_API)
 #if defined(_WIN32) || defined(__CYGWIN__)
@@ -109,24 +52,17 @@ typedef signed char int8;
 #define LIBYUV_API
 #endif  // LIBYUV_BUILDING_SHARED_LIBRARY
 #elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__APPLE__) && \
-    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) || \
-    defined(LIBYUV_USING_SHARED_LIBRARY))
-#define LIBYUV_API __attribute__ ((visibility ("default")))
+    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) ||                      \
+     defined(LIBYUV_USING_SHARED_LIBRARY))
+#define LIBYUV_API __attribute__((visibility("default")))
 #else
 #define LIBYUV_API
 #endif  // __GNUC__
 #endif  // LIBYUV_API
 
+// TODO(fbarchard): Remove bool macros.
 #define LIBYUV_BOOL int
 #define LIBYUV_FALSE 0
 #define LIBYUV_TRUE 1
 
-// Visual C x86 or GCC little endian.
-#if defined(__x86_64__) || defined(_M_X64) || \
-  defined(__i386__) || defined(_M_IX86) || \
-  defined(__arm__) || defined(_M_ARM) || \
-  (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
-#define LIBYUV_LITTLE_ENDIAN
-#endif
-
-#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_
diff --git a/media/libyuv/include/libyuv/compare.h b/media/libyuv/include/libyuv/compare.h
index 08b2bb2ecf..3353ad71c6 100644
--- a/media/libyuv/include/libyuv/compare.h
+++ b/media/libyuv/include/libyuv/compare.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_COMPARE_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_COMPARE_H_
 #define INCLUDE_LIBYUV_COMPARE_H_
 
 #include "libyuv/basic_types.h"
@@ -20,59 +20,92 @@ extern "C" {
 
 // Compute a hash for specified memory. Seed of 5381 recommended.
 LIBYUV_API
-uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed);
+uint32_t HashDjb2(const uint8_t* src, uint64_t count, uint32_t seed);
+
+// Hamming Distance
+LIBYUV_API
+uint64_t ComputeHammingDistance(const uint8_t* src_a,
+                                const uint8_t* src_b,
+                                int count);
 
 // Scan an opaque argb image and return fourcc based on alpha offset.
 // Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
 LIBYUV_API
-uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height);
+uint32_t ARGBDetect(const uint8_t* argb,
+                    int stride_argb,
+                    int width,
+                    int height);
 
 // Sum Square Error - used to compute Mean Square Error or PSNR.
 LIBYUV_API
-uint64 ComputeSumSquareError(const uint8* src_a,
-                             const uint8* src_b, int count);
+uint64_t ComputeSumSquareError(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count);
 
 LIBYUV_API
-uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
-                                  const uint8* src_b, int stride_b,
-                                  int width, int height);
+uint64_t ComputeSumSquareErrorPlane(const uint8_t* src_a,
+                                    int stride_a,
+                                    const uint8_t* src_b,
+                                    int stride_b,
+                                    int width,
+                                    int height);
 
 static const int kMaxPsnr = 128;
 
 LIBYUV_API
-double SumSquareErrorToPsnr(uint64 sse, uint64 count);
+double SumSquareErrorToPsnr(uint64_t sse, uint64_t count);
 
 LIBYUV_API
-double CalcFramePsnr(const uint8* src_a, int stride_a,
-                     const uint8* src_b, int stride_b,
-                     int width, int height);
+double CalcFramePsnr(const uint8_t* src_a,
+                     int stride_a,
+                     const uint8_t* src_b,
+                     int stride_b,
+                     int width,
+                     int height);
 
 LIBYUV_API
-double I420Psnr(const uint8* src_y_a, int stride_y_a,
-                const uint8* src_u_a, int stride_u_a,
-                const uint8* src_v_a, int stride_v_a,
-                const uint8* src_y_b, int stride_y_b,
-                const uint8* src_u_b, int stride_u_b,
-                const uint8* src_v_b, int stride_v_b,
-                int width, int height);
+double I420Psnr(const uint8_t* src_y_a,
+                int stride_y_a,
+                const uint8_t* src_u_a,
+                int stride_u_a,
+                const uint8_t* src_v_a,
+                int stride_v_a,
+                const uint8_t* src_y_b,
+                int stride_y_b,
+                const uint8_t* src_u_b,
+                int stride_u_b,
+                const uint8_t* src_v_b,
+                int stride_v_b,
+                int width,
+                int height);
 
 LIBYUV_API
-double CalcFrameSsim(const uint8* src_a, int stride_a,
-                     const uint8* src_b, int stride_b,
-                     int width, int height);
+double CalcFrameSsim(const uint8_t* src_a,
+                     int stride_a,
+                     const uint8_t* src_b,
+                     int stride_b,
+                     int width,
+                     int height);
 
 LIBYUV_API
-double I420Ssim(const uint8* src_y_a, int stride_y_a,
-                const uint8* src_u_a, int stride_u_a,
-                const uint8* src_v_a, int stride_v_a,
-                const uint8* src_y_b, int stride_y_b,
-                const uint8* src_u_b, int stride_u_b,
-                const uint8* src_v_b, int stride_v_b,
-                int width, int height);
+double I420Ssim(const uint8_t* src_y_a,
+                int stride_y_a,
+                const uint8_t* src_u_a,
+                int stride_u_a,
+                const uint8_t* src_v_a,
+                int stride_v_a,
+                const uint8_t* src_y_b,
+                int stride_y_b,
+                const uint8_t* src_u_b,
+                int stride_u_b,
+                const uint8_t* src_v_b,
+                int stride_v_b,
+                int width,
+                int height);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_COMPARE_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_COMPARE_H_
diff --git a/media/libyuv/include/libyuv/compare_row.h b/media/libyuv/include/libyuv/compare_row.h
index 38a957b2ce..d8e82d721b 100644
--- a/media/libyuv/include/libyuv/compare_row.h
+++ b/media/libyuv/include/libyuv/compare_row.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_COMPARE_ROW_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_COMPARE_ROW_H_
 #define INCLUDE_LIBYUV_COMPARE_ROW_H_
 
 #include "libyuv/basic_types.h"
@@ -18,20 +18,23 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#if defined(__pnacl__) || defined(__CLR_VER) || \
-    (defined(__i386__) && !defined(__SSE2__))
+#if defined(__pnacl__) || defined(__CLR_VER) ||            \
+    (defined(__native_client__) && defined(__x86_64__)) || \
+    (defined(__i386__) && !defined(__SSE__) && !defined(__clang__))
 #define LIBYUV_DISABLE_X86
 #endif
+#if defined(__native_client__)
+#define LIBYUV_DISABLE_NEON
+#endif
 // MemorySanitizer does not support assembly code yet. http://crbug.com/344505
 #if defined(__has_feature)
 #if __has_feature(memory_sanitizer)
 #define LIBYUV_DISABLE_X86
 #endif
 #endif
-
 // Visual C 2012 required for AVX2.
-#if defined(_M_IX86) && !defined(__clang__) && \
-    defined(_MSC_VER) && _MSC_VER >= 1700
+#if defined(_M_IX86) && !defined(__clang__) && defined(_MSC_VER) && \
+    _MSC_VER >= 1700
 #define VISUALC_HAS_AVX2 1
 #endif  // VisualStudio >= 2012
 
@@ -42,43 +45,86 @@ extern "C" {
 #endif  // clang >= 3.4
 #endif  // __clang__
 
-#if !defined(LIBYUV_DISABLE_X86) && \
-    defined(_M_IX86) && (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
-#define HAS_HASHDJB2_AVX2
-#endif
-
 // The following are available for Visual C and GCC:
 #if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) || defined(_M_IX86)))
+    (defined(__x86_64__) || defined(__i386__) || defined(_M_IX86))
 #define HAS_HASHDJB2_SSE41
 #define HAS_SUMSQUAREERROR_SSE2
+#define HAS_HAMMINGDISTANCE_SSE42
 #endif
 
 // The following are available for Visual C and clangcl 32 bit:
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) && \
+    !defined(__clang__) &&                                                   \
     (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
 #define HAS_HASHDJB2_AVX2
 #define HAS_SUMSQUAREERROR_AVX2
 #endif
 
+// The following are available for GCC and clangcl:
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+#define HAS_HAMMINGDISTANCE_SSSE3
+#endif
+
+// The following are available for GCC and clangcl:
+#if !defined(LIBYUV_DISABLE_X86) && defined(CLANG_HAS_AVX2) && \
+    (defined(__x86_64__) || defined(__i386__))
+#define HAS_HAMMINGDISTANCE_AVX2
+#endif
+
 // The following are available for Neon:
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
 #define HAS_SUMSQUAREERROR_NEON
+#define HAS_HAMMINGDISTANCE_NEON
+#endif
+
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#define HAS_HAMMINGDISTANCE_MSA
+#define HAS_SUMSQUAREERROR_MSA
 #endif
 
-uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
-uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
-uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
-uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
+uint32_t HammingDistance_C(const uint8_t* src_a,
+                           const uint8_t* src_b,
+                           int count);
+uint32_t HammingDistance_SSE42(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count);
+uint32_t HammingDistance_SSSE3(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count);
+uint32_t HammingDistance_AVX2(const uint8_t* src_a,
+                              const uint8_t* src_b,
+                              int count);
+uint32_t HammingDistance_NEON(const uint8_t* src_a,
+                              const uint8_t* src_b,
+                              int count);
+uint32_t HammingDistance_MSA(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count);
+uint32_t SumSquareError_C(const uint8_t* src_a,
+                          const uint8_t* src_b,
+                          int count);
+uint32_t SumSquareError_SSE2(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count);
+uint32_t SumSquareError_AVX2(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count);
+uint32_t SumSquareError_NEON(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count);
+uint32_t SumSquareError_MSA(const uint8_t* src_a,
+                            const uint8_t* src_b,
+                            int count);
 
-uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
-uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
-uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
+uint32_t HashDjb2_C(const uint8_t* src, int count, uint32_t seed);
+uint32_t HashDjb2_SSE41(const uint8_t* src, int count, uint32_t seed);
+uint32_t HashDjb2_AVX2(const uint8_t* src, int count, uint32_t seed);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_COMPARE_ROW_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_COMPARE_ROW_H_
diff --git a/media/libyuv/include/libyuv/convert.h b/media/libyuv/include/libyuv/convert.h
index a2cdc5718e..2f1ce4214b 100644
--- a/media/libyuv/include/libyuv/convert.h
+++ b/media/libyuv/include/libyuv/convert.h
@@ -8,13 +8,18 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CONVERT_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_CONVERT_H_
 #define INCLUDE_LIBYUV_CONVERT_H_
 
 #include "libyuv/basic_types.h"
 
 #include "libyuv/rotate.h"  // For enum RotationMode.
 
+// TODO(fbarchard): fix WebRTC source to include following libyuv headers:
+#include "libyuv/convert_argb.h"      // For WebRTC I420ToARGB. b/620
+#include "libyuv/convert_from.h"      // For WebRTC ConvertFromI420. b/620
+#include "libyuv/planar_functions.h"  // For WebRTC I420Rect, CopyPlane. b/618
+
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
@@ -22,185 +27,926 @@ extern "C" {
 
 // Convert I444 to I420.
 LIBYUV_API
-int I444ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int I444ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I444 to NV12.
+LIBYUV_API
+int I444ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I444 to NV21.
+LIBYUV_API
+int I444ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
 
 // Convert I422 to I420.
 LIBYUV_API
-int I422ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int I422ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I422 to I444.
+LIBYUV_API
+int I422ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I422 to I210.
+LIBYUV_API
+int I422ToI210(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
-// Convert I411 to I420.
+// Convert MM21 to NV12.
 LIBYUV_API
-int I411ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int MM21ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert MM21 to I420.
+LIBYUV_API
+int MM21ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert MM21 to YUY2
+LIBYUV_API
+int MM21ToYUY2(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height);
+
+// Convert MT2T to P010
+// Note that src_y and src_uv point to packed 10-bit values, so the Y plane will
+// be 10 / 8 times the dimensions of the image. Also for this reason,
+// src_stride_y and src_stride_uv are given in bytes.
+LIBYUV_API
+int MT2TToP010(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I422 to NV21.
+LIBYUV_API
+int I422ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
 
 // Copy I420 to I420.
 #define I420ToI420 I420Copy
 LIBYUV_API
-int I420Copy(const uint8* src_y, int src_stride_y,
-             const uint8* src_u, int src_stride_u,
-             const uint8* src_v, int src_stride_v,
-             uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int width, int height);
+int I420Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_u,
+             int src_stride_u,
+             const uint8_t* src_v,
+             int src_stride_v,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
+
+// Convert I420 to I444.
+LIBYUV_API
+int I420ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Copy I010 to I010
+#define I010ToI010 I010Copy
+#define H010ToH010 I010Copy
+LIBYUV_API
+int I010Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
+
+// Convert 10 bit YUV to 8 bit
+#define H010ToH420 I010ToI420
+LIBYUV_API
+int I010ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H210ToH420 I210ToI420
+LIBYUV_API
+int I210ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H210ToH422 I210ToI422
+LIBYUV_API
+int I210ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H410ToH420 I410ToI420
+LIBYUV_API
+int I410ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H410ToH444 I410ToI444
+LIBYUV_API
+int I410ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H012ToH420 I012ToI420
+LIBYUV_API
+int I012ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H212ToH422 I212ToI422
+LIBYUV_API
+int I212ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H412ToH444 I412ToI444
+LIBYUV_API
+int I412ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define I412ToI012 I410ToI010
+#define H410ToH010 I410ToI010
+#define H412ToH012 I410ToI010
+LIBYUV_API
+int I410ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define I212ToI012 I210ToI010
+#define H210ToH010 I210ToI010
+#define H212ToH012 I210ToI010
+LIBYUV_API
+int I210ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I010 to I410
+LIBYUV_API
+int I010ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I012 to I412
+#define I012ToI412 I010ToI410
+
+// Convert I210 to I410
+LIBYUV_API
+int I210ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I212 to I412
+#define I212ToI412 I210ToI410
+
+// Convert I010 to P010
+LIBYUV_API
+int I010ToP010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I210 to P210
+LIBYUV_API
+int I210ToP210(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I012 to P012
+LIBYUV_API
+int I012ToP012(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I212 to P212
+LIBYUV_API
+int I212ToP212(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
 
 // Convert I400 (grey) to I420.
 LIBYUV_API
-int I400ToI420(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int I400ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I400 (grey) to NV21.
+LIBYUV_API
+int I400ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
 
 #define J400ToJ420 I400ToI420
 
 // Convert NV12 to I420.
 LIBYUV_API
-int NV12ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_uv, int src_stride_uv,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int NV12ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Convert NV21 to I420.
 LIBYUV_API
-int NV21ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_vu, int src_stride_vu,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int NV21ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert NV12 to NV24.
+LIBYUV_API
+int NV12ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert NV16 to NV24.
+LIBYUV_API
+int NV16ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert P010 to I010.
+LIBYUV_API
+int P010ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert P012 to I012.
+LIBYUV_API
+int P012ToI012(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert P010 to P410.
+LIBYUV_API
+int P010ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert P012 to P412.
+#define P012ToP412 P010ToP410
+
+// Convert P016 to P416.
+#define P016ToP416 P010ToP410
+
+// Convert P210 to P410.
+LIBYUV_API
+int P210ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert P212 to P412.
+#define P212ToP412 P210ToP410
+
+// Convert P216 to P416.
+#define P216ToP416 P210ToP410
 
 // Convert YUY2 to I420.
 LIBYUV_API
-int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int YUY2ToI420(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Convert UYVY to I420.
 LIBYUV_API
-int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int UYVYToI420(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
-// Convert M420 to I420.
+// Convert AYUV to NV12.
 LIBYUV_API
-int M420ToI420(const uint8* src_m420, int src_stride_m420,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int AYUVToNV12(const uint8_t* src_ayuv,
+               int src_stride_ayuv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert AYUV to NV21.
+LIBYUV_API
+int AYUVToNV21(const uint8_t* src_ayuv,
+               int src_stride_ayuv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
+
+// Convert Android420 to I420.
+LIBYUV_API
+int Android420ToI420(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     int src_pixel_stride_uv,
+                     uint8_t* dst_y,
+                     int dst_stride_y,
+                     uint8_t* dst_u,
+                     int dst_stride_u,
+                     uint8_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height);
 
 // ARGB little endian (bgra in memory) to I420.
 LIBYUV_API
-int ARGBToI420(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToI420(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // BGRA little endian (argb in memory) to I420.
 LIBYUV_API
-int BGRAToI420(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int BGRAToI420(const uint8_t* src_bgra,
+               int src_stride_bgra,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // ABGR little endian (rgba in memory) to I420.
 LIBYUV_API
-int ABGRToI420(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ABGRToI420(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // RGBA little endian (abgr in memory) to I420.
 LIBYUV_API
-int RGBAToI420(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int RGBAToI420(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // RGB little endian (bgr in memory) to I420.
 LIBYUV_API
-int RGB24ToI420(const uint8* src_frame, int src_stride_frame,
-                uint8* dst_y, int dst_stride_y,
-                uint8* dst_u, int dst_stride_u,
-                uint8* dst_v, int dst_stride_v,
-                int width, int height);
+int RGB24ToI420(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                uint8_t* dst_u,
+                int dst_stride_u,
+                uint8_t* dst_v,
+                int dst_stride_v,
+                int width,
+                int height);
+
+// RGB little endian (bgr in memory) to J420.
+LIBYUV_API
+int RGB24ToJ420(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                uint8_t* dst_u,
+                int dst_stride_u,
+                uint8_t* dst_v,
+                int dst_stride_v,
+                int width,
+                int height);
 
 // RGB big endian (rgb in memory) to I420.
 LIBYUV_API
-int RAWToI420(const uint8* src_frame, int src_stride_frame,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int width, int height);
+int RAWToI420(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height);
+
+// RGB big endian (rgb in memory) to J420.
+LIBYUV_API
+int RAWToJ420(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height);
 
 // RGB16 (RGBP fourcc) little endian to I420.
 LIBYUV_API
-int RGB565ToI420(const uint8* src_frame, int src_stride_frame,
-                 uint8* dst_y, int dst_stride_y,
-                 uint8* dst_u, int dst_stride_u,
-                 uint8* dst_v, int dst_stride_v,
-                 int width, int height);
+int RGB565ToI420(const uint8_t* src_rgb565,
+                 int src_stride_rgb565,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 uint8_t* dst_u,
+                 int dst_stride_u,
+                 uint8_t* dst_v,
+                 int dst_stride_v,
+                 int width,
+                 int height);
 
 // RGB15 (RGBO fourcc) little endian to I420.
 LIBYUV_API
-int ARGB1555ToI420(const uint8* src_frame, int src_stride_frame,
-                   uint8* dst_y, int dst_stride_y,
-                   uint8* dst_u, int dst_stride_u,
-                   uint8* dst_v, int dst_stride_v,
-                   int width, int height);
+int ARGB1555ToI420(const uint8_t* src_argb1555,
+                   int src_stride_argb1555,
+                   uint8_t* dst_y,
+                   int dst_stride_y,
+                   uint8_t* dst_u,
+                   int dst_stride_u,
+                   uint8_t* dst_v,
+                   int dst_stride_v,
+                   int width,
+                   int height);
 
 // RGB12 (R444 fourcc) little endian to I420.
 LIBYUV_API
-int ARGB4444ToI420(const uint8* src_frame, int src_stride_frame,
-                   uint8* dst_y, int dst_stride_y,
-                   uint8* dst_u, int dst_stride_u,
-                   uint8* dst_v, int dst_stride_v,
-                   int width, int height);
+int ARGB4444ToI420(const uint8_t* src_argb4444,
+                   int src_stride_argb4444,
+                   uint8_t* dst_y,
+                   int dst_stride_y,
+                   uint8_t* dst_u,
+                   int dst_stride_u,
+                   uint8_t* dst_v,
+                   int dst_stride_v,
+                   int width,
+                   int height);
+
+// RGB little endian (bgr in memory) to J400.
+LIBYUV_API
+int RGB24ToJ400(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_yj,
+                int dst_stride_yj,
+                int width,
+                int height);
+
+// RGB big endian (rgb in memory) to J400.
+LIBYUV_API
+int RAWToJ400(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_yj,
+              int dst_stride_yj,
+              int width,
+              int height);
 
-#ifdef HAVE_JPEG
 // src_width/height provided by capture.
 // dst_width/height for clipping determine final size.
 LIBYUV_API
-int MJPGToI420(const uint8* sample, size_t sample_size,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int src_width, int src_height,
-               int dst_width, int dst_height);
+int MJPGToI420(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height);
+
+// JPEG to NV21
+LIBYUV_API
+int MJPGToNV21(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height);
+
+// JPEG to NV12
+LIBYUV_API
+int MJPGToNV12(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height);
 
 // Query size of MJPG in pixels.
 LIBYUV_API
-int MJPGSize(const uint8* sample, size_t sample_size,
-             int* width, int* height);
-#endif
+int MJPGSize(const uint8_t* sample,
+             size_t sample_size,
+             int* width,
+             int* height);
 
 // Convert camera sample to I420 with cropping, rotation and vertical flip.
 // "src_size" is needed to parse MJPG.
@@ -222,22 +968,29 @@ int MJPGSize(const uint8* sample, size_t sample_size,
 //    Must be less than or equal to src_width/src_height
 //    Cropping parameters are pre-rotation.
 // "rotation" can be 0, 90, 180 or 270.
-// "format" is a fourcc. ie 'I420', 'YUY2'
+// "fourcc" is a fourcc. ie 'I420', 'YUY2'
 // Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
 LIBYUV_API
-int ConvertToI420(const uint8* src_frame, size_t src_size,
-                  uint8* dst_y, int dst_stride_y,
-                  uint8* dst_u, int dst_stride_u,
-                  uint8* dst_v, int dst_stride_v,
-                  int crop_x, int crop_y,
-                  int src_width, int src_height,
-                  int crop_width, int crop_height,
+int ConvertToI420(const uint8_t* sample,
+                  size_t sample_size,
+                  uint8_t* dst_y,
+                  int dst_stride_y,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int crop_x,
+                  int crop_y,
+                  int src_width,
+                  int src_height,
+                  int crop_width,
+                  int crop_height,
                   enum RotationMode rotation,
-                  uint32 format);
+                  uint32_t fourcc);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CONVERT_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_CONVERT_H_
diff --git a/media/libyuv/include/libyuv/convert_argb.h b/media/libyuv/include/libyuv/convert_argb.h
index 079d273b11..8e4562efc9 100644
--- a/media/libyuv/include/libyuv/convert_argb.h
+++ b/media/libyuv/include/libyuv/convert_argb.h
@@ -8,272 +8,2269 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CONVERT_ARGB_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_CONVERT_ARGB_H_
 #define INCLUDE_LIBYUV_CONVERT_ARGB_H_
 
 #include "libyuv/basic_types.h"
 
 #include "libyuv/rotate.h"  // For enum RotationMode.
-
-// TODO(fbarchard): This set of functions should exactly match convert.h
-// TODO(fbarchard): Add tests. Create random content of right size and convert
-// with C vs Opt and or to I420 and compare.
-// TODO(fbarchard): Some of these functions lack parameter setting.
+#include "libyuv/scale.h"   // For enum FilterMode.
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
+// Conversion matrix for YUV to RGB
+LIBYUV_API extern const struct YuvConstants kYuvI601Constants;   // BT.601
+LIBYUV_API extern const struct YuvConstants kYuvJPEGConstants;   // BT.601 full
+LIBYUV_API extern const struct YuvConstants kYuvH709Constants;   // BT.709
+LIBYUV_API extern const struct YuvConstants kYuvF709Constants;   // BT.709 full
+LIBYUV_API extern const struct YuvConstants kYuv2020Constants;   // BT.2020
+LIBYUV_API extern const struct YuvConstants kYuvV2020Constants;  // BT.2020 full
+
+// Conversion matrix for YVU to BGR
+LIBYUV_API extern const struct YuvConstants kYvuI601Constants;   // BT.601
+LIBYUV_API extern const struct YuvConstants kYvuJPEGConstants;   // BT.601 full
+LIBYUV_API extern const struct YuvConstants kYvuH709Constants;   // BT.709
+LIBYUV_API extern const struct YuvConstants kYvuF709Constants;   // BT.709 full
+LIBYUV_API extern const struct YuvConstants kYvu2020Constants;   // BT.2020
+LIBYUV_API extern const struct YuvConstants kYvuV2020Constants;  // BT.2020 full
+
+// Macros for end swapped destination Matrix conversions.
+// Swap UV and pass mirrored kYvuJPEGConstants matrix.
+// TODO(fbarchard): Add macro for each Matrix function.
+#define kYuvI601ConstantsVU kYvuI601Constants
+#define kYuvJPEGConstantsVU kYvuJPEGConstants
+#define kYuvH709ConstantsVU kYvuH709Constants
+#define kYuvF709ConstantsVU kYvuF709Constants
+#define kYuv2020ConstantsVU kYvu2020Constants
+#define kYuvV2020ConstantsVU kYvuV2020Constants
+
+#define NV12ToABGRMatrix(a, b, c, d, e, f, g, h, i) \
+  NV21ToARGBMatrix(a, b, c, d, e, f, g##VU, h, i)
+#define NV21ToABGRMatrix(a, b, c, d, e, f, g, h, i) \
+  NV12ToARGBMatrix(a, b, c, d, e, f, g##VU, h, i)
+#define NV12ToRAWMatrix(a, b, c, d, e, f, g, h, i) \
+  NV21ToRGB24Matrix(a, b, c, d, e, f, g##VU, h, i)
+#define NV21ToRAWMatrix(a, b, c, d, e, f, g, h, i) \
+  NV12ToRGB24Matrix(a, b, c, d, e, f, g##VU, h, i)
+#define I010ToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I010ToARGBMatrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I210ToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I210ToARGBMatrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I410ToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I410ToARGBMatrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I010ToAB30Matrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I010ToAR30Matrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I210ToAB30Matrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I210ToAR30Matrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I410ToAB30Matrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I410ToAR30Matrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I420AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I422AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I444AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I010AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I210AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I410AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+
 // Alias.
 #define ARGBToARGB ARGBCopy
 
 // Copy ARGB to ARGB.
 LIBYUV_API
-int ARGBCopy(const uint8* src_argb, int src_stride_argb,
-             uint8* dst_argb, int dst_stride_argb,
-             int width, int height);
+int ARGBCopy(const uint8_t* src_argb,
+             int src_stride_argb,
+             uint8_t* dst_argb,
+             int dst_stride_argb,
+             int width,
+             int height);
 
 // Convert I420 to ARGB.
 LIBYUV_API
-int I420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int I420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I420 to ABGR.
+LIBYUV_API
+int I420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert J420 to ARGB.
+LIBYUV_API
+int J420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert J420 to ABGR.
+LIBYUV_API
+int J420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H420 to ARGB.
+LIBYUV_API
+int H420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert H420 to ABGR.
+LIBYUV_API
+int H420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U420 to ARGB.
+LIBYUV_API
+int U420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U420 to ABGR.
+LIBYUV_API
+int U420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
 
 // Convert I422 to ARGB.
 LIBYUV_API
-int I422ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int I422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert J422 to ABGR.
+LIBYUV_API
+int J422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H422 to ARGB.
+LIBYUV_API
+int H422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert H422 to ABGR.
+LIBYUV_API
+int H422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U422 to ARGB.
+LIBYUV_API
+int U422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U422 to ABGR.
+LIBYUV_API
+int U422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
 
 // Convert I444 to ARGB.
 LIBYUV_API
-int I444ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int I444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I444 to ABGR.
+LIBYUV_API
+int I444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
 
 // Convert J444 to ARGB.
 LIBYUV_API
-int J444ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int J444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
-// Convert I444 to ABGR.
+// Convert J444 to ABGR.
+LIBYUV_API
+int J444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H444 to ARGB.
+LIBYUV_API
+int H444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert H444 to ABGR.
+LIBYUV_API
+int H444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U444 to ARGB.
+LIBYUV_API
+int U444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U444 to ABGR.
+LIBYUV_API
+int U444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I444 to RGB24.
+LIBYUV_API
+int I444ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+// Convert I444 to RAW.
+LIBYUV_API
+int I444ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+// Convert I010 to ARGB.
+LIBYUV_API
+int I010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I010 to ABGR.
+LIBYUV_API
+int I010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H010 to ARGB.
+LIBYUV_API
+int H010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert H010 to ABGR.
+LIBYUV_API
+int H010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U010 to ARGB.
+LIBYUV_API
+int U010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U010 to ABGR.
+LIBYUV_API
+int U010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I210 to ARGB.
+LIBYUV_API
+int I210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I210 to ABGR.
+LIBYUV_API
+int I210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H210 to ARGB.
 LIBYUV_API
-int I444ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
+int H210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
-// Convert I411 to ARGB.
+// Convert H210 to ABGR.
 LIBYUV_API
-int I411ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int H210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U210 to ARGB.
+LIBYUV_API
+int U210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U210 to ABGR.
+LIBYUV_API
+int U210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
 
 // Convert I420 with Alpha to preattenuated ARGB.
 LIBYUV_API
-int I420AlphaToARGB(const uint8* src_y, int src_stride_y,
-                    const uint8* src_u, int src_stride_u,
-                    const uint8* src_v, int src_stride_v,
-                    const uint8* src_a, int src_stride_a,
-                    uint8* dst_argb, int dst_stride_argb,
-                    int width, int height, int attenuate);
+int I420AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate);
 
 // Convert I420 with Alpha to preattenuated ABGR.
 LIBYUV_API
-int I420AlphaToABGR(const uint8* src_y, int src_stride_y,
-                    const uint8* src_u, int src_stride_u,
-                    const uint8* src_v, int src_stride_v,
-                    const uint8* src_a, int src_stride_a,
-                    uint8* dst_abgr, int dst_stride_abgr,
-                    int width, int height, int attenuate);
+int I420AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I422 with Alpha to preattenuated ARGB.
+LIBYUV_API
+int I422AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I422 with Alpha to preattenuated ABGR.
+LIBYUV_API
+int I422AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I444 with Alpha to preattenuated ARGB.
+LIBYUV_API
+int I444AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I444 with Alpha to preattenuated ABGR.
+LIBYUV_API
+int I444AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate);
 
 // Convert I400 (grey) to ARGB.  Reverse of ARGBToI400.
 LIBYUV_API
-int I400ToARGB(const uint8* src_y, int src_stride_y,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int I400ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // Convert J400 (jpeg grey) to ARGB.
 LIBYUV_API
-int J400ToARGB(const uint8* src_y, int src_stride_y,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int J400ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // Alias.
 #define YToARGB I400ToARGB
 
 // Convert NV12 to ARGB.
 LIBYUV_API
-int NV12ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_uv, int src_stride_uv,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int NV12ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // Convert NV21 to ARGB.
 LIBYUV_API
-int NV21ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_vu, int src_stride_vu,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int NV21ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert NV12 to ABGR.
+LIBYUV_API
+int NV12ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert NV21 to ABGR.
+LIBYUV_API
+int NV21ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert NV12 to RGB24.
+LIBYUV_API
+int NV12ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_uv,
+                int src_stride_uv,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+// Convert NV21 to RGB24.
+LIBYUV_API
+int NV21ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_vu,
+                int src_stride_vu,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+// Convert NV21 to YUV24.
+LIBYUV_API
+int NV21ToYUV24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_vu,
+                int src_stride_vu,
+                uint8_t* dst_yuv24,
+                int dst_stride_yuv24,
+                int width,
+                int height);
+
+// Convert NV12 to RAW.
+LIBYUV_API
+int NV12ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_uv,
+              int src_stride_uv,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
 
-// Convert M420 to ARGB.
+// Convert NV21 to RAW.
 LIBYUV_API
-int M420ToARGB(const uint8* src_m420, int src_stride_m420,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int NV21ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_vu,
+              int src_stride_vu,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
 
 // Convert YUY2 to ARGB.
 LIBYUV_API
-int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int YUY2ToARGB(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // Convert UYVY to ARGB.
 LIBYUV_API
-int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int UYVYToARGB(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
-// Convert J420 to ARGB.
+// Convert I010 to AR30.
 LIBYUV_API
-int J420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int I010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
 
-// Convert J422 to ARGB.
+// Convert H010 to AR30.
 LIBYUV_API
-int J422ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int H010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
 
-// Convert J420 to ABGR.
+// Convert I010 to AB30.
 LIBYUV_API
-int J420ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
+int I010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
 
-// Convert J422 to ABGR.
+// Convert H010 to AB30.
 LIBYUV_API
-int J422ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
+int H010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
 
-// Convert H420 to ARGB.
+// Convert U010 to AR30.
 LIBYUV_API
-int H420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int U010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
 
-// Convert H422 to ARGB.
+// Convert U010 to AB30.
 LIBYUV_API
-int H422ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int U010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
 
-// Convert H420 to ABGR.
+// Convert I210 to AR30.
 LIBYUV_API
-int H420ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
+int I210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
 
-// Convert H422 to ABGR.
+// Convert I210 to AB30.
+LIBYUV_API
+int I210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert H210 to AR30.
+LIBYUV_API
+int H210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert H210 to AB30.
+LIBYUV_API
+int H210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert U210 to AR30.
 LIBYUV_API
-int H422ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
+int U210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert U210 to AB30.
+LIBYUV_API
+int U210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
 
 // BGRA little endian (argb in memory) to ARGB.
 LIBYUV_API
-int BGRAToARGB(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int BGRAToARGB(const uint8_t* src_bgra,
+               int src_stride_bgra,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // ABGR little endian (rgba in memory) to ARGB.
 LIBYUV_API
-int ABGRToARGB(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int ABGRToARGB(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // RGBA little endian (abgr in memory) to ARGB.
 LIBYUV_API
-int RGBAToARGB(const uint8* src_frame, int src_stride_frame,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int RGBAToARGB(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // Deprecated function name.
 #define BG24ToARGB RGB24ToARGB
 
 // RGB little endian (bgr in memory) to ARGB.
 LIBYUV_API
-int RGB24ToARGB(const uint8* src_frame, int src_stride_frame,
-                uint8* dst_argb, int dst_stride_argb,
-                int width, int height);
+int RGB24ToARGB(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_argb,
+                int dst_stride_argb,
+                int width,
+                int height);
 
 // RGB big endian (rgb in memory) to ARGB.
 LIBYUV_API
-int RAWToARGB(const uint8* src_frame, int src_stride_frame,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height);
+int RAWToARGB(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height);
+
+// RGB big endian (rgb in memory) to RGBA.
+LIBYUV_API
+int RAWToRGBA(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_rgba,
+              int dst_stride_rgba,
+              int width,
+              int height);
 
 // RGB16 (RGBP fourcc) little endian to ARGB.
 LIBYUV_API
-int RGB565ToARGB(const uint8* src_frame, int src_stride_frame,
-                 uint8* dst_argb, int dst_stride_argb,
-                 int width, int height);
+int RGB565ToARGB(const uint8_t* src_rgb565,
+                 int src_stride_rgb565,
+                 uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int width,
+                 int height);
 
 // RGB15 (RGBO fourcc) little endian to ARGB.
 LIBYUV_API
-int ARGB1555ToARGB(const uint8* src_frame, int src_stride_frame,
-                   uint8* dst_argb, int dst_stride_argb,
-                   int width, int height);
+int ARGB1555ToARGB(const uint8_t* src_argb1555,
+                   int src_stride_argb1555,
+                   uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   int width,
+                   int height);
 
 // RGB12 (R444 fourcc) little endian to ARGB.
 LIBYUV_API
-int ARGB4444ToARGB(const uint8* src_frame, int src_stride_frame,
-                   uint8* dst_argb, int dst_stride_argb,
-                   int width, int height);
+int ARGB4444ToARGB(const uint8_t* src_argb4444,
+                   int src_stride_argb4444,
+                   uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   int width,
+                   int height);
+
+// Aliases
+#define AB30ToARGB AR30ToABGR
+#define AB30ToABGR AR30ToARGB
+#define AB30ToAR30 AR30ToAB30
+
+// Convert AR30 To ARGB.
+LIBYUV_API
+int AR30ToARGB(const uint8_t* src_ar30,
+               int src_stride_ar30,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert AR30 To ABGR.
+LIBYUV_API
+int AR30ToABGR(const uint8_t* src_ar30,
+               int src_stride_ar30,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert AR30 To AB30.
+LIBYUV_API
+int AR30ToAB30(const uint8_t* src_ar30,
+               int src_stride_ar30,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert AR64 to ARGB.
+LIBYUV_API
+int AR64ToARGB(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert AB64 to ABGR.
+#define AB64ToABGR AR64ToARGB
+
+// Convert AB64 to ARGB.
+LIBYUV_API
+int AB64ToARGB(const uint16_t* src_ab64,
+               int src_stride_ab64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert AR64 to ABGR.
+#define AR64ToABGR AB64ToARGB
+
+// Convert AR64 To AB64.
+LIBYUV_API
+int AR64ToAB64(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height);
+
+// Convert AB64 To AR64.
+#define AB64ToAR64 AR64ToAB64
 
-#ifdef HAVE_JPEG
 // src_width/height provided by capture
 // dst_width/height for clipping determine final size.
 LIBYUV_API
-int MJPGToARGB(const uint8* sample, size_t sample_size,
-               uint8* dst_argb, int dst_stride_argb,
-               int src_width, int src_height,
-               int dst_width, int dst_height);
-#endif
+int MJPGToARGB(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height);
+
+// Convert Android420 to ARGB.
+LIBYUV_API
+int Android420ToARGB(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     int src_pixel_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     int width,
+                     int height);
+
+// Convert Android420 to ABGR.
+LIBYUV_API
+int Android420ToABGR(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     int src_pixel_stride_uv,
+                     uint8_t* dst_abgr,
+                     int dst_stride_abgr,
+                     int width,
+                     int height);
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+LIBYUV_API
+int I420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+LIBYUV_API
+int H420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+LIBYUV_API
+int H420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+LIBYUV_API
+int J420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+LIBYUV_API
+int J420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+// Convert I422 to RGB24.
+LIBYUV_API
+int I422ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+// Convert I422 to RAW.
+LIBYUV_API
+int I422ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+LIBYUV_API
+int I420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+LIBYUV_API
+int J420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+LIBYUV_API
+int H420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+LIBYUV_API
+int I422ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+// Convert I420 To RGB565 with 4x4 dither matrix (16 bytes).
+// Values in dither matrix from 0 to 7 recommended.
+// The order of the dither matrix is first byte is upper left.
+
+LIBYUV_API
+int I420ToRGB565Dither(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const uint8_t* dither4x4,
+                       int width,
+                       int height);
+
+LIBYUV_API
+int I420ToARGB1555(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb1555,
+                   int dst_stride_argb1555,
+                   int width,
+                   int height);
+
+LIBYUV_API
+int I420ToARGB4444(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb4444,
+                   int dst_stride_argb4444,
+                   int width,
+                   int height);
+
+// Convert I420 to AR30.
+LIBYUV_API
+int I420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert I420 to AB30.
+LIBYUV_API
+int I420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert H420 to AR30.
+LIBYUV_API
+int H420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert H420 to AB30.
+LIBYUV_API
+int H420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert I420 to ARGB with matrix.
+LIBYUV_API
+int I420ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I422 to ARGB with matrix.
+LIBYUV_API
+int I422ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I444 to ARGB with matrix.
+LIBYUV_API
+int I444ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I444 to RGB24 with matrix.
+LIBYUV_API
+int I444ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert 10 bit 420 YUV to ARGB with matrix.
+LIBYUV_API
+int I010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 420 YUV to ARGB with matrix.
+LIBYUV_API
+int I210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 444 YUV to ARGB with matrix.
+LIBYUV_API
+int I410ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// multiply 12 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I012ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 12 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I012ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 422 YUV to ARGB with matrix.
+LIBYUV_API
+int I210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 444 YUV to ARGB with matrix.
+LIBYUV_API
+int I410ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P010 to ARGB with matrix.
+LIBYUV_API
+int P010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P210 to ARGB with matrix.
+LIBYUV_API
+int P210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P010 to AR30 with matrix.
+LIBYUV_API
+int P010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P210 to AR30 with matrix.
+LIBYUV_API
+int P210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// P012 and P010 use most significant bits so the conversion is the same.
+// Convert P012 to ARGB with matrix.
+#define P012ToARGBMatrix P010ToARGBMatrix
+// Convert P012 to AR30 with matrix.
+#define P012ToAR30Matrix P010ToAR30Matrix
+// Convert P212 to ARGB with matrix.
+#define P212ToARGBMatrix P210ToARGBMatrix
+// Convert P212 to AR30 with matrix.
+#define P212ToAR30Matrix P210ToAR30Matrix
+
+// Convert P016 to ARGB with matrix.
+#define P016ToARGBMatrix P010ToARGBMatrix
+// Convert P016 to AR30 with matrix.
+#define P016ToAR30Matrix P010ToAR30Matrix
+// Convert P216 to ARGB with matrix.
+#define P216ToARGBMatrix P210ToARGBMatrix
+// Convert P216 to AR30 with matrix.
+#define P216ToAR30Matrix P210ToAR30Matrix
+
+// Convert I420 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I420AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I422 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I422AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I444 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I444AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I010 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I010AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I210 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I210AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I410 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I410AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert NV12 to ARGB with matrix.
+LIBYUV_API
+int NV12ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert NV21 to ARGB with matrix.
+LIBYUV_API
+int NV21ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_vu,
+                     int src_stride_vu,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert NV12 to RGB565 with matrix.
+LIBYUV_API
+int NV12ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_uv,
+                       int src_stride_uv,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height);
+
+// Convert NV12 to RGB24 with matrix.
+LIBYUV_API
+int NV12ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_uv,
+                      int src_stride_uv,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert NV21 to RGB24 with matrix.
+LIBYUV_API
+int NV21ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_vu,
+                      int src_stride_vu,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert Android420 to ARGB with matrix.
+LIBYUV_API
+int Android420ToARGBMatrix(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height);
+
+// Convert I422 to RGBA with matrix.
+LIBYUV_API
+int I422ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I420 to RGBA with matrix.
+LIBYUV_API
+int I420ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I420 to RGB24 with matrix.
+LIBYUV_API
+int I420ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert I422 to RGB24 with matrix.
+LIBYUV_API
+int I422ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert I420 to RGB565 with specified color matrix.
+LIBYUV_API
+int I420ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height);
+
+// Convert I422 to RGB565 with specified color matrix.
+LIBYUV_API
+int I422ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height);
+
+// Convert I420 to AR30 with matrix.
+LIBYUV_API
+int I420ToAR30Matrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I400 (grey) to ARGB.  Reverse of ARGBToI400.
+LIBYUV_API
+int I400ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I420 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I420ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I422 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I422ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I422 to RGB24 with matrix and UV filter mode.
+LIBYUV_API
+int I422ToRGB24MatrixFilter(const uint8_t* src_y,
+                            int src_stride_y,
+                            const uint8_t* src_u,
+                            int src_stride_u,
+                            const uint8_t* src_v,
+                            int src_stride_v,
+                            uint8_t* dst_rgb24,
+                            int dst_stride_rgb24,
+                            const struct YuvConstants* yuvconstants,
+                            int width,
+                            int height,
+                            enum FilterMode filter);
+
+// Convert I420 to RGB24 with matrix and UV filter mode.
+LIBYUV_API
+int I420ToRGB24MatrixFilter(const uint8_t* src_y,
+                            int src_stride_y,
+                            const uint8_t* src_u,
+                            int src_stride_u,
+                            const uint8_t* src_v,
+                            int src_stride_v,
+                            uint8_t* dst_rgb24,
+                            int dst_stride_rgb24,
+                            const struct YuvConstants* yuvconstants,
+                            int width,
+                            int height,
+                            enum FilterMode filter);
+
+// Convert I010 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int I010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I210 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int I210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I010 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I210 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I420 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I420AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert I422 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I422AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert I010 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I010AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert I210 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I210AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert P010 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int P010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert P210 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int P210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert P010 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int P010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert P210 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int P210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
 
 // Convert camera sample to ARGB with cropping, rotation and vertical flip.
-// "src_size" is needed to parse MJPG.
+// "sample_size" is needed to parse MJPG.
 // "dst_stride_argb" number of bytes in a row of the dst_argb plane.
 //   Normally this would be the same as dst_width, with recommended alignment
 //   to 16 bytes for better efficiency.
@@ -292,20 +2289,25 @@ int MJPGToARGB(const uint8* sample, size_t sample_size,
 //    Must be less than or equal to src_width/src_height
 //    Cropping parameters are pre-rotation.
 // "rotation" can be 0, 90, 180 or 270.
-// "format" is a fourcc. ie 'I420', 'YUY2'
+// "fourcc" is a fourcc. ie 'I420', 'YUY2'
 // Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
 LIBYUV_API
-int ConvertToARGB(const uint8* src_frame, size_t src_size,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int crop_x, int crop_y,
-                  int src_width, int src_height,
-                  int crop_width, int crop_height,
+int ConvertToARGB(const uint8_t* sample,
+                  size_t sample_size,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int crop_x,
+                  int crop_y,
+                  int src_width,
+                  int src_height,
+                  int crop_width,
+                  int crop_height,
                   enum RotationMode rotation,
-                  uint32 format);
+                  uint32_t fourcc);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CONVERT_ARGB_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_CONVERT_ARGB_H_
diff --git a/media/libyuv/include/libyuv/convert_from.h b/media/libyuv/include/libyuv/convert_from.h
index 39e1578a0e..32f42a6330 100644
--- a/media/libyuv/include/libyuv/convert_from.h
+++ b/media/libyuv/include/libyuv/convert_from.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CONVERT_FROM_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_H_
 #define INCLUDE_LIBYUV_CONVERT_FROM_H_
 
 #include "libyuv/basic_types.h"
@@ -21,159 +21,183 @@ extern "C" {
 
 // See Also convert.h for conversions from formats to I420.
 
-// I420Copy in convert to I420ToI420.
-
-LIBYUV_API
-int I420ToI422(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
-
-LIBYUV_API
-int I420ToI444(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
-
-LIBYUV_API
-int I420ToI411(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+// Convert 8 bit YUV to 10 bit.
+#define H420ToH010 I420ToI010
+LIBYUV_API
+int I420ToI010(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert 8 bit YUV to 12 bit.
+#define H420ToH012 I420ToI012
+LIBYUV_API
+int I420ToI012(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToI422(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Copy to I400. Source can be I420, I422, I444, I400, NV12 or NV21.
 LIBYUV_API
-int I400Copy(const uint8* src_y, int src_stride_y,
-             uint8* dst_y, int dst_stride_y,
-             int width, int height);
-
-LIBYUV_API
-int I420ToNV12(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height);
-
-LIBYUV_API
-int I420ToNV21(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_vu, int dst_stride_vu,
-               int width, int height);
-
-LIBYUV_API
-int I420ToYUY2(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_frame, int dst_stride_frame,
-               int width, int height);
-
-LIBYUV_API
-int I420ToUYVY(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_frame, int dst_stride_frame,
-               int width, int height);
-
-LIBYUV_API
-int I420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
-
-LIBYUV_API
-int I420ToBGRA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
-
-LIBYUV_API
-int I420ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
-
-LIBYUV_API
-int I420ToRGBA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_rgba, int dst_stride_rgba,
-               int width, int height);
-
-LIBYUV_API
-int I420ToRGB24(const uint8* src_y, int src_stride_y,
-                const uint8* src_u, int src_stride_u,
-                const uint8* src_v, int src_stride_v,
-                uint8* dst_frame, int dst_stride_frame,
-                int width, int height);
-
-LIBYUV_API
-int I420ToRAW(const uint8* src_y, int src_stride_y,
-              const uint8* src_u, int src_stride_u,
-              const uint8* src_v, int src_stride_v,
-              uint8* dst_frame, int dst_stride_frame,
-              int width, int height);
-
-LIBYUV_API
-int I420ToRGB565(const uint8* src_y, int src_stride_y,
-                 const uint8* src_u, int src_stride_u,
-                 const uint8* src_v, int src_stride_v,
-                 uint8* dst_frame, int dst_stride_frame,
-                 int width, int height);
-
-// Convert I420 To RGB565 with 4x4 dither matrix (16 bytes).
-// Values in dither matrix from 0 to 7 recommended.
-// The order of the dither matrix is first byte is upper left.
-
-LIBYUV_API
-int I420ToRGB565Dither(const uint8* src_y, int src_stride_y,
-                       const uint8* src_u, int src_stride_u,
-                       const uint8* src_v, int src_stride_v,
-                       uint8* dst_frame, int dst_stride_frame,
-                       const uint8* dither4x4, int width, int height);
-
-LIBYUV_API
-int I420ToARGB1555(const uint8* src_y, int src_stride_y,
-                   const uint8* src_u, int src_stride_u,
-                   const uint8* src_v, int src_stride_v,
-                   uint8* dst_frame, int dst_stride_frame,
-                   int width, int height);
-
-LIBYUV_API
-int I420ToARGB4444(const uint8* src_y, int src_stride_y,
-                   const uint8* src_u, int src_stride_u,
-                   const uint8* src_v, int src_stride_v,
-                   uint8* dst_frame, int dst_stride_frame,
-                   int width, int height);
+int I400Copy(const uint8_t* src_y,
+             int src_stride_y,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             int width,
+             int height);
+
+LIBYUV_API
+int I420ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToYUY2(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToUYVY(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_uyvy,
+               int dst_stride_uyvy,
+               int width,
+               int height);
+
+// The following are from convert_argb.h
+// DEPRECATED: The prototypes will be removed in future.  Use convert_argb.h
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I420 to ABGR.
+LIBYUV_API
+int I420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
 
 // Convert I420 to specified format.
 // "dst_sample_stride" is bytes in a row for the destination. Pass 0 if the
 //    buffer has contiguous rows. Can be negative. A multiple of 16 is optimal.
 LIBYUV_API
-int ConvertFromI420(const uint8* y, int y_stride,
-                    const uint8* u, int u_stride,
-                    const uint8* v, int v_stride,
-                    uint8* dst_sample, int dst_sample_stride,
-                    int width, int height,
-                    uint32 format);
+int ConvertFromI420(const uint8_t* y,
+                    int y_stride,
+                    const uint8_t* u,
+                    int u_stride,
+                    const uint8_t* v,
+                    int v_stride,
+                    uint8_t* dst_sample,
+                    int dst_sample_stride,
+                    int width,
+                    int height,
+                    uint32_t fourcc);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CONVERT_FROM_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_H_
diff --git a/media/libyuv/include/libyuv/convert_from_argb.h b/media/libyuv/include/libyuv/convert_from_argb.h
index 1df53200dd..ff2a581ac4 100644
--- a/media/libyuv/include/libyuv/convert_from_argb.h
+++ b/media/libyuv/include/libyuv/convert_from_argb.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
 #define INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
 
 #include "libyuv/basic_types.h"
@@ -21,170 +21,361 @@ extern "C" {
 // Copy ARGB to ARGB.
 #define ARGBToARGB ARGBCopy
 LIBYUV_API
-int ARGBCopy(const uint8* src_argb, int src_stride_argb,
-             uint8* dst_argb, int dst_stride_argb,
-             int width, int height);
+int ARGBCopy(const uint8_t* src_argb,
+             int src_stride_argb,
+             uint8_t* dst_argb,
+             int dst_stride_argb,
+             int width,
+             int height);
 
 // Convert ARGB To BGRA.
 LIBYUV_API
-int ARGBToBGRA(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_bgra, int dst_stride_bgra,
-               int width, int height);
+int ARGBToBGRA(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height);
 
 // Convert ARGB To ABGR.
 LIBYUV_API
-int ARGBToABGR(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
+int ARGBToABGR(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
 
 // Convert ARGB To RGBA.
 LIBYUV_API
-int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_rgba, int dst_stride_rgba,
-               int width, int height);
+int ARGBToRGBA(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height);
+
+// Aliases
+#define ARGBToAB30 ABGRToAR30
+#define ABGRToAB30 ARGBToAR30
+
+// Convert ABGR To AR30.
+LIBYUV_API
+int ABGRToAR30(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert ARGB To AR30.
+LIBYUV_API
+int ARGBToAR30(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Aliases
+#define ABGRToRGB24 ARGBToRAW
+#define ABGRToRAW ARGBToRGB24
 
 // Convert ARGB To RGB24.
 LIBYUV_API
-int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
-                uint8* dst_rgb24, int dst_stride_rgb24,
-                int width, int height);
+int ARGBToRGB24(const uint8_t* src_argb,
+                int src_stride_argb,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
 
 // Convert ARGB To RAW.
 LIBYUV_API
-int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
-              uint8* dst_rgb, int dst_stride_rgb,
-              int width, int height);
+int ARGBToRAW(const uint8_t* src_argb,
+              int src_stride_argb,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
 
 // Convert ARGB To RGB565.
 LIBYUV_API
-int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
-                 uint8* dst_rgb565, int dst_stride_rgb565,
-                 int width, int height);
+int ARGBToRGB565(const uint8_t* src_argb,
+                 int src_stride_argb,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
 
 // Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes).
 // Values in dither matrix from 0 to 7 recommended.
 // The order of the dither matrix is first byte is upper left.
 // TODO(fbarchard): Consider pointer to 2d array for dither4x4.
-// const uint8(*dither)[4][4];
+// const uint8_t(*dither)[4][4];
 LIBYUV_API
-int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_rgb565, int dst_stride_rgb565,
-                       const uint8* dither4x4, int width, int height);
+int ARGBToRGB565Dither(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const uint8_t* dither4x4,
+                       int width,
+                       int height);
 
 // Convert ARGB To ARGB1555.
 LIBYUV_API
-int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_argb1555, int dst_stride_argb1555,
-                   int width, int height);
+int ARGBToARGB1555(const uint8_t* src_argb,
+                   int src_stride_argb,
+                   uint8_t* dst_argb1555,
+                   int dst_stride_argb1555,
+                   int width,
+                   int height);
 
 // Convert ARGB To ARGB4444.
 LIBYUV_API
-int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_argb4444, int dst_stride_argb4444,
-                   int width, int height);
+int ARGBToARGB4444(const uint8_t* src_argb,
+                   int src_stride_argb,
+                   uint8_t* dst_argb4444,
+                   int dst_stride_argb4444,
+                   int width,
+                   int height);
 
 // Convert ARGB To I444.
 LIBYUV_API
-int ARGBToI444(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToI444(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert ARGB to AR64.
+LIBYUV_API
+int ARGBToAR64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ar64,
+               int dst_stride_ar64,
+               int width,
+               int height);
+
+// Convert ABGR to AB64.
+#define ABGRToAB64 ARGBToAR64
+
+// Convert ARGB to AB64.
+LIBYUV_API
+int ARGBToAB64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height);
+
+// Convert ABGR to AR64.
+#define ABGRToAR64 ARGBToAB64
 
 // Convert ARGB To I422.
 LIBYUV_API
-int ARGBToI422(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToI422(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Convert ARGB To I420. (also in convert.h)
 LIBYUV_API
-int ARGBToI420(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToI420(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Convert ARGB to J420. (JPeg full range I420).
 LIBYUV_API
-int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yj, int dst_stride_yj,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToJ420(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height);
 
 // Convert ARGB to J422.
 LIBYUV_API
-int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yj, int dst_stride_yj,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToJ422(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height);
 
-// Convert ARGB To I411.
+// Convert ARGB to J400. (JPeg full range).
 LIBYUV_API
-int ARGBToI411(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int ARGBToJ400(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height);
 
-// Convert ARGB to J400. (JPeg full range).
+// Convert ABGR to J420. (JPeg full range I420).
+LIBYUV_API
+int ABGRToJ420(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height);
+
+// Convert ABGR to J422.
+LIBYUV_API
+int ABGRToJ422(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height);
+
+// Convert ABGR to J400. (JPeg full range).
+LIBYUV_API
+int ABGRToJ400(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height);
+
+// Convert RGBA to J400. (JPeg full range).
 LIBYUV_API
-int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yj, int dst_stride_yj,
-               int width, int height);
+int RGBAToJ400(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height);
 
 // Convert ARGB to I400.
 LIBYUV_API
-int ARGBToI400(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height);
+int ARGBToI400(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height);
 
 // Convert ARGB to G. (Reverse of J400toARGB, which replicates G back to ARGB)
 LIBYUV_API
-int ARGBToG(const uint8* src_argb, int src_stride_argb,
-            uint8* dst_g, int dst_stride_g,
-            int width, int height);
+int ARGBToG(const uint8_t* src_argb,
+            int src_stride_argb,
+            uint8_t* dst_g,
+            int dst_stride_g,
+            int width,
+            int height);
 
 // Convert ARGB To NV12.
 LIBYUV_API
-int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height);
+int ARGBToNV12(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
 
 // Convert ARGB To NV21.
 LIBYUV_API
-int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_vu, int dst_stride_vu,
-               int width, int height);
+int ARGBToNV21(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
 
-// Convert ARGB To NV21.
+// Convert ABGR To NV12.
+LIBYUV_API
+int ABGRToNV12(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert ABGR To NV21.
 LIBYUV_API
-int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_vu, int dst_stride_vu,
-               int width, int height);
+int ABGRToNV21(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
 
 // Convert ARGB To YUY2.
 LIBYUV_API
-int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yuy2, int dst_stride_yuy2,
-               int width, int height);
+int ARGBToYUY2(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height);
 
 // Convert ARGB To UYVY.
 LIBYUV_API
-int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_uyvy, int dst_stride_uyvy,
-               int width, int height);
+int ARGBToUYVY(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_uyvy,
+               int dst_stride_uyvy,
+               int width,
+               int height);
+
+// RAW to JNV21 full range NV21
+LIBYUV_API
+int RAWToJNV21(const uint8_t* src_raw,
+               int src_stride_raw,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
diff --git a/media/libyuv/include/libyuv/cpu_id.h b/media/libyuv/include/libyuv/cpu_id.h
index dfb7445e2f..fb90c6c737 100644
--- a/media/libyuv/include/libyuv/cpu_id.h
+++ b/media/libyuv/include/libyuv/cpu_id.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CPU_ID_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_CPU_ID_H_
 #define INCLUDE_LIBYUV_CPU_ID_H_
 
 #include "libyuv/basic_types.h"
@@ -31,50 +31,97 @@ static const int kCpuHasX86 = 0x10;
 static const int kCpuHasSSE2 = 0x20;
 static const int kCpuHasSSSE3 = 0x40;
 static const int kCpuHasSSE41 = 0x80;
-static const int kCpuHasSSE42 = 0x100;
+static const int kCpuHasSSE42 = 0x100;  // unused at this time.
 static const int kCpuHasAVX = 0x200;
 static const int kCpuHasAVX2 = 0x400;
 static const int kCpuHasERMS = 0x800;
 static const int kCpuHasFMA3 = 0x1000;
-static const int kCpuHasAVX3 = 0x2000;
-// 0x2000, 0x4000, 0x8000 reserved for future X86 flags.
+static const int kCpuHasF16C = 0x2000;
+static const int kCpuHasGFNI = 0x4000;
+static const int kCpuHasAVX512BW = 0x8000;
+static const int kCpuHasAVX512VL = 0x10000;
+static const int kCpuHasAVX512VNNI = 0x20000;
+static const int kCpuHasAVX512VBMI = 0x40000;
+static const int kCpuHasAVX512VBMI2 = 0x80000;
+static const int kCpuHasAVX512VBITALG = 0x100000;
+static const int kCpuHasAVX512VPOPCNTDQ = 0x200000;
 
 // These flags are only valid on MIPS processors.
-static const int kCpuHasMIPS = 0x10000;
-static const int kCpuHasDSPR2 = 0x20000;
+static const int kCpuHasMIPS = 0x400000;
+static const int kCpuHasMSA = 0x800000;
 
-// Internal function used to auto-init.
-LIBYUV_API
-int InitCpuFlags(void);
+// These flags are only valid on LOONGARCH processors.
+static const int kCpuHasLOONGARCH = 0x2000000;
+static const int kCpuHasLSX = 0x4000000;
+static const int kCpuHasLASX = 0x8000000;
 
-// Internal function for parsing /proc/cpuinfo.
+// Optional init function. TestCpuFlag does an auto-init.
+// Returns cpu_info flags.
 LIBYUV_API
-int ArmCpuCaps(const char* cpuinfo_name);
+int InitCpuFlags(void);
 
 // Detect CPU has SSE2 etc.
 // Test_flag parameter should be one of kCpuHas constants above.
-// returns non-zero if instruction set is detected
+// Returns non-zero if instruction set is detected
 static __inline int TestCpuFlag(int test_flag) {
   LIBYUV_API extern int cpu_info_;
-  return (!cpu_info_ ? InitCpuFlags() : cpu_info_) & test_flag;
+#ifdef __ATOMIC_RELAXED
+  int cpu_info = __atomic_load_n(&cpu_info_, __ATOMIC_RELAXED);
+#else
+  int cpu_info = cpu_info_;
+#endif
+  return (!cpu_info ? InitCpuFlags() : cpu_info) & test_flag;
 }
 
+// Internal function for parsing /proc/cpuinfo.
+LIBYUV_API
+int ArmCpuCaps(const char* cpuinfo_name);
+LIBYUV_API
+int MipsCpuCaps(const char* cpuinfo_name);
+
 // For testing, allow CPU flags to be disabled.
 // ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
 // MaskCpuFlags(-1) to enable all cpu specific optimizations.
 // MaskCpuFlags(1) to disable all cpu specific optimizations.
+// MaskCpuFlags(0) to reset state so next call will auto init.
+// Returns cpu_info flags.
 LIBYUV_API
-void MaskCpuFlags(int enable_flags);
+int MaskCpuFlags(int enable_flags);
+
+// Sets the CPU flags to |cpu_flags|, bypassing the detection code. |cpu_flags|
+// should be a valid combination of the kCpuHas constants above and include
+// kCpuInitialized. Use this method when running in a sandboxed process where
+// the detection code might fail (as it might access /proc/cpuinfo). In such
+// cases the cpu_info can be obtained from a non sandboxed process by calling
+// InitCpuFlags() and passed to the sandboxed process (via command line
+// parameters, IPC...) which can then call this method to initialize the CPU
+// flags.
+// Notes:
+// - when specifying 0 for |cpu_flags|, the auto initialization is enabled
+//   again.
+// - enabling CPU features that are not supported by the CPU will result in
+//   undefined behavior.
+// TODO(fbarchard): consider writing a helper function that translates from
+// other library CPU info to libyuv CPU info and add a .md doc that explains
+// CPU detection.
+static __inline void SetCpuFlags(int cpu_flags) {
+  LIBYUV_API extern int cpu_info_;
+#ifdef __ATOMIC_RELAXED
+  __atomic_store_n(&cpu_info_, cpu_flags, __ATOMIC_RELAXED);
+#else
+  cpu_info_ = cpu_flags;
+#endif
+}
 
 // Low level cpuid for X86. Returns zeros on other CPUs.
 // eax is the info type that you want.
 // ecx is typically the cpu number, and should normally be zero.
 LIBYUV_API
-void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info);
+void CpuId(int info_eax, int info_ecx, int* cpu_info);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CPU_ID_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_CPU_ID_H_
diff --git a/media/libyuv/include/libyuv/loongson_intrinsics.h b/media/libyuv/include/libyuv/loongson_intrinsics.h
new file mode 100644
index 0000000000..1d613defb1
--- /dev/null
+++ b/media/libyuv/include/libyuv/loongson_intrinsics.h
@@ -0,0 +1,1949 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_LOONGSON_INTRINSICS_H
+#define INCLUDE_LIBYUV_LOONGSON_INTRINSICS_H
+
+/*
+ * Copyright (c) 2022 Loongson Technology Corporation Limited
+ * All rights reserved.
+ * Contributed by Shiyou Yin <yinshiyou-hf@loongson.cn>
+ *                Xiwei Gu   <guxiwei-hf@loongson.cn>
+ *                Lu Wang    <wanglu@loongson.cn>
+ *
+ * This file is a header file for loongarch builtin extension.
+ *
+ */
+
+#ifndef LOONGSON_INTRINSICS_H
+#define LOONGSON_INTRINSICS_H
+
+/**
+ * MAJOR version: Macro usage changes.
+ * MINOR version: Add new functions, or bug fixes.
+ * MICRO version: Comment changes or implementation changes.
+ */
+#define LSOM_VERSION_MAJOR 1
+#define LSOM_VERSION_MINOR 1
+#define LSOM_VERSION_MICRO 0
+
+#define DUP2_ARG1(_INS, _IN0, _IN1, _OUT0, _OUT1) \
+  {                                               \
+    _OUT0 = _INS(_IN0);                           \
+    _OUT1 = _INS(_IN1);                           \
+  }
+
+#define DUP2_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1) \
+  {                                                           \
+    _OUT0 = _INS(_IN0, _IN1);                                 \
+    _OUT1 = _INS(_IN2, _IN3);                                 \
+  }
+
+#define DUP2_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _OUT0, _OUT1) \
+  {                                                                       \
+    _OUT0 = _INS(_IN0, _IN1, _IN2);                                       \
+    _OUT1 = _INS(_IN3, _IN4, _IN5);                                       \
+  }
+
+#define DUP4_ARG1(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1, _OUT2, _OUT3) \
+  {                                                                         \
+    DUP2_ARG1(_INS, _IN0, _IN1, _OUT0, _OUT1);                              \
+    DUP2_ARG1(_INS, _IN2, _IN3, _OUT2, _OUT3);                              \
+  }
+
+#define DUP4_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _IN6, _IN7, _OUT0, \
+                  _OUT1, _OUT2, _OUT3)                                         \
+  {                                                                            \
+    DUP2_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1);                     \
+    DUP2_ARG2(_INS, _IN4, _IN5, _IN6, _IN7, _OUT2, _OUT3);                     \
+  }
+
+#define DUP4_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _IN6, _IN7, _IN8, \
+                  _IN9, _IN10, _IN11, _OUT0, _OUT1, _OUT2, _OUT3)             \
+  {                                                                           \
+    DUP2_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _OUT0, _OUT1);        \
+    DUP2_ARG3(_INS, _IN6, _IN7, _IN8, _IN9, _IN10, _IN11, _OUT2, _OUT3);      \
+  }
+
+#ifdef __loongarch_sx
+#include <lsxintrin.h>
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               Then the results plus to signed half-word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_b(__m128i in_c,
+                                        __m128i in_h,
+                                        __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_h_b(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               unsigned byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               The results plus to signed half-word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_bu(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_bu(__m128i in_c,
+                                         __m128i in_h,
+                                         __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_h_bu(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               The results plus to signed half-word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_bu_b(in_c, in_h, in_l)
+ *        in_c : 1,1,1,1, 1,1,1,1
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : -1,-2,-3,-4, -5,-6,-7,-8, 1,2,3,4, 5,6,7,8
+ *         out : -4,-24,-60,-112, 6,26,62,114
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_bu_b(__m128i in_c,
+                                           __m128i in_h,
+                                           __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_h_bu_b(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_h_bu_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of half-word vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - __m128i
+ * Details     : Signed half-word elements from in_h are multiplied by
+ *               signed half-word elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               Then the results plus to signed word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1
+ *         out : 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_w_h(__m128i in_c,
+                                        __m128i in_h,
+                                        __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_w_h(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_h_b(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_b(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_h_b(in_h, in_l);
+  out = __lsx_vmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               unsigned byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_h_bu(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_bu(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_h_bu(in_h, in_l);
+  out = __lsx_vmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_h_bu_b(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,-1
+ *         out : 22,38,38,22, 22,38,38,6
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_bu_b(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_h_bu_b(in_h, in_l);
+  out = __lsx_vmaddwod_h_bu_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_w_h(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_w_h(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_w_h(in_h, in_l);
+  out = __lsx_vmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all halfword elements of input vector between min & max
+ *               out = ((_in) < (min)) ? (min) : (((_in) > (max)) ? (max) :
+ *               (_in))
+ * Arguments   : Inputs  - _in  (input vector)
+ *                       - min  (min threshold)
+ *                       - max  (max threshold)
+ *               Outputs - out  (output vector with clipped elements)
+ *               Return Type - signed halfword
+ * Example     : out = __lsx_vclip_h(_in)
+ *         _in : -8,2,280,249, -8,255,280,249
+ *         min : 1,1,1,1, 1,1,1,1
+ *         max : 9,9,9,9, 9,9,9,9
+ *         out : 1,2,9,9, 1,9,9,9
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip_h(__m128i _in, __m128i min, __m128i max) {
+  __m128i out;
+
+  out = __lsx_vmax_h(min, _in);
+  out = __lsx_vmin_h(max, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Set each element of vector between 0 and 255
+ * Arguments   : Inputs  - _in
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from _in are clamped between 0 and 255.
+ * Example     : out = __lsx_vclip255_h(_in)
+ *         _in : -8,255,280,249, -8,255,280,249
+ *         out : 0,255,255,249, 0,255,255,249
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip255_h(__m128i _in) {
+  __m128i out;
+
+  out = __lsx_vmaxi_h(_in, 0);
+  out = __lsx_vsat_hu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Set each element of vector between 0 and 255
+ * Arguments   : Inputs  - _in
+ *               Outputs - out
+ *               Return Type - word
+ * Details     : Signed byte elements from _in are clamped between 0 and 255.
+ * Example     : out = __lsx_vclip255_w(_in)
+ *         _in : -8,255,280,249
+ *         out : 0,255,255,249
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip255_w(__m128i _in) {
+  __m128i out;
+
+  out = __lsx_vmaxi_w(_in, 0);
+  out = __lsx_vsat_wu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Swap two variables
+ * Arguments   : Inputs  - _in0, _in1
+ *               Outputs - _in0, _in1 (in-place)
+ * Details     : Swapping of two input variables using xor
+ * Example     : LSX_SWAP(_in0, _in1)
+ *        _in0 : 1,2,3,4
+ *        _in1 : 5,6,7,8
+ *   _in0(out) : 5,6,7,8
+ *   _in1(out) : 1,2,3,4
+ * =============================================================================
+ */
+#define LSX_SWAP(_in0, _in1)         \
+  {                                  \
+    _in0 = __lsx_vxor_v(_in0, _in1); \
+    _in1 = __lsx_vxor_v(_in0, _in1); \
+    _in0 = __lsx_vxor_v(_in0, _in1); \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with word elements in vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3
+ *               Outputs - out0, out1, out2, out3
+ * Details     :
+ * Example     :
+ *               1, 2, 3, 4            1, 5, 9,13
+ *               5, 6, 7, 8    to      2, 6,10,14
+ *               9,10,11,12  =====>    3, 7,11,15
+ *              13,14,15,16            4, 8,12,16
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE4x4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    __m128i _t0, _t1, _t2, _t3;                                                \
+                                                                               \
+    _t0 = __lsx_vilvl_w(_in1, _in0);                                           \
+    _t1 = __lsx_vilvh_w(_in1, _in0);                                           \
+    _t2 = __lsx_vilvl_w(_in3, _in2);                                           \
+    _t3 = __lsx_vilvh_w(_in3, _in2);                                           \
+    _out0 = __lsx_vilvl_d(_t2, _t0);                                           \
+    _out1 = __lsx_vilvh_d(_t2, _t0);                                           \
+    _out2 = __lsx_vilvl_d(_t3, _t1);                                           \
+    _out3 = __lsx_vilvh_d(_t3, _t1);                                           \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with byte elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *               _out7
+ * Details     : The rows of the matrix become columns, and the columns
+ *               become rows.
+ * Example     : LSX_TRANSPOSE8x8_B
+ *        _in0 : 00,01,02,03,04,05,06,07, 00,00,00,00,00,00,00,00
+ *        _in1 : 10,11,12,13,14,15,16,17, 00,00,00,00,00,00,00,00
+ *        _in2 : 20,21,22,23,24,25,26,27, 00,00,00,00,00,00,00,00
+ *        _in3 : 30,31,32,33,34,35,36,37, 00,00,00,00,00,00,00,00
+ *        _in4 : 40,41,42,43,44,45,46,47, 00,00,00,00,00,00,00,00
+ *        _in5 : 50,51,52,53,54,55,56,57, 00,00,00,00,00,00,00,00
+ *        _in6 : 60,61,62,63,64,65,66,67, 00,00,00,00,00,00,00,00
+ *        _in7 : 70,71,72,73,74,75,76,77, 00,00,00,00,00,00,00,00
+ *
+ *      _ out0 : 00,10,20,30,40,50,60,70, 00,00,00,00,00,00,00,00
+ *      _ out1 : 01,11,21,31,41,51,61,71, 00,00,00,00,00,00,00,00
+ *      _ out2 : 02,12,22,32,42,52,62,72, 00,00,00,00,00,00,00,00
+ *      _ out3 : 03,13,23,33,43,53,63,73, 00,00,00,00,00,00,00,00
+ *      _ out4 : 04,14,24,34,44,54,64,74, 00,00,00,00,00,00,00,00
+ *      _ out5 : 05,15,25,35,45,55,65,75, 00,00,00,00,00,00,00,00
+ *      _ out6 : 06,16,26,36,46,56,66,76, 00,00,00,00,00,00,00,00
+ *      _ out7 : 07,17,27,37,47,57,67,77, 00,00,00,00,00,00,00,00
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    __m128i zero = {0};                                                     \
+    __m128i shuf8 = {0x0F0E0D0C0B0A0908, 0x1716151413121110};               \
+    __m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;                         \
+                                                                            \
+    _t0 = __lsx_vilvl_b(_in2, _in0);                                        \
+    _t1 = __lsx_vilvl_b(_in3, _in1);                                        \
+    _t2 = __lsx_vilvl_b(_in6, _in4);                                        \
+    _t3 = __lsx_vilvl_b(_in7, _in5);                                        \
+    _t4 = __lsx_vilvl_b(_t1, _t0);                                          \
+    _t5 = __lsx_vilvh_b(_t1, _t0);                                          \
+    _t6 = __lsx_vilvl_b(_t3, _t2);                                          \
+    _t7 = __lsx_vilvh_b(_t3, _t2);                                          \
+    _out0 = __lsx_vilvl_w(_t6, _t4);                                        \
+    _out2 = __lsx_vilvh_w(_t6, _t4);                                        \
+    _out4 = __lsx_vilvl_w(_t7, _t5);                                        \
+    _out6 = __lsx_vilvh_w(_t7, _t5);                                        \
+    _out1 = __lsx_vshuf_b(zero, _out0, shuf8);                              \
+    _out3 = __lsx_vshuf_b(zero, _out2, shuf8);                              \
+    _out5 = __lsx_vshuf_b(zero, _out4, shuf8);                              \
+    _out7 = __lsx_vshuf_b(zero, _out6, shuf8);                              \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with half-word elements in vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+ *               Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+ * Details     :
+ * Example     :
+ *              00,01,02,03,04,05,06,07           00,10,20,30,40,50,60,70
+ *              10,11,12,13,14,15,16,17           01,11,21,31,41,51,61,71
+ *              20,21,22,23,24,25,26,27           02,12,22,32,42,52,62,72
+ *              30,31,32,33,34,35,36,37    to     03,13,23,33,43,53,63,73
+ *              40,41,42,43,44,45,46,47  ======>  04,14,24,34,44,54,64,74
+ *              50,51,52,53,54,55,56,57           05,15,25,35,45,55,65,75
+ *              60,61,62,63,64,65,66,67           06,16,26,36,46,56,66,76
+ *              70,71,72,73,74,75,76,77           07,17,27,37,47,57,67,77
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    __m128i _s0, _s1, _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;               \
+                                                                            \
+    _s0 = __lsx_vilvl_h(_in6, _in4);                                        \
+    _s1 = __lsx_vilvl_h(_in7, _in5);                                        \
+    _t0 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t1 = __lsx_vilvh_h(_s1, _s0);                                          \
+    _s0 = __lsx_vilvh_h(_in6, _in4);                                        \
+    _s1 = __lsx_vilvh_h(_in7, _in5);                                        \
+    _t2 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t3 = __lsx_vilvh_h(_s1, _s0);                                          \
+    _s0 = __lsx_vilvl_h(_in2, _in0);                                        \
+    _s1 = __lsx_vilvl_h(_in3, _in1);                                        \
+    _t4 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t5 = __lsx_vilvh_h(_s1, _s0);                                          \
+    _s0 = __lsx_vilvh_h(_in2, _in0);                                        \
+    _s1 = __lsx_vilvh_h(_in3, _in1);                                        \
+    _t6 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t7 = __lsx_vilvh_h(_s1, _s0);                                          \
+                                                                            \
+    _out0 = __lsx_vpickev_d(_t0, _t4);                                      \
+    _out2 = __lsx_vpickev_d(_t1, _t5);                                      \
+    _out4 = __lsx_vpickev_d(_t2, _t6);                                      \
+    _out6 = __lsx_vpickev_d(_t3, _t7);                                      \
+    _out1 = __lsx_vpickod_d(_t0, _t4);                                      \
+    _out3 = __lsx_vpickod_d(_t1, _t5);                                      \
+    _out5 = __lsx_vpickod_d(_t2, _t6);                                      \
+    _out7 = __lsx_vpickod_d(_t3, _t7);                                      \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 8x4 byte block into 4x8
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3      (input 8x4 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3  (output 4x8 byte block)
+ *               Return Type - as per RTYPE
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : LSX_TRANSPOSE8x4_B
+ *        _in0 : 00,01,02,03,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in1 : 10,11,12,13,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in2 : 20,21,22,23,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in3 : 30,31,32,33,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in4 : 40,41,42,43,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in5 : 50,51,52,53,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in6 : 60,61,62,63,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in7 : 70,71,72,73,00,00,00,00, 00,00,00,00,00,00,00,00
+ *
+ *       _out0 : 00,10,20,30,40,50,60,70, 00,00,00,00,00,00,00,00
+ *       _out1 : 01,11,21,31,41,51,61,71, 00,00,00,00,00,00,00,00
+ *       _out2 : 02,12,22,32,42,52,62,72, 00,00,00,00,00,00,00,00
+ *       _out3 : 03,13,23,33,43,53,63,73, 00,00,00,00,00,00,00,00
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x4_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+                           _out0, _out1, _out2, _out3)                     \
+  {                                                                        \
+    __m128i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                            \
+                                                                           \
+    _tmp0_m = __lsx_vpackev_w(_in4, _in0);                                 \
+    _tmp1_m = __lsx_vpackev_w(_in5, _in1);                                 \
+    _tmp2_m = __lsx_vilvl_b(_tmp1_m, _tmp0_m);                             \
+    _tmp0_m = __lsx_vpackev_w(_in6, _in2);                                 \
+    _tmp1_m = __lsx_vpackev_w(_in7, _in3);                                 \
+                                                                           \
+    _tmp3_m = __lsx_vilvl_b(_tmp1_m, _tmp0_m);                             \
+    _tmp0_m = __lsx_vilvl_h(_tmp3_m, _tmp2_m);                             \
+    _tmp1_m = __lsx_vilvh_h(_tmp3_m, _tmp2_m);                             \
+                                                                           \
+    _out0 = __lsx_vilvl_w(_tmp1_m, _tmp0_m);                               \
+    _out2 = __lsx_vilvh_w(_tmp1_m, _tmp0_m);                               \
+    _out1 = __lsx_vilvh_d(_out2, _out0);                                   \
+    _out3 = __lsx_vilvh_d(_out0, _out2);                                   \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 16x8 block with byte elements in vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7, in8
+ *                         in9, in10, in11, in12, in13, in14, in15
+ *               Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+ * Details     :
+ * Example     :
+ *              000,001,002,003,004,005,006,007
+ *              008,009,010,011,012,013,014,015
+ *              016,017,018,019,020,021,022,023
+ *              024,025,026,027,028,029,030,031
+ *              032,033,034,035,036,037,038,039
+ *              040,041,042,043,044,045,046,047        000,008,...,112,120
+ *              048,049,050,051,052,053,054,055        001,009,...,113,121
+ *              056,057,058,059,060,061,062,063   to   002,010,...,114,122
+ *              064,068,066,067,068,069,070,071 =====> 003,011,...,115,123
+ *              072,073,074,075,076,077,078,079        004,012,...,116,124
+ *              080,081,082,083,084,085,086,087        005,013,...,117,125
+ *              088,089,090,091,092,093,094,095        006,014,...,118,126
+ *              096,097,098,099,100,101,102,103        007,015,...,119,127
+ *              104,105,106,107,108,109,110,111
+ *              112,113,114,115,116,117,118,119
+ *              120,121,122,123,124,125,126,127
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE16x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _in8, _in9, _in10, _in11, _in12, _in13, _in14,   \
+                            _in15, _out0, _out1, _out2, _out3, _out4, _out5, \
+                            _out6, _out7)                                    \
+  {                                                                          \
+    __m128i _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5, _tmp6, _tmp7;          \
+    __m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;                          \
+    DUP4_ARG2(__lsx_vilvl_b, _in2, _in0, _in3, _in1, _in6, _in4, _in7, _in5, \
+              _tmp0, _tmp1, _tmp2, _tmp3);                                   \
+    DUP4_ARG2(__lsx_vilvl_b, _in10, _in8, _in11, _in9, _in14, _in12, _in15,  \
+              _in13, _tmp4, _tmp5, _tmp6, _tmp7);                            \
+    DUP2_ARG2(__lsx_vilvl_b, _tmp1, _tmp0, _tmp3, _tmp2, _t0, _t2);          \
+    DUP2_ARG2(__lsx_vilvh_b, _tmp1, _tmp0, _tmp3, _tmp2, _t1, _t3);          \
+    DUP2_ARG2(__lsx_vilvl_b, _tmp5, _tmp4, _tmp7, _tmp6, _t4, _t6);          \
+    DUP2_ARG2(__lsx_vilvh_b, _tmp5, _tmp4, _tmp7, _tmp6, _t5, _t7);          \
+    DUP2_ARG2(__lsx_vilvl_w, _t2, _t0, _t3, _t1, _tmp0, _tmp4);              \
+    DUP2_ARG2(__lsx_vilvh_w, _t2, _t0, _t3, _t1, _tmp2, _tmp6);              \
+    DUP2_ARG2(__lsx_vilvl_w, _t6, _t4, _t7, _t5, _tmp1, _tmp5);              \
+    DUP2_ARG2(__lsx_vilvh_w, _t6, _t4, _t7, _t5, _tmp3, _tmp7);              \
+    DUP2_ARG2(__lsx_vilvl_d, _tmp1, _tmp0, _tmp3, _tmp2, _out0, _out2);      \
+    DUP2_ARG2(__lsx_vilvh_d, _tmp1, _tmp0, _tmp3, _tmp2, _out1, _out3);      \
+    DUP2_ARG2(__lsx_vilvl_d, _tmp5, _tmp4, _tmp7, _tmp6, _out4, _out6);      \
+    DUP2_ARG2(__lsx_vilvh_d, _tmp5, _tmp4, _tmp7, _tmp6, _out5, _out7);      \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 4 input vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3
+ *               Outputs - out0, out1, out2, out3
+ * Details     : Butterfly operation
+ * Example     :
+ *               out0 = in0 + in3;
+ *               out1 = in1 + in2;
+ *               out2 = in1 - in2;
+ *               out3 = in0 - in3;
+ * =============================================================================
+ */
+#define LSX_BUTTERFLY_4_B(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_b(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_b(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_b(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_b(_in0, _in3);                                         \
+  }
+#define LSX_BUTTERFLY_4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_h(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_h(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_h(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_h(_in0, _in3);                                         \
+  }
+#define LSX_BUTTERFLY_4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_w(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_w(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_w(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_w(_in0, _in3);                                         \
+  }
+#define LSX_BUTTERFLY_4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_d(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_d(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_d(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_d(_in0, _in3);                                         \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 8 input vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, ~
+ *               Outputs - _out0, _out1, _out2, _out3, ~
+ * Details     : Butterfly operation
+ * Example     :
+ *              _out0 = _in0 + _in7;
+ *              _out1 = _in1 + _in6;
+ *              _out2 = _in2 + _in5;
+ *              _out3 = _in3 + _in4;
+ *              _out4 = _in3 - _in4;
+ *              _out5 = _in2 - _in5;
+ *              _out6 = _in1 - _in6;
+ *              _out7 = _in0 - _in7;
+ * =============================================================================
+ */
+#define LSX_BUTTERFLY_8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_b(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_b(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_b(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_b(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_b(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_b(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_b(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_b(_in0, _in7);                                      \
+  }
+
+#define LSX_BUTTERFLY_8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_h(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_h(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_h(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_h(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_h(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_h(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_h(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_h(_in0, _in7);                                      \
+  }
+
+#define LSX_BUTTERFLY_8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_w(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_w(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_w(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_w(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_w(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_w(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_w(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_w(_in0, _in7);                                      \
+  }
+
+#define LSX_BUTTERFLY_8_D(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_d(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_d(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_d(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_d(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_d(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_d(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_d(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_d(_in0, _in7);                                      \
+  }
+
+#endif  // LSX
+
+#ifdef __loongarch_asx
+#include <lasxintrin.h>
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               unsigned byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the out vector
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_h_bu(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_h_bu(in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed halfword
+ * Details     : Signed byte elements from in_h are multiplied with
+ *               signed byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplication results of adjacent odd-even elements
+ *               are added to the out vector
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_h_b(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_h_b(in_h, in_l);
+  out = __lasx_xvmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the out vector.
+ * Example     : out = __lasx_xvdp2_w_h(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of word vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed double
+ * Details     : Signed word elements from in_h are multiplied with
+ *               signed word elements from in_l producing a result
+ *               twice the size of input i.e. signed double-word.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the out vector.
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_d_w(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_d_w(in_h, in_l);
+  out = __lasx_xvmaddwod_d_w(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Unsigned halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. unsigned word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the out vector
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_w_hu_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_hu_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_hu_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied with
+ *               signed byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_b(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_h_b(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               unsigned byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_bu(__m256i in_c,
+                                           __m256i in_h,
+                                           __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_h_bu(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               signed byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_bu_b(__m256i in_c,
+                                             __m256i in_h,
+                                             __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_h_bu_b(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - per RTYPE
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8,
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1,
+ *         out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_w_h(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Unsigned halfword elements from in_h are multiplied with
+ *               unsigned halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_hu(__m256i in_c,
+                                           __m256i in_h,
+                                           __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_w_hu(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_w_hu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Unsigned halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the in_c vector
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_hu_h(__m256i in_c,
+                                             __m256i in_h,
+                                             __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_w_hu_h(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_w_hu_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Vector Unsigned Dot Product and Subtract
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               unsigned byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added together and subtracted from double width elements
+ *               in_c vector.
+ * Example     : See out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2sub_h_bu(__m256i in_c,
+                                           __m256i in_h,
+                                           __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_h_bu(in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+  out = __lasx_xvsub_h(in_c, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Vector Signed Dot Product and Subtract
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               Signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added together and subtracted from double width elements
+ *               in_c vector.
+ * Example     : out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
+ *        in_c : 0,0,0,0, 0,0,0,0
+ *        in_h : 3,1,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1
+ *        in_l : 2,1,1,0, 1,0,0,0, 0,0,1,0, 1,0,0,1
+ *         out : -7,-3,0,0, 0,-1,0,-1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2sub_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  out = __lasx_xvsub_w(in_c, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               four times the size of input i.e. signed doubleword.
+ *               Then this multiplication results of four adjacent elements
+ *               are added together and stored to the out vector.
+ * Example     : out = __lasx_xvdp4_d_h(in_h, in_l)
+ *        in_h :  3,1,3,0, 0,0,0,1, 0,0,1,-1, 0,0,0,1
+ *        in_l : -2,1,1,0, 1,0,0,0, 0,0,1, 0, 1,0,0,1
+ *         out : -2,0,1,1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp4_d_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  out = __lasx_xvhaddw_d_w(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The high half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               higher half of the two-fold sign extension (signed byte
+ *               to signed halfword) and stored to the out vector.
+ * Example     : See out = __lasx_xvaddwh_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwh_h_b(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvh_b(in_h, in_l);
+  out = __lasx_xvhaddw_h_b(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The high half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               higher half of the two-fold sign extension (signed halfword
+ *               to signed word) and stored to the out vector.
+ * Example     : out = __lasx_xvaddwh_w_h(in_h, in_l)
+ *        in_h : 3, 0,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 1,0,1, 0, 1,0,0,1
+ *         out : 1,0,0,-1, 1,0,0, 2
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwh_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvh_h(in_h, in_l);
+  out = __lasx_xvhaddw_w_h(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               lower half of the two-fold sign extension (signed byte
+ *               to signed halfword) and stored to the out vector.
+ * Example     : See out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_h_b(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvl_b(in_h, in_l);
+  out = __lasx_xvhaddw_h_b(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               lower half of the two-fold sign extension (signed halfword
+ *               to signed word) and stored to the out vector.
+ * Example     : out = __lasx_xvaddwl_w_h(in_h, in_l)
+ *        in_h : 3, 0,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 1,0,1, 0, 1,0,0,1
+ *         out : 5,-1,4,2, 1,0,2,-1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvl_h(in_h, in_l);
+  out = __lasx_xvhaddw_w_h(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The out vector and the out vector are added after the
+ *               lower half of the two-fold zero extension (unsigned byte
+ *               to unsigned halfword) and stored to the out vector.
+ * Example     : See out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_h_bu(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvl_b(in_h, in_l);
+  out = __lasx_xvhaddw_hu_bu(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_l vector after double zero extension (unsigned byte to
+ *               signed halfword)，added to the in_h vector.
+ * Example     : See out = __lasx_xvaddw_w_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddw_h_h_bu(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvsllwil_hu_bu(in_l, 0);
+  out = __lasx_xvadd_h(in_h, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_l vector after double sign extension (signed halfword to
+ *               signed word), added to the in_h vector.
+ * Example     : out = __lasx_xvaddw_w_w_h(in_h, in_l)
+ *        in_h : 0, 1,0,0, -1,0,0,1,
+ *        in_l : 2,-1,1,2,  1,0,0,0, 0,0,1,0, 1,0,0,1,
+ *         out : 2, 0,1,2, -1,0,1,1,
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddw_w_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvsllwil_w_h(in_l, 0);
+  out = __lasx_xvadd_w(in_h, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication and addition calculation after expansion
+ *               of the lower half of the vector.
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the lower half of the two-fold sign extension (signed halfword
+ *               to signed word), and the result is added to the vector in_c,
+ *               then stored to the out vector.
+ * Example     : out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 5,6,7,8
+ *        in_h : 1,2,3,4, 1,2,3,4, 5,6,7,8, 5,6,7,8
+ *        in_l : 200, 300, 400, 500,  2000, 3000, 4000, 5000,
+ *              -200,-300,-400,-500, -2000,-3000,-4000,-5000
+ *         out : 201, 602,1203,2004, -995, -1794,-2793,-3992
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmaddwl_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvsllwil_w_h(in_h, 0);
+  tmp1 = __lasx_xvsllwil_w_h(in_l, 0);
+  tmp0 = __lasx_xvmul_w(tmp0, tmp1);
+  out = __lasx_xvadd_w(tmp0, in_c);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication and addition calculation after expansion
+ *               of the higher half of the vector.
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the higher half of the two-fold sign extension (signed
+ *               halfword to signed word), and the result is added to
+ *               the vector in_c, then stored to the out vector.
+ * Example     : See out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmaddwh_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvilvh_h(in_h, in_h);
+  tmp1 = __lasx_xvilvh_h(in_l, in_l);
+  tmp0 = __lasx_xvmulwev_w_h(tmp0, tmp1);
+  out = __lasx_xvadd_w(tmp0, in_c);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication calculation after expansion of the lower
+ *               half of the vector.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the lower half of the two-fold sign extension (signed
+ *               halfword to signed word), then stored to the out vector.
+ * Example     : out = __lasx_xvmulwl_w_h(in_h, in_l)
+ *        in_h : 3,-1,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 0,0,1, 0, 1,0,0,1
+ *         out : 6,1,3,0, 0,0,1,0
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmulwl_w_h(__m256i in_h, __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvsllwil_w_h(in_h, 0);
+  tmp1 = __lasx_xvsllwil_w_h(in_l, 0);
+  out = __lasx_xvmul_w(tmp0, tmp1);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication calculation after expansion of the lower
+ *               half of the vector.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the lower half of the two-fold sign extension (signed
+ *               halfword to signed word), then stored to the out vector.
+ * Example     : out = __lasx_xvmulwh_w_h(in_h, in_l)
+ *        in_h : 3,-1,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 0,0,1, 0, 1,0,0,1
+ *         out : 0,0,0,0, 0,0,0,1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmulwh_w_h(__m256i in_h, __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvilvh_h(in_h, in_h);
+  tmp1 = __lasx_xvilvh_h(in_l, in_l);
+  out = __lasx_xvmulwev_w_h(tmp0, tmp1);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are added to the high half
+ *               after being doubled, then saturated.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector adds the in_l vector after the lower half of
+ *               the two-fold zero extension (unsigned byte to unsigned
+ *               halfword) and then saturated. The results are stored to the out
+ *               vector.
+ * Example     : out = __lasx_xvsaddw_hu_hu_bu(in_h, in_l)
+ *        in_h : 2,65532,1,2, 1,0,0,0, 0,0,1,0, 1,0,0,1
+ *        in_l : 3,6,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1, 3,18,3,0, 0,0,0,1, 0,0,1,1,
+ *               0,0,0,1
+ *        out  : 5,65535,4,2, 1,0,0,1, 3,18,4,0, 1,0,0,2,
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsaddw_hu_hu_bu(__m256i in_h, __m256i in_l) {
+  __m256i tmp1, out;
+  __m256i zero = {0};
+
+  tmp1 = __lasx_xvilvl_b(zero, in_l);
+  out = __lasx_xvsadd_hu(in_h, tmp1);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all halfword elements of input vector between min & max
+ *               out = ((in) < (min)) ? (min) : (((in) > (max)) ? (max) : (in))
+ * Arguments   : Inputs  - in    (input vector)
+ *                       - min   (min threshold)
+ *                       - max   (max threshold)
+ *               Outputs - in    (output vector with clipped elements)
+ *               Return Type - signed halfword
+ * Example     : out = __lasx_xvclip_h(in, min, max)
+ *          in : -8,2,280,249, -8,255,280,249, 4,4,4,4, 5,5,5,5
+ *         min : 1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1
+ *         max : 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9
+ *         out : 1,2,9,9, 1,9,9,9, 4,4,4,4, 5,5,5,5
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip_h(__m256i in, __m256i min, __m256i max) {
+  __m256i out;
+
+  out = __lasx_xvmax_h(min, in);
+  out = __lasx_xvmin_h(max, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all signed halfword elements of input vector
+ *               between 0 & 255
+ * Arguments   : Inputs  - in   (input vector)
+ *               Outputs - out  (output vector with clipped elements)
+ *               Return Type - signed halfword
+ * Example     : See out = __lasx_xvclip255_w(in)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip255_h(__m256i in) {
+  __m256i out;
+
+  out = __lasx_xvmaxi_h(in, 0);
+  out = __lasx_xvsat_hu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all signed word elements of input vector
+ *               between 0 & 255
+ * Arguments   : Inputs - in   (input vector)
+ *               Output - out  (output vector with clipped elements)
+ *               Return Type - signed word
+ * Example     : out = __lasx_xvclip255_w(in)
+ *          in : -8,255,280,249, -8,255,280,249
+ *         out :  0,255,255,249,  0,255,255,249
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip255_w(__m256i in) {
+  __m256i out;
+
+  out = __lasx_xvmaxi_w(in, 0);
+  out = __lasx_xvsat_wu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Indexed halfword element values are replicated to all
+ *               elements in output vector. If 'idx < 8' use xvsplati_l_*,
+ *               if 'idx >= 8' use xvsplati_h_*.
+ * Arguments   : Inputs - in, idx
+ *               Output - out
+ * Details     : Idx element value from in vector is replicated to all
+ *               elements in out vector.
+ *               Valid index range for halfword operation is 0-7
+ * Example     : out = __lasx_xvsplati_l_h(in, idx)
+ *          in : 20,10,11,12, 13,14,15,16, 0,0,2,0, 0,0,0,0
+ *         idx : 0x02
+ *         out : 11,11,11,11, 11,11,11,11, 11,11,11,11, 11,11,11,11
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsplati_l_h(__m256i in, int idx) {
+  __m256i out;
+
+  out = __lasx_xvpermi_q(in, in, 0x02);
+  out = __lasx_xvreplve_h(out, idx);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Indexed halfword element values are replicated to all
+ *               elements in output vector. If 'idx < 8' use xvsplati_l_*,
+ *               if 'idx >= 8' use xvsplati_h_*.
+ * Arguments   : Inputs - in, idx
+ *               Output - out
+ * Details     : Idx element value from in vector is replicated to all
+ *               elements in out vector.
+ *               Valid index range for halfword operation is 0-7
+ * Example     : out = __lasx_xvsplati_h_h(in, idx)
+ *          in : 20,10,11,12, 13,14,15,16, 0,2,0,0, 0,0,0,0
+ *         idx : 0x09
+ *         out : 2,2,2,2, 2,2,2,2, 2,2,2,2, 2,2,2,2
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsplati_h_h(__m256i in, int idx) {
+  __m256i out;
+
+  out = __lasx_xvpermi_q(in, in, 0x13);
+  out = __lasx_xvreplve_h(out, idx);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with double-word elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3
+ *               Outputs - _out0, _out1, _out2, _out3
+ * Example     : LASX_TRANSPOSE4x4_D
+ *        _in0 : 1,2,3,4
+ *        _in1 : 1,2,3,4
+ *        _in2 : 1,2,3,4
+ *        _in3 : 1,2,3,4
+ *
+ *       _out0 : 1,1,1,1
+ *       _out1 : 2,2,2,2
+ *       _out2 : 3,3,3,3
+ *       _out3 : 4,4,4,4
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE4x4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, \
+                            _out3)                                       \
+  {                                                                      \
+    __m256i _tmp0, _tmp1, _tmp2, _tmp3;                                  \
+    _tmp0 = __lasx_xvilvl_d(_in1, _in0);                                 \
+    _tmp1 = __lasx_xvilvh_d(_in1, _in0);                                 \
+    _tmp2 = __lasx_xvilvl_d(_in3, _in2);                                 \
+    _tmp3 = __lasx_xvilvh_d(_in3, _in2);                                 \
+    _out0 = __lasx_xvpermi_q(_tmp2, _tmp0, 0x20);                        \
+    _out2 = __lasx_xvpermi_q(_tmp2, _tmp0, 0x31);                        \
+    _out1 = __lasx_xvpermi_q(_tmp3, _tmp1, 0x20);                        \
+    _out3 = __lasx_xvpermi_q(_tmp3, _tmp1, 0x31);                        \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with word elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *               _out7
+ * Example     : LASX_TRANSPOSE8x8_W
+ *        _in0 : 1,2,3,4,5,6,7,8
+ *        _in1 : 2,2,3,4,5,6,7,8
+ *        _in2 : 3,2,3,4,5,6,7,8
+ *        _in3 : 4,2,3,4,5,6,7,8
+ *        _in4 : 5,2,3,4,5,6,7,8
+ *        _in5 : 6,2,3,4,5,6,7,8
+ *        _in6 : 7,2,3,4,5,6,7,8
+ *        _in7 : 8,2,3,4,5,6,7,8
+ *
+ *       _out0 : 1,2,3,4,5,6,7,8
+ *       _out1 : 2,2,2,2,2,2,2,2
+ *       _out2 : 3,3,3,3,3,3,3,3
+ *       _out3 : 4,4,4,4,4,4,4,4
+ *       _out4 : 5,5,5,5,5,5,5,5
+ *       _out5 : 6,6,6,6,6,6,6,6
+ *       _out6 : 7,7,7,7,7,7,7,7
+ *       _out7 : 8,8,8,8,8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                            _out7)                                           \
+  {                                                                          \
+    __m256i _s0_m, _s1_m;                                                    \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                              \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                              \
+                                                                             \
+    _s0_m = __lasx_xvilvl_w(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvl_w(_in3, _in1);                                     \
+    _tmp0_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp1_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_w(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvh_w(_in3, _in1);                                     \
+    _tmp2_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp3_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvl_w(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvl_w(_in7, _in5);                                     \
+    _tmp4_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp5_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_w(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvh_w(_in7, _in5);                                     \
+    _tmp6_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp7_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _out0 = __lasx_xvpermi_q(_tmp4_m, _tmp0_m, 0x20);                        \
+    _out1 = __lasx_xvpermi_q(_tmp5_m, _tmp1_m, 0x20);                        \
+    _out2 = __lasx_xvpermi_q(_tmp6_m, _tmp2_m, 0x20);                        \
+    _out3 = __lasx_xvpermi_q(_tmp7_m, _tmp3_m, 0x20);                        \
+    _out4 = __lasx_xvpermi_q(_tmp4_m, _tmp0_m, 0x31);                        \
+    _out5 = __lasx_xvpermi_q(_tmp5_m, _tmp1_m, 0x31);                        \
+    _out6 = __lasx_xvpermi_q(_tmp6_m, _tmp2_m, 0x31);                        \
+    _out7 = __lasx_xvpermi_q(_tmp7_m, _tmp3_m, 0x31);                        \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 16x8 byte block
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,
+ *                         _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15
+ *                         (input 16x8 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *                         _out7 (output 8x16 byte block)
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : See LASX_TRANSPOSE16x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE16x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                             _in8, _in9, _in10, _in11, _in12, _in13, _in14,   \
+                             _in15, _out0, _out1, _out2, _out3, _out4, _out5, \
+                             _out6, _out7)                                    \
+  {                                                                           \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                               \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                               \
+                                                                              \
+    _tmp0_m = __lasx_xvilvl_b(_in2, _in0);                                    \
+    _tmp1_m = __lasx_xvilvl_b(_in3, _in1);                                    \
+    _tmp2_m = __lasx_xvilvl_b(_in6, _in4);                                    \
+    _tmp3_m = __lasx_xvilvl_b(_in7, _in5);                                    \
+    _tmp4_m = __lasx_xvilvl_b(_in10, _in8);                                   \
+    _tmp5_m = __lasx_xvilvl_b(_in11, _in9);                                   \
+    _tmp6_m = __lasx_xvilvl_b(_in14, _in12);                                  \
+    _tmp7_m = __lasx_xvilvl_b(_in15, _in13);                                  \
+    _out0 = __lasx_xvilvl_b(_tmp1_m, _tmp0_m);                                \
+    _out1 = __lasx_xvilvh_b(_tmp1_m, _tmp0_m);                                \
+    _out2 = __lasx_xvilvl_b(_tmp3_m, _tmp2_m);                                \
+    _out3 = __lasx_xvilvh_b(_tmp3_m, _tmp2_m);                                \
+    _out4 = __lasx_xvilvl_b(_tmp5_m, _tmp4_m);                                \
+    _out5 = __lasx_xvilvh_b(_tmp5_m, _tmp4_m);                                \
+    _out6 = __lasx_xvilvl_b(_tmp7_m, _tmp6_m);                                \
+    _out7 = __lasx_xvilvh_b(_tmp7_m, _tmp6_m);                                \
+    _tmp0_m = __lasx_xvilvl_w(_out2, _out0);                                  \
+    _tmp2_m = __lasx_xvilvh_w(_out2, _out0);                                  \
+    _tmp4_m = __lasx_xvilvl_w(_out3, _out1);                                  \
+    _tmp6_m = __lasx_xvilvh_w(_out3, _out1);                                  \
+    _tmp1_m = __lasx_xvilvl_w(_out6, _out4);                                  \
+    _tmp3_m = __lasx_xvilvh_w(_out6, _out4);                                  \
+    _tmp5_m = __lasx_xvilvl_w(_out7, _out5);                                  \
+    _tmp7_m = __lasx_xvilvh_w(_out7, _out5);                                  \
+    _out0 = __lasx_xvilvl_d(_tmp1_m, _tmp0_m);                                \
+    _out1 = __lasx_xvilvh_d(_tmp1_m, _tmp0_m);                                \
+    _out2 = __lasx_xvilvl_d(_tmp3_m, _tmp2_m);                                \
+    _out3 = __lasx_xvilvh_d(_tmp3_m, _tmp2_m);                                \
+    _out4 = __lasx_xvilvl_d(_tmp5_m, _tmp4_m);                                \
+    _out5 = __lasx_xvilvh_d(_tmp5_m, _tmp4_m);                                \
+    _out6 = __lasx_xvilvl_d(_tmp7_m, _tmp6_m);                                \
+    _out7 = __lasx_xvilvh_d(_tmp7_m, _tmp6_m);                                \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 16x8 byte block
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,
+ *                         _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15
+ *                         (input 16x8 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *                         _out7 (output 8x16 byte block)
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : LASX_TRANSPOSE16x8_H
+ *        _in0 : 1,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in1 : 2,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in2 : 3,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in3 : 4,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in4 : 5,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in5 : 6,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in6 : 7,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in7 : 8,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in8 : 9,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in9 : 1,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in10 : 0,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in11 : 2,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in12 : 3,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in13 : 7,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in14 : 5,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in15 : 6,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *
+ *       _out0 : 1,2,3,4,5,6,7,8,9,1,0,2,3,7,5,6
+ *       _out1 : 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2
+ *       _out2 : 3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3
+ *       _out3 : 4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4
+ *       _out4 : 5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
+ *       _out5 : 6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6
+ *       _out6 : 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7
+ *       _out7 : 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE16x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                             _in8, _in9, _in10, _in11, _in12, _in13, _in14,   \
+                             _in15, _out0, _out1, _out2, _out3, _out4, _out5, \
+                             _out6, _out7)                                    \
+  {                                                                           \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                               \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                               \
+    __m256i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;                           \
+                                                                              \
+    _tmp0_m = __lasx_xvilvl_h(_in2, _in0);                                    \
+    _tmp1_m = __lasx_xvilvl_h(_in3, _in1);                                    \
+    _tmp2_m = __lasx_xvilvl_h(_in6, _in4);                                    \
+    _tmp3_m = __lasx_xvilvl_h(_in7, _in5);                                    \
+    _tmp4_m = __lasx_xvilvl_h(_in10, _in8);                                   \
+    _tmp5_m = __lasx_xvilvl_h(_in11, _in9);                                   \
+    _tmp6_m = __lasx_xvilvl_h(_in14, _in12);                                  \
+    _tmp7_m = __lasx_xvilvl_h(_in15, _in13);                                  \
+    _t0 = __lasx_xvilvl_h(_tmp1_m, _tmp0_m);                                  \
+    _t1 = __lasx_xvilvh_h(_tmp1_m, _tmp0_m);                                  \
+    _t2 = __lasx_xvilvl_h(_tmp3_m, _tmp2_m);                                  \
+    _t3 = __lasx_xvilvh_h(_tmp3_m, _tmp2_m);                                  \
+    _t4 = __lasx_xvilvl_h(_tmp5_m, _tmp4_m);                                  \
+    _t5 = __lasx_xvilvh_h(_tmp5_m, _tmp4_m);                                  \
+    _t6 = __lasx_xvilvl_h(_tmp7_m, _tmp6_m);                                  \
+    _t7 = __lasx_xvilvh_h(_tmp7_m, _tmp6_m);                                  \
+    _tmp0_m = __lasx_xvilvl_d(_t2, _t0);                                      \
+    _tmp2_m = __lasx_xvilvh_d(_t2, _t0);                                      \
+    _tmp4_m = __lasx_xvilvl_d(_t3, _t1);                                      \
+    _tmp6_m = __lasx_xvilvh_d(_t3, _t1);                                      \
+    _tmp1_m = __lasx_xvilvl_d(_t6, _t4);                                      \
+    _tmp3_m = __lasx_xvilvh_d(_t6, _t4);                                      \
+    _tmp5_m = __lasx_xvilvl_d(_t7, _t5);                                      \
+    _tmp7_m = __lasx_xvilvh_d(_t7, _t5);                                      \
+    _out0 = __lasx_xvpermi_q(_tmp1_m, _tmp0_m, 0x20);                         \
+    _out1 = __lasx_xvpermi_q(_tmp3_m, _tmp2_m, 0x20);                         \
+    _out2 = __lasx_xvpermi_q(_tmp5_m, _tmp4_m, 0x20);                         \
+    _out3 = __lasx_xvpermi_q(_tmp7_m, _tmp6_m, 0x20);                         \
+                                                                              \
+    _tmp0_m = __lasx_xvilvh_h(_in2, _in0);                                    \
+    _tmp1_m = __lasx_xvilvh_h(_in3, _in1);                                    \
+    _tmp2_m = __lasx_xvilvh_h(_in6, _in4);                                    \
+    _tmp3_m = __lasx_xvilvh_h(_in7, _in5);                                    \
+    _tmp4_m = __lasx_xvilvh_h(_in10, _in8);                                   \
+    _tmp5_m = __lasx_xvilvh_h(_in11, _in9);                                   \
+    _tmp6_m = __lasx_xvilvh_h(_in14, _in12);                                  \
+    _tmp7_m = __lasx_xvilvh_h(_in15, _in13);                                  \
+    _t0 = __lasx_xvilvl_h(_tmp1_m, _tmp0_m);                                  \
+    _t1 = __lasx_xvilvh_h(_tmp1_m, _tmp0_m);                                  \
+    _t2 = __lasx_xvilvl_h(_tmp3_m, _tmp2_m);                                  \
+    _t3 = __lasx_xvilvh_h(_tmp3_m, _tmp2_m);                                  \
+    _t4 = __lasx_xvilvl_h(_tmp5_m, _tmp4_m);                                  \
+    _t5 = __lasx_xvilvh_h(_tmp5_m, _tmp4_m);                                  \
+    _t6 = __lasx_xvilvl_h(_tmp7_m, _tmp6_m);                                  \
+    _t7 = __lasx_xvilvh_h(_tmp7_m, _tmp6_m);                                  \
+    _tmp0_m = __lasx_xvilvl_d(_t2, _t0);                                      \
+    _tmp2_m = __lasx_xvilvh_d(_t2, _t0);                                      \
+    _tmp4_m = __lasx_xvilvl_d(_t3, _t1);                                      \
+    _tmp6_m = __lasx_xvilvh_d(_t3, _t1);                                      \
+    _tmp1_m = __lasx_xvilvl_d(_t6, _t4);                                      \
+    _tmp3_m = __lasx_xvilvh_d(_t6, _t4);                                      \
+    _tmp5_m = __lasx_xvilvl_d(_t7, _t5);                                      \
+    _tmp7_m = __lasx_xvilvh_d(_t7, _t5);                                      \
+    _out4 = __lasx_xvpermi_q(_tmp1_m, _tmp0_m, 0x20);                         \
+    _out5 = __lasx_xvpermi_q(_tmp3_m, _tmp2_m, 0x20);                         \
+    _out6 = __lasx_xvpermi_q(_tmp5_m, _tmp4_m, 0x20);                         \
+    _out7 = __lasx_xvpermi_q(_tmp7_m, _tmp6_m, 0x20);                         \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with halfword elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3
+ *               Outputs - _out0, _out1, _out2, _out3
+ *               Return Type - signed halfword
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : See LASX_TRANSPOSE8x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE4x4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, \
+                            _out3)                                       \
+  {                                                                      \
+    __m256i _s0_m, _s1_m;                                                \
+                                                                         \
+    _s0_m = __lasx_xvilvl_h(_in1, _in0);                                 \
+    _s1_m = __lasx_xvilvl_h(_in3, _in2);                                 \
+    _out0 = __lasx_xvilvl_w(_s1_m, _s0_m);                               \
+    _out2 = __lasx_xvilvh_w(_s1_m, _s0_m);                               \
+    _out1 = __lasx_xvilvh_d(_out0, _out0);                               \
+    _out3 = __lasx_xvilvh_d(_out2, _out2);                               \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 8x8 byte block
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ *                         (input 8x8 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *                         _out7 (output 8x8 byte block)
+ * Example     : See LASX_TRANSPOSE8x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                            _out7)                                           \
+  {                                                                          \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                              \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                              \
+    _tmp0_m = __lasx_xvilvl_b(_in2, _in0);                                   \
+    _tmp1_m = __lasx_xvilvl_b(_in3, _in1);                                   \
+    _tmp2_m = __lasx_xvilvl_b(_in6, _in4);                                   \
+    _tmp3_m = __lasx_xvilvl_b(_in7, _in5);                                   \
+    _tmp4_m = __lasx_xvilvl_b(_tmp1_m, _tmp0_m);                             \
+    _tmp5_m = __lasx_xvilvh_b(_tmp1_m, _tmp0_m);                             \
+    _tmp6_m = __lasx_xvilvl_b(_tmp3_m, _tmp2_m);                             \
+    _tmp7_m = __lasx_xvilvh_b(_tmp3_m, _tmp2_m);                             \
+    _out0 = __lasx_xvilvl_w(_tmp6_m, _tmp4_m);                               \
+    _out2 = __lasx_xvilvh_w(_tmp6_m, _tmp4_m);                               \
+    _out4 = __lasx_xvilvl_w(_tmp7_m, _tmp5_m);                               \
+    _out6 = __lasx_xvilvh_w(_tmp7_m, _tmp5_m);                               \
+    _out1 = __lasx_xvbsrl_v(_out0, 8);                                       \
+    _out3 = __lasx_xvbsrl_v(_out2, 8);                                       \
+    _out5 = __lasx_xvbsrl_v(_out4, 8);                                       \
+    _out7 = __lasx_xvbsrl_v(_out6, 8);                                       \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with halfword elements in vectors.
+ * Arguments   : Inputs  - _in0, _in1, ~
+ *               Outputs - _out0, _out1, ~
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : LASX_TRANSPOSE8x8_H
+ *        _in0 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in1 : 8,2,3,4, 5,6,7,8, 8,2,3,4, 5,6,7,8
+ *        _in2 : 8,2,3,4, 5,6,7,8, 8,2,3,4, 5,6,7,8
+ *        _in3 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in4 : 9,2,3,4, 5,6,7,8, 9,2,3,4, 5,6,7,8
+ *        _in5 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in6 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in7 : 9,2,3,4, 5,6,7,8, 9,2,3,4, 5,6,7,8
+ *
+ *       _out0 : 1,8,8,1, 9,1,1,9, 1,8,8,1, 9,1,1,9
+ *       _out1 : 2,2,2,2, 2,2,2,2, 2,2,2,2, 2,2,2,2
+ *       _out2 : 3,3,3,3, 3,3,3,3, 3,3,3,3, 3,3,3,3
+ *       _out3 : 4,4,4,4, 4,4,4,4, 4,4,4,4, 4,4,4,4
+ *       _out4 : 5,5,5,5, 5,5,5,5, 5,5,5,5, 5,5,5,5
+ *       _out5 : 6,6,6,6, 6,6,6,6, 6,6,6,6, 6,6,6,6
+ *       _out6 : 7,7,7,7, 7,7,7,7, 7,7,7,7, 7,7,7,7
+ *       _out7 : 8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                            _out7)                                           \
+  {                                                                          \
+    __m256i _s0_m, _s1_m;                                                    \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                              \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                              \
+                                                                             \
+    _s0_m = __lasx_xvilvl_h(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvl_h(_in7, _in5);                                     \
+    _tmp0_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp1_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_h(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvh_h(_in7, _in5);                                     \
+    _tmp2_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp3_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+                                                                             \
+    _s0_m = __lasx_xvilvl_h(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvl_h(_in3, _in1);                                     \
+    _tmp4_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp5_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_h(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvh_h(_in3, _in1);                                     \
+    _tmp6_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp7_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+                                                                             \
+    _out0 = __lasx_xvpickev_d(_tmp0_m, _tmp4_m);                             \
+    _out2 = __lasx_xvpickev_d(_tmp1_m, _tmp5_m);                             \
+    _out4 = __lasx_xvpickev_d(_tmp2_m, _tmp6_m);                             \
+    _out6 = __lasx_xvpickev_d(_tmp3_m, _tmp7_m);                             \
+    _out1 = __lasx_xvpickod_d(_tmp0_m, _tmp4_m);                             \
+    _out3 = __lasx_xvpickod_d(_tmp1_m, _tmp5_m);                             \
+    _out5 = __lasx_xvpickod_d(_tmp2_m, _tmp6_m);                             \
+    _out7 = __lasx_xvpickod_d(_tmp3_m, _tmp7_m);                             \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 4 input vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3
+ *               Outputs - _out0, _out1, _out2, _out3
+ * Details     : Butterfly operation
+ * Example     : LASX_BUTTERFLY_4
+ *               _out0 = _in0 + _in3;
+ *               _out1 = _in1 + _in2;
+ *               _out2 = _in1 - _in2;
+ *               _out3 = _in0 - _in3;
+ * =============================================================================
+ */
+#define LASX_BUTTERFLY_4_B(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_b(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_b(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_b(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_b(_in0, _in3);                                        \
+  }
+#define LASX_BUTTERFLY_4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_h(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_h(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_h(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_h(_in0, _in3);                                        \
+  }
+#define LASX_BUTTERFLY_4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_w(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_w(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_w(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_w(_in0, _in3);                                        \
+  }
+#define LASX_BUTTERFLY_4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_d(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_d(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_d(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_d(_in0, _in3);                                        \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 8 input vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, ~
+ *               Outputs - _out0, _out1, _out2, _out3, ~
+ * Details     : Butterfly operation
+ * Example     : LASX_BUTTERFLY_8
+ *               _out0 = _in0 + _in7;
+ *               _out1 = _in1 + _in6;
+ *               _out2 = _in2 + _in5;
+ *               _out3 = _in3 + _in4;
+ *               _out4 = _in3 - _in4;
+ *               _out5 = _in2 - _in5;
+ *               _out6 = _in1 - _in6;
+ *               _out7 = _in0 - _in7;
+ * =============================================================================
+ */
+#define LASX_BUTTERFLY_8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_b(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_b(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_b(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_b(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_b(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_b(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_b(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_b(_in0, _in7);                                     \
+  }
+
+#define LASX_BUTTERFLY_8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_h(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_h(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_h(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_h(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_h(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_h(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_h(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_h(_in0, _in7);                                     \
+  }
+
+#define LASX_BUTTERFLY_8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_w(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_w(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_w(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_w(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_w(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_w(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_w(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_w(_in0, _in7);                                     \
+  }
+
+#define LASX_BUTTERFLY_8_D(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_d(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_d(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_d(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_d(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_d(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_d(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_d(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_d(_in0, _in7);                                     \
+  }
+
+#endif  // LASX
+
+/*
+ * =============================================================================
+ * Description : Print out elements in vector.
+ * Arguments   : Inputs  - RTYPE, _element_num, _in0, _enter
+ *               Outputs -
+ * Details     : Print out '_element_num' elements in 'RTYPE' vector '_in0', if
+ *               '_enter' is TRUE, prefix "\nVP:" will be added first.
+ * Example     : VECT_PRINT(v4i32,4,in0,1); // in0: 1,2,3,4
+ *               VP:1,2,3,4,
+ * =============================================================================
+ */
+#define VECT_PRINT(RTYPE, element_num, in0, enter) \
+  {                                                \
+    RTYPE _tmp0 = (RTYPE)in0;                      \
+    int _i = 0;                                    \
+    if (enter)                                     \
+      printf("\nVP:");                             \
+    for (_i = 0; _i < element_num; _i++)           \
+      printf("%d,", _tmp0[_i]);                    \
+  }
+
+#endif /* LOONGSON_INTRINSICS_H */
+#endif /* INCLUDE_LIBYUV_LOONGSON_INTRINSICS_H */
diff --git a/media/libyuv/include/libyuv/macros_msa.h b/media/libyuv/include/libyuv/macros_msa.h
new file mode 100644
index 0000000000..b9a44fcced
--- /dev/null
+++ b/media/libyuv/include/libyuv/macros_msa.h
@@ -0,0 +1,246 @@
+/*
+ *  Copyright 2016 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_MACROS_MSA_H_
+#define INCLUDE_LIBYUV_MACROS_MSA_H_
+
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include <msa.h>
+#include <stdint.h>
+
+#if (__mips_isa_rev >= 6)
+#define LW(psrc)                                       \
+  ({                                                   \
+    const uint8_t* psrc_lw_m = (const uint8_t*)(psrc); \
+    uint32_t val_m;                                    \
+    asm volatile("lw  %[val_m],  %[psrc_lw_m]  \n"     \
+                 : [val_m] "=r"(val_m)                 \
+                 : [psrc_lw_m] "m"(*psrc_lw_m));       \
+    val_m;                                             \
+  })
+
+#if (__mips == 64)
+#define LD(psrc)                                       \
+  ({                                                   \
+    const uint8_t* psrc_ld_m = (const uint8_t*)(psrc); \
+    uint64_t val_m = 0;                                \
+    asm volatile("ld  %[val_m],  %[psrc_ld_m]  \n"     \
+                 : [val_m] "=r"(val_m)                 \
+                 : [psrc_ld_m] "m"(*psrc_ld_m));       \
+    val_m;                                             \
+  })
+#else  // !(__mips == 64)
+#define LD(psrc)                                                         \
+  ({                                                                     \
+    const uint8_t* psrc_ld_m = (const uint8_t*)(psrc);                   \
+    uint32_t val0_m, val1_m;                                             \
+    uint64_t val_m = 0;                                                  \
+    val0_m = LW(psrc_ld_m);                                              \
+    val1_m = LW(psrc_ld_m + 4);                                          \
+    val_m = (uint64_t)(val1_m);                             /* NOLINT */ \
+    val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000); /* NOLINT */ \
+    val_m = (uint64_t)(val_m | (uint64_t)val0_m);           /* NOLINT */ \
+    val_m;                                                               \
+  })
+#endif  // (__mips == 64)
+
+#define SW(val, pdst)                                   \
+  ({                                                    \
+    uint8_t* pdst_sw_m = (uint8_t*)(pdst); /* NOLINT */ \
+    uint32_t val_m = (val);                             \
+    asm volatile("sw  %[val_m],  %[pdst_sw_m]  \n"      \
+                 : [pdst_sw_m] "=m"(*pdst_sw_m)         \
+                 : [val_m] "r"(val_m));                 \
+  })
+
+#if (__mips == 64)
+#define SD(val, pdst)                                   \
+  ({                                                    \
+    uint8_t* pdst_sd_m = (uint8_t*)(pdst); /* NOLINT */ \
+    uint64_t val_m = (val);                             \
+    asm volatile("sd  %[val_m],  %[pdst_sd_m]  \n"      \
+                 : [pdst_sd_m] "=m"(*pdst_sd_m)         \
+                 : [val_m] "r"(val_m));                 \
+  })
+#else  // !(__mips == 64)
+#define SD(val, pdst)                                        \
+  ({                                                         \
+    uint8_t* pdst_sd_m = (uint8_t*)(pdst); /* NOLINT */      \
+    uint32_t val0_m, val1_m;                                 \
+    val0_m = (uint32_t)((val)&0x00000000FFFFFFFF);           \
+    val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \
+    SW(val0_m, pdst_sd_m);                                   \
+    SW(val1_m, pdst_sd_m + 4);                               \
+  })
+#endif  // !(__mips == 64)
+#else   // !(__mips_isa_rev >= 6)
+#define LW(psrc)                                \
+  ({                                            \
+    uint8_t* psrc_lw_m = (uint8_t*)(psrc);      \
+    uint32_t val_lw_m;                          \
+                                                \
+    __asm__ volatile(                           \
+        "lwr %[val_lw_m], 0(%[psrc_lw_m]) \n\t" \
+        "lwl %[val_lw_m], 3(%[psrc_lw_m]) \n\t" \
+                                                \
+        : [val_lw_m] "=&r"(val_lw_m)            \
+        : [psrc_lw_m] "r"(psrc_lw_m));          \
+                                                \
+    val_lw_m;                                   \
+  })
+
+#if (__mips == 64)
+#define LD(psrc)                                \
+  ({                                            \
+    uint8_t* psrc_ld_m = (uint8_t*)(psrc);      \
+    uint64_t val_ld_m = 0;                      \
+                                                \
+    __asm__ volatile(                           \
+        "ldr %[val_ld_m], 0(%[psrc_ld_m]) \n\t" \
+        "ldl %[val_ld_m], 7(%[psrc_ld_m]) \n\t" \
+                                                \
+        : [val_ld_m] "=&r"(val_ld_m)            \
+        : [psrc_ld_m] "r"(psrc_ld_m));          \
+                                                \
+    val_ld_m;                                   \
+  })
+#else  // !(__mips == 64)
+#define LD(psrc)                                                         \
+  ({                                                                     \
+    const uint8_t* psrc_ld_m = (const uint8_t*)(psrc);                   \
+    uint32_t val0_m, val1_m;                                             \
+    uint64_t val_m = 0;                                                  \
+    val0_m = LW(psrc_ld_m);                                              \
+    val1_m = LW(psrc_ld_m + 4);                                          \
+    val_m = (uint64_t)(val1_m);                             /* NOLINT */ \
+    val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000); /* NOLINT */ \
+    val_m = (uint64_t)(val_m | (uint64_t)val0_m);           /* NOLINT */ \
+    val_m;                                                               \
+  })
+#endif  // (__mips == 64)
+
+#define SW(val, pdst)                                   \
+  ({                                                    \
+    uint8_t* pdst_sw_m = (uint8_t*)(pdst); /* NOLINT */ \
+    uint32_t val_m = (val);                             \
+    asm volatile("usw  %[val_m],  %[pdst_sw_m]  \n"     \
+                 : [pdst_sw_m] "=m"(*pdst_sw_m)         \
+                 : [val_m] "r"(val_m));                 \
+  })
+
+#define SD(val, pdst)                                        \
+  ({                                                         \
+    uint8_t* pdst_sd_m = (uint8_t*)(pdst); /* NOLINT */      \
+    uint32_t val0_m, val1_m;                                 \
+    val0_m = (uint32_t)((val)&0x00000000FFFFFFFF);           \
+    val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \
+    SW(val0_m, pdst_sd_m);                                   \
+    SW(val1_m, pdst_sd_m + 4);                               \
+  })
+#endif  // (__mips_isa_rev >= 6)
+
+// TODO(fbarchard): Consider removing __VAR_ARGS versions.
+#define LD_B(RTYPE, psrc) *((RTYPE*)(psrc)) /* NOLINT */
+#define LD_UB(...) LD_B(const v16u8, __VA_ARGS__)
+
+#define LD_H(RTYPE, psrc) *((RTYPE*)(psrc)) /* NOLINT */
+#define LD_UH(...) LD_H(const v8u16, __VA_ARGS__)
+
+#define ST_B(RTYPE, in, pdst) *((RTYPE*)(pdst)) = (in) /* NOLINT */
+#define ST_UB(...) ST_B(v16u8, __VA_ARGS__)
+
+#define ST_H(RTYPE, in, pdst) *((RTYPE*)(pdst)) = (in) /* NOLINT */
+#define ST_UH(...) ST_H(v8u16, __VA_ARGS__)
+
+/* Description : Load two vectors with 16 'byte' sized elements
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Load 16 byte elements in 'out0' from (psrc)
+                 Load 16 byte elements in 'out1' from (psrc + stride)
+*/
+#define LD_B2(RTYPE, psrc, stride, out0, out1) \
+  {                                            \
+    out0 = LD_B(RTYPE, (psrc));                \
+    out1 = LD_B(RTYPE, (psrc) + stride);       \
+  }
+#define LD_UB2(...) LD_B2(const v16u8, __VA_ARGS__)
+
+#define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) \
+  {                                                        \
+    LD_B2(RTYPE, (psrc), stride, out0, out1);              \
+    LD_B2(RTYPE, (psrc) + 2 * stride, stride, out2, out3); \
+  }
+#define LD_UB4(...) LD_B4(const v16u8, __VA_ARGS__)
+
+/* Description : Store two vectors with stride each having 16 'byte' sized
+                 elements
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 16 byte elements from 'in0' to (pdst)
+                 Store 16 byte elements from 'in1' to (pdst + stride)
+*/
+#define ST_B2(RTYPE, in0, in1, pdst, stride) \
+  {                                          \
+    ST_B(RTYPE, in0, (pdst));                \
+    ST_B(RTYPE, in1, (pdst) + stride);       \
+  }
+#define ST_UB2(...) ST_B2(v16u8, __VA_ARGS__)
+
+#define ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride)   \
+  {                                                      \
+    ST_B2(RTYPE, in0, in1, (pdst), stride);              \
+    ST_B2(RTYPE, in2, in3, (pdst) + 2 * stride, stride); \
+  }
+#define ST_UB4(...) ST_B4(v16u8, __VA_ARGS__)
+
+/* Description : Store vectors of 8 halfword elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 8 halfword elements from 'in0' to (pdst)
+                 Store 8 halfword elements from 'in1' to (pdst + stride)
+*/
+#define ST_H2(RTYPE, in0, in1, pdst, stride) \
+  {                                          \
+    ST_H(RTYPE, in0, (pdst));                \
+    ST_H(RTYPE, in1, (pdst) + stride);       \
+  }
+#define ST_UH2(...) ST_H2(v8u16, __VA_ARGS__)
+
+// TODO(fbarchard): Consider using __msa_vshf_b and __msa_ilvr_b directly.
+/* Description : Shuffle byte vector elements as per mask vector
+   Arguments   : Inputs  - in0, in1, in2, in3, mask0, mask1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0' & 'in1' are copied selectively to
+                 'out0' as per control vector 'mask0'
+*/
+#define VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1)  \
+  {                                                                   \
+    out0 = (RTYPE)__msa_vshf_b((v16i8)mask0, (v16i8)in1, (v16i8)in0); \
+    out1 = (RTYPE)__msa_vshf_b((v16i8)mask1, (v16i8)in3, (v16i8)in2); \
+  }
+#define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__)
+
+/* Description : Interleave both left and right half of input vectors
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements from 'in0' and 'in1' are
+                 interleaved and written to 'out0'
+*/
+#define ILVRL_B2(RTYPE, in0, in1, out0, out1)           \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1); \
+    out1 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1); \
+  }
+#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__)
+
+#endif /* !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa) */
+
+#endif  // INCLUDE_LIBYUV_MACROS_MSA_H_
diff --git a/media/libyuv/include/libyuv/mjpeg_decoder.h b/media/libyuv/include/libyuv/mjpeg_decoder.h
index 8423121d11..275f8d4c18 100644
--- a/media/libyuv/include/libyuv/mjpeg_decoder.h
+++ b/media/libyuv/include/libyuv/mjpeg_decoder.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_MJPEG_DECODER_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_MJPEG_DECODER_H_
 #define INCLUDE_LIBYUV_MJPEG_DECODER_H_
 
 #include "libyuv/basic_types.h"
@@ -26,25 +26,24 @@ namespace libyuv {
 extern "C" {
 #endif
 
-LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size);
+LIBYUV_BOOL ValidateJpeg(const uint8_t* sample, size_t sample_size);
 
 #ifdef __cplusplus
 }  // extern "C"
 #endif
 
-static const uint32 kUnknownDataSize = 0xFFFFFFFF;
+static const uint32_t kUnknownDataSize = 0xFFFFFFFF;
 
 enum JpegSubsamplingType {
   kJpegYuv420,
   kJpegYuv422,
-  kJpegYuv411,
   kJpegYuv444,
   kJpegYuv400,
   kJpegUnknown
 };
 
 struct Buffer {
-  const uint8* data;
+  const uint8_t* data;
   int len;
 };
 
@@ -66,7 +65,7 @@ struct SetJmpErrorMgr;
 class LIBYUV_API MJpegDecoder {
  public:
   typedef void (*CallbackFunction)(void* opaque,
-                                   const uint8* const* data,
+                                   const uint8_t* const* data,
                                    const int* strides,
                                    int rows);
 
@@ -86,7 +85,7 @@ class LIBYUV_API MJpegDecoder {
   // If return value is LIBYUV_TRUE, then the values for all the following
   // getters are populated.
   // src_len is the size of the compressed mjpeg frame in bytes.
-  LIBYUV_BOOL LoadFrame(const uint8* src, size_t src_len);
+  LIBYUV_BOOL LoadFrame(const uint8_t* src, size_t src_len);
 
   // Returns width of the last loaded frame in pixels.
   int GetWidth();
@@ -139,18 +138,22 @@ class LIBYUV_API MJpegDecoder {
   // at least GetComponentSize(i). The pointers in planes are incremented
   // to point to after the end of the written data.
   // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
-  LIBYUV_BOOL DecodeToBuffers(uint8** planes, int dst_width, int dst_height);
+  LIBYUV_BOOL DecodeToBuffers(uint8_t** planes, int dst_width, int dst_height);
 
   // Decodes the entire image and passes the data via repeated calls to a
   // callback function. Each call will get the data for a whole number of
   // image scanlines.
   // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
-  LIBYUV_BOOL DecodeToCallback(CallbackFunction fn, void* opaque,
-                        int dst_width, int dst_height);
+  LIBYUV_BOOL DecodeToCallback(CallbackFunction fn,
+                               void* opaque,
+                               int dst_width,
+                               int dst_height);
 
   // The helper function which recognizes the jpeg sub-sampling type.
   static JpegSubsamplingType JpegSubsamplingTypeHelper(
-     int* subsample_x, int* subsample_y, int number_of_components);
+      int* subsample_x,
+      int* subsample_y,
+      int number_of_components);
 
  private:
   void AllocOutputBuffers(int num_outbufs);
@@ -159,7 +162,7 @@ class LIBYUV_API MJpegDecoder {
   LIBYUV_BOOL StartDecode();
   LIBYUV_BOOL FinishDecode();
 
-  void SetScanlinePointers(uint8** data);
+  void SetScanlinePointers(uint8_t** data);
   LIBYUV_BOOL DecodeImcuRow();
 
   int GetComponentScanlinePadding(int component);
@@ -178,15 +181,15 @@ class LIBYUV_API MJpegDecoder {
 
   // Temporaries used to point to scanline outputs.
   int num_outbufs_;  // Outermost size of all arrays below.
-  uint8*** scanlines_;
+  uint8_t*** scanlines_;
   int* scanlines_sizes_;
   // Temporary buffer used for decoding when we can't decode directly to the
   // output buffers. Large enough for just one iMCU row.
-  uint8** databuf_;
+  uint8_t** databuf_;
   int* databuf_strides_;
 };
 
 }  // namespace libyuv
 
 #endif  //  __cplusplus
-#endif  // INCLUDE_LIBYUV_MJPEG_DECODER_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_MJPEG_DECODER_H_
diff --git a/media/libyuv/include/libyuv/planar_functions.h b/media/libyuv/include/libyuv/planar_functions.h
index 881b0c5c64..154f2f2134 100644
--- a/media/libyuv/include/libyuv/planar_functions.h
+++ b/media/libyuv/include/libyuv/planar_functions.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
 #define INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
 
 #include "libyuv/basic_types.h"
@@ -22,88 +22,507 @@ namespace libyuv {
 extern "C" {
 #endif
 
+// TODO(fbarchard): Move cpu macros to row.h
+#if defined(__pnacl__) || defined(__CLR_VER) ||            \
+    (defined(__native_client__) && defined(__x86_64__)) || \
+    (defined(__i386__) && !defined(__SSE__) && !defined(__clang__))
+#define LIBYUV_DISABLE_X86
+#endif
+// MemorySanitizer does not support assembly code yet. http://crbug.com/344505
+#if defined(__has_feature)
+#if __has_feature(memory_sanitizer)
+#define LIBYUV_DISABLE_X86
+#endif
+#endif
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_ARGBAFFINEROW_SSE2
+#endif
+
 // Copy a plane of data.
 LIBYUV_API
-void CopyPlane(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height);
+void CopyPlane(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height);
+
+LIBYUV_API
+void CopyPlane_16(const uint16_t* src_y,
+                  int src_stride_y,
+                  uint16_t* dst_y,
+                  int dst_stride_y,
+                  int width,
+                  int height);
+
+LIBYUV_API
+void Convert16To8Plane(const uint16_t* src_y,
+                       int src_stride_y,
+                       uint8_t* dst_y,
+                       int dst_stride_y,
+                       int scale,  // 16384 for 10 bits
+                       int width,
+                       int height);
 
 LIBYUV_API
-void CopyPlane_16(const uint16* src_y, int src_stride_y,
-                  uint16* dst_y, int dst_stride_y,
-                  int width, int height);
+void Convert8To16Plane(const uint8_t* src_y,
+                       int src_stride_y,
+                       uint16_t* dst_y,
+                       int dst_stride_y,
+                       int scale,  // 1024 for 10 bits
+                       int width,
+                       int height);
 
 // Set a plane of data to a 32 bit value.
 LIBYUV_API
-void SetPlane(uint8* dst_y, int dst_stride_y,
-              int width, int height,
-              uint32 value);
+void SetPlane(uint8_t* dst_y,
+              int dst_stride_y,
+              int width,
+              int height,
+              uint32_t value);
+
+// Convert a plane of tiles of 16 x H to linear.
+LIBYUV_API
+int DetilePlane(const uint8_t* src_y,
+                int src_stride_y,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                int width,
+                int height,
+                int tile_height);
+
+// Convert a plane of 16 bit tiles of 16 x H to linear.
+LIBYUV_API
+int DetilePlane_16(const uint16_t* src_y,
+                   int src_stride_y,
+                   uint16_t* dst_y,
+                   int dst_stride_y,
+                   int width,
+                   int height,
+                   int tile_height);
+
+// Convert a UV plane of tiles of 16 x H into linear U and V planes.
+LIBYUV_API
+void DetileSplitUVPlane(const uint8_t* src_uv,
+                        int src_stride_uv,
+                        uint8_t* dst_u,
+                        int dst_stride_u,
+                        uint8_t* dst_v,
+                        int dst_stride_v,
+                        int width,
+                        int height,
+                        int tile_height);
+
+// Convert a Y and UV plane of tiles into interlaced YUY2.
+LIBYUV_API
+void DetileToYUY2(const uint8_t* src_y,
+                  int src_stride_y,
+                  const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_yuy2,
+                  int dst_stride_yuy2,
+                  int width,
+                  int height,
+                  int tile_height);
+
+// Split interleaved UV plane into separate U and V planes.
+LIBYUV_API
+void SplitUVPlane(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int width,
+                  int height);
+
+// Merge separate U and V planes into one interleaved UV plane.
+LIBYUV_API
+void MergeUVPlane(const uint8_t* src_u,
+                  int src_stride_u,
+                  const uint8_t* src_v,
+                  int src_stride_v,
+                  uint8_t* dst_uv,
+                  int dst_stride_uv,
+                  int width,
+                  int height);
+
+// Split interleaved msb UV plane into separate lsb U and V planes.
+LIBYUV_API
+void SplitUVPlane_16(const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint16_t* dst_u,
+                     int dst_stride_u,
+                     uint16_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
+                     int depth);
+
+// Merge separate lsb U and V planes into one interleaved msb UV plane.
+LIBYUV_API
+void MergeUVPlane_16(const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint16_t* dst_uv,
+                     int dst_stride_uv,
+                     int width,
+                     int height,
+                     int depth);
+
+// Convert lsb plane to msb plane
+LIBYUV_API
+void ConvertToMSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth);
+
+// Convert msb plane to lsb plane
+LIBYUV_API
+void ConvertToLSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth);
+
+// Scale U and V to half width and height and merge into interleaved UV plane.
+// width and height are source size, allowing odd sizes.
+// Use for converting I444 or I422 to NV12.
+LIBYUV_API
+void HalfMergeUVPlane(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int dst_stride_uv,
+                      int width,
+                      int height);
+
+// Swap U and V channels in interleaved UV plane.
+LIBYUV_API
+void SwapUVPlane(const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_vu,
+                 int dst_stride_vu,
+                 int width,
+                 int height);
+
+// Split interleaved RGB plane into separate R, G and B planes.
+LIBYUV_API
+void SplitRGBPlane(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_r,
+                   int dst_stride_r,
+                   uint8_t* dst_g,
+                   int dst_stride_g,
+                   uint8_t* dst_b,
+                   int dst_stride_b,
+                   int width,
+                   int height);
+
+// Merge separate R, G and B planes into one interleaved RGB plane.
+LIBYUV_API
+void MergeRGBPlane(const uint8_t* src_r,
+                   int src_stride_r,
+                   const uint8_t* src_g,
+                   int src_stride_g,
+                   const uint8_t* src_b,
+                   int src_stride_b,
+                   uint8_t* dst_rgb,
+                   int dst_stride_rgb,
+                   int width,
+                   int height);
+
+// Split interleaved ARGB plane into separate R, G, B and A planes.
+// dst_a can be NULL to discard alpha plane.
+LIBYUV_API
+void SplitARGBPlane(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_r,
+                    int dst_stride_r,
+                    uint8_t* dst_g,
+                    int dst_stride_g,
+                    uint8_t* dst_b,
+                    int dst_stride_b,
+                    uint8_t* dst_a,
+                    int dst_stride_a,
+                    int width,
+                    int height);
+
+// Merge separate R, G, B and A planes into one interleaved ARGB plane.
+// src_a can be NULL to fill opaque value to alpha.
+LIBYUV_API
+void MergeARGBPlane(const uint8_t* src_r,
+                    int src_stride_r,
+                    const uint8_t* src_g,
+                    int src_stride_g,
+                    const uint8_t* src_b,
+                    int src_stride_b,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height);
+
+// Merge separate 'depth' bit R, G and B planes stored in lsb
+// into one interleaved XR30 plane.
+// depth should in range [10, 16]
+LIBYUV_API
+void MergeXR30Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    uint8_t* dst_ar30,
+                    int dst_stride_ar30,
+                    int width,
+                    int height,
+                    int depth);
+
+// Merge separate 'depth' bit R, G, B and A planes stored in lsb
+// into one interleaved AR64 plane.
+// src_a can be NULL to fill opaque value to alpha.
+// depth should in range [1, 16]
+LIBYUV_API
+void MergeAR64Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    const uint16_t* src_a,
+                    int src_stride_a,
+                    uint16_t* dst_ar64,
+                    int dst_stride_ar64,
+                    int width,
+                    int height,
+                    int depth);
+
+// Merge separate 'depth' bit R, G, B and A planes stored in lsb
+// into one interleaved ARGB plane.
+// src_a can be NULL to fill opaque value to alpha.
+// depth should in range [8, 16]
+LIBYUV_API
+void MergeARGB16To8Plane(const uint16_t* src_r,
+                         int src_stride_r,
+                         const uint16_t* src_g,
+                         int src_stride_g,
+                         const uint16_t* src_b,
+                         int src_stride_b,
+                         const uint16_t* src_a,
+                         int src_stride_a,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height,
+                         int depth);
 
 // Copy I400.  Supports inverting.
 LIBYUV_API
-int I400ToI400(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height);
+int I400ToI400(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height);
 
 #define J400ToJ400 I400ToI400
 
 // Copy I422 to I422.
 #define I422ToI422 I422Copy
 LIBYUV_API
-int I422Copy(const uint8* src_y, int src_stride_y,
-             const uint8* src_u, int src_stride_u,
-             const uint8* src_v, int src_stride_v,
-             uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int width, int height);
+int I422Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_u,
+             int src_stride_u,
+             const uint8_t* src_v,
+             int src_stride_v,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
 
 // Copy I444 to I444.
 #define I444ToI444 I444Copy
 LIBYUV_API
-int I444Copy(const uint8* src_y, int src_stride_y,
-             const uint8* src_u, int src_stride_u,
-             const uint8* src_v, int src_stride_v,
-             uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int width, int height);
+int I444Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_u,
+             int src_stride_u,
+             const uint8_t* src_v,
+             int src_stride_v,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
+
+// Copy I210 to I210.
+#define I210ToI210 I210Copy
+LIBYUV_API
+int I210Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
+
+// Copy I410 to I410.
+#define I410ToI410 I410Copy
+LIBYUV_API
+int I410Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
+
+// Copy NV12. Supports inverting.
+LIBYUV_API
+int NV12Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_uv,
+             int src_stride_uv,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_uv,
+             int dst_stride_uv,
+             int width,
+             int height);
+
+// Copy NV21. Supports inverting.
+LIBYUV_API
+int NV21Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_vu,
+             int src_stride_vu,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_vu,
+             int dst_stride_vu,
+             int width,
+             int height);
 
 // Convert YUY2 to I422.
 LIBYUV_API
-int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int YUY2ToI422(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Convert UYVY to I422.
 LIBYUV_API
-int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int UYVYToI422(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+LIBYUV_API
+int YUY2ToNV12(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+LIBYUV_API
+int UYVYToNV12(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert NV21 to NV12.
+LIBYUV_API
+int NV21ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
 
 LIBYUV_API
-int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height);
+int YUY2ToY(const uint8_t* src_yuy2,
+            int src_stride_yuy2,
+            uint8_t* dst_y,
+            int dst_stride_y,
+            int width,
+            int height);
 
 LIBYUV_API
-int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height);
+int UYVYToY(const uint8_t* src_uyvy,
+            int src_stride_uyvy,
+            uint8_t* dst_y,
+            int dst_stride_y,
+            int width,
+            int height);
 
 // Convert I420 to I400. (calls CopyPlane ignoring u/v).
 LIBYUV_API
-int I420ToI400(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height);
+int I420ToI400(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height);
 
 // Alias
 #define J420ToJ400 I420ToI400
@@ -111,13 +530,20 @@ int I420ToI400(const uint8* src_y, int src_stride_y,
 
 // I420 mirror.
 LIBYUV_API
-int I420Mirror(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height);
+int I420Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
 
 // Alias
 #define I400ToI400Mirror I400Mirror
@@ -125,87 +551,134 @@ int I420Mirror(const uint8* src_y, int src_stride_y,
 // I400 mirror.  A single plane is mirrored horizontally.
 // Pass negative height to achieve 180 degree rotation.
 LIBYUV_API
-int I400Mirror(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height);
+int I400Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height);
+
+// Alias
+#define NV12ToNV12Mirror NV12Mirror
+
+// NV12 mirror.
+LIBYUV_API
+int NV12Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
 
 // Alias
 #define ARGBToARGBMirror ARGBMirror
 
 // ARGB mirror.
 LIBYUV_API
-int ARGBMirror(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int ARGBMirror(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
-// Convert NV12 to RGB565.
-LIBYUV_API
-int NV12ToRGB565(const uint8* src_y, int src_stride_y,
-                 const uint8* src_uv, int src_stride_uv,
-                 uint8* dst_rgb565, int dst_stride_rgb565,
-                 int width, int height);
+// Alias
+#define RGB24ToRGB24Mirror RGB24Mirror
 
-// I422ToARGB is in convert_argb.h
-// Convert I422 to BGRA.
+// RGB24 mirror.
 LIBYUV_API
-int I422ToBGRA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_bgra, int dst_stride_bgra,
-               int width, int height);
-
-// Convert I422 to ABGR.
+int RGB24Mirror(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+// Mirror a plane of data.
 LIBYUV_API
-int I422ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height);
-
-// Convert I422 to RGBA.
+void MirrorPlane(const uint8_t* src_y,
+                 int src_stride_y,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 int width,
+                 int height);
+
+// Mirror a plane of UV data.
 LIBYUV_API
-int I422ToRGBA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_rgba, int dst_stride_rgba,
-               int width, int height);
+void MirrorUVPlane(const uint8_t* src_uv,
+                   int src_stride_uv,
+                   uint8_t* dst_uv,
+                   int dst_stride_uv,
+                   int width,
+                   int height);
 
 // Alias
 #define RGB24ToRAW RAWToRGB24
 
 LIBYUV_API
-int RAWToRGB24(const uint8* src_raw, int src_stride_raw,
-               uint8* dst_rgb24, int dst_stride_rgb24,
-               int width, int height);
+int RAWToRGB24(const uint8_t* src_raw,
+               int src_stride_raw,
+               uint8_t* dst_rgb24,
+               int dst_stride_rgb24,
+               int width,
+               int height);
 
 // Draw a rectangle into I420.
 LIBYUV_API
-int I420Rect(uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int x, int y, int width, int height,
-             int value_y, int value_u, int value_v);
+int I420Rect(uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int x,
+             int y,
+             int width,
+             int height,
+             int value_y,
+             int value_u,
+             int value_v);
 
 // Draw a rectangle into ARGB.
 LIBYUV_API
-int ARGBRect(uint8* dst_argb, int dst_stride_argb,
-             int x, int y, int width, int height, uint32 value);
+int ARGBRect(uint8_t* dst_argb,
+             int dst_stride_argb,
+             int dst_x,
+             int dst_y,
+             int width,
+             int height,
+             uint32_t value);
 
 // Convert ARGB to gray scale ARGB.
 LIBYUV_API
-int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height);
+int ARGBGrayTo(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
 
 // Make a rectangle of ARGB gray scale.
 LIBYUV_API
-int ARGBGray(uint8* dst_argb, int dst_stride_argb,
-             int x, int y, int width, int height);
+int ARGBGray(uint8_t* dst_argb,
+             int dst_stride_argb,
+             int dst_x,
+             int dst_y,
+             int width,
+             int height);
 
 // Make a rectangle of ARGB Sepia tone.
 LIBYUV_API
-int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
-              int x, int y, int width, int height);
+int ARGBSepia(uint8_t* dst_argb,
+              int dst_stride_argb,
+              int dst_x,
+              int dst_y,
+              int width,
+              int height);
 
 // Apply a matrix rotation to each ARGB pixel.
 // matrix_argb is 4 signed ARGB values. -128 to 127 representing -2 to 2.
@@ -214,10 +687,13 @@ int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
 // The next 4 coefficients apply to B, G, R, A and produce R of the output.
 // The last 4 coefficients apply to B, G, R, A and produce A of the output.
 LIBYUV_API
-int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
-                    uint8* dst_argb, int dst_stride_argb,
-                    const int8* matrix_argb,
-                    int width, int height);
+int ARGBColorMatrix(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    const int8_t* matrix_argb,
+                    int width,
+                    int height);
 
 // Deprecated. Use ARGBColorMatrix instead.
 // Apply a matrix rotation to each ARGB pixel.
@@ -226,32 +702,47 @@ int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
 // The next 4 coefficients apply to B, G, R, A and produce G of the output.
 // The last 4 coefficients apply to B, G, R, A and produce R of the output.
 LIBYUV_API
-int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
-                   const int8* matrix_rgb,
-                   int x, int y, int width, int height);
+int RGBColorMatrix(uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   const int8_t* matrix_rgb,
+                   int dst_x,
+                   int dst_y,
+                   int width,
+                   int height);
 
 // Apply a color table each ARGB pixel.
 // Table contains 256 ARGB values.
 LIBYUV_API
-int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
-                   const uint8* table_argb,
-                   int x, int y, int width, int height);
+int ARGBColorTable(uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   const uint8_t* table_argb,
+                   int dst_x,
+                   int dst_y,
+                   int width,
+                   int height);
 
 // Apply a color table each ARGB pixel but preserve destination alpha.
 // Table contains 256 ARGB values.
 LIBYUV_API
-int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
-                  const uint8* table_argb,
-                  int x, int y, int width, int height);
+int RGBColorTable(uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  const uint8_t* table_argb,
+                  int dst_x,
+                  int dst_y,
+                  int width,
+                  int height);
 
 // Apply a luma/color table each ARGB pixel but preserve destination alpha.
 // Table contains 32768 values indexed by [Y][C] where 7 it 7 bit luma from
 // RGB (YJ style) and C is an 8 bit color component (R, G or B).
 LIBYUV_API
-int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_argb, int dst_stride_argb,
-                       const uint8* luma_rgb_table,
-                       int width, int height);
+int ARGBLumaColorTable(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_argb,
+                       int dst_stride_argb,
+                       const uint8_t* luma,
+                       int width,
+                       int height);
 
 // Apply a 3 term polynomial to ARGB values.
 // poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
@@ -262,46 +753,84 @@ int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
 // A polynomial approximation can be dirived using software such as 'R'.
 
 LIBYUV_API
-int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_argb, int dst_stride_argb,
+int ARGBPolynomial(const uint8_t* src_argb,
+                   int src_stride_argb,
+                   uint8_t* dst_argb,
+                   int dst_stride_argb,
                    const float* poly,
-                   int width, int height);
+                   int width,
+                   int height);
+
+// Convert plane of 16 bit shorts to half floats.
+// Source values are multiplied by scale before storing as half float.
+LIBYUV_API
+int HalfFloatPlane(const uint16_t* src_y,
+                   int src_stride_y,
+                   uint16_t* dst_y,
+                   int dst_stride_y,
+                   float scale,
+                   int width,
+                   int height);
+
+// Convert a buffer of bytes to floats, scale the values and store as floats.
+LIBYUV_API
+int ByteToFloat(const uint8_t* src_y, float* dst_y, float scale, int width);
 
 // Quantize a rectangle of ARGB. Alpha unaffected.
 // scale is a 16 bit fractional fixed point scaler between 0 and 65535.
 // interval_size should be a value between 1 and 255.
 // interval_offset should be a value between 0 and 255.
 LIBYUV_API
-int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
-                 int scale, int interval_size, int interval_offset,
-                 int x, int y, int width, int height);
+int ARGBQuantize(uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int scale,
+                 int interval_size,
+                 int interval_offset,
+                 int dst_x,
+                 int dst_y,
+                 int width,
+                 int height);
 
 // Copy ARGB to ARGB.
 LIBYUV_API
-int ARGBCopy(const uint8* src_argb, int src_stride_argb,
-             uint8* dst_argb, int dst_stride_argb,
-             int width, int height);
+int ARGBCopy(const uint8_t* src_argb,
+             int src_stride_argb,
+             uint8_t* dst_argb,
+             int dst_stride_argb,
+             int width,
+             int height);
 
 // Copy Alpha channel of ARGB to alpha of ARGB.
 LIBYUV_API
-int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int width, int height);
+int ARGBCopyAlpha(const uint8_t* src_argb,
+                  int src_stride_argb,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int width,
+                  int height);
 
 // Extract the alpha channel from ARGB.
 LIBYUV_API
-int ARGBExtractAlpha(const uint8* src_argb, int src_stride_argb,
-                     uint8* dst_a, int dst_stride_a,
-                     int width, int height);
+int ARGBExtractAlpha(const uint8_t* src_argb,
+                     int src_stride_argb,
+                     uint8_t* dst_a,
+                     int dst_stride_a,
+                     int width,
+                     int height);
 
 // Copy Y channel to Alpha of ARGB.
 LIBYUV_API
-int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
-                     uint8* dst_argb, int dst_stride_argb,
-                     int width, int height);
-
-typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
-                             uint8* dst_argb, int width);
+int ARGBCopyYToAlpha(const uint8_t* src_y,
+                     int src_stride_y,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     int width,
+                     int height);
+
+typedef void (*ARGBBlendRow)(const uint8_t* src_argb0,
+                             const uint8_t* src_argb1,
+                             uint8_t* dst_argb,
+                             int width);
 
 // Get function to Alpha Blend ARGB pixels and store to destination.
 LIBYUV_API
@@ -311,92 +840,143 @@ ARGBBlendRow GetARGBBlend();
 // Source is pre-multiplied by alpha using ARGBAttenuate.
 // Alpha of destination is set to 255.
 LIBYUV_API
-int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
-              const uint8* src_argb1, int src_stride_argb1,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height);
+int ARGBBlend(const uint8_t* src_argb0,
+              int src_stride_argb0,
+              const uint8_t* src_argb1,
+              int src_stride_argb1,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height);
 
 // Alpha Blend plane and store to destination.
 // Source is not pre-multiplied by alpha.
 LIBYUV_API
-int BlendPlane(const uint8* src_y0, int src_stride_y0,
-               const uint8* src_y1, int src_stride_y1,
-               const uint8* alpha, int alpha_stride,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height);
+int BlendPlane(const uint8_t* src_y0,
+               int src_stride_y0,
+               const uint8_t* src_y1,
+               int src_stride_y1,
+               const uint8_t* alpha,
+               int alpha_stride,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height);
 
 // Alpha Blend YUV images and store to destination.
 // Source is not pre-multiplied by alpha.
 // Alpha is full width x height and subsampled to half size to apply to UV.
 LIBYUV_API
-int I420Blend(const uint8* src_y0, int src_stride_y0,
-              const uint8* src_u0, int src_stride_u0,
-              const uint8* src_v0, int src_stride_v0,
-              const uint8* src_y1, int src_stride_y1,
-              const uint8* src_u1, int src_stride_u1,
-              const uint8* src_v1, int src_stride_v1,
-              const uint8* alpha, int alpha_stride,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int width, int height);
+int I420Blend(const uint8_t* src_y0,
+              int src_stride_y0,
+              const uint8_t* src_u0,
+              int src_stride_u0,
+              const uint8_t* src_v0,
+              int src_stride_v0,
+              const uint8_t* src_y1,
+              int src_stride_y1,
+              const uint8_t* src_u1,
+              int src_stride_u1,
+              const uint8_t* src_v1,
+              int src_stride_v1,
+              const uint8_t* alpha,
+              int alpha_stride,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height);
 
 // Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
 LIBYUV_API
-int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
-                 const uint8* src_argb1, int src_stride_argb1,
-                 uint8* dst_argb, int dst_stride_argb,
-                 int width, int height);
+int ARGBMultiply(const uint8_t* src_argb0,
+                 int src_stride_argb0,
+                 const uint8_t* src_argb1,
+                 int src_stride_argb1,
+                 uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int width,
+                 int height);
 
 // Add ARGB image with ARGB image. Saturates to 255.
 LIBYUV_API
-int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
-            const uint8* src_argb1, int src_stride_argb1,
-            uint8* dst_argb, int dst_stride_argb,
-            int width, int height);
+int ARGBAdd(const uint8_t* src_argb0,
+            int src_stride_argb0,
+            const uint8_t* src_argb1,
+            int src_stride_argb1,
+            uint8_t* dst_argb,
+            int dst_stride_argb,
+            int width,
+            int height);
 
 // Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
 LIBYUV_API
-int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
-                 const uint8* src_argb1, int src_stride_argb1,
-                 uint8* dst_argb, int dst_stride_argb,
-                 int width, int height);
+int ARGBSubtract(const uint8_t* src_argb0,
+                 int src_stride_argb0,
+                 const uint8_t* src_argb1,
+                 int src_stride_argb1,
+                 uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int width,
+                 int height);
 
 // Convert I422 to YUY2.
 LIBYUV_API
-int I422ToYUY2(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_frame, int dst_stride_frame,
-               int width, int height);
+int I422ToYUY2(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height);
 
 // Convert I422 to UYVY.
 LIBYUV_API
-int I422ToUYVY(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_frame, int dst_stride_frame,
-               int width, int height);
+int I422ToUYVY(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_uyvy,
+               int dst_stride_uyvy,
+               int width,
+               int height);
 
 // Convert unattentuated ARGB to preattenuated ARGB.
 LIBYUV_API
-int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int width, int height);
+int ARGBAttenuate(const uint8_t* src_argb,
+                  int src_stride_argb,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int width,
+                  int height);
 
 // Convert preattentuated ARGB to unattenuated ARGB.
 LIBYUV_API
-int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
-                    uint8* dst_argb, int dst_stride_argb,
-                    int width, int height);
+int ARGBUnattenuate(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height);
 
 // Internal function - do not call directly.
 // Computes table of cumulative sum for image where the value is the sum
 // of all values above and to the left of the entry. Used by ARGBBlur.
 LIBYUV_API
-int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
-                             int32* dst_cumsum, int dst_stride32_cumsum,
-                             int width, int height);
+int ARGBComputeCumulativeSum(const uint8_t* src_argb,
+                             int src_stride_argb,
+                             int32_t* dst_cumsum,
+                             int dst_stride32_cumsum,
+                             int width,
+                             int height);
 
 // Blur ARGB image.
 // dst_cumsum table of width * (height + 1) * 16 bytes aligned to
@@ -405,103 +985,176 @@ int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
 // radius is number of pixels around the center.  e.g. 1 = 3x3. 2=5x5.
 // Blur is optimized for radius of 5 (11x11) or less.
 LIBYUV_API
-int ARGBBlur(const uint8* src_argb, int src_stride_argb,
-             uint8* dst_argb, int dst_stride_argb,
-             int32* dst_cumsum, int dst_stride32_cumsum,
-             int width, int height, int radius);
+int ARGBBlur(const uint8_t* src_argb,
+             int src_stride_argb,
+             uint8_t* dst_argb,
+             int dst_stride_argb,
+             int32_t* dst_cumsum,
+             int dst_stride32_cumsum,
+             int width,
+             int height,
+             int radius);
+
+// Gaussian 5x5 blur a float plane.
+// Coefficients of 1, 4, 6, 4, 1.
+// Each destination pixel is a blur of the 5x5
+// pixels from the source.
+// Source edges are clamped.
+LIBYUV_API
+int GaussPlane_F32(const float* src,
+                   int src_stride,
+                   float* dst,
+                   int dst_stride,
+                   int width,
+                   int height);
 
 // Multiply ARGB image by ARGB value.
 LIBYUV_API
-int ARGBShade(const uint8* src_argb, int src_stride_argb,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height, uint32 value);
+int ARGBShade(const uint8_t* src_argb,
+              int src_stride_argb,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height,
+              uint32_t value);
 
 // Interpolate between two images using specified amount of interpolation
 // (0 to 255) and store to destination.
 // 'interpolation' is specified as 8 bit fraction where 0 means 100% src0
 // and 255 means 1% src0 and 99% src1.
 LIBYUV_API
-int InterpolatePlane(const uint8* src0, int src_stride0,
-                     const uint8* src1, int src_stride1,
-                     uint8* dst, int dst_stride,
-                     int width, int height, int interpolation);
+int InterpolatePlane(const uint8_t* src0,
+                     int src_stride0,
+                     const uint8_t* src1,
+                     int src_stride1,
+                     uint8_t* dst,
+                     int dst_stride,
+                     int width,
+                     int height,
+                     int interpolation);
+
+// Interpolate between two images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src0
+// and 255 means 1% src0 and 99% src1.
+LIBYUV_API
+int InterpolatePlane_16(const uint16_t* src0,
+                        int src_stride0,  // measured in 16 bit pixels
+                        const uint16_t* src1,
+                        int src_stride1,
+                        uint16_t* dst,
+                        int dst_stride,
+                        int width,
+                        int height,
+                        int interpolation);
 
 // Interpolate between two ARGB images using specified amount of interpolation
 // Internally calls InterpolatePlane with width * 4 (bpp).
 LIBYUV_API
-int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
-                    const uint8* src_argb1, int src_stride_argb1,
-                    uint8* dst_argb, int dst_stride_argb,
-                    int width, int height, int interpolation);
+int ARGBInterpolate(const uint8_t* src_argb0,
+                    int src_stride_argb0,
+                    const uint8_t* src_argb1,
+                    int src_stride_argb1,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int interpolation);
 
 // Interpolate between two YUV images using specified amount of interpolation
 // Internally calls InterpolatePlane on each plane where the U and V planes
 // are half width and half height.
 LIBYUV_API
-int I420Interpolate(const uint8* src0_y, int src0_stride_y,
-                    const uint8* src0_u, int src0_stride_u,
-                    const uint8* src0_v, int src0_stride_v,
-                    const uint8* src1_y, int src1_stride_y,
-                    const uint8* src1_u, int src1_stride_u,
-                    const uint8* src1_v, int src1_stride_v,
-                    uint8* dst_y, int dst_stride_y,
-                    uint8* dst_u, int dst_stride_u,
-                    uint8* dst_v, int dst_stride_v,
-                    int width, int height, int interpolation);
-
-#if defined(__pnacl__) || defined(__CLR_VER) || \
-    (defined(__i386__) && !defined(__SSE2__))
-#define LIBYUV_DISABLE_X86
-#endif
-// MemorySanitizer does not support assembly code yet. http://crbug.com/344505
-#if defined(__has_feature)
-#if __has_feature(memory_sanitizer)
-#define LIBYUV_DISABLE_X86
-#endif
-#endif
-// The following are available on all x86 platforms:
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
-#define HAS_ARGBAFFINEROW_SSE2
-#endif
+int I420Interpolate(const uint8_t* src0_y,
+                    int src0_stride_y,
+                    const uint8_t* src0_u,
+                    int src0_stride_u,
+                    const uint8_t* src0_v,
+                    int src0_stride_v,
+                    const uint8_t* src1_y,
+                    int src1_stride_y,
+                    const uint8_t* src1_u,
+                    int src1_stride_u,
+                    const uint8_t* src1_v,
+                    int src1_stride_v,
+                    uint8_t* dst_y,
+                    int dst_stride_y,
+                    uint8_t* dst_u,
+                    int dst_stride_u,
+                    uint8_t* dst_v,
+                    int dst_stride_v,
+                    int width,
+                    int height,
+                    int interpolation);
 
 // Row function for copying pixels from a source with a slope to a row
 // of destination. Useful for scaling, rotation, mirror, texture mapping.
 LIBYUV_API
-void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
-                     uint8* dst_argb, const float* uv_dudv, int width);
+void ARGBAffineRow_C(const uint8_t* src_argb,
+                     int src_argb_stride,
+                     uint8_t* dst_argb,
+                     const float* uv_dudv,
+                     int width);
+// TODO(fbarchard): Move ARGBAffineRow_SSE2 to row.h
 LIBYUV_API
-void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
-                        uint8* dst_argb, const float* uv_dudv, int width);
+void ARGBAffineRow_SSE2(const uint8_t* src_argb,
+                        int src_argb_stride,
+                        uint8_t* dst_argb,
+                        const float* uv_dudv,
+                        int width);
 
 // Shuffle ARGB channel order.  e.g. BGRA to ARGB.
-// shuffler is 16 bytes and must be aligned.
+// shuffler is 16 bytes.
+LIBYUV_API
+int ARGBShuffle(const uint8_t* src_bgra,
+                int src_stride_bgra,
+                uint8_t* dst_argb,
+                int dst_stride_argb,
+                const uint8_t* shuffler,
+                int width,
+                int height);
+
+// Shuffle AR64 channel order.  e.g. AR64 to AB64.
+// shuffler is 16 bytes.
 LIBYUV_API
-int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
-                uint8* dst_argb, int dst_stride_argb,
-                const uint8* shuffler, int width, int height);
+int AR64Shuffle(const uint16_t* src_ar64,
+                int src_stride_ar64,
+                uint16_t* dst_ar64,
+                int dst_stride_ar64,
+                const uint8_t* shuffler,
+                int width,
+                int height);
 
 // Sobel ARGB effect with planar output.
 LIBYUV_API
-int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
-                     uint8* dst_y, int dst_stride_y,
-                     int width, int height);
+int ARGBSobelToPlane(const uint8_t* src_argb,
+                     int src_stride_argb,
+                     uint8_t* dst_y,
+                     int dst_stride_y,
+                     int width,
+                     int height);
 
 // Sobel ARGB effect.
 LIBYUV_API
-int ARGBSobel(const uint8* src_argb, int src_stride_argb,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height);
+int ARGBSobel(const uint8_t* src_argb,
+              int src_stride_argb,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height);
 
 // Sobel ARGB effect w/ Sobel X, Sobel, Sobel Y in ARGB.
 LIBYUV_API
-int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
-                uint8* dst_argb, int dst_stride_argb,
-                int width, int height);
+int ARGBSobelXY(const uint8_t* src_argb,
+                int src_stride_argb,
+                uint8_t* dst_argb,
+                int dst_stride_argb,
+                int width,
+                int height);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
diff --git a/media/libyuv/include/libyuv/rotate.h b/media/libyuv/include/libyuv/rotate.h
index 8af60b8955..37460c4ac5 100644
--- a/media/libyuv/include/libyuv/rotate.h
+++ b/media/libyuv/include/libyuv/rotate.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_ROTATE_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_ROTATE_H_
 #define INCLUDE_LIBYUV_ROTATE_H_
 
 #include "libyuv/basic_types.h"
@@ -20,8 +20,8 @@ extern "C" {
 
 // Supported rotation.
 typedef enum RotationMode {
-  kRotate0 = 0,  // No rotation.
-  kRotate90 = 90,  // Rotate 90 degrees clockwise.
+  kRotate0 = 0,      // No rotation.
+  kRotate90 = 90,    // Rotate 90 degrees clockwise.
   kRotate180 = 180,  // Rotate 180 degrees.
   kRotate270 = 270,  // Rotate 270 degrees clockwise.
 
@@ -33,85 +33,264 @@ typedef enum RotationMode {
 
 // Rotate I420 frame.
 LIBYUV_API
-int I420Rotate(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int src_width, int src_height, enum RotationMode mode);
+int I420Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I422 frame.
+LIBYUV_API
+int I422Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I444 frame.
+LIBYUV_API
+int I444Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I010 frame.
+LIBYUV_API
+int I010Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I210 frame.
+LIBYUV_API
+int I210Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I410 frame.
+LIBYUV_API
+int I410Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
 
 // Rotate NV12 input and store in I420.
 LIBYUV_API
-int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
-                     const uint8* src_uv, int src_stride_uv,
-                     uint8* dst_y, int dst_stride_y,
-                     uint8* dst_u, int dst_stride_u,
-                     uint8* dst_v, int dst_stride_v,
-                     int src_width, int src_height, enum RotationMode mode);
+int NV12ToI420Rotate(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_y,
+                     int dst_stride_y,
+                     uint8_t* dst_u,
+                     int dst_stride_u,
+                     uint8_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
+                     enum RotationMode mode);
+
+// Convert Android420 to I420 with rotation.
+// "rotation" can be 0, 90, 180 or 270.
+LIBYUV_API
+int Android420ToI420Rotate(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_y,
+                           int dst_stride_y,
+                           uint8_t* dst_u,
+                           int dst_stride_u,
+                           uint8_t* dst_v,
+                           int dst_stride_v,
+                           int width,
+                           int height,
+                           enum RotationMode rotation);
 
 // Rotate a plane by 0, 90, 180, or 270.
 LIBYUV_API
-int RotatePlane(const uint8* src, int src_stride,
-                uint8* dst, int dst_stride,
-                int src_width, int src_height, enum RotationMode mode);
+int RotatePlane(const uint8_t* src,
+                int src_stride,
+                uint8_t* dst,
+                int dst_stride,
+                int width,
+                int height,
+                enum RotationMode mode);
 
 // Rotate planes by 90, 180, 270. Deprecated.
 LIBYUV_API
-void RotatePlane90(const uint8* src, int src_stride,
-                   uint8* dst, int dst_stride,
-                   int width, int height);
+void RotatePlane90(const uint8_t* src,
+                   int src_stride,
+                   uint8_t* dst,
+                   int dst_stride,
+                   int width,
+                   int height);
 
 LIBYUV_API
-void RotatePlane180(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height);
+void RotatePlane180(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height);
 
 LIBYUV_API
-void RotatePlane270(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height);
+void RotatePlane270(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height);
 
+// Rotate a plane by 0, 90, 180, or 270.
 LIBYUV_API
-void RotateUV90(const uint8* src, int src_stride,
-                uint8* dst_a, int dst_stride_a,
-                uint8* dst_b, int dst_stride_b,
-                int width, int height);
+int RotatePlane_16(const uint16_t* src,
+                   int src_stride,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int width,
+                   int height,
+                   enum RotationMode mode);
 
 // Rotations for when U and V are interleaved.
-// These functions take one input pointer and
+// These functions take one UV input pointer and
 // split the data into two buffers while
-// rotating them. Deprecated.
+// rotating them.
+// width and height expected to be half size for NV12.
+LIBYUV_API
+int SplitRotateUV(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int width,
+                  int height,
+                  enum RotationMode mode);
+
+LIBYUV_API
+void SplitRotateUV90(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst_a,
+                     int dst_stride_a,
+                     uint8_t* dst_b,
+                     int dst_stride_b,
+                     int width,
+                     int height);
+
 LIBYUV_API
-void RotateUV180(const uint8* src, int src_stride,
-                 uint8* dst_a, int dst_stride_a,
-                 uint8* dst_b, int dst_stride_b,
-                 int width, int height);
+void SplitRotateUV180(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
 
 LIBYUV_API
-void RotateUV270(const uint8* src, int src_stride,
-                 uint8* dst_a, int dst_stride_a,
-                 uint8* dst_b, int dst_stride_b,
-                 int width, int height);
+void SplitRotateUV270(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
 
 // The 90 and 270 functions are based on transposes.
 // Doing a transpose with reversing the read/write
 // order will result in a rotation by +- 90 degrees.
 // Deprecated.
 LIBYUV_API
-void TransposePlane(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height);
+void TransposePlane(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height);
 
 LIBYUV_API
-void TransposeUV(const uint8* src, int src_stride,
-                 uint8* dst_a, int dst_stride_a,
-                 uint8* dst_b, int dst_stride_b,
-                 int width, int height);
+void SplitTransposeUV(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_ROTATE_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_ROTATE_H_
diff --git a/media/libyuv/include/libyuv/rotate_argb.h b/media/libyuv/include/libyuv/rotate_argb.h
index 660ff5573e..20432949ab 100644
--- a/media/libyuv/include/libyuv/rotate_argb.h
+++ b/media/libyuv/include/libyuv/rotate_argb.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_ROTATE_ARGB_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_ROTATE_ARGB_H_
 #define INCLUDE_LIBYUV_ROTATE_ARGB_H_
 
 #include "libyuv/basic_types.h"
@@ -21,13 +21,17 @@ extern "C" {
 
 // Rotate ARGB frame
 LIBYUV_API
-int ARGBRotate(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_argb, int dst_stride_argb,
-               int src_width, int src_height, enum RotationMode mode);
+int ARGBRotate(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int src_width,
+               int src_height,
+               enum RotationMode mode);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_ROTATE_ARGB_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_ROTATE_ARGB_H_
diff --git a/media/libyuv/include/libyuv/rotate_row.h b/media/libyuv/include/libyuv/rotate_row.h
index ebc487f9ab..64d0b59f7f 100644
--- a/media/libyuv/include/libyuv/rotate_row.h
+++ b/media/libyuv/include/libyuv/rotate_row.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_ROTATE_ROW_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_ROTATE_ROW_H_
 #define INCLUDE_LIBYUV_ROTATE_ROW_H_
 
 #include "libyuv/basic_types.h"
@@ -18,104 +18,244 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#if defined(__pnacl__) || defined(__CLR_VER) || \
-    (defined(__i386__) && !defined(__SSE2__))
+#if defined(__pnacl__) || defined(__CLR_VER) ||            \
+    (defined(__native_client__) && defined(__x86_64__)) || \
+    (defined(__i386__) && !defined(__SSE__) && !defined(__clang__))
 #define LIBYUV_DISABLE_X86
 #endif
+#if defined(__native_client__)
+#define LIBYUV_DISABLE_NEON
+#endif
 // MemorySanitizer does not support assembly code yet. http://crbug.com/344505
 #if defined(__has_feature)
 #if __has_feature(memory_sanitizer)
 #define LIBYUV_DISABLE_X86
 #endif
 #endif
-// The following are available for Visual C and clangcl 32 bit:
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+// The following are available for Visual C 32 bit:
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) && \
+    !defined(__clang__)
 #define HAS_TRANSPOSEWX8_SSSE3
 #define HAS_TRANSPOSEUVWX8_SSE2
 #endif
 
-// The following are available for GCC 32 or 64 bit but not NaCL for 64 bit:
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+// The following are available for GCC 32 or 64 bit:
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__i386__) || defined(__x86_64__))
 #define HAS_TRANSPOSEWX8_SSSE3
 #endif
 
-// The following are available for 64 bit GCC but not NaCL:
-#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
-    defined(__x86_64__)
+// The following are available for 64 bit GCC:
+#if !defined(LIBYUV_DISABLE_X86) && defined(__x86_64__)
 #define HAS_TRANSPOSEWX8_FAST_SSSE3
 #define HAS_TRANSPOSEUVWX8_SSE2
 #endif
 
-#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+#if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
 #define HAS_TRANSPOSEWX8_NEON
 #define HAS_TRANSPOSEUVWX8_NEON
 #endif
 
-#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
-    defined(__mips__) && \
-    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-#define HAS_TRANSPOSEWX8_DSPR2
-#define HAS_TRANSPOSEUVWX8_DSPR2
-#endif  // defined(__mips__)
-
-void TransposeWxH_C(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride, int width, int height);
-
-void TransposeWx8_C(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride, int width);
-void TransposeWx8_NEON(const uint8* src, int src_stride,
-                       uint8* dst, int dst_stride, int width);
-void TransposeWx8_SSSE3(const uint8* src, int src_stride,
-                        uint8* dst, int dst_stride, int width);
-void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride,
-                             uint8* dst, int dst_stride, int width);
-void TransposeWx8_DSPR2(const uint8* src, int src_stride,
-                        uint8* dst, int dst_stride, int width);
-void TransposeWx8_Fast_DSPR2(const uint8* src, int src_stride,
-                             uint8* dst, int dst_stride, int width);
-
-void TransposeWx8_Any_NEON(const uint8* src, int src_stride,
-                           uint8* dst, int dst_stride, int width);
-void TransposeWx8_Any_SSSE3(const uint8* src, int src_stride,
-                            uint8* dst, int dst_stride, int width);
-void TransposeWx8_Fast_Any_SSSE3(const uint8* src, int src_stride,
-                                 uint8* dst, int dst_stride, int width);
-void TransposeWx8_Any_DSPR2(const uint8* src, int src_stride,
-                            uint8* dst, int dst_stride, int width);
-
-void TransposeUVWxH_C(const uint8* src, int src_stride,
-                      uint8* dst_a, int dst_stride_a,
-                      uint8* dst_b, int dst_stride_b,
-                      int width, int height);
-
-void TransposeUVWx8_C(const uint8* src, int src_stride,
-                      uint8* dst_a, int dst_stride_a,
-                      uint8* dst_b, int dst_stride_b, int width);
-void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b, int width);
-void TransposeUVWx8_NEON(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b, int width);
-void TransposeUVWx8_DSPR2(const uint8* src, int src_stride,
-                          uint8* dst_a, int dst_stride_a,
-                          uint8* dst_b, int dst_stride_b, int width);
-
-void TransposeUVWx8_Any_SSE2(const uint8* src, int src_stride,
-                             uint8* dst_a, int dst_stride_a,
-                             uint8* dst_b, int dst_stride_b, int width);
-void TransposeUVWx8_Any_NEON(const uint8* src, int src_stride,
-                             uint8* dst_a, int dst_stride_a,
-                             uint8* dst_b, int dst_stride_b, int width);
-void TransposeUVWx8_Any_DSPR2(const uint8* src, int src_stride,
-                              uint8* dst_a, int dst_stride_a,
-                              uint8* dst_b, int dst_stride_b, int width);
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#define HAS_TRANSPOSEWX16_MSA
+#define HAS_TRANSPOSEUVWX16_MSA
+#endif
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#define HAS_TRANSPOSEWX16_LSX
+#define HAS_TRANSPOSEUVWX16_LSX
+#endif
+
+void TransposeWxH_C(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height);
+
+void TransposeWx8_C(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width);
+void TransposeWx16_C(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst,
+                     int dst_stride,
+                     int width);
+void TransposeWx8_NEON(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width);
+void TransposeWx8_SSSE3(const uint8_t* src,
+                        int src_stride,
+                        uint8_t* dst,
+                        int dst_stride,
+                        int width);
+void TransposeWx8_Fast_SSSE3(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst,
+                             int dst_stride,
+                             int width);
+void TransposeWx16_MSA(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width);
+void TransposeWx16_LSX(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width);
+
+void TransposeWx8_Any_NEON(const uint8_t* src,
+                           int src_stride,
+                           uint8_t* dst,
+                           int dst_stride,
+                           int width);
+void TransposeWx8_Any_SSSE3(const uint8_t* src,
+                            int src_stride,
+                            uint8_t* dst,
+                            int dst_stride,
+                            int width);
+void TransposeWx8_Fast_Any_SSSE3(const uint8_t* src,
+                                 int src_stride,
+                                 uint8_t* dst,
+                                 int dst_stride,
+                                 int width);
+void TransposeWx16_Any_MSA(const uint8_t* src,
+                           int src_stride,
+                           uint8_t* dst,
+                           int dst_stride,
+                           int width);
+void TransposeWx16_Any_LSX(const uint8_t* src,
+                           int src_stride,
+                           uint8_t* dst,
+                           int dst_stride,
+                           int width);
+
+void TransposeUVWxH_C(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
+
+void TransposeUVWx8_C(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width);
+void TransposeUVWx16_C(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst_a,
+                       int dst_stride_a,
+                       uint8_t* dst_b,
+                       int dst_stride_b,
+                       int width);
+void TransposeUVWx8_SSE2(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width);
+void TransposeUVWx8_NEON(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width);
+void TransposeUVWx16_MSA(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width);
+void TransposeUVWx16_LSX(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width);
+
+void TransposeUVWx8_Any_SSE2(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst_a,
+                             int dst_stride_a,
+                             uint8_t* dst_b,
+                             int dst_stride_b,
+                             int width);
+void TransposeUVWx8_Any_NEON(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst_a,
+                             int dst_stride_a,
+                             uint8_t* dst_b,
+                             int dst_stride_b,
+                             int width);
+void TransposeUVWx16_Any_MSA(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst_a,
+                             int dst_stride_a,
+                             uint8_t* dst_b,
+                             int dst_stride_b,
+                             int width);
+void TransposeUVWx16_Any_LSX(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst_a,
+                             int dst_stride_a,
+                             uint8_t* dst_b,
+                             int dst_stride_b,
+                             int width);
+void TransposeWxH_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width,
+                       int height);
+
+void TransposeWx8_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width);
+void TransposeWx1_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width);
+
+// Transpose 32 bit values (ARGB)
+void Transpose4x4_32_NEON(const uint8_t* src,
+                          int src_stride,
+                          uint8_t* dst,
+                          int dst_stride,
+                          int width);
+
+void Transpose4x4_32_C(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width);
+
+// Transpose 32 bit values (ARGB)
+void Transpose8x8_32_NEON(const uint8_t* src,
+                          int src_stride,
+                          uint8_t* dst,
+                          int dst_stride,
+                          int width);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_ROTATE_ROW_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_ROTATE_ROW_H_
diff --git a/media/libyuv/include/libyuv/row.h b/media/libyuv/include/libyuv/row.h
index 069d73bd3f..ff6ffe47cb 100644
--- a/media/libyuv/include/libyuv/row.h
+++ b/media/libyuv/include/libyuv/row.h
@@ -8,10 +8,11 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_ROW_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_ROW_H_
 #define INCLUDE_LIBYUV_ROW_H_
 
-#include <stdlib.h>  // For malloc.
+#include <stddef.h>  // For NULL
+#include <stdlib.h>  // For malloc
 
 #include "libyuv/basic_types.h"
 
@@ -20,41 +21,20 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
-
-#ifdef __cplusplus
-#define align_buffer_64(var, size)                                             \
-  uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63));            \
-  uint8* var = reinterpret_cast<uint8*>                                        \
-      ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
-#else
-#define align_buffer_64(var, size)                                             \
-  uint8* var##_mem = (uint8*)(malloc((size) + 63));               /* NOLINT */ \
-  uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63)       /* NOLINT */
-#endif
-
-#define free_aligned_buffer_64(var) \
-  free(var##_mem);  \
-  var = 0
-
-#if defined(__pnacl__) || defined(__CLR_VER) || \
-    (defined(__i386__) && !defined(__SSE2__))
+#if defined(__pnacl__) || defined(__CLR_VER) ||            \
+    (defined(__native_client__) && defined(__x86_64__)) || \
+    (defined(__i386__) && !defined(__SSE__) && !defined(__clang__))
 #define LIBYUV_DISABLE_X86
 #endif
+#if defined(__native_client__)
+#define LIBYUV_DISABLE_NEON
+#endif
 // MemorySanitizer does not support assembly code yet. http://crbug.com/344505
 #if defined(__has_feature)
 #if __has_feature(memory_sanitizer)
 #define LIBYUV_DISABLE_X86
 #endif
 #endif
-// True if compiling for SSSE3 as a requirement.
-#if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
-#define LIBYUV_SSSE3_ONLY
-#endif
-
-#if defined(__native_client__)
-#define LIBYUV_DISABLE_NEON
-#endif
 // clang >= 3.5.0 required for Arm64.
 #if defined(__clang__) && defined(__aarch64__) && !defined(LIBYUV_DISABLE_NEON)
 #if (__clang_major__ < 3) || (__clang_major__ == 3 && (__clang_minor__ < 5))
@@ -76,9 +56,18 @@ extern "C" {
 #endif  // clang >= 3.4
 #endif  // __clang__
 
+// clang >= 6.0.0 required for AVX512.
+#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
+// clang in xcode follows a different versioning scheme.
+// TODO(fbarchard): fix xcode 9 ios b/789.
+#if (__clang_major__ >= 7) && !defined(__APPLE__)
+#define CLANG_HAS_AVX512 1
+#endif  // clang >= 7
+#endif  // __clang__
+
 // Visual C 2012 required for AVX2.
-#if defined(_M_IX86) && !defined(__clang__) && \
-    defined(_MSC_VER) && _MSC_VER >= 1700
+#if defined(_M_IX86) && !defined(__clang__) && defined(_MSC_VER) && \
+    _MSC_VER >= 1700
 #define VISUALC_HAS_AVX2 1
 #endif  // VisualStudio >= 2012
 
@@ -86,12 +75,11 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && \
     (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
 // Conversions:
-#define HAS_ABGRTOUVROW_SSSE3
 #define HAS_ABGRTOYROW_SSSE3
 #define HAS_ARGB1555TOARGBROW_SSE2
 #define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBEXTRACTALPHAROW_SSE2
 #define HAS_ARGBSETROW_X86
-#define HAS_ARGBSHUFFLEROW_SSE2
 #define HAS_ARGBSHUFFLEROW_SSSE3
 #define HAS_ARGBTOARGB1555ROW_SSE2
 #define HAS_ARGBTOARGB4444ROW_SSE2
@@ -99,18 +87,13 @@ extern "C" {
 #define HAS_ARGBTORGB24ROW_SSSE3
 #define HAS_ARGBTORGB565DITHERROW_SSE2
 #define HAS_ARGBTORGB565ROW_SSE2
-#define HAS_ARGBTOUV444ROW_SSSE3
-#define HAS_ARGBTOUVJROW_SSSE3
-#define HAS_ARGBTOUVROW_SSSE3
 #define HAS_ARGBTOYJROW_SSSE3
 #define HAS_ARGBTOYROW_SSSE3
-#define HAS_ARGBEXTRACTALPHAROW_SSE2
-#define HAS_BGRATOUVROW_SSSE3
 #define HAS_BGRATOYROW_SSSE3
 #define HAS_COPYROW_ERMS
 #define HAS_COPYROW_SSE2
 #define HAS_H422TOARGBROW_SSSE3
-#define HAS_I400TOARGBROW_SSE2
+#define HAS_HALFFLOATROW_SSE2
 #define HAS_I422TOARGB1555ROW_SSSE3
 #define HAS_I422TOARGB4444ROW_SSSE3
 #define HAS_I422TOARGBROW_SSSE3
@@ -120,21 +103,26 @@ extern "C" {
 #define HAS_I422TOUYVYROW_SSE2
 #define HAS_I422TOYUY2ROW_SSE2
 #define HAS_I444TOARGBROW_SSSE3
+#define HAS_I444TORGB24ROW_SSSE3
+#define HAS_INTERPOLATEROW_SSSE3
 #define HAS_J400TOARGBROW_SSE2
 #define HAS_J422TOARGBROW_SSSE3
 #define HAS_MERGEUVROW_SSE2
 #define HAS_MIRRORROW_SSSE3
-#define HAS_MIRRORUVROW_SSSE3
+#define HAS_MIRRORSPLITUVROW_SSSE3
 #define HAS_NV12TOARGBROW_SSSE3
+#define HAS_NV12TORGB24ROW_SSSE3
 #define HAS_NV12TORGB565ROW_SSSE3
 #define HAS_NV21TOARGBROW_SSSE3
+#define HAS_NV21TORGB24ROW_SSSE3
 #define HAS_RAWTOARGBROW_SSSE3
 #define HAS_RAWTORGB24ROW_SSSE3
+#define HAS_RAWTOYJROW_SSSE3
 #define HAS_RAWTOYROW_SSSE3
 #define HAS_RGB24TOARGBROW_SSSE3
+#define HAS_RGB24TOYJROW_SSSE3
 #define HAS_RGB24TOYROW_SSSE3
 #define HAS_RGB565TOARGBROW_SSE2
-#define HAS_RGBATOUVROW_SSSE3
 #define HAS_RGBATOYROW_SSSE3
 #define HAS_SETROW_ERMS
 #define HAS_SETROW_X86
@@ -147,11 +135,18 @@ extern "C" {
 #define HAS_YUY2TOUV422ROW_SSE2
 #define HAS_YUY2TOUVROW_SSE2
 #define HAS_YUY2TOYROW_SSE2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ARGBTOUV444ROW_SSSE3
+#define HAS_ARGBTOUVJROW_SSSE3
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_RGBATOUVROW_SSSE3
+#endif
 
 // Effects:
 #define HAS_ARGBADDROW_SSE2
 #define HAS_ARGBAFFINEROW_SSE2
-#define HAS_ARGBATTENUATEROW_SSSE3
 #define HAS_ARGBBLENDROW_SSSE3
 #define HAS_ARGBCOLORMATRIXROW_SSSE3
 #define HAS_ARGBCOLORTABLEROW_X86
@@ -170,51 +165,44 @@ extern "C" {
 #define HAS_BLENDPLANEROW_SSSE3
 #define HAS_COMPUTECUMULATIVESUMROW_SSE2
 #define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
-#define HAS_INTERPOLATEROW_SSSE3
 #define HAS_RGBCOLORTABLEROW_X86
 #define HAS_SOBELROW_SSE2
 #define HAS_SOBELTOPLANEROW_SSE2
 #define HAS_SOBELXROW_SSE2
 #define HAS_SOBELXYROW_SSE2
 #define HAS_SOBELYROW_SSE2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ARGBATTENUATEROW_SSSE3
+#endif
 
 // The following functions fail on gcc/clang 32 bit with fpic and framepointer.
 // caveat: clangcl uses row_win.cc which works.
-#if !((defined(_DEBUG) || defined(MOZ_PROFILING)) && defined(__i386__)) || \
-    !defined(__i386__) || defined(_MSC_VER)
-// TODO(fbarchard): fix build error on x86 debug
-// https://code.google.com/p/libyuv/issues/detail?id=524
-#define HAS_I411TOARGBROW_SSSE3
+#if defined(__x86_64__) || !defined(__pic__) || defined(__clang__) || \
+    defined(_MSC_VER)
 // TODO(fbarchard): fix build error on android_full_debug=1
 // https://code.google.com/p/libyuv/issues/detail?id=517
 #define HAS_I422ALPHATOARGBROW_SSSE3
+#define HAS_I444ALPHATOARGBROW_SSSE3
 #endif
 #endif
 
 // The following are available on all x86 platforms, but
 // require VS2012, clang 3.4 or gcc 4.7.
-// The code supports NaCL but requires a new compiler and validator.
-#if !defined(LIBYUV_DISABLE_AVX2) && !defined(LIBYUV_DISABLE_X86) &&    \
-    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#if !defined(LIBYUV_DISABLE_X86) &&                          \
+    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2) || \
+     defined(GCC_HAS_AVX2))
 #define HAS_ARGBCOPYALPHAROW_AVX2
 #define HAS_ARGBCOPYYTOALPHAROW_AVX2
+#define HAS_ARGBEXTRACTALPHAROW_AVX2
 #define HAS_ARGBMIRRORROW_AVX2
 #define HAS_ARGBPOLYNOMIALROW_AVX2
 #define HAS_ARGBSHUFFLEROW_AVX2
 #define HAS_ARGBTORGB565DITHERROW_AVX2
-#define HAS_ARGBTOUVJROW_AVX2
-#define HAS_ARGBTOUVROW_AVX2
 #define HAS_ARGBTOYJROW_AVX2
 #define HAS_ARGBTOYROW_AVX2
 #define HAS_COPYROW_AVX
 #define HAS_H422TOARGBROW_AVX2
-#define HAS_I400TOARGBROW_AVX2
-#if !((defined(_DEBUG) || defined(MOZ_PROFILING)) && defined(__i386__))
-// TODO(fbarchard): fix build error on android_full_debug=1
-// https://code.google.com/p/libyuv/issues/detail?id=517
-#define HAS_I422ALPHATOARGBROW_AVX2
-#endif
-#define HAS_I411TOARGBROW_AVX2
+#define HAS_HALFFLOATROW_AVX2
 #define HAS_I422TOARGB1555ROW_AVX2
 #define HAS_I422TOARGB4444ROW_AVX2
 #define HAS_I422TOARGBROW_AVX2
@@ -222,13 +210,18 @@ extern "C" {
 #define HAS_I422TORGB565ROW_AVX2
 #define HAS_I422TORGBAROW_AVX2
 #define HAS_I444TOARGBROW_AVX2
+#define HAS_I444TORGB24ROW_AVX2
 #define HAS_INTERPOLATEROW_AVX2
 #define HAS_J422TOARGBROW_AVX2
 #define HAS_MERGEUVROW_AVX2
 #define HAS_MIRRORROW_AVX2
 #define HAS_NV12TOARGBROW_AVX2
+#define HAS_NV12TORGB24ROW_AVX2
 #define HAS_NV12TORGB565ROW_AVX2
 #define HAS_NV21TOARGBROW_AVX2
+#define HAS_NV21TORGB24ROW_AVX2
+#define HAS_RAWTOYJROW_AVX2
+#define HAS_RGB24TOYJROW_AVX2
 #define HAS_SPLITUVROW_AVX2
 #define HAS_UYVYTOARGBROW_AVX2
 #define HAS_UYVYTOUV422ROW_AVX2
@@ -238,20 +231,35 @@ extern "C" {
 #define HAS_YUY2TOUV422ROW_AVX2
 #define HAS_YUY2TOUVROW_AVX2
 #define HAS_YUY2TOYROW_AVX2
+//  #define HAS_HALFFLOATROW_F16C  // Enable to test half float cast
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ARGBTOUVJROW_AVX2
+#define HAS_ARGBTOUVROW_AVX2
+#endif
 
 // Effects:
 #define HAS_ARGBADDROW_AVX2
-#define HAS_ARGBATTENUATEROW_AVX2
 #define HAS_ARGBMULTIPLYROW_AVX2
 #define HAS_ARGBSUBTRACTROW_AVX2
 #define HAS_ARGBUNATTENUATEROW_AVX2
 #define HAS_BLENDPLANEROW_AVX2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ARGBATTENUATEROW_AVX2
 #endif
 
-// The following are available for AVX2 Visual C and clangcl 32 bit:
+#if defined(__x86_64__) || !defined(__pic__) || defined(__clang__) || \
+    defined(_MSC_VER)
+// TODO(fbarchard): fix build error on android_full_debug=1
+// https://code.google.com/p/libyuv/issues/detail?id=517
+#define HAS_I422ALPHATOARGBROW_AVX2
+#define HAS_I444ALPHATOARGBROW_AVX2
+#endif
+#endif
+
+// The following are available for AVX2 Visual C 32 bit:
 // TODO(fbarchard): Port to gcc.
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
-    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(VISUALC_HAS_AVX2)
 #define HAS_ARGB1555TOARGBROW_AVX2
 #define HAS_ARGB4444TOARGBROW_AVX2
 #define HAS_ARGBTOARGB1555ROW_AVX2
@@ -264,40 +272,196 @@ extern "C" {
 // The following are also available on x64 Visual C.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && defined(_M_X64) && \
     (!defined(__clang__) || defined(__SSSE3__))
+#define HAS_I444ALPHATOARGBROW_SSSE3
+#define HAS_I444TOARGBROW_SSSE3
 #define HAS_I422ALPHATOARGBROW_SSSE3
 #define HAS_I422TOARGBROW_SSSE3
 #endif
 
+// The following are available for gcc/clang x86 platforms:
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+#define HAS_AB64TOARGBROW_SSSE3
+#define HAS_ABGRTOAR30ROW_SSSE3
+#define HAS_ABGRTOYJROW_SSSE3
+#define HAS_AR64TOARGBROW_SSSE3
+#define HAS_ARGBTOAB64ROW_SSSE3
+#define HAS_ARGBTOAR30ROW_SSSE3
+#define HAS_ARGBTOAR64ROW_SSSE3
+#define HAS_CONVERT16TO8ROW_SSSE3
+#define HAS_CONVERT8TO16ROW_SSE2
+#define HAS_DETILEROW_SSE2
+#define HAS_DETILEROW_16_SSE2
+#define HAS_DETILEROW_16_AVX
+#define HAS_DETILESPLITUVROW_SSSE3
+#define HAS_DETILETOYUY2_SSE2
+#define HAS_HALFMERGEUVROW_SSSE3
+#define HAS_I210TOAR30ROW_SSSE3
+#define HAS_I210TOARGBROW_SSSE3
+#define HAS_I212TOAR30ROW_SSSE3
+#define HAS_I212TOARGBROW_SSSE3
+#define HAS_I400TOARGBROW_SSE2
+#define HAS_I410TOAR30ROW_SSSE3
+#define HAS_I410TOARGBROW_SSSE3
+#define HAS_I422TOAR30ROW_SSSE3
+#define HAS_MERGEARGBROW_SSE2
+#define HAS_MERGERGBROW_SSSE3
+#define HAS_MERGEXRGBROW_SSE2
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV21TOYUV24ROW_SSSE3
+#define HAS_P210TOAR30ROW_SSSE3
+#define HAS_P210TOARGBROW_SSSE3
+#define HAS_P410TOAR30ROW_SSSE3
+#define HAS_P410TOARGBROW_SSSE3
+#define HAS_RAWTORGBAROW_SSSE3
+#define HAS_RGB24MIRRORROW_SSSE3
+#define HAS_RGBATOYJROW_SSSE3
+#define HAS_SPLITARGBROW_SSE2
+#define HAS_SPLITARGBROW_SSSE3
+#define HAS_SPLITRGBROW_SSSE3
+#define HAS_SPLITXRGBROW_SSE2
+#define HAS_SPLITXRGBROW_SSSE3
+#define HAS_SWAPUVROW_SSSE3
+#define HAS_YUY2TONVUVROW_SSE2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ABGRTOUVJROW_SSSE3
+#endif
+
+#if defined(__x86_64__) || !defined(__pic__)
+// TODO(fbarchard): fix build error on android_full_debug=1
+// https://code.google.com/p/libyuv/issues/detail?id=517
+#define HAS_I210ALPHATOARGBROW_SSSE3
+#define HAS_I410ALPHATOARGBROW_SSSE3
+#endif
+#endif
+
+// The following are available for AVX2 gcc/clang x86 platforms:
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) &&               \
+    (defined(__x86_64__) || defined(__i386__)) && \
+    (defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#define HAS_AB64TOARGBROW_AVX2
+#define HAS_ABGRTOAR30ROW_AVX2
+#define HAS_ABGRTOYJROW_AVX2
+#define HAS_ABGRTOYROW_AVX2
+#define HAS_AR64TOARGBROW_AVX2
+#define HAS_ARGBTOAB64ROW_AVX2
+#define HAS_ARGBTOAR30ROW_AVX2
+#define HAS_ARGBTOAR64ROW_AVX2
+#define HAS_ARGBTORAWROW_AVX2
+#define HAS_ARGBTORGB24ROW_AVX2
+#define HAS_CONVERT16TO8ROW_AVX2
+#define HAS_CONVERT8TO16ROW_AVX2
+#define HAS_DIVIDEROW_16_AVX2
+#define HAS_HALFMERGEUVROW_AVX2
+#define HAS_I210TOAR30ROW_AVX2
+#define HAS_I210TOARGBROW_AVX2
+#define HAS_I212TOAR30ROW_AVX2
+#define HAS_I212TOARGBROW_AVX2
+#define HAS_I400TOARGBROW_AVX2
+#define HAS_I410TOAR30ROW_AVX2
+#define HAS_I410TOARGBROW_AVX2
+#define HAS_I422TOAR30ROW_AVX2
+#define HAS_I422TOUYVYROW_AVX2
+#define HAS_I422TOYUY2ROW_AVX2
+#define HAS_INTERPOLATEROW_16TO8_AVX2
+#define HAS_MERGEAR64ROW_AVX2
+#define HAS_MERGEARGB16TO8ROW_AVX2
+#define HAS_MERGEARGBROW_AVX2
+#define HAS_MERGEUVROW_16_AVX2
+#define HAS_MERGEXR30ROW_AVX2
+#define HAS_MERGEXR64ROW_AVX2
+#define HAS_MERGEXRGB16TO8ROW_AVX2
+#define HAS_MERGEXRGBROW_AVX2
+#define HAS_MIRRORUVROW_AVX2
+#define HAS_MULTIPLYROW_16_AVX2
+#define HAS_NV21TOYUV24ROW_AVX2
+#define HAS_P210TOAR30ROW_AVX2
+#define HAS_P210TOARGBROW_AVX2
+#define HAS_P410TOAR30ROW_AVX2
+#define HAS_P410TOARGBROW_AVX2
+#define HAS_RGBATOYJROW_AVX2
+#define HAS_SPLITARGBROW_AVX2
+#define HAS_SPLITUVROW_16_AVX2
+#define HAS_SPLITXRGBROW_AVX2
+#define HAS_SWAPUVROW_AVX2
+#define HAS_YUY2TONVUVROW_AVX2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ABGRTOUVJROW_AVX2
+#define HAS_ABGRTOUVROW_AVX2
+#endif
+
+#if defined(__x86_64__) || !defined(__pic__)
+// TODO(fbarchard): fix build error on android_full_debug=1
+// https://code.google.com/p/libyuv/issues/detail?id=517
+#define HAS_I210ALPHATOARGBROW_AVX2
+#define HAS_I410ALPHATOARGBROW_AVX2
+#endif
+#endif
+
+// The following are available for AVX512 clang x86 platforms:
+// TODO(fbarchard): Port to GCC and Visual C
+// TODO(fbarchard): re-enable HAS_ARGBTORGB24ROW_AVX512VBMI. Issue libyuv:789
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || defined(__i386__)) && defined(CLANG_HAS_AVX512)
+#define HAS_ARGBTORGB24ROW_AVX512VBMI
+#define HAS_MERGEUVROW_AVX512BW
+#endif
+
+// The following are available for AVX512 clang x64 platforms:
+// TODO(fbarchard): Port to x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(__x86_64__) && \
+    (defined(CLANG_HAS_AVX512))
+#define HAS_I422TOARGBROW_AVX512BW
+#endif
+
 // The following are available on Neon platforms:
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_AB64TOARGBROW_NEON
+#define HAS_ABGRTOUVJROW_NEON
 #define HAS_ABGRTOUVROW_NEON
+#define HAS_ABGRTOYJROW_NEON
 #define HAS_ABGRTOYROW_NEON
+#define HAS_AR64TOARGBROW_NEON
 #define HAS_ARGB1555TOARGBROW_NEON
 #define HAS_ARGB1555TOUVROW_NEON
 #define HAS_ARGB1555TOYROW_NEON
 #define HAS_ARGB4444TOARGBROW_NEON
 #define HAS_ARGB4444TOUVROW_NEON
 #define HAS_ARGB4444TOYROW_NEON
+#define HAS_ARGBEXTRACTALPHAROW_NEON
 #define HAS_ARGBSETROW_NEON
+#define HAS_ARGBTOAB64ROW_NEON
+#define HAS_ARGBTOAR64ROW_NEON
 #define HAS_ARGBTOARGB1555ROW_NEON
 #define HAS_ARGBTOARGB4444ROW_NEON
 #define HAS_ARGBTORAWROW_NEON
 #define HAS_ARGBTORGB24ROW_NEON
 #define HAS_ARGBTORGB565DITHERROW_NEON
 #define HAS_ARGBTORGB565ROW_NEON
-#define HAS_ARGBTOUV411ROW_NEON
 #define HAS_ARGBTOUV444ROW_NEON
 #define HAS_ARGBTOUVJROW_NEON
 #define HAS_ARGBTOUVROW_NEON
 #define HAS_ARGBTOYJROW_NEON
 #define HAS_ARGBTOYROW_NEON
-#define HAS_ARGBEXTRACTALPHAROW_NEON
+#define HAS_AYUVTOUVROW_NEON
+#define HAS_AYUVTOVUROW_NEON
+#define HAS_AYUVTOYROW_NEON
 #define HAS_BGRATOUVROW_NEON
 #define HAS_BGRATOYROW_NEON
+#define HAS_BYTETOFLOATROW_NEON
+#define HAS_CONVERT16TO8ROW_NEON
 #define HAS_COPYROW_NEON
+#define HAS_DETILEROW_16_NEON
+#define HAS_DETILEROW_NEON
+#define HAS_DETILESPLITUVROW_NEON
+#define HAS_DETILETOYUY2_NEON
+#define HAS_UNPACKMT2T_NEON
+#define HAS_DIVIDEROW_16_NEON
+#define HAS_HALFFLOATROW_NEON
+#define HAS_HALFMERGEUVROW_NEON
 #define HAS_I400TOARGBROW_NEON
-#define HAS_I411TOARGBROW_NEON
 #define HAS_I422ALPHATOARGBROW_NEON
 #define HAS_I422TOARGB1555ROW_NEON
 #define HAS_I422TOARGB4444ROW_NEON
@@ -307,33 +471,62 @@ extern "C" {
 #define HAS_I422TORGBAROW_NEON
 #define HAS_I422TOUYVYROW_NEON
 #define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444ALPHATOARGBROW_NEON
 #define HAS_I444TOARGBROW_NEON
+#define HAS_I444TORGB24ROW_NEON
+#define HAS_INTERPOLATEROW_16_NEON
+#define HAS_INTERPOLATEROW_NEON
 #define HAS_J400TOARGBROW_NEON
+#define HAS_MERGEAR64ROW_NEON
+#define HAS_MERGEARGB16TO8ROW_NEON
+#define HAS_MERGEARGBROW_NEON
+#define HAS_MERGEUVROW_16_NEON
 #define HAS_MERGEUVROW_NEON
+#define HAS_MERGEXR30ROW_NEON
+#define HAS_MERGEXR64ROW_NEON
+#define HAS_MERGEXRGB16TO8ROW_NEON
+#define HAS_MERGEXRGBROW_NEON
 #define HAS_MIRRORROW_NEON
+#define HAS_MIRRORSPLITUVROW_NEON
 #define HAS_MIRRORUVROW_NEON
+#define HAS_MULTIPLYROW_16_NEON
 #define HAS_NV12TOARGBROW_NEON
+#define HAS_NV12TORGB24ROW_NEON
 #define HAS_NV12TORGB565ROW_NEON
 #define HAS_NV21TOARGBROW_NEON
+#define HAS_NV21TORGB24ROW_NEON
+#define HAS_NV21TOYUV24ROW_NEON
 #define HAS_RAWTOARGBROW_NEON
 #define HAS_RAWTORGB24ROW_NEON
+#define HAS_RAWTORGBAROW_NEON
+#define HAS_RAWTOUVJROW_NEON
 #define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYJROW_NEON
 #define HAS_RAWTOYROW_NEON
 #define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVJROW_NEON
 #define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYJROW_NEON
 #define HAS_RGB24TOYROW_NEON
 #define HAS_RGB565TOARGBROW_NEON
 #define HAS_RGB565TOUVROW_NEON
 #define HAS_RGB565TOYROW_NEON
 #define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYJROW_NEON
 #define HAS_RGBATOYROW_NEON
 #define HAS_SETROW_NEON
+#define HAS_SPLITARGBROW_NEON
+#define HAS_SPLITRGBROW_NEON
+#define HAS_SPLITUVROW_16_NEON
 #define HAS_SPLITUVROW_NEON
+#define HAS_SPLITXRGBROW_NEON
+#define HAS_SWAPUVROW_NEON
 #define HAS_UYVYTOARGBROW_NEON
 #define HAS_UYVYTOUV422ROW_NEON
 #define HAS_UYVYTOUVROW_NEON
 #define HAS_UYVYTOYROW_NEON
 #define HAS_YUY2TOARGBROW_NEON
+#define HAS_YUY2TONVUVROW_NEON
 #define HAS_YUY2TOUV422ROW_NEON
 #define HAS_YUY2TOUVROW_NEON
 #define HAS_YUY2TOYROW_NEON
@@ -351,7 +544,7 @@ extern "C" {
 #define HAS_ARGBSHADEROW_NEON
 #define HAS_ARGBSHUFFLEROW_NEON
 #define HAS_ARGBSUBTRACTROW_NEON
-#define HAS_INTERPOLATEROW_NEON
+#define HAS_RGB24MIRRORROW_NEON
 #define HAS_SOBELROW_NEON
 #define HAS_SOBELTOPLANEROW_NEON
 #define HAS_SOBELXROW_NEON
@@ -359,116 +552,312 @@ extern "C" {
 #define HAS_SOBELYROW_NEON
 #endif
 
-// The following are available on Mips platforms:
-#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
-#define HAS_COPYROW_MIPS
-#if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-#define HAS_I422TOARGBROW_DSPR2
-#define HAS_INTERPOLATEROW_DSPR2
-#define HAS_MIRRORROW_DSPR2
-#define HAS_MIRRORUVROW_DSPR2
-#define HAS_SPLITUVROW_DSPR2
+// The following are available on AArch64 platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+#define HAS_GAUSSCOL_F32_NEON
+#define HAS_GAUSSROW_F32_NEON
+#define HAS_INTERPOLATEROW_16TO8_NEON
+#define HAS_SCALESUMSAMPLES_NEON
 #endif
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#define HAS_ABGRTOUVJROW_MSA
+#define HAS_ABGRTOUVROW_MSA
+#define HAS_ABGRTOYROW_MSA
+#define HAS_ARGB1555TOARGBROW_MSA
+#define HAS_ARGB1555TOUVROW_MSA
+#define HAS_ARGB1555TOYROW_MSA
+#define HAS_ARGB4444TOARGBROW_MSA
+#define HAS_ARGBADDROW_MSA
+#define HAS_ARGBATTENUATEROW_MSA
+#define HAS_ARGBBLENDROW_MSA
+#define HAS_ARGBCOLORMATRIXROW_MSA
+#define HAS_ARGBEXTRACTALPHAROW_MSA
+#define HAS_ARGBGRAYROW_MSA
+#define HAS_ARGBMIRRORROW_MSA
+#define HAS_ARGBMULTIPLYROW_MSA
+#define HAS_ARGBQUANTIZEROW_MSA
+#define HAS_ARGBSEPIAROW_MSA
+#define HAS_ARGBSETROW_MSA
+#define HAS_ARGBSHADEROW_MSA
+#define HAS_ARGBSHUFFLEROW_MSA
+#define HAS_ARGBSUBTRACTROW_MSA
+#define HAS_ARGBTOARGB1555ROW_MSA
+#define HAS_ARGBTOARGB4444ROW_MSA
+#define HAS_ARGBTORAWROW_MSA
+#define HAS_ARGBTORGB24ROW_MSA
+#define HAS_ARGBTORGB565DITHERROW_MSA
+#define HAS_ARGBTORGB565ROW_MSA
+#define HAS_ARGBTOUV444ROW_MSA
+#define HAS_ARGBTOUVJROW_MSA
+#define HAS_ARGBTOUVROW_MSA
+#define HAS_ARGBTOYJROW_MSA
+#define HAS_ARGBTOYROW_MSA
+#define HAS_BGRATOUVROW_MSA
+#define HAS_BGRATOYROW_MSA
+#define HAS_HALFFLOATROW_MSA
+#define HAS_I400TOARGBROW_MSA
+#define HAS_I422ALPHATOARGBROW_MSA
+#define HAS_I422TOARGB1555ROW_MSA
+#define HAS_I422TOARGB4444ROW_MSA
+#define HAS_I422TOARGBROW_MSA
+#define HAS_I422TORGB24ROW_MSA
+#define HAS_I422TORGB565ROW_MSA
+#define HAS_I422TORGBAROW_MSA
+#define HAS_I422TOUYVYROW_MSA
+#define HAS_I422TOYUY2ROW_MSA
+#define HAS_I444TOARGBROW_MSA
+#define HAS_INTERPOLATEROW_MSA
+#define HAS_J400TOARGBROW_MSA
+#define HAS_MERGEUVROW_MSA
+#define HAS_MIRRORROW_MSA
+#define HAS_MIRRORSPLITUVROW_MSA
+#define HAS_MIRRORUVROW_MSA
+#define HAS_NV12TOARGBROW_MSA
+#define HAS_NV12TORGB565ROW_MSA
+#define HAS_NV21TOARGBROW_MSA
+#define HAS_RAWTOARGBROW_MSA
+#define HAS_RAWTORGB24ROW_MSA
+#define HAS_RAWTOUVROW_MSA
+#define HAS_RAWTOYROW_MSA
+#define HAS_RGB24TOARGBROW_MSA
+#define HAS_RGB24TOUVROW_MSA
+#define HAS_RGB24TOYROW_MSA
+#define HAS_RGB565TOARGBROW_MSA
+#define HAS_RGB565TOUVROW_MSA
+#define HAS_RGB565TOYROW_MSA
+#define HAS_RGBATOUVROW_MSA
+#define HAS_RGBATOYROW_MSA
+#define HAS_SETROW_MSA
+#define HAS_SOBELROW_MSA
+#define HAS_SOBELTOPLANEROW_MSA
+#define HAS_SOBELXROW_MSA
+#define HAS_SOBELXYROW_MSA
+#define HAS_SOBELYROW_MSA
+#define HAS_SPLITUVROW_MSA
+#define HAS_UYVYTOARGBROW_MSA
+#define HAS_UYVYTOUVROW_MSA
+#define HAS_UYVYTOYROW_MSA
+#define HAS_YUY2TOARGBROW_MSA
+#define HAS_YUY2TOUV422ROW_MSA
+#define HAS_YUY2TOUVROW_MSA
+#define HAS_YUY2TOYROW_MSA
+#endif
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#define HAS_ABGRTOUVROW_LSX
+#define HAS_ABGRTOYROW_LSX
+#define HAS_ARGB1555TOARGBROW_LSX
+#define HAS_ARGB1555TOUVROW_LSX
+#define HAS_ARGB1555TOYROW_LSX
+#define HAS_ARGB4444TOARGBROW_LSX
+#define HAS_ARGBBLENDROW_LSX
+#define HAS_ARGBCOLORMATRIXROW_LSX
+#define HAS_ARGBEXTRACTALPHAROW_LSX
+#define HAS_ARGBQUANTIZEROW_LSX
+#define HAS_ARGBSETROW_LSX
+#define HAS_ARGBTOUVJROW_LSX
+#define HAS_ARGBTOYJROW_LSX
+#define HAS_BGRATOUVROW_LSX
+#define HAS_BGRATOYROW_LSX
+#define HAS_I400TOARGBROW_LSX
+#define HAS_I444TOARGBROW_LSX
+#define HAS_INTERPOLATEROW_LSX
+#define HAS_J400TOARGBROW_LSX
+#define HAS_MERGEUVROW_LSX
+#define HAS_MIRRORSPLITUVROW_LSX
+#define HAS_NV12TOARGBROW_LSX
+#define HAS_NV12TORGB565ROW_LSX
+#define HAS_NV21TOARGBROW_LSX
+#define HAS_RAWTOARGBROW_LSX
+#define HAS_RAWTORGB24ROW_LSX
+#define HAS_RAWTOUVROW_LSX
+#define HAS_RAWTOYROW_LSX
+#define HAS_RGB24TOARGBROW_LSX
+#define HAS_RGB24TOUVROW_LSX
+#define HAS_RGB24TOYROW_LSX
+#define HAS_RGB565TOARGBROW_LSX
+#define HAS_RGB565TOUVROW_LSX
+#define HAS_RGB565TOYROW_LSX
+#define HAS_RGBATOUVROW_LSX
+#define HAS_RGBATOYROW_LSX
+#define HAS_SETROW_LSX
+#define HAS_SOBELROW_LSX
+#define HAS_SOBELTOPLANEROW_LSX
+#define HAS_SOBELXYROW_LSX
+#define HAS_SPLITUVROW_LSX
+#define HAS_UYVYTOARGBROW_LSX
+#define HAS_YUY2TOARGBROW_LSX
+#define HAS_ARGBTOYROW_LSX
+#define HAS_ABGRTOYJROW_LSX
+#define HAS_RGBATOYJROW_LSX
+#define HAS_RGB24TOYJROW_LSX
+#define HAS_RAWTOYJROW_LSX
+#endif
+
+#if !defined(LIBYUV_DISABLE_LASX) && defined(__loongarch_asx)
+#define HAS_ARGB1555TOARGBROW_LASX
+#define HAS_ARGB1555TOUVROW_LASX
+#define HAS_ARGB1555TOYROW_LASX
+#define HAS_ARGB4444TOARGBROW_LASX
+#define HAS_ARGBADDROW_LASX
+#define HAS_ARGBATTENUATEROW_LASX
+#define HAS_ARGBGRAYROW_LASX
+#define HAS_ARGBMIRRORROW_LASX
+#define HAS_ARGBMULTIPLYROW_LASX
+#define HAS_ARGBSEPIAROW_LASX
+#define HAS_ARGBSHADEROW_LASX
+#define HAS_ARGBSHUFFLEROW_LASX
+#define HAS_ARGBSUBTRACTROW_LASX
+#define HAS_ARGBTOARGB1555ROW_LASX
+#define HAS_ARGBTOARGB4444ROW_LASX
+#define HAS_ARGBTORAWROW_LASX
+#define HAS_ARGBTORGB24ROW_LASX
+#define HAS_ARGBTORGB565DITHERROW_LASX
+#define HAS_ARGBTORGB565ROW_LASX
+#define HAS_ARGBTOUV444ROW_LASX
+#define HAS_ARGBTOUVJROW_LASX
+#define HAS_ARGBTOUVROW_LASX
+#define HAS_ARGBTOYJROW_LASX
+#define HAS_ARGBTOYROW_LASX
+#define HAS_ABGRTOYJROW_LASX
+#define HAS_ABGRTOYROW_LASX
+#define HAS_I422ALPHATOARGBROW_LASX
+#define HAS_I422TOARGB1555ROW_LASX
+#define HAS_I422TOARGB4444ROW_LASX
+#define HAS_I422TOARGBROW_LASX
+#define HAS_I422TORGB24ROW_LASX
+#define HAS_I422TORGB565ROW_LASX
+#define HAS_I422TORGBAROW_LASX
+#define HAS_I422TOUYVYROW_LASX
+#define HAS_I422TOYUY2ROW_LASX
+#define HAS_MIRRORROW_LASX
+#define HAS_MIRRORUVROW_LASX
+#define HAS_NV12TOARGBROW_LASX
+#define HAS_NV12TORGB565ROW_LASX
+#define HAS_NV21TOARGBROW_LASX
+#define HAS_RAWTOARGBROW_LASX
+#define HAS_RAWTOUVROW_LASX
+#define HAS_RAWTOYROW_LASX
+#define HAS_RGB24TOARGBROW_LASX
+#define HAS_RGB24TOUVROW_LASX
+#define HAS_RGB24TOYROW_LASX
+#define HAS_RGB565TOARGBROW_LASX
+#define HAS_RGB565TOUVROW_LASX
+#define HAS_RGB565TOYROW_LASX
+#define HAS_UYVYTOUV422ROW_LASX
+#define HAS_UYVYTOUVROW_LASX
+#define HAS_UYVYTOYROW_LASX
+#define HAS_YUY2TOUV422ROW_LASX
+#define HAS_YUY2TOUVROW_LASX
+#define HAS_YUY2TOYROW_LASX
+#define HAS_RGBATOYROW_LASX
+#define HAS_RGBATOYJROW_LASX
+#define HAS_BGRATOYROW_LASX
+#define HAS_RGB24TOYJROW_LASX
+#define HAS_RAWTOYJROW_LASX
 #endif
 
 #if defined(_MSC_VER) && !defined(__CLR_VER) && !defined(__clang__)
+#if defined(VISUALC_HAS_AVX2)
+#define SIMD_ALIGNED(var) __declspec(align(32)) var
+#else
 #define SIMD_ALIGNED(var) __declspec(align(16)) var
-#define SIMD_ALIGNED32(var) __declspec(align(64)) var
-typedef __declspec(align(16)) int16 vec16[8];
-typedef __declspec(align(16)) int32 vec32[4];
-typedef __declspec(align(16)) int8 vec8[16];
-typedef __declspec(align(16)) uint16 uvec16[8];
-typedef __declspec(align(16)) uint32 uvec32[4];
-typedef __declspec(align(16)) uint8 uvec8[16];
-typedef __declspec(align(32)) int16 lvec16[16];
-typedef __declspec(align(32)) int32 lvec32[8];
-typedef __declspec(align(32)) int8 lvec8[32];
-typedef __declspec(align(32)) uint16 ulvec16[16];
-typedef __declspec(align(32)) uint32 ulvec32[8];
-typedef __declspec(align(32)) uint8 ulvec8[32];
+#endif
+#define LIBYUV_NOINLINE __declspec(noinline)
+typedef __declspec(align(16)) int16_t vec16[8];
+typedef __declspec(align(16)) int32_t vec32[4];
+typedef __declspec(align(16)) float vecf32[4];
+typedef __declspec(align(16)) int8_t vec8[16];
+typedef __declspec(align(16)) uint16_t uvec16[8];
+typedef __declspec(align(16)) uint32_t uvec32[4];
+typedef __declspec(align(16)) uint8_t uvec8[16];
+typedef __declspec(align(32)) int16_t lvec16[16];
+typedef __declspec(align(32)) int32_t lvec32[8];
+typedef __declspec(align(32)) int8_t lvec8[32];
+typedef __declspec(align(32)) uint16_t ulvec16[16];
+typedef __declspec(align(32)) uint32_t ulvec32[8];
+typedef __declspec(align(32)) uint8_t ulvec8[32];
 #elif !defined(__pnacl__) && (defined(__GNUC__) || defined(__clang__))
 // Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
+#if defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2)
+#define SIMD_ALIGNED(var) var __attribute__((aligned(32)))
+#else
 #define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
-#define SIMD_ALIGNED32(var) var __attribute__((aligned(64)))
-typedef int16 __attribute__((vector_size(16))) vec16;
-typedef int32 __attribute__((vector_size(16))) vec32;
-typedef int8 __attribute__((vector_size(16))) vec8;
-typedef uint16 __attribute__((vector_size(16))) uvec16;
-typedef uint32 __attribute__((vector_size(16))) uvec32;
-typedef uint8 __attribute__((vector_size(16))) uvec8;
-typedef int16 __attribute__((vector_size(32))) lvec16;
-typedef int32 __attribute__((vector_size(32))) lvec32;
-typedef int8 __attribute__((vector_size(32))) lvec8;
-typedef uint16 __attribute__((vector_size(32))) ulvec16;
-typedef uint32 __attribute__((vector_size(32))) ulvec32;
-typedef uint8 __attribute__((vector_size(32))) ulvec8;
+#endif
+#define LIBYUV_NOINLINE __attribute__((noinline))
+typedef int16_t __attribute__((vector_size(16))) vec16;
+typedef int32_t __attribute__((vector_size(16))) vec32;
+typedef float __attribute__((vector_size(16))) vecf32;
+typedef int8_t __attribute__((vector_size(16))) vec8;
+typedef uint16_t __attribute__((vector_size(16))) uvec16;
+typedef uint32_t __attribute__((vector_size(16))) uvec32;
+typedef uint8_t __attribute__((vector_size(16))) uvec8;
+typedef int16_t __attribute__((vector_size(32))) lvec16;
+typedef int32_t __attribute__((vector_size(32))) lvec32;
+typedef int8_t __attribute__((vector_size(32))) lvec8;
+typedef uint16_t __attribute__((vector_size(32))) ulvec16;
+typedef uint32_t __attribute__((vector_size(32))) ulvec32;
+typedef uint8_t __attribute__((vector_size(32))) ulvec8;
 #else
 #define SIMD_ALIGNED(var) var
-#define SIMD_ALIGNED32(var) var
-typedef int16 vec16[8];
-typedef int32 vec32[4];
-typedef int8 vec8[16];
-typedef uint16 uvec16[8];
-typedef uint32 uvec32[4];
-typedef uint8 uvec8[16];
-typedef int16 lvec16[16];
-typedef int32 lvec32[8];
-typedef int8 lvec8[32];
-typedef uint16 ulvec16[16];
-typedef uint32 ulvec32[8];
-typedef uint8 ulvec8[32];
+#define LIBYUV_NOINLINE
+typedef int16_t vec16[8];
+typedef int32_t vec32[4];
+typedef float vecf32[4];
+typedef int8_t vec8[16];
+typedef uint16_t uvec16[8];
+typedef uint32_t uvec32[4];
+typedef uint8_t uvec8[16];
+typedef int16_t lvec16[16];
+typedef int32_t lvec32[8];
+typedef int8_t lvec8[32];
+typedef uint16_t ulvec16[16];
+typedef uint32_t ulvec32[8];
+typedef uint8_t ulvec8[32];
 #endif
 
-#if defined(__aarch64__)
-// This struct is for Arm64 color conversion.
+#if defined(__aarch64__) || defined(__arm__)
+// This struct is for ARM color conversion.
 struct YuvConstants {
-  uvec16 kUVToRB;
-  uvec16 kUVToRB2;
-  uvec16 kUVToG;
-  uvec16 kUVToG2;
-  vec16 kUVBiasBGR;
-  vec32 kYToRgb;
-};
-#elif defined(__arm__)
-// This struct is for ArmV7 color conversion.
-struct YuvConstants {
-  uvec8 kUVToRB;
-  uvec8 kUVToG;
-  vec16 kUVBiasBGR;
-  vec32 kYToRgb;
+  uvec8 kUVCoeff;
+  vec16 kRGBCoeffBias;
 };
 #else
 // This struct is for Intel color conversion.
 struct YuvConstants {
-  lvec8 kUVToB;
-  lvec8 kUVToG;
-  lvec8 kUVToR;
-  lvec16 kUVBiasB;
-  lvec16 kUVBiasG;
-  lvec16 kUVBiasR;
-  lvec16 kYToRgb;
+  uint8_t kUVToB[32];
+  uint8_t kUVToG[32];
+  uint8_t kUVToR[32];
+  int16_t kYToRgb[16];
+  int16_t kYBiasToRgb[16];
 };
 
 // Offsets into YuvConstants structure
-#define KUVTOB   0
-#define KUVTOG   32
-#define KUVTOR   64
-#define KUVBIASB 96
-#define KUVBIASG 128
-#define KUVBIASR 160
-#define KYTORGB  192
+#define KUVTOB 0
+#define KUVTOG 32
+#define KUVTOR 64
+#define KYTORGB 96
+#define KYBIASTORGB 128
+
 #endif
 
-// Conversion matrix for YUV to RGB
-extern const struct YuvConstants kYuvI601Constants;  // BT.601
-extern const struct YuvConstants kYuvJPEGConstants;  // JPeg color space
-extern const struct YuvConstants kYuvH709Constants;  // BT.709
+#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a)-1)))
+
+#define align_buffer_64(var, size)                                         \
+  void* var##_mem = malloc((size) + 63);                      /* NOLINT */ \
+  uint8_t* var = (uint8_t*)(((intptr_t)var##_mem + 63) & ~63) /* NOLINT */
+
+#define free_aligned_buffer_64(var) \
+  free(var##_mem);                  \
+  var = NULL
+
+#define align_buffer_64_16(var, size)                                        \
+  void* var##_mem = malloc((size)*2 + 63);                      /* NOLINT */ \
+  uint16_t* var = (uint16_t*)(((intptr_t)var##_mem + 63) & ~63) /* NOLINT */
 
-// Conversion matrix for YVU to BGR
-extern const struct YuvConstants kYvuI601Constants;  // BT.601
-extern const struct YuvConstants kYvuJPEGConstants;  // JPeg color space
-extern const struct YuvConstants kYvuH709Constants;  // BT.709
+#define free_aligned_buffer_64_16(var) \
+  free(var##_mem);                     \
+  var = NULL
 
 #if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
 #define OMITFP
@@ -482,1458 +871,4929 @@ extern const struct YuvConstants kYvuH709Constants;  // BT.709
 #else
 #define LABELALIGN
 #endif
-#if defined(__native_client__) && defined(__x86_64__)
-// r14 is used for MEMOP macros.
-#define NACL_R14 "r14",
-#define BUNDLELOCK ".bundle_lock\n"
-#define BUNDLEUNLOCK ".bundle_unlock\n"
-#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
-#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
-#define MEMLEA(offset, base) #offset "(%q" #base ")"
-#define MEMLEA3(offset, index, scale) \
-    #offset "(,%q" #index "," #scale ")"
-#define MEMLEA4(offset, base, index, scale) \
-    #offset "(%q" #base ",%q" #index "," #scale ")"
-#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
-#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
-#define MEMOPREG(opcode, offset, base, index, scale, reg) \
-    BUNDLELOCK \
-    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
-    #opcode " (%%r15,%%r14),%%" #reg "\n" \
-    BUNDLEUNLOCK
-#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
-    BUNDLELOCK \
-    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
-    #opcode " %%" #reg ",(%%r15,%%r14)\n" \
-    BUNDLEUNLOCK
-#define MEMOPARG(opcode, offset, base, index, scale, arg) \
-    BUNDLELOCK \
-    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
-    #opcode " (%%r15,%%r14),%" #arg "\n" \
-    BUNDLEUNLOCK
-#define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \
-    BUNDLELOCK \
-    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
-    #opcode " (%%r15,%%r14),%%" #reg1 ",%%" #reg2 "\n" \
-    BUNDLEUNLOCK
-#define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \
-    BUNDLELOCK \
-    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
-    #op " $" #sel ",%%" #reg ",(%%r15,%%r14)\n" \
-    BUNDLEUNLOCK
-#else  // defined(__native_client__) && defined(__x86_64__)
-#define NACL_R14
-#define BUNDLEALIGN
-#define MEMACCESS(base) "(%" #base ")"
-#define MEMACCESS2(offset, base) #offset "(%" #base ")"
-#define MEMLEA(offset, base) #offset "(%" #base ")"
-#define MEMLEA3(offset, index, scale) \
-    #offset "(,%" #index "," #scale ")"
-#define MEMLEA4(offset, base, index, scale) \
-    #offset "(%" #base ",%" #index "," #scale ")"
-#define MEMMOVESTRING(s, d)
-#define MEMSTORESTRING(reg, d)
-#define MEMOPREG(opcode, offset, base, index, scale, reg) \
-    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
-#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
-    #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
-#define MEMOPARG(opcode, offset, base, index, scale, arg) \
-    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
-#define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \
-    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg1 ",%%" \
-    #reg2 "\n"
-#define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \
-    #op " $" #sel ",%%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
-#endif  // defined(__native_client__) && defined(__x86_64__)
-
-#if defined(__arm__) || defined(__aarch64__)
-#undef MEMACCESS
-#if defined(__native_client__)
-#define MEMACCESS(base) ".p2align 3\nbic %" #base ", #0xc0000000\n"
-#else
-#define MEMACCESS(base)
-#endif
+
+// Intel Code Analizer markers.  Insert IACA_START IACA_END around code to be
+// measured and then run with iaca -64 libyuv_unittest.
+// IACA_ASM_START amd IACA_ASM_END are equivalents that can be used within
+// inline assembly blocks.
+// example of iaca:
+// ~/iaca-lin64/bin/iaca.sh -64 -analysis LATENCY out/Release/libyuv_unittest
+
+#if defined(__x86_64__) || defined(__i386__)
+
+#define IACA_ASM_START  \
+  ".byte 0x0F, 0x0B\n"  \
+  " movl $111, %%ebx\n" \
+  ".byte 0x64, 0x67, 0x90\n"
+
+#define IACA_ASM_END         \
+  " movl $222, %%ebx\n"      \
+  ".byte 0x64, 0x67, 0x90\n" \
+  ".byte 0x0F, 0x0B\n"
+
+#define IACA_SSC_MARK(MARK_ID)                        \
+  __asm__ __volatile__("\n\t  movl $" #MARK_ID        \
+                       ", %%ebx"                      \
+                       "\n\t  .byte 0x64, 0x67, 0x90" \
+                       :                              \
+                       :                              \
+                       : "memory");
+
+#define IACA_UD_BYTES __asm__ __volatile__("\n\t .byte 0x0F, 0x0B");
+
+#else /* Visual C */
+#define IACA_UD_BYTES \
+  { __asm _emit 0x0F __asm _emit 0x0B }
+
+#define IACA_SSC_MARK(x) \
+  { __asm mov ebx, x __asm _emit 0x64 __asm _emit 0x67 __asm _emit 0x90 }
+
+#define IACA_VC64_START __writegsbyte(111, 111);
+#define IACA_VC64_END __writegsbyte(222, 222);
 #endif
 
-void I444ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+#define IACA_START     \
+  {                    \
+    IACA_UD_BYTES      \
+    IACA_SSC_MARK(111) \
+  }
+#define IACA_END       \
+  {                    \
+    IACA_SSC_MARK(222) \
+    IACA_UD_BYTES      \
+  }
+
+void I444ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I422ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I444ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I422ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I422AlphaToARGBRow_NEON(const uint8* y_buf,
-                             const uint8* u_buf,
-                             const uint8* v_buf,
-                             const uint8* a_buf,
-                             uint8* dst_argb,
+void I444AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422ToRGBARow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_rgba,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I411ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I422ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I422ToRGB565Row_NEON(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I422ToARGB1555Row_NEON(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422ToARGB4444Row_NEON(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void NV12ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I422ToRGBARow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_rgba,
+void NV12ToRGB565Row_NEON(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void NV21ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I422ToRGB24Row_NEON(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_rgb24,
+void NV12ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I422ToRGB565Row_NEON(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          uint8* dst_rgb565,
+void NV21ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV21ToYUV24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width);
+void YUY2ToARGBRow_NEON(const uint8_t* src_yuy2,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void UYVYToARGBRow_NEON(const uint8_t* src_uyvy,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I444ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void I444ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+
+void I422ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void I422ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I422ToRGBARow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void I422ToRGBARow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I422AlphaToARGBRow_MSA(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            const uint8_t* src_a,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422AlphaToARGBRow_LASX(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422ToRGB24Row_MSA(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I422ToRGB24Row_LASX(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I422ToRGB565Row_MSA(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I422ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width);
-void I422ToARGB1555Row_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb1555,
+void I422ToARGB4444Row_MSA(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb4444,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I422ToARGB4444Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToARGB4444Row_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb4444,
+void I422ToARGB1555Row_MSA(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb1555,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I422ToARGB1555Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void NV12ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_uv,
-                        uint8* dst_argb,
+void NV12ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void NV12ToRGB565Row_MSA(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV21ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void YUY2ToARGBRow_MSA(const uint8_t* src_yuy2,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void UYVYToARGBRow_MSA(const uint8_t* src_uyvy,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+
+void NV12ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void NV12ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void NV12ToRGB565Row_NEON(const uint8* src_y,
-                          const uint8* src_uv,
-                          uint8* dst_rgb565,
+void NV12ToRGB565Row_LSX(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV12ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width);
-void NV21ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_vu,
-                        uint8* dst_argb,
+void NV21ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void NV21ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
+void YUY2ToARGBRow_LSX(const uint8_t* src_yuy2,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void UYVYToARGBRow_LSX(const uint8_t* src_uyvy,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+
+void ARGBToYRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ARGBToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void ABGRToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ARGBToYJRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ARGBToYJRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ARGBToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void ABGRToYJRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void ABGRToYJRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void ABGRToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_AVX2(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void RGBAToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width);
+void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void RGBAToYRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void RGB24ToYRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
+void RAWToYRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_yj, int width);
+void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
+void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_yj, int width);
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width);
+void ABGRToYJRow_NEON(const uint8_t* src_abgr, uint8_t* dst_yj, int width);
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width);
+void ARGBToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYJRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYRow_LSX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYRow_LASX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYJRow_LSX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ABGRToYJRow_LSX(const uint8_t* src_abgr, uint8_t* dst_yj, int width);
+void RGBAToYJRow_LSX(const uint8_t* src_rgba, uint8_t* dst_yj, int width);
+void ARGBToYJRow_LASX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ABGRToYJRow_LASX(const uint8_t* src_abgr, uint8_t* dst_yj, int width);
+void RGBAToYJRow_LASX(const uint8_t* src_rgba, uint8_t* dst_yj, int width);
+void ARGBToUV444Row_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGBToUVRow_NEON(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ARGBToUV444Row_MSA(const uint8_t* src_argb,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
                         int width);
-void UYVYToARGBRow_NEON(const uint8* src_uyvy,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
+void ARGBToUVRow_MSA(const uint8_t* src_argb,
+                     int src_stride_argb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void ARGBToUVRow_LASX(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ARGBToUV444Row_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGBToUVJRow_NEON(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ABGRToUVJRow_NEON(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_uj,
+                       uint8_t* dst_vj,
+                       int width);
+void BGRAToUVRow_NEON(const uint8_t* src_bgra,
+                      int src_stride_bgra,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ABGRToUVRow_NEON(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void RGBAToUVRow_NEON(const uint8_t* src_rgba,
+                      int src_stride_rgba,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void RGB24ToUVRow_NEON(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void RAWToUVRow_NEON(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void RGB24ToUVJRow_NEON(const uint8_t* src_rgb24,
+                        int src_stride_rgb24,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RAWToUVJRow_NEON(const uint8_t* src_raw,
+                      int src_stride_raw,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
                         int width);
+void ARGB1555ToUVRow_NEON(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
+                          int src_stride_argb4444,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUVJRow_MSA(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ABGRToUVJRow_MSA(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void BGRAToUVRow_MSA(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void ABGRToUVRow_MSA(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void RGBAToUVRow_MSA(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void RGB24ToUVRow_MSA(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void RAWToUVRow_MSA(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void RGB565ToUVRow_MSA(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ARGB1555ToUVRow_MSA(const uint8_t* src_argb1555,
+                         int src_stride_argb1555,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void BGRAToUVRow_LSX(const uint8_t* src_bgra,
+                     int src_stride_bgra,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void ABGRToUVRow_LSX(const uint8_t* src_abgr,
+                     int src_stride_abgr,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void RGBAToUVRow_LSX(const uint8_t* src_rgba,
+                     int src_stride_rgba,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void ARGBToUVJRow_LSX(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ARGBToUVJRow_LASX(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ARGB1555ToUVRow_LSX(const uint8_t* src_argb1555,
+                         int src_stride_argb1555,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGB1555ToUVRow_LASX(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGB565ToUVRow_LSX(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void RGB565ToUVRow_LASX(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RGB24ToUVRow_LSX(const uint8_t* src_rgb24,
+                      int src_stride_rgb24,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void RAWToUVRow_LSX(const uint8_t* src_raw,
+                    int src_stride_raw,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void RAWToUVRow_LASX(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width);
+void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
+void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width);
+void RGB565ToYRow_NEON(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width);
+void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
+                         uint8_t* dst_y,
+                         int width);
+void BGRAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ABGRToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGBAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGB24ToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RAWToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGB565ToYRow_MSA(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_MSA(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
+
+void BGRAToYRow_LSX(const uint8_t* src_bgra, uint8_t* dst_y, int width);
+void ABGRToYRow_LSX(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void RGBAToYRow_LSX(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_LSX(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
+void RGB24ToYJRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
+void ABGRToYRow_LASX(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width);
+void RGB565ToYRow_LSX(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void RGB565ToYRow_LASX(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void RGB24ToYRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RAWToYRow_LSX(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RGBAToYRow_LASX(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void BGRAToYRow_LASX(const uint8_t* src_bgra, uint8_t* dst_y, int width);
+void RGB24ToYJRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
+void RAWToYJRow_LSX(const uint8_t* src_raw, uint8_t* dst_yj, int width);
+void RAWToYJRow_LASX(const uint8_t* src_raw, uint8_t* dst_yj, int width);
+
+void ARGBToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void ARGBToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void ABGRToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGBAToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void BGRAToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void ABGRToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGBAToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGB24ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGB24ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RAWToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RAWToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGB565ToYRow_C(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_C(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
+void ARGB4444ToYRow_C(const uint8_t* src_argb4444, uint8_t* dst_y, int width);
+void ARGBToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void BGRAToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RAWToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void BGRAToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB565ToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGB1555ToYRow_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGB4444ToYRow_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void BGRAToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB565ToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGB1555ToYRow_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
 
-void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width);
-void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int width);
-void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int width);
-void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int width);
-void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int width);
-void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int width);
-void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-                         int width);
-void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-                         int width);
-void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
-                      uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
-                      uint8* dst_u, uint8* dst_v, int width);
-void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
-                      uint8* dst_u, uint8* dst_v, int width);
-void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
-                       uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
-                     uint8* dst_u, uint8* dst_v, int width);
-void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
-                        uint8* dst_u, uint8* dst_v, int width);
-void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
-                          uint8* dst_u, uint8* dst_v, int width);
-void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
-                          uint8* dst_u, uint8* dst_v, int width);
-void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int width);
-void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int width);
-void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int width);
-void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int width);
-void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int width);
-void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int width);
-void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int width);
-void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int width);
-void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int width);
-void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int width);
-void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int width);
-void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int width);
-void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int width);
-void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int width);
-void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width);
-void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width);
-void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width);
-void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int width);
-void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int width);
-void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int width);
-void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int width);
-void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int width);
-void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int width);
-void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int width);
-void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int width);
-void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int width);
-void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int width);
-void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int width);
-void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int width);
-void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int width);
-void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int width);
-void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y,
-                             int width);
-void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y,
-                             int width);
-
-void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_AVX2(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
-                        uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
-                       uint8* dst_u, uint8* dst_v, int width);
-void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
-                          uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
-                           uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
-                           uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
-                            uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
-                           uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
-                           uint8* dst_u, uint8* dst_v, int width);
-void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
-                           uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-                             int width);
-void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-                             int width);
-void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
-                          uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
-                           uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
-                          uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
-                          uint8* dst_u, uint8* dst_v, int width);
-void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
-                          uint8* dst_u, uint8* dst_v, int width);
-void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
-                           uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
-                         uint8* dst_u, uint8* dst_v, int width);
-void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
-                            uint8* dst_u, uint8* dst_v, int width);
-void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
-                              int src_stride_argb1555,
-                              uint8* dst_u, uint8* dst_v, int width);
-void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
-                              int src_stride_argb4444,
-                              uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
-                    uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
-                   uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
-                   uint8* dst_u, uint8* dst_v, int width);
-void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
-                   uint8* dst_u, uint8* dst_v, int width);
-void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
-                    uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
-                  uint8* dst_u, uint8* dst_v, int width);
-void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
-                     uint8* dst_u, uint8* dst_v, int width);
-void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
-                       uint8* dst_u, uint8* dst_v, int width);
-
-void ARGBToUV444Row_SSSE3(const uint8* src_argb,
-                          uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
-                              uint8* dst_u, uint8* dst_v, int width);
-
-void ARGBToUV444Row_C(const uint8* src_argb,
-                      uint8* dst_u, uint8* dst_v, int width);
-void ARGBToUV411Row_C(const uint8* src_argb,
-                      uint8* dst_u, uint8* dst_v, int width);
-
-void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
-void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
-void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
-void MirrorRow_DSPR2(const uint8* src, uint8* dst, int width);
-void MirrorRow_C(const uint8* src, uint8* dst, int width);
-void MirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width);
-void MirrorRow_Any_SSSE3(const uint8* src, uint8* dst, int width);
-void MirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width);
-void MirrorRow_Any_NEON(const uint8* src, uint8* dst, int width);
-
-void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                       int width);
-void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void BGRAToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB565ToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYJRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGB1555ToYRow_Any_LSX(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+
+void RGB565ToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYJRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void BGRAToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGB1555ToYRow_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+
+void ARGBToUVRow_AVX2(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
                       int width);
-void MirrorUVRow_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void ABGRToUVRow_AVX2(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ABGRToUVJRow_AVX2(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
+                        int src_stride_argb,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void ABGRToUVJRow_SSSE3(const uint8_t* src_abgr,
+                        int src_stride_abgr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void BGRAToUVRow_SSSE3(const uint8_t* src_bgra,
+                       int src_stride_bgra,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void RGBAToUVRow_SSSE3(const uint8_t* src_rgba,
+                       int src_stride_rgba,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void ARGBToUVRow_Any_AVX2(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ABGRToUVRow_Any_AVX2(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUVJRow_Any_AVX2(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ABGRToUVJRow_Any_AVX2(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ARGBToUVRow_Any_SSSE3(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ARGBToUVJRow_Any_SSSE3(const uint8_t* src_ptr,
+                            int src_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void ABGRToUVJRow_Any_SSSE3(const uint8_t* src_ptr,
+                            int src_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void BGRAToUVRow_Any_SSSE3(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ABGRToUVRow_Any_SSSE3(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void RGBAToUVRow_Any_SSSE3(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ARGBToUV444Row_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void ARGBToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUV444Row_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void ARGBToUVRow_Any_MSA(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGBToUVRow_Any_LASX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUV444Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void ARGBToUVJRow_Any_NEON(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ABGRToUVJRow_Any_NEON(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void BGRAToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ABGRToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGBAToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGB24ToUVRow_Any_NEON(const uint8_t* src_ptr,
+                           int src_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void RAWToUVRow_Any_NEON(const uint8_t* src_ptr,
+                         int src_stride,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void RGB24ToUVJRow_Any_NEON(const uint8_t* src_ptr,
+                            int src_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void RAWToUVJRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGB565ToUVRow_Any_NEON(const uint8_t* src_ptr,
+                            int src_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void ARGB1555ToUVRow_Any_NEON(const uint8_t* src_ptr,
+                              int src_stride,
+                              uint8_t* dst_u,
+                              uint8_t* dst_v,
+                              int width);
+void ARGB4444ToUVRow_Any_NEON(const uint8_t* src_ptr,
+                              int src_stride,
+                              uint8_t* dst_u,
+                              uint8_t* dst_v,
+                              int width);
+void ARGBToUVJRow_Any_MSA(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void BGRAToUVRow_Any_MSA(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ABGRToUVRow_Any_MSA(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void RGBAToUVRow_Any_MSA(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void RGB24ToUVRow_Any_MSA(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RAWToUVRow_Any_MSA(const uint8_t* src_ptr,
+                        int src_stride_ptr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RGB565ToUVRow_Any_MSA(const uint8_t* src_ptr,
+                           int src_stride_ptr,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ARGB1555ToUVRow_Any_MSA(const uint8_t* src_ptr,
+                             int src_stride_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void ABGRToUVRow_Any_LSX(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void BGRAToUVRow_Any_LSX(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void RGBAToUVRow_Any_LSX(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGBToUVJRow_Any_LSX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUVJRow_Any_LASX(const uint8_t* src_ptr,
+                           int src_stride_ptr,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void ARGB1555ToUVRow_Any_LSX(const uint8_t* src_ptr,
+                             int src_stride_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void ARGB1555ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                              int src_stride_ptr,
+                              uint8_t* dst_u,
+                              uint8_t* dst_v,
+                              int width);
+void RGB565ToUVRow_Any_LSX(const uint8_t* src_ptr,
+                           int src_stride_ptr,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void RGB565ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                            int src_stride_ptr,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void RGB24ToUVRow_Any_LSX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGB24ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                           int src_stride_ptr,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void RAWToUVRow_Any_LSX(const uint8_t* src_ptr,
+                        int src_stride_ptr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RAWToUVRow_Any_LASX(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGBToUVRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void ARGBToUVJRow_C(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void ABGRToUVJRow_C(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void ARGBToUVRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void BGRAToUVRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void ABGRToUVRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void RGBAToUVRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void RGBAToUVJRow_C(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void RGB24ToUVRow_C(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void RAWToUVRow_C(const uint8_t* src_rgb,
+                  int src_stride_rgb,
+                  uint8_t* dst_u,
+                  uint8_t* dst_v,
+                  int width);
+void RGB24ToUVJRow_C(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void RAWToUVJRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void RGB565ToUVRow_C(const uint8_t* src_rgb565,
+                     int src_stride_rgb565,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void ARGB1555ToUVRow_C(const uint8_t* src_argb1555,
+                       int src_stride_argb1555,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
                        int width);
-void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width);
+void ARGB4444ToUVRow_C(const uint8_t* src_argb4444,
+                       int src_stride_argb4444,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+
+void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUV444Row_Any_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_u,
+                              uint8_t* dst_v,
+                              int width);
+
+void ARGBToUV444Row_C(const uint8_t* src_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+
+void MirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_SSSE3(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_C(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorRow_Any_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_MSA(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_LASX(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_C(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+
+void MirrorSplitUVRow_SSSE3(const uint8_t* src,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void MirrorSplitUVRow_NEON(const uint8_t* src_uv,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void MirrorSplitUVRow_MSA(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void MirrorSplitUVRow_LSX(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void MirrorSplitUVRow_C(const uint8_t* src_uv,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
 
-void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
-void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width);
-void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
-void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
-void ARGBMirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width);
-void ARGBMirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width);
-void ARGBMirrorRow_Any_NEON(const uint8* src, uint8* dst, int width);
+void MirrorRow_16_C(const uint16_t* src, uint16_t* dst, int width);
 
-void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width);
-void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBMirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_C(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBMirrorRow_Any_SSE2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBMirrorRow_Any_NEON(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBMirrorRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBMirrorRow_Any_LASX(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+
+void RGB24MirrorRow_SSSE3(const uint8_t* src_rgb24,
+                          uint8_t* dst_rgb24,
+                          int width);
+void RGB24MirrorRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_rgb24,
+                         int width);
+void RGB24MirrorRow_C(const uint8_t* src_rgb24, uint8_t* dst_rgb24, int width);
+void RGB24MirrorRow_Any_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void RGB24MirrorRow_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+
+void SplitUVRow_C(const uint8_t* src_uv,
+                  uint8_t* dst_u,
+                  uint8_t* dst_v,
+                  int width);
+void SplitUVRow_SSE2(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width);
-void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_AVX2(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width);
-void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_NEON(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width);
-void SplitUVRow_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                      int width);
-void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_MSA(const uint8_t* src_uv,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void SplitUVRow_LSX(const uint8_t* src_uv,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void SplitUVRow_Any_SSE2(const uint8_t* src_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width);
-void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_Any_AVX2(const uint8_t* src_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width);
-void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_Any_NEON(const uint8_t* src_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width);
-void SplitUVRow_Any_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_Any_MSA(const uint8_t* src_ptr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void SplitUVRow_Any_LSX(const uint8_t* src_ptr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void DetileRow_C(const uint8_t* src,
+                 ptrdiff_t src_tile_stride,
+                 uint8_t* dst,
+                 int width);
+void DetileRow_NEON(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width);
+void DetileRow_Any_NEON(const uint8_t* src,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst,
+                        int width);
+void DetileRow_SSE2(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width);
+void DetileRow_Any_SSE2(const uint8_t* src,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst,
+                        int width);
+void DetileRow_AVX(const uint8_t* src,
+                   ptrdiff_t src_tile_stride,
+                   uint8_t* dst,
+                   int width);
+void DetileRow_Any_AVX(const uint8_t* src,
+                       ptrdiff_t src_tile_stride,
+                       uint8_t* dst,
+                       int width);
+void DetileRow_16_C(const uint16_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint16_t* dst,
+                    int width);
+void DetileRow_16_NEON(const uint16_t* src,
+                       ptrdiff_t src_tile_stride,
+                       uint16_t* dst,
+                       int width);
+void DetileRow_16_Any_NEON(const uint16_t* src,
+                           ptrdiff_t src_tile_stride,
+                           uint16_t* dst,
+                           int width);
+void DetileRow_16_SSE2(const uint16_t* src,
+                       ptrdiff_t src_tile_stride,
+                       uint16_t* dst,
+                       int width);
+void DetileRow_16_Any_SSE2(const uint16_t* src,
+                           ptrdiff_t src_tile_stride,
+                           uint16_t* dst,
+                           int width);
+void DetileRow_16_AVX(const uint16_t* src,
+                      ptrdiff_t src_tile_stride,
+                      uint16_t* dst,
+                      int width);
+void DetileRow_16_Any_AVX(const uint16_t* src,
+                          ptrdiff_t src_tile_stride,
+                          uint16_t* dst,
                           int width);
-
-void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void DetileSplitUVRow_C(const uint8_t* src_uv,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void DetileSplitUVRow_SSSE3(const uint8_t* src_uv,
+                            ptrdiff_t src_tile_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void DetileSplitUVRow_Any_SSSE3(const uint8_t* src_uv,
+                                ptrdiff_t src_tile_stride,
+                                uint8_t* dst_u,
+                                uint8_t* dst_v,
+                                int width);
+void DetileSplitUVRow_NEON(const uint8_t* src_uv,
+                           ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void DetileSplitUVRow_Any_NEON(const uint8_t* src_uv,
+                               ptrdiff_t src_tile_stride,
+                               uint8_t* dst_u,
+                               uint8_t* dst_v,
+                               int width);
+void DetileToYUY2_C(const uint8_t* src_y,
+                    ptrdiff_t src_y_tile_stride,
+                    const uint8_t* src_uv,
+                    ptrdiff_t src_uv_tile_stride,
+                    uint8_t* dst_yuy2,
+                    int width);
+void DetileToYUY2_SSE2(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width);
+void DetileToYUY2_Any_SSE2(const uint8_t* src_y,
+                           ptrdiff_t src_y_tile_stride,
+                           const uint8_t* src_uv,
+                           ptrdiff_t src_uv_tile_stride,
+                           uint8_t* dst_yuy2,
+                           int width);
+void DetileToYUY2_NEON(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width);
+void DetileToYUY2_Any_NEON(const uint8_t* src_y,
+                           ptrdiff_t src_y_tile_stride,
+                           const uint8_t* src_uv,
+                           ptrdiff_t src_uv_tile_stride,
+                           uint8_t* dst_yuy2,
+                           int width);
+void UnpackMT2T_C(const uint8_t* src, uint16_t* dst, size_t size);
+void UnpackMT2T_NEON(const uint8_t* src, uint16_t* dst, size_t size);
+void MergeUVRow_C(const uint8_t* src_u,
+                  const uint8_t* src_v,
+                  uint8_t* dst_uv,
                   int width);
-void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_SSE2(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
                      int width);
-void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_AVX2(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
                      int width);
-void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_AVX512BW(const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_uv,
+                         int width);
+void MergeUVRow_NEON(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
                      int width);
-void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_MSA(const uint8_t* src_u,
+                    const uint8_t* src_v,
+                    uint8_t* dst_uv,
+                    int width);
+void MergeUVRow_LSX(const uint8_t* src_u,
+                    const uint8_t* src_v,
+                    uint8_t* dst_uv,
+                    int width);
+void MergeUVRow_Any_SSE2(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
                          int width);
-void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_Any_AVX2(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
                          int width);
-void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_Any_AVX512BW(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void MergeUVRow_Any_NEON(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
+void MergeUVRow_Any_MSA(const uint8_t* y_buf,
+                        const uint8_t* uv_buf,
+                        uint8_t* dst_ptr,
+                        int width);
+void MergeUVRow_Any_LSX(const uint8_t* y_buf,
+                        const uint8_t* uv_buf,
+                        uint8_t* dst_ptr,
+                        int width);
+
+void HalfMergeUVRow_C(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int width);
+
+void HalfMergeUVRow_NEON(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
                          int width);
 
-void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
-void CopyRow_AVX(const uint8* src, uint8* dst, int count);
-void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
-void CopyRow_NEON(const uint8* src, uint8* dst, int count);
-void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
-void CopyRow_C(const uint8* src, uint8* dst, int count);
-void CopyRow_Any_SSE2(const uint8* src, uint8* dst, int count);
-void CopyRow_Any_AVX(const uint8* src, uint8* dst, int count);
-void CopyRow_Any_NEON(const uint8* src, uint8* dst, int count);
+void HalfMergeUVRow_SSSE3(const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          uint8_t* dst_uv,
+                          int width);
+
+void HalfMergeUVRow_AVX2(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width);
+
+void SplitRGBRow_C(const uint8_t* src_rgb,
+                   uint8_t* dst_r,
+                   uint8_t* dst_g,
+                   uint8_t* dst_b,
+                   int width);
+void SplitRGBRow_SSSE3(const uint8_t* src_rgb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitRGBRow_NEON(const uint8_t* src_rgb,
+                      uint8_t* dst_r,
+                      uint8_t* dst_g,
+                      uint8_t* dst_b,
+                      int width);
+void SplitRGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
+void SplitRGBRow_Any_NEON(const uint8_t* src_ptr,
+                          uint8_t* dst_r,
+                          uint8_t* dst_g,
+                          uint8_t* dst_b,
+                          int width);
 
-void CopyRow_16_C(const uint16* src, uint16* dst, int count);
+void MergeRGBRow_C(const uint8_t* src_r,
+                   const uint8_t* src_g,
+                   const uint8_t* src_b,
+                   uint8_t* dst_rgb,
+                   int width);
+void MergeRGBRow_SSSE3(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_rgb,
+                       int width);
+void MergeRGBRow_NEON(const uint8_t* src_r,
+                      const uint8_t* src_g,
+                      const uint8_t* src_b,
+                      uint8_t* dst_rgb,
+                      int width);
+void MergeRGBRow_Any_SSSE3(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeRGBRow_Any_NEON(const uint8_t* src_r,
+                          const uint8_t* src_g,
+                          const uint8_t* src_b,
+                          uint8_t* dst_rgb,
+                          int width);
+void MergeARGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    const uint8_t* src_a,
+                    uint8_t* dst_argb,
+                    int width);
+void MergeARGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeARGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeARGBRow_Any_SSE2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           const uint8_t* a_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeARGBRow_Any_AVX2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           const uint8_t* a_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeARGBRow_Any_NEON(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           const uint8_t* a_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void SplitARGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    uint8_t* dst_a,
+                    int width);
+void SplitARGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width);
+void SplitARGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        uint8_t* dst_a,
+                        int width);
+void SplitARGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width);
+void SplitARGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width);
+void SplitARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           uint8_t* dst_a,
+                           int width);
+void SplitARGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_r,
+                            uint8_t* dst_g,
+                            uint8_t* dst_b,
+                            uint8_t* dst_a,
+                            int width);
+void SplitARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           uint8_t* dst_a,
+                           int width);
+void SplitARGBRow_Any_NEON(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           uint8_t* dst_a,
+                           int width);
+void MergeXRGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    uint8_t* dst_argb,
+                    int width);
+void MergeXRGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeXRGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeXRGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeXRGBRow_Any_SSE2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeXRGBRow_Any_AVX2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeXRGBRow_Any_NEON(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void SplitXRGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    int width);
+void SplitXRGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitXRGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        int width);
+void SplitXRGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitXRGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitXRGBRow_Any_SSE2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
+void SplitXRGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_r,
+                            uint8_t* dst_g,
+                            uint8_t* dst_b,
+                            int width);
+void SplitXRGBRow_Any_AVX2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
+void SplitXRGBRow_Any_NEON(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
 
-void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBCopyAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+void MergeXR30Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint8_t* dst_ar30,
+                    int depth,
+                    int width);
+void MergeAR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    const uint16_t* src_a,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width);
+void MergeARGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         const uint16_t* src_a,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width);
+void MergeXR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width);
+void MergeXRGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width);
+void MergeXR30Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width);
+void MergeAR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeARGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeXRGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR30Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width);
+void MergeXR30Row_10_NEON(const uint16_t* src_r,
+                          const uint16_t* src_g,
+                          const uint16_t* src_b,
+                          uint8_t* dst_ar30,
+                          int /* depth */,
+                          int width);
+void MergeAR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeARGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeXRGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR30Row_Any_AVX2(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint8_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeAR64Row_Any_AVX2(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           const uint16_t* a_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeXR64Row_Any_AVX2(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeARGB16To8Row_Any_AVX2(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                const uint16_t* a_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+void MergeXRGB16To8Row_Any_AVX2(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+void MergeXR30Row_Any_NEON(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint8_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeXR30Row_10_Any_NEON(const uint16_t* r_buf,
+                              const uint16_t* g_buf,
+                              const uint16_t* b_buf,
+                              uint8_t* dst_ptr,
+                              int depth,
+                              int width);
+void MergeAR64Row_Any_NEON(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           const uint16_t* a_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeARGB16To8Row_Any_NEON(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                const uint16_t* a_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+void MergeXR64Row_Any_NEON(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeXRGB16To8Row_Any_NEON(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+
+void MergeUVRow_16_C(const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint16_t* dst_uv,
+                     int depth,
+                     int width);
+void MergeUVRow_16_AVX2(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width);
+void MergeUVRow_16_Any_AVX2(const uint16_t* src_u,
+                            const uint16_t* src_v,
+                            uint16_t* dst_uv,
+                            int depth,
+                            int width);
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width);
+void MergeUVRow_16_Any_NEON(const uint16_t* src_u,
+                            const uint16_t* src_v,
+                            uint16_t* dst_uv,
+                            int depth,
+                            int width);
+
+void SplitUVRow_16_C(const uint16_t* src_uv,
+                     uint16_t* dst_u,
+                     uint16_t* dst_v,
+                     int depth,
+                     int width);
+void SplitUVRow_16_AVX2(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width);
+void SplitUVRow_16_Any_AVX2(const uint16_t* src_uv,
+                            uint16_t* dst_u,
+                            uint16_t* dst_v,
+                            int depth,
+                            int width);
+void SplitUVRow_16_NEON(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width);
+void SplitUVRow_16_Any_NEON(const uint16_t* src_uv,
+                            uint16_t* dst_u,
+                            uint16_t* dst_v,
+                            int depth,
+                            int width);
+
+void MultiplyRow_16_C(const uint16_t* src_y,
+                      uint16_t* dst_y,
+                      int scale,
+                      int width);
+void MultiplyRow_16_AVX2(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width);
+void MultiplyRow_16_Any_AVX2(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int scale,
+                             int width);
+void MultiplyRow_16_NEON(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width);
+void MultiplyRow_16_Any_NEON(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int scale,
+                             int width);
+
+void DivideRow_16_C(const uint16_t* src_y,
+                    uint16_t* dst_y,
+                    int scale,
+                    int width);
+void DivideRow_16_AVX2(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width);
+void DivideRow_16_Any_AVX2(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           int scale,
+                           int width);
+void DivideRow_16_NEON(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width);
+void DivideRow_16_Any_NEON(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           int scale,
+                           int width);
+
+void Convert8To16Row_C(const uint8_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width);
+void Convert8To16Row_SSE2(const uint8_t* src_y,
+                          uint16_t* dst_y,
+                          int scale,
+                          int width);
+void Convert8To16Row_AVX2(const uint8_t* src_y,
+                          uint16_t* dst_y,
+                          int scale,
+                          int width);
+void Convert8To16Row_Any_SSE2(const uint8_t* src_ptr,
+                              uint16_t* dst_ptr,
+                              int scale,
+                              int width);
+void Convert8To16Row_Any_AVX2(const uint8_t* src_ptr,
+                              uint16_t* dst_ptr,
+                              int scale,
+                              int width);
+
+void Convert16To8Row_C(const uint16_t* src_y,
+                       uint8_t* dst_y,
+                       int scale,
+                       int width);
+void Convert16To8Row_SSSE3(const uint16_t* src_y,
+                           uint8_t* dst_y,
+                           int scale,
+                           int width);
+void Convert16To8Row_AVX2(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width);
+void Convert16To8Row_Any_SSSE3(const uint16_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int scale,
                                int width);
-void ARGBCopyAlphaRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+void Convert16To8Row_Any_AVX2(const uint16_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int scale,
+                              int width);
+void Convert16To8Row_NEON(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width);
+void Convert16To8Row_Any_NEON(const uint16_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int scale,
+                              int width);
+
+void CopyRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void CopyRow_AVX(const uint8_t* src, uint8_t* dst, int width);
+void CopyRow_ERMS(const uint8_t* src, uint8_t* dst, int width);
+void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width);
+void CopyRow_MIPS(const uint8_t* src, uint8_t* dst, int count);
+void CopyRow_C(const uint8_t* src, uint8_t* dst, int count);
+void CopyRow_Any_SSE2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void CopyRow_Any_AVX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void CopyRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+
+void CopyRow_16_C(const uint16_t* src, uint16_t* dst, int count);
+
+void ARGBCopyAlphaRow_C(const uint8_t* src, uint8_t* dst, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBCopyAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBCopyAlphaRow_Any_SSE2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGBCopyAlphaRow_Any_AVX2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
                                int width);
 
-void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width);
-void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width);
-void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width);
-void ARGBExtractAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_a,
+void ARGBExtractAlphaRow_C(const uint8_t* src_argb, uint8_t* dst_a, int width);
+void ARGBExtractAlphaRow_SSE2(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width);
+void ARGBExtractAlphaRow_AVX2(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width);
+void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width);
+void ARGBExtractAlphaRow_MSA(const uint8_t* src_argb,
+                             uint8_t* dst_a,
+                             int width);
+void ARGBExtractAlphaRow_LSX(const uint8_t* src_argb,
+                             uint8_t* dst_a,
+                             int width);
+void ARGBExtractAlphaRow_Any_SSE2(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
+                                  int width);
+void ARGBExtractAlphaRow_Any_AVX2(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
                                   int width);
-void ARGBExtractAlphaRow_Any_NEON(const uint8* src_argb, uint8* dst_a,
+void ARGBExtractAlphaRow_Any_NEON(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
                                   int width);
+void ARGBExtractAlphaRow_Any_MSA(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int width);
+void ARGBExtractAlphaRow_Any_LSX(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int width);
 
-void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
-void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
-void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
-void ARGBCopyYToAlphaRow_Any_SSE2(const uint8* src_y, uint8* dst_argb,
+void ARGBCopyYToAlphaRow_C(const uint8_t* src, uint8_t* dst, int width);
+void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBCopyYToAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBCopyYToAlphaRow_Any_SSE2(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
                                   int width);
-void ARGBCopyYToAlphaRow_Any_AVX2(const uint8* src_y, uint8* dst_argb,
+void ARGBCopyYToAlphaRow_Any_AVX2(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
                                   int width);
 
-void SetRow_C(uint8* dst, uint8 v8, int count);
-void SetRow_X86(uint8* dst, uint8 v8, int count);
-void SetRow_ERMS(uint8* dst, uint8 v8, int count);
-void SetRow_NEON(uint8* dst, uint8 v8, int count);
-void SetRow_Any_X86(uint8* dst, uint8 v8, int count);
-void SetRow_Any_NEON(uint8* dst, uint8 v8, int count);
+void SetRow_C(uint8_t* dst, uint8_t v8, int width);
+void SetRow_MSA(uint8_t* dst, uint8_t v8, int width);
+void SetRow_X86(uint8_t* dst, uint8_t v8, int width);
+void SetRow_ERMS(uint8_t* dst, uint8_t v8, int width);
+void SetRow_NEON(uint8_t* dst, uint8_t v8, int width);
+void SetRow_LSX(uint8_t* dst, uint8_t v8, int width);
+void SetRow_Any_X86(uint8_t* dst_ptr, uint8_t v32, int width);
+void SetRow_Any_NEON(uint8_t* dst_ptr, uint8_t v32, int width);
+void SetRow_Any_LSX(uint8_t* dst_ptr, uint8_t v32, int width);
 
-void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int count);
-void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count);
-void ARGBSetRow_NEON(uint8* dst_argb, uint32 v32, int count);
-void ARGBSetRow_Any_NEON(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_C(uint8_t* dst_argb, uint32_t v32, int width);
+void ARGBSetRow_X86(uint8_t* dst_argb, uint32_t v32, int width);
+void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width);
+void ARGBSetRow_Any_NEON(uint8_t* dst_ptr, uint32_t v32, int width);
+void ARGBSetRow_MSA(uint8_t* dst_argb, uint32_t v32, int width);
+void ARGBSetRow_Any_MSA(uint8_t* dst_ptr, uint32_t v32, int width);
+void ARGBSetRow_LSX(uint8_t* dst_argb, uint32_t v32, int width);
+void ARGBSetRow_Any_LSX(uint8_t* dst_ptr, uint32_t v32, int width);
 
 // ARGBShufflers for BGRAToARGB etc.
-void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
-                      const uint8* shuffler, int width);
-void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width);
-void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                          const uint8* shuffler, int width);
-void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width);
-void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width);
-void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
-                             const uint8* shuffler, int width);
-void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                              const uint8* shuffler, int width);
-void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
-                             const uint8* shuffler, int width);
-void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
-                             const uint8* shuffler, int width);
-
-void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int width);
-void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int width);
-void RAWToRGB24Row_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width);
-void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int width);
-void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
-                            int width);
-void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
-                            int width);
-void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb, int width);
-void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
-                            int width);
-void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
-                            int width);
-
-void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int width);
-void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int width);
-void RAWToRGB24Row_NEON(const uint8* src_raw, uint8* dst_rgb24, int width);
-void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int width);
-void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
-                            int width);
-void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
-                            int width);
-void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width);
-void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width);
-void RAWToRGB24Row_C(const uint8* src_raw, uint8* dst_rgb24, int width);
-void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int width);
-void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int width);
-void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb,
+void ARGBShuffleRow_C(const uint8_t* src_argb,
+                      uint8_t* dst_argb,
+                      const uint8_t* shuffler,
+                      int width);
+void ARGBShuffleRow_SSSE3(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          const uint8_t* shuffler,
+                          int width);
+void ARGBShuffleRow_AVX2(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width);
+void ARGBShuffleRow_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width);
+void ARGBShuffleRow_MSA(const uint8_t* src_argb,
+                        uint8_t* dst_argb,
+                        const uint8_t* shuffler,
+                        int width);
+void ARGBShuffleRow_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width);
+void ARGBShuffleRow_Any_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              const uint8_t* param,
                               int width);
-void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int width);
-void RAWToRGB24Row_Any_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width);
+void ARGBShuffleRow_Any_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             const uint8_t* param,
+                             int width);
+void ARGBShuffleRow_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             const uint8_t* param,
+                             int width);
+void ARGBShuffleRow_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            const uint8_t* param,
+                            int width);
+void ARGBShuffleRow_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             const uint8_t* param,
+                             int width);
+
+void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
+                          uint8_t* dst_argb,
+                          int width);
+void RAWToARGBRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGBARow_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
+void RAWToRGB24Row_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
+void RGB565ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGB1555ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGB4444ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void RGB565ToARGBRow_AVX2(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGB1555ToARGBRow_AVX2(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
+                            int width);
+void ARGB4444ToARGBRow_AVX2(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
+                            int width);
 
-void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width);
+void RGB24ToARGBRow_MSA(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
+void RGB24ToARGBRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
+void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width);
+void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
+void RAWToARGBRow_MSA(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToARGBRow_LSX(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGB24Row_NEON(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
+void RAWToRGB24Row_MSA(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
+void RAWToRGB24Row_LSX(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
+void RGB565ToARGBRow_NEON(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width);
+void RGB565ToARGBRow_MSA(const uint8_t* src_rgb565,
+                         uint8_t* dst_argb,
+                         int width);
+void RGB565ToARGBRow_LSX(const uint8_t* src_rgb565,
+                         uint8_t* dst_argb,
+                         int width);
+void RGB565ToARGBRow_LASX(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
+                            int width);
+void ARGB1555ToARGBRow_MSA(const uint8_t* src_argb1555,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGB1555ToARGBRow_LSX(const uint8_t* src_argb1555,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGB1555ToARGBRow_LASX(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
+                            int width);
+void ARGB4444ToARGBRow_NEON(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
+                            int width);
+void ARGB4444ToARGBRow_MSA(const uint8_t* src_argb4444,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGB4444ToARGBRow_LSX(const uint8_t* src_argb4444,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGB4444ToARGBRow_LASX(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
+                            int width);
+void RGB24ToARGBRow_C(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
+void RAWToARGBRow_C(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGBARow_C(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
+void RAWToRGB24Row_C(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
+void RGB565ToARGBRow_C(const uint8_t* src_rgb565, uint8_t* dst_argb, int width);
+void ARGB1555ToARGBRow_C(const uint8_t* src_argb1555,
+                         uint8_t* dst_argb,
+                         int width);
+void ARGB4444ToARGBRow_C(const uint8_t* src_argb4444,
+                         uint8_t* dst_argb,
+                         int width);
+void AR30ToARGBRow_C(const uint8_t* src_ar30, uint8_t* dst_argb, int width);
+void AR30ToABGRRow_C(const uint8_t* src_ar30, uint8_t* dst_abgr, int width);
+void ARGBToAR30Row_C(const uint8_t* src_argb, uint8_t* dst_ar30, int width);
+void AR30ToAB30Row_C(const uint8_t* src_ar30, uint8_t* dst_ab30, int width);
+
+void RGB24ToARGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
                               int width);
-void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+void RAWToARGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RAWToRGBARow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RAWToRGB24Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+
+void RGB565ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGB1555ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+void ARGB4444ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void RGB565ToARGBRow_Any_AVX2(const uint8* src_rgb565, uint8* dst_argb,
+void RGB565ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
                               int width);
-void ARGB1555ToARGBRow_Any_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+void ARGB1555ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGB4444ToARGBRow_Any_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+void ARGB4444ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
 
-void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb,
+void RGB24ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
                              int width);
-void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int width);
-void RAWToRGB24Row_Any_NEON(const uint8* src_raw, uint8* dst_rgb24, int width);
-void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
+void RGB24ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RGB24ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RGB24ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void RAWToARGBRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToRGBARow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToARGBRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToARGBRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToARGBRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToRGB24Row_Any_NEON(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RAWToRGB24Row_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToRGB24Row_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB565ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
                               int width);
-void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
+void RGB565ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void RGB565ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void RGB565ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGB1555ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
+void ARGB1555ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGB4444ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
+void ARGB1555ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGB1555ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
 
-void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width);
-
-void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
-                             const uint32 dither4, int width);
-void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width);
-void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width);
-
-void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width);
-
-void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width);
-
-void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width);
-
-void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
-void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
-void J400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int width);
-void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width);
-void J400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int width);
-void J400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int width);
-void J400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width);
-
-void I444ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_argb,
+void ARGB4444ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGB4444ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGB4444ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
+
+void ARGBToRGB24Row_SSSE3(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToRAWRow_SSSE3(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToRGB565Row_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToARGB1555Row_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToARGB4444Row_SSE2(const uint8_t* src, uint8_t* dst, int width);
+void ABGRToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width);
+
+void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width);
+
+void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width);
+
+void ARGBToRGB565DitherRow_C(const uint8_t* src_argb,
+                             uint8_t* dst_rgb,
+                             const uint32_t dither4,
+                             int width);
+void ARGBToRGB565DitherRow_SSE2(const uint8_t* src,
+                                uint8_t* dst,
+                                const uint32_t dither4,
+                                int width);
+void ARGBToRGB565DitherRow_AVX2(const uint8_t* src,
+                                uint8_t* dst,
+                                const uint32_t dither4,
+                                int width);
+
+void ARGBToRGB565Row_AVX2(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToARGB1555Row_AVX2(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width);
+void ARGBToARGB4444Row_AVX2(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width);
+void ABGRToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width);
+void ARGBToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width);
+
+void ARGBToRGB24Row_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_rgb24,
+                         int width);
+void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width);
+void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
+                          uint8_t* dst_rgb565,
+                          int width);
+void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
+                            uint8_t* dst_argb1555,
+                            int width);
+void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
+                            uint8_t* dst_argb4444,
+                            int width);
+void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width);
+void ARGBToRGB24Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRAWRow_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRGB565Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToARGB1555Row_MSA(const uint8_t* src_argb,
+                           uint8_t* dst_rgb,
+                           int width);
+void ARGBToARGB4444Row_MSA(const uint8_t* src_argb,
+                           uint8_t* dst_rgb,
+                           int width);
+void ARGBToRGB565DitherRow_MSA(const uint8_t* src_argb,
+                               uint8_t* dst_rgb,
+                               const uint32_t dither4,
+                               int width);
+void ARGBToRGB565DitherRow_LASX(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width);
+
+void ARGBToRGB24Row_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRAWRow_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRGB565Row_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToARGB1555Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width);
+void ARGBToARGB4444Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width);
+
+void ARGBToRGBARow_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRGB24Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRAWRow_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRGB565Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToARGB1555Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToARGB4444Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ABGRToAR30Row_C(const uint8_t* src_abgr, uint8_t* dst_ar30, int width);
+void ARGBToAR30Row_C(const uint8_t* src_argb, uint8_t* dst_ar30, int width);
+
+void ARGBToAR64Row_C(const uint8_t* src_argb, uint16_t* dst_ar64, int width);
+void ARGBToAB64Row_C(const uint8_t* src_argb, uint16_t* dst_ab64, int width);
+void AR64ToARGBRow_C(const uint16_t* src_ar64, uint8_t* dst_argb, int width);
+void AB64ToARGBRow_C(const uint16_t* src_ab64, uint8_t* dst_argb, int width);
+void AR64ShuffleRow_C(const uint8_t* src_ar64,
+                      uint8_t* dst_ar64,
+                      const uint8_t* shuffler,
+                      int width);
+void ARGBToAR64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ar64,
+                         int width);
+void ARGBToAB64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ab64,
+                         int width);
+void AR64ToARGBRow_SSSE3(const uint16_t* src_ar64,
+                         uint8_t* dst_argb,
+                         int width);
+void AB64ToARGBRow_SSSE3(const uint16_t* src_ab64,
+                         uint8_t* dst_argb,
+                         int width);
+void ARGBToAR64Row_AVX2(const uint8_t* src_argb, uint16_t* dst_ar64, int width);
+void ARGBToAB64Row_AVX2(const uint8_t* src_argb, uint16_t* dst_ab64, int width);
+void AR64ToARGBRow_AVX2(const uint16_t* src_ar64, uint8_t* dst_argb, int width);
+void AB64ToARGBRow_AVX2(const uint16_t* src_ab64, uint8_t* dst_argb, int width);
+void ARGBToAR64Row_NEON(const uint8_t* src_argb, uint16_t* dst_ar64, int width);
+void ARGBToAB64Row_NEON(const uint8_t* src_argb, uint16_t* dst_ab64, int width);
+void AR64ToARGBRow_NEON(const uint16_t* src_ar64, uint8_t* dst_argb, int width);
+void AB64ToARGBRow_NEON(const uint16_t* src_ab64, uint8_t* dst_argb, int width);
+void ARGBToAR64Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int width);
+void ARGBToAB64Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int width);
+void AR64ToARGBRow_Any_SSSE3(const uint16_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void AB64ToARGBRow_Any_SSSE3(const uint16_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToAR64Row_Any_AVX2(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void ARGBToAB64Row_Any_AVX2(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void AR64ToARGBRow_Any_AVX2(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void AB64ToARGBRow_Any_AVX2(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBToAR64Row_Any_NEON(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void ARGBToAB64Row_Any_NEON(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void AR64ToARGBRow_Any_NEON(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void AB64ToARGBRow_Any_NEON(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+
+void J400ToARGBRow_SSE2(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_MSA(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_LSX(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void J400ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void J400ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void J400ToARGBRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void J400ToARGBRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+
+void I444ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I444ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_u,
+                      const uint8_t* src_v,
+                      uint8_t* rgb_buf,
+                      const struct YuvConstants* yuvconstants,
+                      int width);
+void I422ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void I422ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_argb,
+void I422ToAR30Row_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void I422ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_argb,
+void I210ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void I422AlphaToARGBRow_C(const uint8* y_buf,
-                          const uint8* u_buf,
-                          const uint8* v_buf,
-                          const uint8* a_buf,
-                          uint8* dst_argb,
-                          const struct YuvConstants* yuvconstants,
-                          int width);
-void I411ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_argb,
+void I210ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void NV12ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_uv,
-                     uint8* dst_argb,
+void I212ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I212ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I210AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I410AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I444AlphaToARGBRow_C(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          const uint8_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I422AlphaToARGBRow_C(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          const uint8_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void NV12ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_uv,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void NV12ToRGB565Row_C(const uint8* src_y,
-                       const uint8* src_uv,
-                       uint8* dst_argb,
+void NV12ToRGB565Row_C(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_rgb565,
                        const struct YuvConstants* yuvconstants,
                        int width);
-void NV21ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_uv,
-                     uint8* dst_argb,
+void NV21ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_vu,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void YUY2ToARGBRow_C(const uint8* src_yuy2,
-                     uint8* dst_argb,
+void NV12ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_uv,
+                      uint8_t* rgb_buf,
+                      const struct YuvConstants* yuvconstants,
+                      int width);
+void NV21ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_vu,
+                      uint8_t* rgb_buf,
+                      const struct YuvConstants* yuvconstants,
+                      int width);
+void NV21ToYUV24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_vu,
+                      uint8_t* dst_yuv24,
+                      int width);
+void YUY2ToARGBRow_C(const uint8_t* src_yuy2,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void UYVYToARGBRow_C(const uint8* src_uyvy,
-                     uint8* dst_argb,
+void UYVYToARGBRow_C(const uint8_t* src_uyvy,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void I422ToRGBARow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_rgba,
+void P210ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void P410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void P210ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void P410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+
+void I422ToRGBARow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
-void I422ToRGB24Row_C(const uint8* src_y,
-                      const uint8* src_u,
-                      const uint8* src_v,
-                      uint8* dst_rgb24,
+void I422ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_u,
+                      const uint8_t* src_v,
+                      uint8_t* rgb_buf,
                       const struct YuvConstants* yuvconstants,
                       int width);
-void I422ToARGB4444Row_C(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb4444,
+void I422ToARGB4444Row_C(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb4444,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I422ToARGB1555Row_C(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb4444,
+void I422ToARGB1555Row_C(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb1555,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I422ToRGB565Row_C(const uint8* src_y,
-                       const uint8* src_u,
-                       const uint8* src_v,
-                       uint8* dst_rgb565,
+void I422ToRGB565Row_C(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_rgb565,
                        const struct YuvConstants* yuvconstants,
                        int width);
-void I422ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I422ToARGBRow_AVX2(const uint8_t* y_buf,
+                        const uint8_t* u_buf,
+                        const uint8_t* v_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I422ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width);
-void I422ToRGBARow_AVX2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I422ToARGBRow_AVX512BW(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422ToRGBARow_AVX2(const uint8_t* y_buf,
+                        const uint8_t* u_buf,
+                        const uint8_t* v_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I444ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
+void I444ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I444ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I444ToARGBRow_AVX2(const uint8_t* y_buf,
+                        const uint8_t* u_buf,
+                        const uint8_t* v_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I444ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
+void I444ToRGB24Row_SSSE3(const uint8_t* y_buf,
+                          const uint8_t* u_buf,
+                          const uint8_t* v_buf,
+                          uint8_t* dst_rgb24,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I444ToRGB24Row_AVX2(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I444ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width);
-void I422ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
+void I422ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
-                              const uint8* u_buf,
-                              const uint8* v_buf,
-                              const uint8* a_buf,
-                              uint8* dst_argb,
-                              const struct YuvConstants* yuvconstants,
-                              int width);
-void I422AlphaToARGBRow_AVX2(const uint8* y_buf,
-                             const uint8* u_buf,
-                             const uint8* v_buf,
-                             const uint8* a_buf,
-                             uint8* dst_argb,
-                             const struct YuvConstants* yuvconstants,
-                             int width);
-void I422ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
+
+void I422ToAR30Row_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I210ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I210ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I411ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
+void I212ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_ar30,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I411ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I212ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I210AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                              const uint16_t* u_buf,
+                              const uint16_t* v_buf,
+                              const uint16_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I410AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                              const uint16_t* u_buf,
+                              const uint16_t* v_buf,
+                              const uint16_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I422ToAR30Row_AVX2(const uint8_t* y_buf,
+                        const uint8_t* u_buf,
+                        const uint8_t* v_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I210ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I210ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I212ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I212ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I410ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I410ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void NV12ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_uv,
-                         uint8* dst_argb,
+void I210AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I410AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              const uint8_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I444AlphaToARGBRow_AVX2(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              const uint8_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I422AlphaToARGBRow_AVX2(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV12ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void NV12ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_uv,
-                        uint8* dst_argb,
+void NV12ToARGBRow_AVX2(const uint8_t* y_buf,
+                        const uint8_t* uv_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void NV12ToRGB565Row_SSSE3(const uint8* src_y,
-                           const uint8* src_uv,
-                           uint8* dst_argb,
+void NV12ToRGB24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb24,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void NV21ToRGB24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_vu,
+                          uint8_t* dst_rgb24,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void NV12ToRGB565Row_SSSE3(const uint8_t* src_y,
+                           const uint8_t* src_uv,
+                           uint8_t* dst_rgb565,
                            const struct YuvConstants* yuvconstants,
                            int width);
-void NV12ToRGB565Row_AVX2(const uint8* src_y,
-                          const uint8* src_uv,
-                          uint8* dst_argb,
+void NV12ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV21ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV21ToYUV24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_vu,
+                          uint8_t* dst_yuv24,
+                          int width);
+void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width);
+void NV12ToRGB565Row_AVX2(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width);
-void NV21ToARGBRow_SSSE3(const uint8* src_y,
-                         const uint8* src_uv,
-                         uint8* dst_argb,
+void NV21ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* vu_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV21ToARGBRow_AVX2(const uint8_t* y_buf,
+                        const uint8_t* vu_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void YUY2ToARGBRow_SSSE3(const uint8_t* yuy2_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void UYVYToARGBRow_SSSE3(const uint8_t* uyvy_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void NV21ToARGBRow_AVX2(const uint8* src_y,
-                        const uint8* src_uv,
-                        uint8* dst_argb,
+void YUY2ToARGBRow_AVX2(const uint8_t* yuy2_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void UYVYToARGBRow_AVX2(const uint8_t* uyvy_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
-                         uint8* dst_argb,
+
+void P210ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
-                         uint8* dst_argb,
+void P410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void YUY2ToARGBRow_AVX2(const uint8* src_yuy2,
-                        uint8* dst_argb,
+void P210ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void P410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void P210ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void UYVYToARGBRow_AVX2(const uint8* src_uyvy,
-                        uint8* dst_argb,
+void P210ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I422ToRGBARow_SSSE3(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_rgba,
+void P410ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+
+void I422ToRGBARow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_rgba,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I422ToARGB4444Row_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I422ToARGB4444Row_SSSE3(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             uint8_t* dst_argb4444,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422ToARGB4444Row_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I422ToARGB4444Row_AVX2(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToARGB1555Row_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I422ToARGB1555Row_SSSE3(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             uint8_t* dst_argb1555,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422ToARGB1555Row_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I422ToARGB1555Row_AVX2(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToRGB565Row_SSSE3(const uint8* src_y,
-                           const uint8* src_u,
-                           const uint8* src_v,
-                           uint8* dst_argb,
+void I422ToRGB565Row_SSSE3(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_rgb565,
                            const struct YuvConstants* yuvconstants,
                            int width);
-void I422ToRGB565Row_AVX2(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          uint8* dst_argb,
+void I422ToRGB565Row_AVX2(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width);
-void I422ToRGB24Row_SSSE3(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          uint8* dst_rgb24,
+void I422ToRGB24Row_SSSE3(const uint8_t* y_buf,
+                          const uint8_t* u_buf,
+                          const uint8_t* v_buf,
+                          uint8_t* dst_rgb24,
                           const struct YuvConstants* yuvconstants,
                           int width);
-void I422ToRGB24Row_AVX2(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_rgb24,
+void I422ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width);
-void I422ToARGBRow_Any_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I422ToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToRGBARow_Any_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I422ToARGBRow_Any_AVX512BW(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
+void I422ToRGBARow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I444ToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I444ToARGBRow_Any_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I444ToRGB24Row_Any_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I444ToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I444ToRGB24Row_Any_AVX2(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422ToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422AlphaToARGBRow_Any_SSSE3(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  const uint8* a_buf,
-                                  uint8* dst_argb,
+void I422ToAR30Row_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I210ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I210ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I212ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I212ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I410ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I410ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I210AlphaToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                                  const uint16_t* u_buf,
+                                  const uint16_t* v_buf,
+                                  const uint16_t* a_buf,
+                                  uint8_t* dst_ptr,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width);
+void I410AlphaToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                                  const uint16_t* u_buf,
+                                  const uint16_t* v_buf,
+                                  const uint16_t* a_buf,
+                                  uint8_t* dst_ptr,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width);
+void I422ToAR30Row_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I210ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I210ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I212ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I212ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I410ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I410ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I210AlphaToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                                 const uint16_t* u_buf,
+                                 const uint16_t* v_buf,
+                                 const uint16_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void I410AlphaToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                                 const uint16_t* u_buf,
+                                 const uint16_t* v_buf,
+                                 const uint16_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void I444AlphaToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                                  const uint8_t* u_buf,
+                                  const uint8_t* v_buf,
+                                  const uint8_t* a_buf,
+                                  uint8_t* dst_ptr,
                                   const struct YuvConstants* yuvconstants,
                                   int width);
-void I422AlphaToARGBRow_Any_AVX2(const uint8* y_buf,
-                                 const uint8* u_buf,
-                                 const uint8* v_buf,
-                                 const uint8* a_buf,
-                                 uint8* dst_argb,
+void I444AlphaToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
                                  const struct YuvConstants* yuvconstants,
                                  int width);
-void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I422AlphaToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                                  const uint8_t* u_buf,
+                                  const uint8_t* v_buf,
+                                  const uint8_t* a_buf,
+                                  uint8_t* dst_ptr,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width);
+void I422AlphaToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void NV12ToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I411ToARGBRow_Any_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void NV12ToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
-                             const uint8* src_uv,
-                             uint8* dst_argb,
+void NV21ToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void NV12ToARGBRow_Any_AVX2(const uint8* src_y,
-                            const uint8* src_uv,
-                            uint8* dst_argb,
+void NV21ToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
-                             const uint8* src_vu,
-                             uint8* dst_argb,
+void NV12ToRGB24Row_Any_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void NV21ToRGB24Row_Any_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void NV12ToRGB24Row_Any_AVX2(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void NV21ToARGBRow_Any_AVX2(const uint8* src_y,
-                            const uint8* src_vu,
-                            uint8* dst_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width);
-void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
-                               const uint8* src_uv,
-                               uint8* dst_argb,
+void NV21ToRGB24Row_Any_AVX2(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV21ToYUV24Row_Any_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void NV21ToYUV24Row_Any_AVX2(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void NV12ToRGB565Row_Any_SSSE3(const uint8_t* y_buf,
+                               const uint8_t* uv_buf,
+                               uint8_t* dst_ptr,
                                const struct YuvConstants* yuvconstants,
                                int width);
-void NV12ToRGB565Row_Any_AVX2(const uint8* src_y,
-                              const uint8* src_uv,
-                              uint8* dst_argb,
+void NV12ToRGB565Row_Any_AVX2(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
                               const struct YuvConstants* yuvconstants,
                               int width);
-void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
-                             uint8* dst_argb,
+void YUY2ToARGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void UYVYToARGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void YUY2ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void UYVYToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P210ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void P410ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
-                             uint8* dst_argb,
+void P210ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void YUY2ToARGBRow_Any_AVX2(const uint8* src_yuy2,
-                            uint8* dst_argb,
+void P410ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void P210ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void UYVYToARGBRow_Any_AVX2(const uint8* src_uyvy,
-                            uint8* dst_argb,
+void P210ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_rgba,
+void P410ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422ToRGBARow_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
-                                 const uint8* src_u,
-                                 const uint8* src_v,
-                                 uint8* dst_rgba,
+void I422ToARGB4444Row_Any_SSSE3(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 uint8_t* dst_ptr,
                                  const struct YuvConstants* yuvconstants,
                                  int width);
-void I422ToARGB4444Row_Any_AVX2(const uint8* src_y,
-                                const uint8* src_u,
-                                const uint8* src_v,
-                                uint8* dst_rgba,
+void I422ToARGB4444Row_Any_AVX2(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
                                 const struct YuvConstants* yuvconstants,
                                 int width);
-void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
-                                 const uint8* src_u,
-                                 const uint8* src_v,
-                                 uint8* dst_rgba,
+void I422ToARGB1555Row_Any_SSSE3(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 uint8_t* dst_ptr,
                                  const struct YuvConstants* yuvconstants,
                                  int width);
-void I422ToARGB1555Row_Any_AVX2(const uint8* src_y,
-                                const uint8* src_u,
-                                const uint8* src_v,
-                                uint8* dst_rgba,
+void I422ToARGB1555Row_Any_AVX2(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
                                 const struct YuvConstants* yuvconstants,
                                 int width);
-void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
-                               const uint8* src_u,
-                               const uint8* src_v,
-                               uint8* dst_rgba,
+void I422ToRGB565Row_Any_SSSE3(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_ptr,
                                const struct YuvConstants* yuvconstants,
                                int width);
-void I422ToRGB565Row_Any_AVX2(const uint8* src_y,
-                              const uint8* src_u,
-                              const uint8* src_v,
-                              uint8* dst_rgba,
+void I422ToRGB565Row_Any_AVX2(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              uint8_t* dst_ptr,
                               const struct YuvConstants* yuvconstants,
                               int width);
-void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
-                              const uint8* src_u,
-                              const uint8* src_v,
-                              uint8* dst_argb,
+void I422ToRGB24Row_Any_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              uint8_t* dst_ptr,
                               const struct YuvConstants* yuvconstants,
                               int width);
-void I422ToRGB24Row_Any_AVX2(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I422ToRGB24Row_Any_AVX2(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
 
-void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width);
-void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
-void I400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
-void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int width);
-void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int width);
-void I400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int width);
-void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_C(const uint8_t* src_y,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I400ToARGBRow_SSE2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I400ToARGBRow_AVX2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I400ToARGBRow_NEON(const uint8_t* src_y,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I400ToARGBRow_MSA(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void I400ToARGBRow_LSX(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void I400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* param,
+                            int width);
+void I400ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* param,
+                            int width);
+void I400ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* param,
+                            int width);
+void I400ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I400ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
 
 // ARGB preattenuated alpha blend.
-void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
-                        uint8* dst_argb, int width);
-void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
-                       uint8* dst_argb, int width);
-void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
-                    uint8* dst_argb, int width);
+void ARGBBlendRow_SSSE3(const uint8_t* src_argb,
+                        const uint8_t* src_argb1,
+                        uint8_t* dst_argb,
+                        int width);
+void ARGBBlendRow_NEON(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width);
+void ARGBBlendRow_MSA(const uint8_t* src_argb0,
+                      const uint8_t* src_argb1,
+                      uint8_t* dst_argb,
+                      int width);
+void ARGBBlendRow_LSX(const uint8_t* src_argb0,
+                      const uint8_t* src_argb1,
+                      uint8_t* dst_argb,
+                      int width);
+void ARGBBlendRow_C(const uint8_t* src_argb,
+                    const uint8_t* src_argb1,
+                    uint8_t* dst_argb,
+                    int width);
 
 // Unattenuated planar alpha blend.
-void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
-                         const uint8* alpha, uint8* dst, int width);
-void BlendPlaneRow_Any_SSSE3(const uint8* src0, const uint8* src1,
-                             const uint8* alpha, uint8* dst, int width);
-void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
-                        const uint8* alpha, uint8* dst, int width);
-void BlendPlaneRow_Any_AVX2(const uint8* src0, const uint8* src1,
-                            const uint8* alpha, uint8* dst, int width);
-void BlendPlaneRow_C(const uint8* src0, const uint8* src1,
-                     const uint8* alpha, uint8* dst, int width);
+void BlendPlaneRow_SSSE3(const uint8_t* src0,
+                         const uint8_t* src1,
+                         const uint8_t* alpha,
+                         uint8_t* dst,
+                         int width);
+void BlendPlaneRow_Any_SSSE3(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void BlendPlaneRow_AVX2(const uint8_t* src0,
+                        const uint8_t* src1,
+                        const uint8_t* alpha,
+                        uint8_t* dst,
+                        int width);
+void BlendPlaneRow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void BlendPlaneRow_C(const uint8_t* src0,
+                     const uint8_t* src1,
+                     const uint8_t* alpha,
+                     uint8_t* dst,
+                     int width);
 
 // ARGB multiply images. Same API as Blend, but these require
 // pointer and width alignment for SSE2.
-void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
-                       uint8* dst_argb, int width);
-void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
-                          uint8* dst_argb, int width);
-void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
-                              uint8* dst_argb, int width);
-void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
-                          uint8* dst_argb, int width);
-void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
-                              uint8* dst_argb, int width);
-void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
-                          uint8* dst_argb, int width);
-void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
-                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_C(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width);
+void ARGBMultiplyRow_SSE2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBMultiplyRow_Any_SSE2(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBMultiplyRow_AVX2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBMultiplyRow_Any_AVX2(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBMultiplyRow_NEON(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBMultiplyRow_Any_NEON(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBMultiplyRow_MSA(const uint8_t* src_argb0,
+                         const uint8_t* src_argb1,
+                         uint8_t* dst_argb,
+                         int width);
+void ARGBMultiplyRow_Any_MSA(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBMultiplyRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBMultiplyRow_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
 
 // ARGB add images.
-void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
-                  uint8* dst_argb, int width);
-void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
-                     uint8* dst_argb, int width);
-void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
-                         uint8* dst_argb, int width);
-void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
-                     uint8* dst_argb, int width);
-void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
-                         uint8* dst_argb, int width);
-void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
-                     uint8* dst_argb, int width);
-void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
-                         uint8* dst_argb, int width);
+void ARGBAddRow_C(const uint8_t* src_argb,
+                  const uint8_t* src_argb1,
+                  uint8_t* dst_argb,
+                  int width);
+void ARGBAddRow_SSE2(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width);
+void ARGBAddRow_Any_SSE2(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
+void ARGBAddRow_AVX2(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width);
+void ARGBAddRow_Any_AVX2(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
+void ARGBAddRow_NEON(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width);
+void ARGBAddRow_Any_NEON(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
+void ARGBAddRow_MSA(const uint8_t* src_argb0,
+                    const uint8_t* src_argb1,
+                    uint8_t* dst_argb,
+                    int width);
+void ARGBAddRow_Any_MSA(const uint8_t* y_buf,
+                        const uint8_t* uv_buf,
+                        uint8_t* dst_ptr,
+                        int width);
+void ARGBAddRow_LASX(const uint8_t* src_argb0,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width);
+void ARGBAddRow_Any_LASX(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
 
 // ARGB subtract images. Same API as Blend, but these require
 // pointer and width alignment for SSE2.
-void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
-                       uint8* dst_argb, int width);
-void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
-                          uint8* dst_argb, int width);
-void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
-                              uint8* dst_argb, int width);
-void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
-                          uint8* dst_argb, int width);
-void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
-                              uint8* dst_argb, int width);
-void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
-                          uint8* dst_argb, int width);
-void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
-                              uint8* dst_argb, int width);
-
-void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb,
+void ARGBSubtractRow_C(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width);
+void ARGBSubtractRow_SSE2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBSubtractRow_Any_SSE2(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBSubtractRow_AVX2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBSubtractRow_Any_AVX2(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBSubtractRow_NEON(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBSubtractRow_Any_NEON(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBSubtractRow_MSA(const uint8_t* src_argb0,
+                         const uint8_t* src_argb1,
+                         uint8_t* dst_argb,
+                         int width);
+void ARGBSubtractRow_Any_MSA(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBSubtractRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBSubtractRow_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+
+void ARGBToRGB24Row_Any_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBToRAWRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBToRGB565Row_Any_SSE2(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBToARGB1555Row_Any_SSE2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb,
+void ARGBToARGB4444Row_Any_SSE2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
+void ABGRToAR30Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToAR30Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToRAWRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToRGB24Row_Any_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToRGB24Row_Any_AVX512VBMI(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   int width);
+void ARGBToRGB565DitherRow_Any_SSE2(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    const uint32_t param,
+                                    int width);
+void ARGBToRGB565DitherRow_Any_AVX2(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    const uint32_t param,
+                                    int width);
 
-void ARGBToRGB565DitherRow_Any_SSE2(const uint8* src_argb, uint8* dst_rgb,
-                                    const uint32 dither4, int width);
-void ARGBToRGB565DitherRow_Any_AVX2(const uint8* src_argb, uint8* dst_rgb,
-                                    const uint32 dither4, int width);
+void ARGBToRGB565Row_Any_AVX2(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBToARGB1555Row_Any_AVX2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
+void ARGBToARGB4444Row_Any_AVX2(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
+void ABGRToAR30Row_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBToAR30Row_Any_AVX2(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
 
-void ARGBToRGB565Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb,
+void ARGBToRGB24Row_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToRAWRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToRGB565Row_Any_NEON(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBToARGB1555Row_Any_NEON(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGBToARGB4444Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb,
+void ARGBToARGB4444Row_Any_NEON(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
+void ARGBToRGB565DitherRow_Any_NEON(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    const uint32_t param,
+                                    int width);
+void ARGBToRGB24Row_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBToRAWRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToRGB565Row_Any_MSA(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToARGB1555Row_Any_MSA(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGBToARGB4444Row_Any_MSA(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGBToRGB565DitherRow_Any_MSA(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   const uint32_t param,
+                                   int width);
+void ARGBToRGB565DitherRow_Any_LASX(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    const uint32_t param,
+                                    int width);
 
-void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int width);
-void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb,
+void ARGBToRGB24Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToRAWRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToRGB565Row_Any_LASX(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBToARGB1555Row_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb,
+void ARGBToARGB4444Row_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGBToRGB565DitherRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb,
-                                    const uint32 dither4, int width);
 
-void I444ToARGBRow_Any_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I444ToARGBRow_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToARGBRow_Any_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I444ToRGB24Row_Any_NEON(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422ToARGBRow_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422AlphaToARGBRow_Any_NEON(const uint8* src_y,
-                                 const uint8* src_u,
-                                 const uint8* src_v,
-                                 const uint8* src_a,
-                                 uint8* dst_argb,
+void I444AlphaToARGBRow_Any_NEON(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
                                  const struct YuvConstants* yuvconstants,
                                  int width);
-void I411ToARGBRow_Any_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width);
-void I422ToRGBARow_Any_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb,
+void I422AlphaToARGBRow_Any_NEON(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void I422ToRGBARow_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToRGB24Row_Any_NEON(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb,
+void I422ToRGB24Row_Any_NEON(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
-                                const uint8* src_u,
-                                const uint8* src_v,
-                                uint8* dst_argb,
+void I422ToARGB4444Row_Any_NEON(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
                                 const struct YuvConstants* yuvconstants,
                                 int width);
-void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
-                                const uint8* src_u,
-                                const uint8* src_v,
-                                uint8* dst_argb,
+void I422ToARGB1555Row_Any_NEON(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
                                 const struct YuvConstants* yuvconstants,
                                 int width);
-void I422ToRGB565Row_Any_NEON(const uint8* src_y,
-                              const uint8* src_u,
-                              const uint8* src_v,
-                              uint8* dst_argb,
+void I422ToRGB565Row_Any_NEON(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              uint8_t* dst_ptr,
                               const struct YuvConstants* yuvconstants,
                               int width);
-void NV12ToARGBRow_Any_NEON(const uint8* src_y,
-                            const uint8* src_uv,
-                            uint8* dst_argb,
+void NV12ToARGBRow_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void NV21ToARGBRow_Any_NEON(const uint8* src_y,
-                            const uint8* src_vu,
-                            uint8* dst_argb,
+void NV21ToARGBRow_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
-                              const uint8* src_uv,
-                              uint8* dst_argb,
+void NV12ToRGB24Row_Any_NEON(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV21ToRGB24Row_Any_NEON(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV21ToYUV24Row_Any_NEON(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void NV12ToRGB565Row_Any_NEON(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
                               const struct YuvConstants* yuvconstants,
                               int width);
-void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
-                            uint8* dst_argb,
+void YUY2ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
-                            uint8* dst_argb,
+void UYVYToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
-void I422ToARGBRow_DSPR2(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
+void P210ToARGBRow_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToARGBRow_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P210ToAR30Row_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToAR30Row_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P210ToARGBRow_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToARGBRow_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P210ToAR30Row_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ar30,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToAR30Row_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ar30,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I444ToARGBRow_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I444ToARGBRow_Any_LSX(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I422ToARGBRow_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I422ToARGBRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422ToRGBARow_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I422ToRGBARow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422AlphaToARGBRow_Any_MSA(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                const uint8_t* a_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
+void I422AlphaToARGBRow_Any_LASX(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void I422ToRGB24Row_Any_MSA(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I422ToRGB24Row_Any_LASX(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422ToRGB565Row_Any_MSA(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I422ToRGB565Row_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I422ToARGB4444Row_Any_MSA(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_ptr,
+                               const struct YuvConstants* yuvconstants,
+                               int width);
+void I422ToARGB4444Row_Any_LASX(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
+void I422ToARGB1555Row_Any_MSA(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_ptr,
+                               const struct YuvConstants* yuvconstants,
+                               int width);
+void I422ToARGB1555Row_Any_LASX(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
+void NV12ToARGBRow_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* uv_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void NV12ToRGB565Row_Any_MSA(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV21ToARGBRow_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* uv_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void YUY2ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void UYVYToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+
+void NV12ToARGBRow_Any_LSX(const uint8_t* y_buf,
+                           const uint8_t* uv_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void NV12ToARGBRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void NV12ToRGB565Row_Any_LSX(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV12ToRGB565Row_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void NV21ToARGBRow_Any_LSX(const uint8_t* y_buf,
+                           const uint8_t* uv_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void NV21ToARGBRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void YUY2ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void UYVYToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+
+void YUY2ToYRow_AVX2(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToUVRow_AVX2(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void YUY2ToNVUVRow_AVX2(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width);
+void YUY2ToUV422Row_AVX2(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width);
-void I422ToARGBRow_DSPR2(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
+void YUY2ToYRow_SSE2(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToUVRow_SSE2(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void YUY2ToNVUVRow_SSE2(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width);
+void YUY2ToUV422Row_SSE2(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToUVRow_NEON(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void YUY2ToNVUVRow_NEON(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width);
+void YUY2ToUV422Row_NEON(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void YUY2ToYRow_MSA(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToYRow_LASX(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToUVRow_MSA(const uint8_t* src_yuy2,
+                     int src_stride_yuy2,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void YUY2ToUVRow_LASX(const uint8_t* src_yuy2,
+                      int src_stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void YUY2ToUV422Row_MSA(const uint8_t* src_yuy2,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void YUY2ToUV422Row_LASX(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void YUY2ToYRow_C(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToUVRow_C(const uint8_t* src_yuy2,
+                   int src_stride_yuy2,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void YUY2ToNVUVRow_C(const uint8_t* src_yuy2,
+                     int src_stride_yuy2,
+                     uint8_t* dst_uv,
+                     int width);
+void YUY2ToUV422Row_C(const uint8_t* src_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void YUY2ToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void YUY2ToUVRow_Any_AVX2(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void YUY2ToNVUVRow_Any_AVX2(const uint8_t* src_yuy2,
+                            int stride_yuy2,
+                            uint8_t* dst_uv,
+                            int width);
+void YUY2ToUV422Row_Any_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void YUY2ToYRow_Any_SSE2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void YUY2ToUVRow_Any_SSE2(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void YUY2ToNVUVRow_Any_SSE2(const uint8_t* src_yuy2,
+                            int stride_yuy2,
+                            uint8_t* dst_uv,
+                            int width);
+void YUY2ToUV422Row_Any_SSE2(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void YUY2ToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void YUY2ToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void YUY2ToNVUVRow_Any_NEON(const uint8_t* src_yuy2,
+                            int stride_yuy2,
+                            uint8_t* dst_uv,
+                            int width);
+void YUY2ToUV422Row_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void YUY2ToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void YUY2ToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void YUY2ToUVRow_Any_MSA(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void YUY2ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void YUY2ToUV422Row_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void YUY2ToUV422Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void UYVYToYRow_AVX2(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToUVRow_AVX2(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void UYVYToUV422Row_AVX2(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void UYVYToYRow_SSE2(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToUVRow_SSE2(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void UYVYToUV422Row_SSE2(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void UYVYToYRow_AVX2(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToUVRow_AVX2(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void UYVYToUV422Row_AVX2(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void UYVYToYRow_NEON(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToUVRow_NEON(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void UYVYToUV422Row_NEON(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void UYVYToYRow_MSA(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToYRow_LASX(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToUVRow_MSA(const uint8_t* src_uyvy,
+                     int src_stride_uyvy,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void UYVYToUVRow_LASX(const uint8_t* src_uyvy,
+                      int src_stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void UYVYToUV422Row_MSA(const uint8_t* src_uyvy,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void UYVYToUV422Row_LASX(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width);
 
-void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
-                   uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_C(const uint8* src_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
-                          uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
-                             uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
-                          uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
-                             uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int width);
-void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
-                          uint8* dst_u, uint8* dst_v, int width);
-void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
-                             uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_NEON(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width);
-
-void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
-                   uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_C(const uint8* src_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
-                          uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
-                             uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                          uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
-                             uint8* dst_u, uint8* dst_v, int width);
-void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int width);
-void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
-                          uint8* dst_u, uint8* dst_v, int width);
-void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
-                             uint8* dst_u, uint8* dst_v, int width);
-
-void I422ToYUY2Row_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_yuy2, int width);
-void I422ToUYVYRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_uyvy, int width);
-void I422ToYUY2Row_SSE2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_yuy2, int width);
-void I422ToUYVYRow_SSE2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_uyvy, int width);
-void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_yuy2, int width);
-void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_uyvy, int width);
-void I422ToYUY2Row_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_yuy2, int width);
-void I422ToUYVYRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_uyvy, int width);
-void I422ToYUY2Row_Any_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_yuy2, int width);
-void I422ToUYVYRow_Any_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_uyvy, int width);
+void UYVYToYRow_C(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToUVRow_C(const uint8_t* src_uyvy,
+                   int src_stride_uyvy,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
+void UYVYToUV422Row_C(const uint8_t* src_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void UYVYToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void UYVYToUVRow_Any_AVX2(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void UYVYToUV422Row_Any_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void UYVYToYRow_Any_SSE2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void UYVYToUVRow_Any_SSE2(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void UYVYToUV422Row_Any_SSE2(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void UYVYToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void UYVYToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void UYVYToUV422Row_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void UYVYToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void UYVYToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void UYVYToUVRow_Any_MSA(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void UYVYToUVRow_Any_LASX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void UYVYToUV422Row_Any_MSA(const uint8_t* src_ptr,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void UYVYToUV422Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void SwapUVRow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void SwapUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void SwapUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void AYUVToYRow_C(const uint8_t* src_ayuv, uint8_t* dst_y, int width);
+void AYUVToUVRow_C(const uint8_t* src_ayuv,
+                   int src_stride_ayuv,
+                   uint8_t* dst_uv,
+                   int width);
+void AYUVToVURow_C(const uint8_t* src_ayuv,
+                   int src_stride_ayuv,
+                   uint8_t* dst_vu,
+                   int width);
+void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width);
+void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
+                      int src_stride_ayuv,
+                      uint8_t* dst_uv,
+                      int width);
+void AYUVToVURow_NEON(const uint8_t* src_ayuv,
+                      int src_stride_ayuv,
+                      uint8_t* dst_vu,
+                      int width);
+void AYUVToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void AYUVToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_vu,
+                          int width);
+void AYUVToVURow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_vu,
+                          int width);
+
+void I422ToYUY2Row_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_frame,
+                     int width);
+void I422ToUYVYRow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_frame,
+                     int width);
+void I422ToYUY2Row_SSE2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width);
+void I422ToUYVYRow_SSE2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width);
+void I422ToYUY2Row_Any_SSE2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToUYVYRow_Any_SSE2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToYUY2Row_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width);
+void I422ToUYVYRow_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width);
+void I422ToYUY2Row_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToUYVYRow_Any_AVX2(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToYUY2Row_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width);
+void I422ToUYVYRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width);
+void I422ToYUY2Row_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToUYVYRow_Any_NEON(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToYUY2Row_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_yuy2,
+                       int width);
+void I422ToYUY2Row_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width);
+void I422ToUYVYRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_uyvy,
+                       int width);
+void I422ToUYVYRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width);
+void I422ToYUY2Row_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void I422ToYUY2Row_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
+void I422ToUYVYRow_Any_MSA(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void I422ToUYVYRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
 
 // Effects related row functions.
-void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
-                               int width);
-void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+void ARGBAttenuateRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBAttenuateRow_SSSE3(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            int width);
+void ARGBAttenuateRow_AVX2(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGBAttenuateRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGBAttenuateRow_MSA(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBAttenuateRow_LASX(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width);
+void ARGBAttenuateRow_Any_SSSE3(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
                                 int width);
-void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+void ARGBAttenuateRow_Any_AVX2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
                                int width);
-void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+void ARGBAttenuateRow_Any_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
+void ARGBAttenuateRow_Any_MSA(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBAttenuateRow_Any_LASX(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
                                int width);
 
 // Inverse table for unattenuate, shared by C and SSE2.
-extern const uint32 fixed_invtbl8[256];
-void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+extern const uint32_t fixed_invtbl8[256];
+void ARGBUnattenuateRow_C(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBUnattenuateRow_SSE2(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             int width);
+void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
                                  int width);
-void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+void ARGBUnattenuateRow_Any_AVX2(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
                                  int width);
 
-void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
-void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
-
-void ARGBSepiaRow_C(uint8* dst_argb, int width);
-void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
-void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
-
-void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
-                          const int8* matrix_argb, int width);
-void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                              const int8* matrix_argb, int width);
-void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
-                             const int8* matrix_argb, int width);
-
-void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
-void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
-
-void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
-void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
-
-void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
-                       int interval_offset, int width);
-void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
-                          int interval_offset, int width);
-void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
-                          int interval_offset, int width);
-
-void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
-                    uint32 value);
-void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value);
-void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value);
+void ARGBGrayRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBGrayRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBGrayRow_MSA(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBGrayRow_LASX(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+
+void ARGBSepiaRow_C(uint8_t* dst_argb, int width);
+void ARGBSepiaRow_SSSE3(uint8_t* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8_t* dst_argb, int width);
+void ARGBSepiaRow_MSA(uint8_t* dst_argb, int width);
+void ARGBSepiaRow_LASX(uint8_t* dst_argb, int width);
+
+void ARGBColorMatrixRow_C(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          const int8_t* matrix_argb,
+                          int width);
+void ARGBColorMatrixRow_SSSE3(const uint8_t* src_argb,
+                              uint8_t* dst_argb,
+                              const int8_t* matrix_argb,
+                              int width);
+void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             const int8_t* matrix_argb,
+                             int width);
+void ARGBColorMatrixRow_MSA(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const int8_t* matrix_argb,
+                            int width);
+void ARGBColorMatrixRow_LSX(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const int8_t* matrix_argb,
+                            int width);
+
+void ARGBColorTableRow_C(uint8_t* dst_argb,
+                         const uint8_t* table_argb,
+                         int width);
+void ARGBColorTableRow_X86(uint8_t* dst_argb,
+                           const uint8_t* table_argb,
+                           int width);
+
+void RGBColorTableRow_C(uint8_t* dst_argb,
+                        const uint8_t* table_argb,
+                        int width);
+void RGBColorTableRow_X86(uint8_t* dst_argb,
+                          const uint8_t* table_argb,
+                          int width);
+
+void ARGBQuantizeRow_C(uint8_t* dst_argb,
+                       int scale,
+                       int interval_size,
+                       int interval_offset,
+                       int width);
+void ARGBQuantizeRow_SSE2(uint8_t* dst_argb,
+                          int scale,
+                          int interval_size,
+                          int interval_offset,
+                          int width);
+void ARGBQuantizeRow_NEON(uint8_t* dst_argb,
+                          int scale,
+                          int interval_size,
+                          int interval_offset,
+                          int width);
+void ARGBQuantizeRow_MSA(uint8_t* dst_argb,
+                         int scale,
+                         int interval_size,
+                         int interval_offset,
+                         int width);
+void ARGBQuantizeRow_LSX(uint8_t* dst_argb,
+                         int scale,
+                         int interval_size,
+                         int interval_offset,
+                         int width);
+
+void ARGBShadeRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_argb,
+                    int width,
+                    uint32_t value);
+void ARGBShadeRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value);
+void ARGBShadeRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value);
+void ARGBShadeRow_MSA(const uint8_t* src_argb,
+                      uint8_t* dst_argb,
+                      int width,
+                      uint32_t value);
+void ARGBShadeRow_LASX(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value);
 
 // Used for blur.
-void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
-                                    int width, int area, uint8* dst, int count);
-void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
-                                  const int32* previous_cumsum, int width);
+void CumulativeSumToAverageRow_SSE2(const int32_t* topleft,
+                                    const int32_t* botleft,
+                                    int width,
+                                    int area,
+                                    uint8_t* dst,
+                                    int count);
+void ComputeCumulativeSumRow_SSE2(const uint8_t* row,
+                                  int32_t* cumsum,
+                                  const int32_t* previous_cumsum,
+                                  int width);
 
-void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
-                                 int width, int area, uint8* dst, int count);
-void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
-                               const int32* previous_cumsum, int width);
+void CumulativeSumToAverageRow_C(const int32_t* tl,
+                                 const int32_t* bl,
+                                 int w,
+                                 int area,
+                                 uint8_t* dst,
+                                 int count);
+void ComputeCumulativeSumRow_C(const uint8_t* row,
+                               int32_t* cumsum,
+                               const int32_t* previous_cumsum,
+                               int width);
 
 LIBYUV_API
-void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
-                     uint8* dst_argb, const float* uv_dudv, int width);
+void ARGBAffineRow_C(const uint8_t* src_argb,
+                     int src_argb_stride,
+                     uint8_t* dst_argb,
+                     const float* uv_dudv,
+                     int width);
 LIBYUV_API
-void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
-                        uint8* dst_argb, const float* uv_dudv, int width);
+void ARGBAffineRow_SSE2(const uint8_t* src_argb,
+                        int src_argb_stride,
+                        uint8_t* dst_argb,
+                        const float* src_dudv,
+                        int width);
 
 // Used for I420Scale, ARGBScale, and ARGBInterpolate.
-void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
-                      ptrdiff_t src_stride_ptr,
-                      int width, int source_y_fraction);
-void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                          ptrdiff_t src_stride_ptr, int width,
+void InterpolateRow_C(uint8_t* dst_ptr,
+                      const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      int width,
+                      int source_y_fraction);
+void InterpolateRow_SSSE3(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          int dst_width,
                           int source_y_fraction);
-void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
-                         ptrdiff_t src_stride_ptr, int width,
+void InterpolateRow_AVX2(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         int dst_width,
                          int source_y_fraction);
-void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
-                         ptrdiff_t src_stride_ptr, int width,
+void InterpolateRow_NEON(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         int dst_width,
                          int source_y_fraction);
-void InterpolateRow_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
-                          ptrdiff_t src_stride_ptr, int width,
-                          int source_y_fraction);
-void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
-                             ptrdiff_t src_stride_ptr, int width,
+void InterpolateRow_MSA(uint8_t* dst_ptr,
+                        const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int source_y_fraction);
+void InterpolateRow_LSX(uint8_t* dst_ptr,
+                        const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int source_y_fraction);
+void InterpolateRow_Any_NEON(uint8_t* dst_ptr,
+                             const uint8_t* src_ptr,
+                             ptrdiff_t src_stride_ptr,
+                             int width,
                              int source_y_fraction);
-void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                              ptrdiff_t src_stride_ptr, int width,
+void InterpolateRow_Any_SSSE3(uint8_t* dst_ptr,
+                              const uint8_t* src_ptr,
+                              ptrdiff_t src_stride_ptr,
+                              int width,
                               int source_y_fraction);
-void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
-                             ptrdiff_t src_stride_ptr, int width,
+void InterpolateRow_Any_AVX2(uint8_t* dst_ptr,
+                             const uint8_t* src_ptr,
+                             ptrdiff_t src_stride_ptr,
+                             int width,
                              int source_y_fraction);
-void InterpolateRow_Any_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
-                              ptrdiff_t src_stride_ptr, int width,
-                              int source_y_fraction);
+void InterpolateRow_Any_MSA(uint8_t* dst_ptr,
+                            const uint8_t* src_ptr,
+                            ptrdiff_t src_stride_ptr,
+                            int width,
+                            int source_y_fraction);
+void InterpolateRow_Any_LSX(uint8_t* dst_ptr,
+                            const uint8_t* src_ptr,
+                            ptrdiff_t src_stride_ptr,
+                            int width,
+                            int source_y_fraction);
+
+void InterpolateRow_16_C(uint16_t* dst_ptr,
+                         const uint16_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         int width,
+                         int source_y_fraction);
+void InterpolateRow_16_NEON(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int width,
+                            int source_y_fraction);
+void InterpolateRow_16_Any_NEON(uint16_t* dst_ptr,
+                                const uint16_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                int width,
+                                int source_y_fraction);
 
-void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                         ptrdiff_t src_stride_ptr,
-                         int width, int source_y_fraction);
+void InterpolateRow_16To8_C(uint8_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int scale,
+                            int width,
+                            int source_y_fraction);
+void InterpolateRow_16To8_NEON(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int width,
+                               int source_y_fraction);
+void InterpolateRow_16To8_Any_NEON(uint8_t* dst_ptr,
+                                   const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   int scale,
+                                   int width,
+                                   int source_y_fraction);
+void InterpolateRow_16To8_AVX2(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int width,
+                               int source_y_fraction);
+void InterpolateRow_16To8_Any_AVX2(uint8_t* dst_ptr,
+                                   const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   int scale,
+                                   int width,
+                                   int source_y_fraction);
 
 // Sobel images.
-void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
-                 uint8* dst_sobelx, int width);
-void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobelx, int width);
-void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobelx, int width);
-void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
-                 uint8* dst_sobely, int width);
-void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width);
-void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width);
-void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
-                uint8* dst_argb, int width);
-void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                   uint8* dst_argb, int width);
-void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                   uint8* dst_argb, int width);
-void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
-                       uint8* dst_y, int width);
-void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_y, int width);
-void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_y, int width);
-void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
-                  uint8* dst_argb, int width);
-void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width);
-void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width);
-void SobelRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                       uint8* dst_argb, int width);
-void SobelRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                       uint8* dst_argb, int width);
-void SobelToPlaneRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                              uint8* dst_y, int width);
-void SobelToPlaneRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                              uint8* dst_y, int width);
-void SobelXYRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                         uint8* dst_argb, int width);
-void SobelXYRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                         uint8* dst_argb, int width);
-
-void ARGBPolynomialRow_C(const uint8* src_argb,
-                         uint8* dst_argb, const float* poly,
-                         int width);
-void ARGBPolynomialRow_SSE2(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
-                            int width);
-void ARGBPolynomialRow_AVX2(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
-                            int width);
-
-void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
-                             const uint8* luma, uint32 lumacoeff);
-void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+void SobelXRow_C(const uint8_t* src_y0,
+                 const uint8_t* src_y1,
+                 const uint8_t* src_y2,
+                 uint8_t* dst_sobelx,
+                 int width);
+void SobelXRow_SSE2(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    const uint8_t* src_y2,
+                    uint8_t* dst_sobelx,
+                    int width);
+void SobelXRow_NEON(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    const uint8_t* src_y2,
+                    uint8_t* dst_sobelx,
+                    int width);
+void SobelXRow_MSA(const uint8_t* src_y0,
+                   const uint8_t* src_y1,
+                   const uint8_t* src_y2,
+                   uint8_t* dst_sobelx,
+                   int width);
+void SobelYRow_C(const uint8_t* src_y0,
+                 const uint8_t* src_y1,
+                 uint8_t* dst_sobely,
+                 int width);
+void SobelYRow_SSE2(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    uint8_t* dst_sobely,
+                    int width);
+void SobelYRow_NEON(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    uint8_t* dst_sobely,
+                    int width);
+void SobelYRow_MSA(const uint8_t* src_y0,
+                   const uint8_t* src_y1,
+                   uint8_t* dst_sobely,
+                   int width);
+void SobelRow_C(const uint8_t* src_sobelx,
+                const uint8_t* src_sobely,
+                uint8_t* dst_argb,
+                int width);
+void SobelRow_SSE2(const uint8_t* src_sobelx,
+                   const uint8_t* src_sobely,
+                   uint8_t* dst_argb,
+                   int width);
+void SobelRow_NEON(const uint8_t* src_sobelx,
+                   const uint8_t* src_sobely,
+                   uint8_t* dst_argb,
+                   int width);
+void SobelRow_MSA(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width);
+void SobelRow_LSX(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width);
+void SobelToPlaneRow_C(const uint8_t* src_sobelx,
+                       const uint8_t* src_sobely,
+                       uint8_t* dst_y,
+                       int width);
+void SobelToPlaneRow_SSE2(const uint8_t* src_sobelx,
+                          const uint8_t* src_sobely,
+                          uint8_t* dst_y,
+                          int width);
+void SobelToPlaneRow_NEON(const uint8_t* src_sobelx,
+                          const uint8_t* src_sobely,
+                          uint8_t* dst_y,
+                          int width);
+void SobelToPlaneRow_MSA(const uint8_t* src_sobelx,
+                         const uint8_t* src_sobely,
+                         uint8_t* dst_y,
+                         int width);
+void SobelToPlaneRow_LSX(const uint8_t* src_sobelx,
+                         const uint8_t* src_sobely,
+                         uint8_t* dst_y,
+                         int width);
+void SobelXYRow_C(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width);
+void SobelXYRow_SSE2(const uint8_t* src_sobelx,
+                     const uint8_t* src_sobely,
+                     uint8_t* dst_argb,
+                     int width);
+void SobelXYRow_NEON(const uint8_t* src_sobelx,
+                     const uint8_t* src_sobely,
+                     uint8_t* dst_argb,
+                     int width);
+void SobelXYRow_MSA(const uint8_t* src_sobelx,
+                    const uint8_t* src_sobely,
+                    uint8_t* dst_argb,
+                    int width);
+void SobelXYRow_LSX(const uint8_t* src_sobelx,
+                    const uint8_t* src_sobely,
+                    uint8_t* dst_argb,
+                    int width);
+void SobelRow_Any_SSE2(const uint8_t* y_buf,
+                       const uint8_t* uv_buf,
+                       uint8_t* dst_ptr,
+                       int width);
+void SobelRow_Any_NEON(const uint8_t* y_buf,
+                       const uint8_t* uv_buf,
+                       uint8_t* dst_ptr,
+                       int width);
+void SobelRow_Any_MSA(const uint8_t* y_buf,
+                      const uint8_t* uv_buf,
+                      uint8_t* dst_ptr,
+                      int width);
+void SobelRow_Any_LSX(const uint8_t* y_buf,
+                      const uint8_t* uv_buf,
+                      uint8_t* dst_ptr,
+                      int width);
+void SobelToPlaneRow_Any_SSE2(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void SobelToPlaneRow_Any_NEON(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
+void SobelToPlaneRow_Any_MSA(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void SobelToPlaneRow_Any_LSX(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             int width);
+void SobelXYRow_Any_SSE2(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
+void SobelXYRow_Any_NEON(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
+void SobelXYRow_Any_MSA(const uint8_t* y_buf,
+                        const uint8_t* uv_buf,
+                        uint8_t* dst_ptr,
+                        int width);
+void SobelXYRow_Any_LSX(const uint8_t* y_buf,
+                        const uint8_t* uv_buf,
+                        uint8_t* dst_ptr,
+                        int width);
+
+void ARGBPolynomialRow_C(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const float* poly,
+                         int width);
+void ARGBPolynomialRow_SSE2(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const float* poly,
+                            int width);
+void ARGBPolynomialRow_AVX2(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const float* poly,
+                            int width);
+
+// Scale and convert to half float.
+void HalfFloatRow_C(const uint16_t* src, uint16_t* dst, float scale, int width);
+void HalfFloatRow_SSE2(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width);
+void HalfFloatRow_Any_SSE2(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           float param,
+                           int width);
+void HalfFloatRow_AVX2(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width);
+void HalfFloatRow_Any_AVX2(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           float param,
+                           int width);
+void HalfFloatRow_F16C(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width);
+void HalfFloatRow_Any_F16C(const uint16_t* src,
+                           uint16_t* dst,
+                           float scale,
+                           int width);
+void HalfFloat1Row_F16C(const uint16_t* src,
+                        uint16_t* dst,
+                        float scale,
+                        int width);
+void HalfFloat1Row_Any_F16C(const uint16_t* src,
+                            uint16_t* dst,
+                            float scale,
+                            int width);
+void HalfFloatRow_NEON(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width);
+void HalfFloatRow_Any_NEON(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           float param,
+                           int width);
+void HalfFloat1Row_NEON(const uint16_t* src,
+                        uint16_t* dst,
+                        float scale,
+                        int width);
+void HalfFloat1Row_Any_NEON(const uint16_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            float param,
+                            int width);
+void HalfFloatRow_MSA(const uint16_t* src,
+                      uint16_t* dst,
+                      float scale,
+                      int width);
+void HalfFloatRow_Any_MSA(const uint16_t* src_ptr,
+                          uint16_t* dst_ptr,
+                          float param,
+                          int width);
+void HalfFloatRow_LSX(const uint16_t* src,
+                      uint16_t* dst,
+                      float scale,
+                      int width);
+void HalfFloatRow_Any_LSX(const uint16_t* src_ptr,
+                          uint16_t* dst_ptr,
+                          float param,
+                          int width);
+void ByteToFloatRow_C(const uint8_t* src, float* dst, float scale, int width);
+void ByteToFloatRow_NEON(const uint8_t* src,
+                         float* dst,
+                         float scale,
+                         int width);
+void ByteToFloatRow_Any_NEON(const uint8_t* src_ptr,
+                             float* dst_ptr,
+                             float param,
+                             int width);
+
+void ARGBLumaColorTableRow_C(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             int width,
+                             const uint8_t* luma,
+                             uint32_t lumacoeff);
+void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb,
+                                 uint8_t* dst_argb,
                                  int width,
-                                 const uint8* luma, uint32 lumacoeff);
+                                 const uint8_t* luma,
+                                 uint32_t lumacoeff);
+
+float ScaleMaxSamples_C(const float* src, float* dst, float scale, int width);
+float ScaleMaxSamples_NEON(const float* src,
+                           float* dst,
+                           float scale,
+                           int width);
+float ScaleSumSamples_C(const float* src, float* dst, float scale, int width);
+float ScaleSumSamples_NEON(const float* src,
+                           float* dst,
+                           float scale,
+                           int width);
+void ScaleSamples_C(const float* src, float* dst, float scale, int width);
+void ScaleSamples_NEON(const float* src, float* dst, float scale, int width);
+
+void GaussRow_F32_NEON(const float* src, float* dst, int width);
+void GaussRow_F32_C(const float* src, float* dst, int width);
+
+void GaussCol_F32_NEON(const float* src0,
+                       const float* src1,
+                       const float* src2,
+                       const float* src3,
+                       const float* src4,
+                       float* dst,
+                       int width);
+
+void GaussCol_F32_C(const float* src0,
+                    const float* src1,
+                    const float* src2,
+                    const float* src3,
+                    const float* src4,
+                    float* dst,
+                    int width);
+
+void GaussRow_C(const uint32_t* src, uint16_t* dst, int width);
+void GaussCol_C(const uint16_t* src0,
+                const uint16_t* src1,
+                const uint16_t* src2,
+                const uint16_t* src3,
+                const uint16_t* src4,
+                uint32_t* dst,
+                int width);
+
+void ClampFloatToZero_SSE2(const float* src_x, float* dst_y, int width);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_ROW_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_ROW_H_
diff --git a/media/libyuv/include/libyuv/scale.h b/media/libyuv/include/libyuv/scale.h
index 102158d1ab..443f89c2f9 100644
--- a/media/libyuv/include/libyuv/scale.h
+++ b/media/libyuv/include/libyuv/scale.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_SCALE_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_SCALE_H_
 #define INCLUDE_LIBYUV_SCALE_H_
 
 #include "libyuv/basic_types.h"
@@ -20,25 +20,45 @@ extern "C" {
 
 // Supported filtering.
 typedef enum FilterMode {
-  kFilterNone = 0,  // Point sample; Fastest.
-  kFilterLinear = 1,  // Filter horizontally only.
+  kFilterNone = 0,      // Point sample; Fastest.
+  kFilterLinear = 1,    // Filter horizontally only.
   kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
-  kFilterBox = 3  // Highest quality.
+  kFilterBox = 3        // Highest quality.
 } FilterModeEnum;
 
 // Scale a YUV plane.
 LIBYUV_API
-void ScalePlane(const uint8* src, int src_stride,
-                int src_width, int src_height,
-                uint8* dst, int dst_stride,
-                int dst_width, int dst_height,
+void ScalePlane(const uint8_t* src,
+                int src_stride,
+                int src_width,
+                int src_height,
+                uint8_t* dst,
+                int dst_stride,
+                int dst_width,
+                int dst_height,
                 enum FilterMode filtering);
 
 LIBYUV_API
-void ScalePlane_16(const uint16* src, int src_stride,
-                   int src_width, int src_height,
-                   uint16* dst, int dst_stride,
-                   int dst_width, int dst_height,
+void ScalePlane_16(const uint16_t* src,
+                   int src_stride,
+                   int src_width,
+                   int src_height,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int dst_width,
+                   int dst_height,
+                   enum FilterMode filtering);
+
+// Sample is expected to be in the low 12 bits.
+LIBYUV_API
+void ScalePlane_12(const uint16_t* src,
+                   int src_stride,
+                   int src_width,
+                   int src_height,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int dst_width,
+                   int dst_height,
                    enum FilterMode filtering);
 
 // Scales a YUV 4:2:0 image from the src width and height to the
@@ -52,43 +72,240 @@ void ScalePlane_16(const uint16* src, int src_stride,
 // Returns 0 if successful.
 
 LIBYUV_API
-int I420Scale(const uint8* src_y, int src_stride_y,
-              const uint8* src_u, int src_stride_u,
-              const uint8* src_v, int src_stride_v,
-              int src_width, int src_height,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int dst_width, int dst_height,
+int I420Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
               enum FilterMode filtering);
 
 LIBYUV_API
-int I420Scale_16(const uint16* src_y, int src_stride_y,
-                 const uint16* src_u, int src_stride_u,
-                 const uint16* src_v, int src_stride_v,
-                 int src_width, int src_height,
-                 uint16* dst_y, int dst_stride_y,
-                 uint16* dst_u, int dst_stride_u,
-                 uint16* dst_v, int dst_stride_v,
-                 int dst_width, int dst_height,
+int I420Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
                  enum FilterMode filtering);
 
-#ifdef __cplusplus
-// Legacy API.  Deprecated.
 LIBYUV_API
-int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
-          int src_stride_y, int src_stride_u, int src_stride_v,
-          int src_width, int src_height,
-          uint8* dst_y, uint8* dst_u, uint8* dst_v,
-          int dst_stride_y, int dst_stride_u, int dst_stride_v,
-          int dst_width, int dst_height,
-          LIBYUV_BOOL interpolate);
+int I420Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+// Scales a YUV 4:4:4 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// If filtering is kFilterBox, averaging is used to produce ever better
+// quality image, at further expense of speed.
+// Returns 0 if successful.
+
+LIBYUV_API
+int I444Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering);
+
+LIBYUV_API
+int I444Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+LIBYUV_API
+int I444Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+// Scales a YUV 4:2:2 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// If filtering is kFilterBox, averaging is used to produce ever better
+// quality image, at further expense of speed.
+// Returns 0 if successful.
+LIBYUV_API
+int I422Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering);
 
+LIBYUV_API
+int I422Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+LIBYUV_API
+int I422Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+// Scales an NV12 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// kFilterBox is not supported for the UV channel and will be treated as
+// bilinear.
+// Returns 0 if successful.
+
+LIBYUV_API
+int NV12Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_uv,
+              int src_stride_uv,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_uv,
+              int dst_stride_uv,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering);
+
+#ifdef __cplusplus
 // Legacy API.  Deprecated.
 LIBYUV_API
-int ScaleOffset(const uint8* src_i420, int src_width, int src_height,
-                uint8* dst_i420, int dst_width, int dst_height, int dst_yoffset,
-                LIBYUV_BOOL interpolate);
+int Scale(const uint8_t* src_y,
+          const uint8_t* src_u,
+          const uint8_t* src_v,
+          int src_stride_y,
+          int src_stride_u,
+          int src_stride_v,
+          int src_width,
+          int src_height,
+          uint8_t* dst_y,
+          uint8_t* dst_u,
+          uint8_t* dst_v,
+          int dst_stride_y,
+          int dst_stride_u,
+          int dst_stride_v,
+          int dst_width,
+          int dst_height,
+          LIBYUV_BOOL interpolate);
 
 // For testing, allow disabling of specialized scalers.
 LIBYUV_API
@@ -100,4 +317,4 @@ void SetUseReferenceImpl(LIBYUV_BOOL use);
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_SCALE_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_SCALE_H_
diff --git a/media/libyuv/include/libyuv/scale_argb.h b/media/libyuv/include/libyuv/scale_argb.h
index b56cf52099..7641f18e34 100644
--- a/media/libyuv/include/libyuv/scale_argb.h
+++ b/media/libyuv/include/libyuv/scale_argb.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_SCALE_ARGB_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_SCALE_ARGB_H_
 #define INCLUDE_LIBYUV_SCALE_ARGB_H_
 
 #include "libyuv/basic_types.h"
@@ -20,32 +20,52 @@ extern "C" {
 #endif
 
 LIBYUV_API
-int ARGBScale(const uint8* src_argb, int src_stride_argb,
-              int src_width, int src_height,
-              uint8* dst_argb, int dst_stride_argb,
-              int dst_width, int dst_height,
+int ARGBScale(const uint8_t* src_argb,
+              int src_stride_argb,
+              int src_width,
+              int src_height,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int dst_width,
+              int dst_height,
               enum FilterMode filtering);
 
 // Clipped scale takes destination rectangle coordinates for clip values.
 LIBYUV_API
-int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
-                  int src_width, int src_height,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int dst_width, int dst_height,
-                  int clip_x, int clip_y, int clip_width, int clip_height,
+int ARGBScaleClip(const uint8_t* src_argb,
+                  int src_stride_argb,
+                  int src_width,
+                  int src_height,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int dst_width,
+                  int dst_height,
+                  int clip_x,
+                  int clip_y,
+                  int clip_width,
+                  int clip_height,
                   enum FilterMode filtering);
 
 // Scale with YUV conversion to ARGB and clipping.
 LIBYUV_API
-int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
-                       const uint8* src_u, int src_stride_u,
-                       const uint8* src_v, int src_stride_v,
-                       uint32 src_fourcc,
-                       int src_width, int src_height,
-                       uint8* dst_argb, int dst_stride_argb,
-                       uint32 dst_fourcc,
-                       int dst_width, int dst_height,
-                       int clip_x, int clip_y, int clip_width, int clip_height,
+int YUVToARGBScaleClip(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint32_t src_fourcc,
+                       int src_width,
+                       int src_height,
+                       uint8_t* dst_argb,
+                       int dst_stride_argb,
+                       uint32_t dst_fourcc,
+                       int dst_width,
+                       int dst_height,
+                       int clip_x,
+                       int clip_y,
+                       int clip_width,
+                       int clip_height,
                        enum FilterMode filtering);
 
 #ifdef __cplusplus
@@ -53,4 +73,4 @@ int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_SCALE_ARGB_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_SCALE_ARGB_H_
diff --git a/media/libyuv/include/libyuv/scale_rgb.h b/media/libyuv/include/libyuv/scale_rgb.h
new file mode 100644
index 0000000000..d17c39fd6e
--- /dev/null
+++ b/media/libyuv/include/libyuv/scale_rgb.h
@@ -0,0 +1,42 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_RGB_H_
+#define INCLUDE_LIBYUV_SCALE_RGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// RGB can be RAW, RGB24 or YUV24
+// RGB scales 24 bit images by converting a row at a time to ARGB
+// and using ARGB row functions to scale, then convert to RGB.
+// TODO(fbarchard): Allow input/output formats to be specified.
+LIBYUV_API
+int RGBScale(const uint8_t* src_rgb,
+             int src_stride_rgb,
+             int src_width,
+             int src_height,
+             uint8_t* dst_rgb,
+             int dst_stride_rgb,
+             int dst_width,
+             int dst_height,
+             enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_UV_H_
diff --git a/media/libyuv/include/libyuv/scale_row.h b/media/libyuv/include/libyuv/scale_row.h
index df699e6c22..7996ea05d6 100644
--- a/media/libyuv/include/libyuv/scale_row.h
+++ b/media/libyuv/include/libyuv/scale_row.h
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_
 #define INCLUDE_LIBYUV_SCALE_ROW_H_
 
 #include "libyuv/basic_types.h"
@@ -19,17 +19,20 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#if defined(__pnacl__) || defined(__CLR_VER) || \
-    (defined(__i386__) && !defined(__SSE2__))
+#if defined(__pnacl__) || defined(__CLR_VER) ||            \
+    (defined(__native_client__) && defined(__x86_64__)) || \
+    (defined(__i386__) && !defined(__SSE__) && !defined(__clang__))
 #define LIBYUV_DISABLE_X86
 #endif
+#if defined(__native_client__)
+#define LIBYUV_DISABLE_NEON
+#endif
 // MemorySanitizer does not support assembly code yet. http://crbug.com/344505
 #if defined(__has_feature)
 #if __has_feature(memory_sanitizer)
 #define LIBYUV_DISABLE_X86
 #endif
 #endif
-
 // GCC >= 4.7.0 required for AVX2.
 #if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
 #if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
@@ -45,8 +48,8 @@ extern "C" {
 #endif  // __clang__
 
 // Visual C 2012 required for AVX2.
-#if defined(_M_IX86) && !defined(__clang__) && \
-    defined(_MSC_VER) && _MSC_VER >= 1700
+#if defined(_M_IX86) && !defined(__clang__) && defined(_MSC_VER) && \
+    _MSC_VER >= 1700
 #define VISUALC_HAS_AVX2 1
 #endif  // VisualStudio >= 2012
 
@@ -55,6 +58,7 @@ extern "C" {
     (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
 #define HAS_FIXEDDIV1_X86
 #define HAS_FIXEDDIV_X86
+#define HAS_SCALEADDROW_SSE2
 #define HAS_SCALEARGBCOLS_SSE2
 #define HAS_SCALEARGBCOLSUP2_SSE2
 #define HAS_SCALEARGBFILTERCOLS_SSSE3
@@ -66,23 +70,62 @@ extern "C" {
 #define HAS_SCALEROWDOWN34_SSSE3
 #define HAS_SCALEROWDOWN38_SSSE3
 #define HAS_SCALEROWDOWN4_SSSE3
-#define HAS_SCALEADDROW_SSE2
+#endif
+
+// The following are available for gcc/clang x86 platforms:
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+#define HAS_SCALEUVROWDOWN2BOX_SSSE3
+#define HAS_SCALEROWUP2_LINEAR_SSE2
+#define HAS_SCALEROWUP2_LINEAR_SSSE3
+#define HAS_SCALEROWUP2_BILINEAR_SSE2
+#define HAS_SCALEROWUP2_BILINEAR_SSSE3
+#define HAS_SCALEROWUP2_LINEAR_12_SSSE3
+#define HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+#define HAS_SCALEROWUP2_LINEAR_16_SSE2
+#define HAS_SCALEROWUP2_BILINEAR_16_SSE2
+#define HAS_SCALEUVROWUP2_LINEAR_SSSE3
+#define HAS_SCALEUVROWUP2_BILINEAR_SSSE3
+#define HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+#define HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+#endif
+
+// The following are available for gcc/clang x86 platforms, but
+// require clang 3.4 or gcc 4.7.
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) &&               \
+    (defined(__x86_64__) || defined(__i386__)) && \
+    (defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#define HAS_SCALEUVROWDOWN2BOX_AVX2
+#define HAS_SCALEROWUP2_LINEAR_AVX2
+#define HAS_SCALEROWUP2_BILINEAR_AVX2
+#define HAS_SCALEROWUP2_LINEAR_12_AVX2
+#define HAS_SCALEROWUP2_BILINEAR_12_AVX2
+#define HAS_SCALEROWUP2_LINEAR_16_AVX2
+#define HAS_SCALEROWUP2_BILINEAR_16_AVX2
+#define HAS_SCALEUVROWUP2_LINEAR_AVX2
+#define HAS_SCALEUVROWUP2_BILINEAR_AVX2
+#define HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+#define HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
 #endif
 
 // The following are available on all x86 platforms, but
 // require VS2012, clang 3.4 or gcc 4.7.
 // The code supports NaCL but requires a new compiler and validator.
-#if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
-    defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#if !defined(LIBYUV_DISABLE_X86) &&                          \
+    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2) || \
+     defined(GCC_HAS_AVX2))
 #define HAS_SCALEADDROW_AVX2
 #define HAS_SCALEROWDOWN2_AVX2
 #define HAS_SCALEROWDOWN4_AVX2
 #endif
 
 // The following are available on Neon platforms:
-#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+#if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SCALEADDROW_NEON
 #define HAS_SCALEARGBCOLS_NEON
+#define HAS_SCALEARGBFILTERCOLS_NEON
 #define HAS_SCALEARGBROWDOWN2_NEON
 #define HAS_SCALEARGBROWDOWNEVEN_NEON
 #define HAS_SCALEFILTERCOLS_NEON
@@ -90,414 +133,1618 @@ extern "C" {
 #define HAS_SCALEROWDOWN34_NEON
 #define HAS_SCALEROWDOWN38_NEON
 #define HAS_SCALEROWDOWN4_NEON
-#define HAS_SCALEARGBFILTERCOLS_NEON
+#define HAS_SCALEUVROWDOWN2BOX_NEON
+#define HAS_SCALEUVROWDOWNEVEN_NEON
+#define HAS_SCALEROWUP2_LINEAR_NEON
+#define HAS_SCALEROWUP2_BILINEAR_NEON
+#define HAS_SCALEROWUP2_LINEAR_12_NEON
+#define HAS_SCALEROWUP2_BILINEAR_12_NEON
+#define HAS_SCALEROWUP2_LINEAR_16_NEON
+#define HAS_SCALEROWUP2_BILINEAR_16_NEON
+#define HAS_SCALEUVROWUP2_LINEAR_NEON
+#define HAS_SCALEUVROWUP2_BILINEAR_NEON
+#define HAS_SCALEUVROWUP2_LINEAR_16_NEON
+#define HAS_SCALEUVROWUP2_BILINEAR_16_NEON
 #endif
 
-// The following are available on Mips platforms:
-#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
-    defined(__mips__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-#define HAS_SCALEROWDOWN2_DSPR2
-#define HAS_SCALEROWDOWN4_DSPR2
-#define HAS_SCALEROWDOWN34_DSPR2
-#define HAS_SCALEROWDOWN38_DSPR2
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#define HAS_SCALEADDROW_MSA
+#define HAS_SCALEARGBCOLS_MSA
+#define HAS_SCALEARGBFILTERCOLS_MSA
+#define HAS_SCALEARGBROWDOWN2_MSA
+#define HAS_SCALEARGBROWDOWNEVEN_MSA
+#define HAS_SCALEFILTERCOLS_MSA
+#define HAS_SCALEROWDOWN2_MSA
+#define HAS_SCALEROWDOWN34_MSA
+#define HAS_SCALEROWDOWN38_MSA
+#define HAS_SCALEROWDOWN4_MSA
+#endif
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#define HAS_SCALEARGBROWDOWN2_LSX
+#define HAS_SCALEARGBROWDOWNEVEN_LSX
+#define HAS_SCALEROWDOWN2_LSX
+#define HAS_SCALEROWDOWN4_LSX
+#define HAS_SCALEROWDOWN38_LSX
+#define HAS_SCALEFILTERCOLS_LSX
+#define HAS_SCALEADDROW_LSX
+#define HAS_SCALEARGBCOLS_LSX
+#define HAS_SCALEARGBFILTERCOLS_LSX
+#define HAS_SCALEROWDOWN34_LSX
 #endif
 
 // Scale ARGB vertically with bilinear interpolation.
 void ScalePlaneVertical(int src_height,
-                        int dst_width, int dst_height,
-                        int src_stride, int dst_stride,
-                        const uint8* src_argb, uint8* dst_argb,
-                        int x, int y, int dy,
-                        int bpp, enum FilterMode filtering);
+                        int dst_width,
+                        int dst_height,
+                        int src_stride,
+                        int dst_stride,
+                        const uint8_t* src_argb,
+                        uint8_t* dst_argb,
+                        int x,
+                        int y,
+                        int dy,
+                        int bpp,
+                        enum FilterMode filtering);
 
 void ScalePlaneVertical_16(int src_height,
-                           int dst_width, int dst_height,
-                           int src_stride, int dst_stride,
-                           const uint16* src_argb, uint16* dst_argb,
-                           int x, int y, int dy,
-                           int wpp, enum FilterMode filtering);
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint16_t* src_argb,
+                           uint16_t* dst_argb,
+                           int x,
+                           int y,
+                           int dy,
+                           int wpp,
+                           enum FilterMode filtering);
+
+void ScalePlaneVertical_16To8(int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint16_t* src_argb,
+                              uint8_t* dst_argb,
+                              int x,
+                              int y,
+                              int dy,
+                              int wpp,
+                              int scale,
+                              enum FilterMode filtering);
+
+void ScalePlaneDown2_16To8(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint16_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           int scale,
+                           enum FilterMode filtering);
 
 // Simplify the filtering based on scale factors.
-enum FilterMode ScaleFilterReduce(int src_width, int src_height,
-                                  int dst_width, int dst_height,
+enum FilterMode ScaleFilterReduce(int src_width,
+                                  int src_height,
+                                  int dst_width,
+                                  int dst_height,
                                   enum FilterMode filtering);
 
 // Divide num by div and return as 16.16 fixed point result.
 int FixedDiv_C(int num, int div);
 int FixedDiv_X86(int num, int div);
+int FixedDiv_MIPS(int num, int div);
 // Divide num - 1 by div - 1 and return as 16.16 fixed point result.
 int FixedDiv1_C(int num, int div);
 int FixedDiv1_X86(int num, int div);
+int FixedDiv1_MIPS(int num, int div);
 #ifdef HAS_FIXEDDIV_X86
 #define FixedDiv FixedDiv_X86
 #define FixedDiv1 FixedDiv1_X86
+#elif defined HAS_FIXEDDIV_MIPS
+#define FixedDiv FixedDiv_MIPS
+#define FixedDiv1 FixedDiv1_MIPS
 #else
 #define FixedDiv FixedDiv_C
 #define FixedDiv1 FixedDiv1_C
 #endif
 
 // Compute slope values for stepping.
-void ScaleSlope(int src_width, int src_height,
-                int dst_width, int dst_height,
+void ScaleSlope(int src_width,
+                int src_height,
+                int dst_width,
+                int dst_height,
                 enum FilterMode filtering,
-                int* x, int* y, int* dx, int* dy);
-
-void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                     uint8* dst, int dst_width);
-void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                        uint16* dst, int dst_width);
-void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width);
-void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                              uint16* dst, int dst_width);
-void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width);
-void ScaleRowDown2Box_Odd_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width);
-void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst, int dst_width);
-void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                     uint8* dst, int dst_width);
-void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                        uint16* dst, int dst_width);
-void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width);
-void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst, int dst_width);
-void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                      uint8* dst, int dst_width);
-void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                         uint16* dst, int dst_width);
-void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* d, int dst_width);
-void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                               uint16* d, int dst_width);
-void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* d, int dst_width);
-void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                               uint16* d, int dst_width);
-void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
-                 int dst_width, int x, int dx);
-void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                    int dst_width, int x, int dx);
-void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
-                    int dst_width, int, int);
-void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                       int dst_width, int, int);
-void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
-                       int dst_width, int x, int dx);
-void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                          int dst_width, int x, int dx);
-void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
-                         int dst_width, int x, int dx);
-void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                            int dst_width, int x, int dx);
-void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                      uint8* dst, int dst_width);
-void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                         uint16* dst, int dst_width);
-void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                int* x,
+                int* y,
+                int* dx,
+                int* dy);
+
+void ScaleRowDown2_C(const uint8_t* src_ptr,
+                     ptrdiff_t src_stride,
+                     uint8_t* dst,
+                     int dst_width);
+void ScaleRowDown2_16_C(const uint16_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint16_t* dst,
+                        int dst_width);
+void ScaleRowDown2_16To8_C(const uint16_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width,
+                           int scale);
+void ScaleRowDown2_16To8_Odd_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst,
+                               int dst_width,
+                               int scale);
+void ScaleRowDown2Linear_C(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width);
+void ScaleRowDown2Linear_16_C(const uint16_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint16_t* dst,
+                              int dst_width);
+void ScaleRowDown2Linear_16To8_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst,
+                                 int dst_width,
+                                 int scale);
+void ScaleRowDown2Linear_16To8_Odd_C(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst,
+                                     int dst_width,
+                                     int scale);
+void ScaleRowDown2Box_C(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown2Box_Odd_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst,
+                            int dst_width);
+void ScaleRowDown2Box_16_C(const uint16_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint16_t* dst,
+                           int dst_width);
+void ScaleRowDown2Box_16To8_C(const uint16_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst,
+                              int dst_width,
+                              int scale);
+void ScaleRowDown2Box_16To8_Odd_C(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst,
+                                  int dst_width,
+                                  int scale);
+void ScaleRowDown4_C(const uint8_t* src_ptr,
+                     ptrdiff_t src_stride,
+                     uint8_t* dst,
+                     int dst_width);
+void ScaleRowDown4_16_C(const uint16_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint16_t* dst,
+                        int dst_width);
+void ScaleRowDown4Box_C(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown4Box_16_C(const uint16_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint16_t* dst,
+                           int dst_width);
+void ScaleRowDown34_C(const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint8_t* dst,
+                      int dst_width);
+void ScaleRowDown34_16_C(const uint16_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint16_t* dst,
+                         int dst_width);
+void ScaleRowDown34_0_Box_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* d,
+                            int dst_width);
+void ScaleRowDown34_0_Box_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* d,
+                               int dst_width);
+void ScaleRowDown34_1_Box_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* d,
+                            int dst_width);
+void ScaleRowDown34_1_Box_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* d,
+                               int dst_width);
+
+void ScaleRowUp2_Linear_C(const uint8_t* src_ptr,
+                          uint8_t* dst_ptr,
+                          int dst_width);
+void ScaleRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            ptrdiff_t dst_stride,
+                            int dst_width);
+void ScaleRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_Any_C(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowUp2_Bilinear_Any_C(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                ptrdiff_t dst_stride,
+                                int dst_width);
+void ScaleRowUp2_Linear_16_Any_C(const uint16_t* src_ptr,
+                                 uint16_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_C(const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint16_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+
+void ScaleCols_C(uint8_t* dst_ptr,
+                 const uint8_t* src_ptr,
+                 int dst_width,
+                 int x,
+                 int dx);
+void ScaleCols_16_C(uint16_t* dst_ptr,
+                    const uint16_t* src_ptr,
+                    int dst_width,
+                    int x,
+                    int dx);
+void ScaleColsUp2_C(uint8_t* dst_ptr,
+                    const uint8_t* src_ptr,
+                    int dst_width,
+                    int,
+                    int);
+void ScaleColsUp2_16_C(uint16_t* dst_ptr,
+                       const uint16_t* src_ptr,
+                       int dst_width,
+                       int,
+                       int);
+void ScaleFilterCols_C(uint8_t* dst_ptr,
+                       const uint8_t* src_ptr,
+                       int dst_width,
+                       int x,
+                       int dx);
+void ScaleFilterCols_16_C(uint16_t* dst_ptr,
+                          const uint16_t* src_ptr,
+                          int dst_width,
+                          int x,
+                          int dx);
+void ScaleFilterCols64_C(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x32,
+                         int dx);
+void ScaleFilterCols64_16_C(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            int dst_width,
+                            int x32,
+                            int dx);
+void ScaleRowDown38_C(const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint8_t* dst,
+                      int dst_width);
+void ScaleRowDown38_16_C(const uint16_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint16_t* dst,
+                         int dst_width);
+void ScaleRowDown38_3_Box_C(const uint8_t* src_ptr,
                             ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown38_3_Box_16_C(const uint16_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint16* dst_ptr, int dst_width);
-void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                               uint16* dst_ptr, int dst_width);
-void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width);
-void ScaleARGBRowDown2_C(const uint8* src_argb,
+                               uint16_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown38_2_Box_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown38_2_Box_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* dst_ptr,
+                               int dst_width);
+void ScaleAddRow_C(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleAddRow_16_C(const uint16_t* src_ptr,
+                      uint32_t* dst_ptr,
+                      int src_width);
+void ScaleARGBRowDown2_C(const uint8_t* src_argb,
                          ptrdiff_t src_stride,
-                         uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                         uint8_t* dst_argb,
+                         int dst_width);
+void ScaleARGBRowDown2Linear_C(const uint8_t* src_argb,
                                ptrdiff_t src_stride,
-                               uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
-                            uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                               uint8_t* dst_argb,
+                               int dst_width);
+void ScaleARGBRowDown2Box_C(const uint8_t* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_argb,
+                            int dst_width);
+void ScaleARGBRowDownEven_C(const uint8_t* src_argb,
+                            ptrdiff_t src_stride,
                             int src_stepx,
-                            uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                            uint8_t* dst_argb,
+                            int dst_width);
+void ScaleARGBRowDownEvenBox_C(const uint8_t* src_argb,
                                ptrdiff_t src_stride,
                                int src_stepx,
-                               uint8* dst_argb, int dst_width);
-void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
-                     int dst_width, int x, int dx);
-void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
-                       int dst_width, int x, int dx);
-void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int, int);
-void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
-                           int dst_width, int x, int dx);
-void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
-                             int dst_width, int x, int dx);
+                               uint8_t* dst_argb,
+                               int dst_width);
+void ScaleARGBCols_C(uint8_t* dst_argb,
+                     const uint8_t* src_argb,
+                     int dst_width,
+                     int x,
+                     int dx);
+void ScaleARGBCols64_C(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x32,
+                       int dx);
+void ScaleARGBColsUp2_C(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int,
+                        int);
+void ScaleARGBFilterCols_C(uint8_t* dst_argb,
+                           const uint8_t* src_argb,
+                           int dst_width,
+                           int x,
+                           int dx);
+void ScaleARGBFilterCols64_C(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x32,
+                             int dx);
+void ScaleUVRowDown2_C(const uint8_t* src_uv,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst_uv,
+                       int dst_width);
+void ScaleUVRowDown2Linear_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_uv,
+                             int dst_width);
+void ScaleUVRowDown2Box_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_uv,
+                          int dst_width);
+void ScaleUVRowDownEven_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          int src_stepx,
+                          uint8_t* dst_uv,
+                          int dst_width);
+void ScaleUVRowDownEvenBox_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             int src_stepx,
+                             uint8_t* dst_uv,
+                             int dst_width);
+
+void ScaleUVRowUp2_Linear_C(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleUVRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              ptrdiff_t dst_stride,
+                              int dst_width);
+void ScaleUVRowUp2_Linear_Any_C(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_C(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleUVRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint16_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_C(const uint16_t* src_ptr,
+                                   uint16_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_C(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint16_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+
+void ScaleUVCols_C(uint8_t* dst_uv,
+                   const uint8_t* src_uv,
+                   int dst_width,
+                   int x,
+                   int dx);
+void ScaleUVCols64_C(uint8_t* dst_uv,
+                     const uint8_t* src_uv,
+                     int dst_width,
+                     int x32,
+                     int dx);
+void ScaleUVColsUp2_C(uint8_t* dst_uv,
+                      const uint8_t* src_uv,
+                      int dst_width,
+                      int,
+                      int);
+void ScaleUVFilterCols_C(uint8_t* dst_uv,
+                         const uint8_t* src_uv,
+                         int dst_width,
+                         int x,
+                         int dx);
+void ScaleUVFilterCols64_C(uint8_t* dst_uv,
+                           const uint8_t* src_uv,
+                           int dst_width,
+                           int x32,
+                           int dx);
 
 // Specialized scalers for x86.
-void ScaleRowDown2_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Linear_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
-void ScaleRowDown4_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width);
-void ScaleRowDown4Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width);
-void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
-
-void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+void ScaleRowDown2_SSSE3(const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_ptr,
+                         int dst_width);
+void ScaleRowDown2Linear_SSSE3(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown2_AVX2(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width);
+void ScaleRowDown2Linear_AVX2(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown2Box_AVX2(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
+void ScaleRowDown4_SSSE3(const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_ptr,
+                         int dst_width);
+void ScaleRowDown4Box_SSSE3(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown4_AVX2(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width);
+void ScaleRowDown4Box_AVX2(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
+
+void ScaleRowDown34_SSSE3(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_ptr,
+                          int dst_width);
+void ScaleRowDown34_1_Box_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleRowDown34_0_Box_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleRowDown38_SSSE3(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_ptr,
+                          int dst_width);
+void ScaleRowDown38_3_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleRowDown38_2_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+
+void ScaleRowUp2_Linear_SSE2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_SSE2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_12_SSSE3(const uint16_t* src_ptr,
+                                 uint16_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_12_SSSE3(const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint16_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_16_SSE2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_16_SSE2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                ptrdiff_t dst_stride,
+                                int dst_width);
+void ScaleRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_12_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_12_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_Any_SSE2(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_Any_SSE2(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_12_Any_SSSE3(const uint16_t* src_ptr,
+                                     uint16_t* dst_ptr,
+                                     int dst_width);
+void ScaleRowUp2_Bilinear_12_Any_SSSE3(const uint16_t* src_ptr,
+                                       ptrdiff_t src_stride,
+                                       uint16_t* dst_ptr,
+                                       ptrdiff_t dst_stride,
+                                       int dst_width);
+void ScaleRowUp2_Linear_16_Any_SSE2(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_SSE2(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleRowUp2_Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowUp2_Bilinear_Any_SSSE3(const uint8_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width);
+void ScaleRowUp2_Linear_Any_AVX2(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_Any_AVX2(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_12_Any_AVX2(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_12_Any_AVX2(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleRowUp2_Linear_16_Any_AVX2(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_AVX2(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+
+void ScaleRowDown2_Any_SSSE3(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowDown2Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleRowDown2Box_Any_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleRowDown2Box_Odd_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown2_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                             uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Linear_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                                   uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Box_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Box_Odd_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown2_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Linear_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                  uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-void ScaleRowDown2Box_Odd_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-void ScaleRowDown4_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                             uint8* dst_ptr, int dst_width);
-void ScaleRowDown4Box_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown4_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown4Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-
-void ScaleRowDown34_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_1_Box_Any_SSSE3(const uint8* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleRowDown2_Any_AVX2(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown2Linear_Any_AVX2(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown2Box_Any_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown2Box_Odd_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown4_Any_SSSE3(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowDown4Box_Any_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleRowDown4_Any_AVX2(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown4Box_Any_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+
+void ScaleRowDown34_Any_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown34_1_Box_Any_SSSE3(const uint8_t* src_ptr,
                                     ptrdiff_t src_stride,
-                                    uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_0_Box_Any_SSSE3(const uint8* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowDown34_0_Box_Any_SSSE3(const uint8_t* src_ptr,
                                     ptrdiff_t src_stride,
-                                    uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_3_Box_Any_SSSE3(const uint8* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowDown38_Any_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown38_3_Box_Any_SSSE3(const uint8_t* src_ptr,
                                     ptrdiff_t src_stride,
-                                    uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_2_Box_Any_SSSE3(const uint8* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowDown38_2_Box_Any_SSSE3(const uint8_t* src_ptr,
                                     ptrdiff_t src_stride,
-                                    uint8* dst_ptr, int dst_width);
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
 
-void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-void ScaleAddRow_Any_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-void ScaleAddRow_Any_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-
-void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                           int dst_width, int x, int dx);
-void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                       int dst_width, int x, int dx);
+void ScaleAddRow_SSE2(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleAddRow_AVX2(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleAddRow_Any_SSE2(const uint8_t* src_ptr,
+                          uint16_t* dst_ptr,
+                          int src_width);
+void ScaleAddRow_Any_AVX2(const uint8_t* src_ptr,
+                          uint16_t* dst_ptr,
+                          int src_width);
 
+void ScaleFilterCols_SSSE3(uint8_t* dst_ptr,
+                           const uint8_t* src_ptr,
+                           int dst_width,
+                           int x,
+                           int dx);
+void ScaleColsUp2_SSE2(uint8_t* dst_ptr,
+                       const uint8_t* src_ptr,
+                       int dst_width,
+                       int x,
+                       int dx);
 
 // ARGB Column functions
-void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx);
-void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
-                               int dst_width, int x, int dx);
-void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
-                           int dst_width, int x, int dx);
-void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
-                              int dst_width, int x, int dx);
-void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx);
-void ScaleARGBFilterCols_Any_NEON(uint8* dst_argb, const uint8* src_argb,
-                                  int dst_width, int x, int dx);
-void ScaleARGBCols_Any_NEON(uint8* dst_argb, const uint8* src_argb,
-                            int dst_width, int x, int dx);
+void ScaleARGBCols_SSE2(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx);
+void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
+                               const uint8_t* src_argb,
+                               int dst_width,
+                               int x,
+                               int dx);
+void ScaleARGBColsUp2_SSE2(uint8_t* dst_argb,
+                           const uint8_t* src_argb,
+                           int dst_width,
+                           int x,
+                           int dx);
+void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
+                              const uint8_t* src_argb,
+                              int dst_width,
+                              int x,
+                              int dx);
+void ScaleARGBCols_NEON(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx);
+void ScaleARGBFilterCols_Any_NEON(uint8_t* dst_ptr,
+                                  const uint8_t* src_ptr,
+                                  int dst_width,
+                                  int x,
+                                  int dx);
+void ScaleARGBCols_Any_NEON(uint8_t* dst_ptr,
+                            const uint8_t* src_ptr,
+                            int dst_width,
+                            int x,
+                            int dx);
+void ScaleARGBFilterCols_MSA(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleARGBCols_MSA(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx);
+void ScaleARGBFilterCols_Any_MSA(uint8_t* dst_ptr,
+                                 const uint8_t* src_ptr,
+                                 int dst_width,
+                                 int x,
+                                 int dx);
+void ScaleARGBCols_Any_MSA(uint8_t* dst_ptr,
+                           const uint8_t* src_ptr,
+                           int dst_width,
+                           int x,
+                           int dx);
 
 // ARGB Row functions
-void ScaleARGBRowDown2_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                            uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                                  uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                               uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width);
-void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
-                                  uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst, int dst_width);
-void ScaleARGBRowDown2_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                                uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Linear_Any_SSE2(const uint8* src_argb,
+void ScaleARGBRowDown2_SSE2(const uint8_t* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_argb,
+                            int dst_width);
+void ScaleARGBRowDown2Linear_SSE2(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_argb,
+                                  int dst_width);
+void ScaleARGBRowDown2Box_SSE2(const uint8_t* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_argb,
+                               int dst_width);
+void ScaleARGBRowDown2_NEON(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst,
+                            int dst_width);
+void ScaleARGBRowDown2Linear_NEON(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_argb,
+                                  int dst_width);
+void ScaleARGBRowDown2Box_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst,
+                               int dst_width);
+void ScaleARGBRowDown2_MSA(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width);
+void ScaleARGBRowDown2Linear_MSA(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width);
+void ScaleARGBRowDown2Box_MSA(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width);
+void ScaleARGBRowDown2_LSX(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width);
+void ScaleARGBRowDown2Linear_LSX(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width);
+void ScaleARGBRowDown2Box_LSX(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width);
+void ScaleARGBRowDown2_Any_SSE2(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleARGBRowDown2Linear_Any_SSE2(const uint8_t* src_ptr,
                                       ptrdiff_t src_stride,
-                                      uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Box_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                                   uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst, int dst_width);
-void ScaleARGBRowDown2Linear_Any_NEON(const uint8* src_argb,
+                                      uint8_t* dst_ptr,
+                                      int dst_width);
+void ScaleARGBRowDown2Box_Any_SSE2(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleARGBRowDown2_Any_NEON(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleARGBRowDown2Linear_Any_NEON(const uint8_t* src_ptr,
                                       ptrdiff_t src_stride,
-                                      uint8* dst_argb, int dst_width);
-void ScaleARGBRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                                   uint8* dst, int dst_width);
-
-void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                               int src_stepx, uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                      uint8_t* dst_ptr,
+                                      int dst_width);
+void ScaleARGBRowDown2Box_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleARGBRowDown2_Any_MSA(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleARGBRowDown2Linear_Any_MSA(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+void ScaleARGBRowDown2Box_Any_MSA(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleARGBRowDown2_Any_LSX(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleARGBRowDown2Linear_Any_LSX(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+void ScaleARGBRowDown2Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleARGBRowDownEven_SSE2(const uint8_t* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8_t* dst_argb,
+                               int dst_width);
+void ScaleARGBRowDownEvenBox_SSE2(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
                                   int src_stepx,
-                                  uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8_t* dst_argb,
+                                  int dst_width);
+void ScaleARGBRowDownEven_NEON(const uint8_t* src_argb,
+                               ptrdiff_t src_stride,
                                int src_stepx,
-                               uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                               uint8_t* dst_argb,
+                               int dst_width);
+void ScaleARGBRowDownEvenBox_NEON(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
                                   int src_stepx,
-                                  uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEven_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8_t* dst_argb,
+                                  int dst_width);
+void ScaleARGBRowDownEven_MSA(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int32_t src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width);
+void ScaleARGBRowDownEvenBox_MSA(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_argb,
+                                 int dst_width);
+void ScaleARGBRowDownEven_LSX(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int32_t src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width);
+void ScaleARGBRowDownEvenBox_LSX(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_argb,
+                                 int dst_width);
+void ScaleARGBRowDownEven_Any_SSE2(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
                                    int src_stepx,
-                                   uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEvenBox_Any_SSE2(const uint8* src_argb,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleARGBRowDownEvenBox_Any_SSE2(const uint8_t* src_ptr,
                                       ptrdiff_t src_stride,
                                       int src_stepx,
-                                      uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEven_Any_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                      uint8_t* dst_ptr,
+                                      int dst_width);
+void ScaleARGBRowDownEven_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
                                    int src_stepx,
-                                   uint8* dst_argb, int dst_width);
-void ScaleARGBRowDownEvenBox_Any_NEON(const uint8* src_argb,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleARGBRowDownEvenBox_Any_NEON(const uint8_t* src_ptr,
                                       ptrdiff_t src_stride,
                                       int src_stepx,
-                                      uint8* dst_argb, int dst_width);
+                                      uint8_t* dst_ptr,
+                                      int dst_width);
+void ScaleARGBRowDownEven_Any_MSA(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  int32_t src_stepx,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleARGBRowDownEvenBox_Any_MSA(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     int src_stepx,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+void ScaleARGBRowDownEven_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  int32_t src_stepx,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleARGBRowDownEvenBox_Any_LSX(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     int src_stepx,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+
+// UV Row functions
+void ScaleUVRowDown2_SSSE3(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_uv,
+                           int dst_width);
+void ScaleUVRowDown2Linear_SSSE3(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_uv,
+                                 int dst_width);
+void ScaleUVRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_uv,
+                              int dst_width);
+void ScaleUVRowDown2Box_AVX2(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_uv,
+                             int dst_width);
+void ScaleUVRowDown2_NEON(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width);
+void ScaleUVRowDown2Linear_NEON(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_uv,
+                                int dst_width);
+void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width);
+void ScaleUVRowDown2_MSA(const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_uv,
+                         int dst_width);
+void ScaleUVRowDown2Linear_MSA(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_uv,
+                               int dst_width);
+void ScaleUVRowDown2Box_MSA(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_uv,
+                            int dst_width);
+void ScaleUVRowDown2_Any_SSSE3(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowDown2Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+void ScaleUVRowDown2Box_Any_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowDown2Box_Any_AVX2(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleUVRowDown2_Any_NEON(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleUVRowDown2Linear_Any_NEON(const uint8_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleUVRowDown2Box_Any_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleUVRowDown2_Any_MSA(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleUVRowDown2Linear_Any_MSA(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowDown2Box_Any_MSA(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowDownEven_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              int src_stepx,
+                              uint8_t* dst_uv,
+                              int dst_width);
+void ScaleUVRowDownEvenBox_SSSE3(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_uv,
+                                 int dst_width);
+void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             int src_stepx,
+                             uint8_t* dst_uv,
+                             int dst_width);
+void ScaleUVRowDownEvenBox_NEON(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                int src_stepx,
+                                uint8_t* dst_uv,
+                                int dst_width);
+void ScaleUVRowDownEven_MSA(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int32_t src_stepx,
+                            uint8_t* dst_uv,
+                            int dst_width);
+void ScaleUVRowDownEvenBox_MSA(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8_t* dst_uv,
+                               int dst_width);
+void ScaleUVRowDownEven_Any_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowDownEvenBox_Any_SSSE3(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     int src_stepx,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+void ScaleUVRowDownEven_Any_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleUVRowDownEvenBox_Any_NEON(const uint8_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    int src_stepx,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleUVRowDownEven_Any_MSA(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                int32_t src_stepx,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowDownEvenBox_Any_MSA(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   int src_stepx,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+
+void ScaleUVRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleUVRowUp2_Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_SSSE3(const uint8_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleUVRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width);
+void ScaleUVRowUp2_Linear_Any_AVX2(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_AVX2(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+void ScaleUVRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width);
+void ScaleUVRowUp2_Linear_Any_NEON(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_NEON(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+void ScaleUVRowUp2_Linear_16_SSE41(const uint16_t* src_ptr,
+                                   uint16_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_16_SSE41(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint16_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_SSE41(const uint16_t* src_ptr,
+                                       uint16_t* dst_ptr,
+                                       int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_SSE41(const uint16_t* src_ptr,
+                                         ptrdiff_t src_stride,
+                                         uint16_t* dst_ptr,
+                                         ptrdiff_t dst_stride,
+                                         int dst_width);
+void ScaleUVRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_AVX2(const uint16_t* src_ptr,
+                                      uint16_t* dst_ptr,
+                                      int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_AVX2(const uint16_t* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint16_t* dst_ptr,
+                                        ptrdiff_t dst_stride,
+                                        int dst_width);
+void ScaleUVRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_NEON(const uint16_t* src_ptr,
+                                      uint16_t* dst_ptr,
+                                      int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_NEON(const uint16_t* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint16_t* dst_ptr,
+                                        ptrdiff_t dst_stride,
+                                        int dst_width);
 
 // ScaleRowDown2Box also used by planar functions
 // NEON downscalers with interpolation.
 
 // Note - not static due to reuse in convert for 444 to 420.
-void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width);
-void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst, int dst_width);
-void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width);
-
-void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width);
-void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_NEON(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst,
+                              int dst_width);
+void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width);
+
+void ScaleRowDown4_NEON(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width);
+void ScaleRowDown4Box_NEON(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
 
 // Down scale from 4 to 3 pixels. Use the neon multilane read/write
 //  to load up the every 4th pixel into a 4 different registers.
 // Point samples 32 pixels to 24 pixels.
-void ScaleRowDown34_NEON(const uint8* src_ptr,
+void ScaleRowDown34_NEON(const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                         uint8_t* dst_ptr,
+                         int dst_width);
+void ScaleRowDown34_0_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown34_1_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
+                               uint8_t* dst_ptr,
+                               int dst_width);
 
 // 32 -> 12
-void ScaleRowDown38_NEON(const uint8* src_ptr,
+void ScaleRowDown38_NEON(const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width);
+                         uint8_t* dst_ptr,
+                         int dst_width);
 // 32x3 -> 12x1
-void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
+                               uint8_t* dst_ptr,
+                               int dst_width);
 // 32x2 -> 12x1
-void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+
+void ScaleRowDown2_Any_NEON(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown2Linear_Any_NEON(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown2Box_Any_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown2Box_Odd_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown4_Any_NEON(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown4Box_Any_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-
-void ScaleRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width);
-void ScaleRowDown2Linear_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                                  uint8* dst, int dst_width);
-void ScaleRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst, int dst_width);
-void ScaleRowDown2Box_Odd_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst, int dst_width);
-void ScaleRowDown4_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width);
-void ScaleRowDown4Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                             uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_0_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                                   uint8* dst_ptr, int dst_width);
-void ScaleRowDown34_1_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                                   uint8* dst_ptr, int dst_width);
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleRowDown34_Any_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowDown34_0_Box_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleRowDown34_1_Box_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
 // 32 -> 12
-void ScaleRowDown38_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                             uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_Any_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width);
 // 32x3 -> 12x1
-void ScaleRowDown38_3_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
 // 32x2 -> 12x1
-void ScaleRowDown38_2_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width);
-
-void ScaleAddRow_NEON(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-void ScaleAddRow_Any_NEON(const uint8* src_ptr, uint16* dst_ptr, int src_width);
-
-void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
-                          int dst_width, int x, int dx);
-
-void ScaleFilterCols_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
-                              int dst_width, int x, int dx);
-
-void ScaleRowDown2_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst, int dst_width);
-void ScaleRowDown2Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width);
-void ScaleRowDown4_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst, int dst_width);
-void ScaleRowDown4Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width);
-void ScaleRowDown34_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst, int dst_width);
-void ScaleRowDown34_0_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* d, int dst_width);
-void ScaleRowDown34_1_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* d, int dst_width);
-void ScaleRowDown38_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst, int dst_width);
-void ScaleRowDown38_2_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
-void ScaleRowDown38_3_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+
+void ScaleRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_12_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_12_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_Any_NEON(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_12_Any_NEON(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_12_Any_NEON(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleRowUp2_Linear_16_Any_NEON(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_NEON(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+
+void ScaleAddRow_NEON(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleAddRow_Any_NEON(const uint8_t* src_ptr,
+                          uint16_t* dst_ptr,
+                          int src_width);
+
+void ScaleFilterCols_NEON(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          int dst_width,
+                          int x,
+                          int dx);
+
+void ScaleFilterCols_Any_NEON(uint8_t* dst_ptr,
+                              const uint8_t* src_ptr,
+                              int dst_width,
+                              int x,
+                              int dx);
+
+void ScaleRowDown2_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width);
+void ScaleRowDown2Linear_MSA(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width);
+void ScaleRowDown2Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width);
+void ScaleRowDown4_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width);
+void ScaleRowDown4Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width);
+void ScaleRowDown38_MSA(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown38_2_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown38_3_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleAddRow_MSA(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleFilterCols_MSA(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx);
+void ScaleRowDown34_MSA(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown34_0_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width);
+void ScaleRowDown34_1_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width);
+
+void ScaleRowDown2_Any_MSA(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
+void ScaleRowDown2Linear_Any_MSA(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowDown2Box_Any_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown4_Any_MSA(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
+void ScaleRowDown4Box_Any_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown38_Any_MSA(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown38_2_Box_Any_MSA(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown38_3_Box_Any_MSA(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleAddRow_Any_MSA(const uint8_t* src_ptr,
+                         uint16_t* dst_ptr,
+                         int src_width);
+void ScaleFilterCols_Any_MSA(uint8_t* dst_ptr,
+                             const uint8_t* src_ptr,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleRowDown34_Any_MSA(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown34_0_Box_Any_MSA(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown34_1_Box_Any_MSA(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+
+void ScaleRowDown2_LSX(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width);
+void ScaleRowDown2Linear_LSX(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width);
+void ScaleRowDown2Box_LSX(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width);
+void ScaleRowDown4_LSX(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width);
+void ScaleRowDown4Box_LSX(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width);
+void ScaleRowDown38_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown38_2_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown38_3_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleAddRow_LSX(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleFilterCols_LSX(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx);
+void ScaleARGBFilterCols_LSX(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleARGBCols_LSX(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx);
+void ScaleRowDown34_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown34_0_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width);
+void ScaleRowDown34_1_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width);
+void ScaleRowDown2_Any_LSX(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
+void ScaleRowDown2Linear_Any_LSX(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowDown2Box_Any_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown4_Any_LSX(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width);
+void ScaleRowDown4Box_Any_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown38_Any_LSX(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown38_2_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown38_3_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleAddRow_Any_LSX(const uint8_t* src_ptr,
+                         uint16_t* dst_ptr,
+                         int src_width);
+void ScaleFilterCols_Any_LSX(uint8_t* dst_ptr,
+                             const uint8_t* src_ptr,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleARGBCols_Any_LSX(uint8_t* dst_ptr,
+                           const uint8_t* src_ptr,
+                           int dst_width,
+                           int x,
+                           int dx);
+void ScaleARGBFilterCols_Any_LSX(uint8_t* dst_ptr,
+                                 const uint8_t* src_ptr,
+                                 int dst_width,
+                                 int x,
+                                 int dx);
+void ScaleRowDown34_Any_LSX(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown34_0_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown34_1_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_
diff --git a/media/libyuv/include/libyuv/scale_uv.h b/media/libyuv/include/libyuv/scale_uv.h
new file mode 100644
index 0000000000..8e74e3195b
--- /dev/null
+++ b/media/libyuv/include/libyuv/scale_uv.h
@@ -0,0 +1,51 @@
+/*
+ *  Copyright 2020 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_UV_H_
+#define INCLUDE_LIBYUV_SCALE_UV_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+int UVScale(const uint8_t* src_uv,
+            int src_stride_uv,
+            int src_width,
+            int src_height,
+            uint8_t* dst_uv,
+            int dst_stride_uv,
+            int dst_width,
+            int dst_height,
+            enum FilterMode filtering);
+
+// Scale a 16 bit UV image.
+// This function is currently incomplete, it can't handle all cases.
+LIBYUV_API
+int UVScale_16(const uint16_t* src_uv,
+               int src_stride_uv,
+               int src_width,
+               int src_height,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int dst_width,
+               int dst_height,
+               enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_UV_H_
diff --git a/media/libyuv/include/libyuv/version.h b/media/libyuv/include/libyuv/version.h
index ca0c062eb0..42f8166265 100644
--- a/media/libyuv/include/libyuv/version.h
+++ b/media/libyuv/include/libyuv/version.h
@@ -8,9 +8,9 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1602
+#define LIBYUV_VERSION 1861
 
-#endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_VERSION_H_
diff --git a/media/libyuv/include/libyuv/video_common.h b/media/libyuv/include/libyuv/video_common.h
index ad934e4241..32b8a5210b 100644
--- a/media/libyuv/include/libyuv/video_common.h
+++ b/media/libyuv/include/libyuv/video_common.h
@@ -10,7 +10,7 @@
 
 // Common definitions for video, including fourcc and VideoFormat.
 
-#ifndef INCLUDE_LIBYUV_VIDEO_COMMON_H_  // NOLINT
+#ifndef INCLUDE_LIBYUV_VIDEO_COMMON_H_
 #define INCLUDE_LIBYUV_VIDEO_COMMON_H_
 
 #include "libyuv/basic_types.h"
@@ -28,13 +28,14 @@ extern "C" {
 // Needs to be a macro otherwise the OS X compiler complains when the kFormat*
 // constants are used in a switch.
 #ifdef __cplusplus
-#define FOURCC(a, b, c, d) ( \
-    (static_cast<uint32>(a)) | (static_cast<uint32>(b) << 8) | \
-    (static_cast<uint32>(c) << 16) | (static_cast<uint32>(d) << 24))
+#define FOURCC(a, b, c, d)                                        \
+  ((static_cast<uint32_t>(a)) | (static_cast<uint32_t>(b) << 8) | \
+   (static_cast<uint32_t>(c) << 16) | /* NOLINT */                \
+   (static_cast<uint32_t>(d) << 24))  /* NOLINT */
 #else
-#define FOURCC(a, b, c, d) ( \
-    ((uint32)(a)) | ((uint32)(b) << 8) | /* NOLINT */ \
-    ((uint32)(c) << 16) | ((uint32)(d) << 24))  /* NOLINT */
+#define FOURCC(a, b, c, d)                                     \
+  (((uint32_t)(a)) | ((uint32_t)(b) << 8) |       /* NOLINT */ \
+   ((uint32_t)(c) << 16) | ((uint32_t)(d) << 24)) /* NOLINT */
 #endif
 
 // Some pages discussing FourCC codes:
@@ -49,49 +50,69 @@ extern "C" {
 // Secondary formats are converted in 2 steps.
 // Auxilliary formats call primary converters.
 enum FourCC {
-  // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+  // 10 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
   FOURCC_I420 = FOURCC('I', '4', '2', '0'),
   FOURCC_I422 = FOURCC('I', '4', '2', '2'),
   FOURCC_I444 = FOURCC('I', '4', '4', '4'),
-  FOURCC_I411 = FOURCC('I', '4', '1', '1'),
   FOURCC_I400 = FOURCC('I', '4', '0', '0'),
   FOURCC_NV21 = FOURCC('N', 'V', '2', '1'),
   FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
   FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
   FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
+  FOURCC_I010 = FOURCC('I', '0', '1', '0'),  // bt.601 10 bit 420
+  FOURCC_I210 = FOURCC('I', '2', '1', '0'),  // bt.601 10 bit 422
 
-  // 2 Secondary YUV formats: row biplanar.
+  // 1 Secondary YUV format: row biplanar.  deprecated.
   FOURCC_M420 = FOURCC('M', '4', '2', '0'),
-  FOURCC_Q420 = FOURCC('Q', '4', '2', '0'),  // deprecated.
 
-  // 9 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp.
+  // 13 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp, 1 10 bpc 2 64 bpp
   FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
   FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
   FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
+  FOURCC_AR30 = FOURCC('A', 'R', '3', '0'),  // 10 bit per channel. 2101010.
+  FOURCC_AB30 = FOURCC('A', 'B', '3', '0'),  // ABGR version of 10 bit
+  FOURCC_AR64 = FOURCC('A', 'R', '6', '4'),  // 16 bit per channel.
+  FOURCC_AB64 = FOURCC('A', 'B', '6', '4'),  // ABGR version of 16 bit
   FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
-  FOURCC_RAW  = FOURCC('r', 'a', 'w', ' '),
+  FOURCC_RAW = FOURCC('r', 'a', 'w', ' '),
   FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
   FOURCC_RGBP = FOURCC('R', 'G', 'B', 'P'),  // rgb565 LE.
   FOURCC_RGBO = FOURCC('R', 'G', 'B', 'O'),  // argb1555 LE.
   FOURCC_R444 = FOURCC('R', '4', '4', '4'),  // argb4444 LE.
 
-  // 4 Secondary RGB formats: 4 Bayer Patterns. deprecated.
-  FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
-  FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
-  FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
-  FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
-
   // 1 Primary Compressed YUV format.
   FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
 
-  // 5 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+  // 14 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
   FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
   FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
   FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
   FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
-  FOURCC_J420 = FOURCC('J', '4', '2', '0'),
-  FOURCC_J400 = FOURCC('J', '4', '0', '0'),  // unofficial fourcc
-  FOURCC_H420 = FOURCC('H', '4', '2', '0'),  // unofficial fourcc
+  FOURCC_J420 =
+      FOURCC('J', '4', '2', '0'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_J422 =
+      FOURCC('J', '4', '2', '2'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_J444 =
+      FOURCC('J', '4', '4', '4'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_J400 =
+      FOURCC('J', '4', '0', '0'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_F420 = FOURCC('F', '4', '2', '0'),  // bt.709 full, unofficial fourcc
+  FOURCC_F422 = FOURCC('F', '4', '2', '2'),  // bt.709 full, unofficial fourcc
+  FOURCC_F444 = FOURCC('F', '4', '4', '4'),  // bt.709 full, unofficial fourcc
+  FOURCC_H420 = FOURCC('H', '4', '2', '0'),  // bt.709, unofficial fourcc
+  FOURCC_H422 = FOURCC('H', '4', '2', '2'),  // bt.709, unofficial fourcc
+  FOURCC_H444 = FOURCC('H', '4', '4', '4'),  // bt.709, unofficial fourcc
+  FOURCC_U420 = FOURCC('U', '4', '2', '0'),  // bt.2020, unofficial fourcc
+  FOURCC_U422 = FOURCC('U', '4', '2', '2'),  // bt.2020, unofficial fourcc
+  FOURCC_U444 = FOURCC('U', '4', '4', '4'),  // bt.2020, unofficial fourcc
+  FOURCC_F010 = FOURCC('F', '0', '1', '0'),  // bt.709 full range 10 bit 420
+  FOURCC_H010 = FOURCC('H', '0', '1', '0'),  // bt.709 10 bit 420
+  FOURCC_U010 = FOURCC('U', '0', '1', '0'),  // bt.2020 10 bit 420
+  FOURCC_F210 = FOURCC('F', '2', '1', '0'),  // bt.709 full range 10 bit 422
+  FOURCC_H210 = FOURCC('H', '2', '1', '0'),  // bt.709 10 bit 422
+  FOURCC_U210 = FOURCC('U', '2', '1', '0'),  // bt.2020 10 bit 422
+  FOURCC_P010 = FOURCC('P', '0', '1', '0'),
+  FOURCC_P210 = FOURCC('P', '2', '1', '0'),
 
   // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
   FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
@@ -112,7 +133,13 @@ enum FourCC {
   FOURCC_L565 = FOURCC('L', '5', '6', '5'),  // Alias for RGBP.
   FOURCC_5551 = FOURCC('5', '5', '5', '1'),  // Alias for RGBO.
 
-  // 1 Auxiliary compressed YUV format set aside for capturer.
+  // deprecated formats.  Not supported, but defined for backward compatibility.
+  FOURCC_I411 = FOURCC('I', '4', '1', '1'),
+  FOURCC_Q420 = FOURCC('Q', '4', '2', '0'),
+  FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
+  FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
+  FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
+  FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
   FOURCC_H264 = FOURCC('H', '2', '6', '4'),
 
   // Match any fourcc.
@@ -130,14 +157,18 @@ enum FourCCBpp {
   FOURCC_BPP_NV12 = 12,
   FOURCC_BPP_YUY2 = 16,
   FOURCC_BPP_UYVY = 16,
-  FOURCC_BPP_M420 = 12,
+  FOURCC_BPP_M420 = 12,  // deprecated
   FOURCC_BPP_Q420 = 12,
   FOURCC_BPP_ARGB = 32,
   FOURCC_BPP_BGRA = 32,
   FOURCC_BPP_ABGR = 32,
   FOURCC_BPP_RGBA = 32,
+  FOURCC_BPP_AR30 = 32,
+  FOURCC_BPP_AB30 = 32,
+  FOURCC_BPP_AR64 = 64,
+  FOURCC_BPP_AB64 = 64,
   FOURCC_BPP_24BG = 24,
-  FOURCC_BPP_RAW  = 24,
+  FOURCC_BPP_RAW = 24,
   FOURCC_BPP_RGBP = 16,
   FOURCC_BPP_RGBO = 16,
   FOURCC_BPP_R444 = 16,
@@ -152,6 +183,13 @@ enum FourCCBpp {
   FOURCC_BPP_J420 = 12,
   FOURCC_BPP_J400 = 8,
   FOURCC_BPP_H420 = 12,
+  FOURCC_BPP_H422 = 16,
+  FOURCC_BPP_I010 = 15,
+  FOURCC_BPP_I210 = 20,
+  FOURCC_BPP_H010 = 15,
+  FOURCC_BPP_H210 = 20,
+  FOURCC_BPP_P010 = 15,
+  FOURCC_BPP_P210 = 20,
   FOURCC_BPP_MJPG = 0,  // 0 means unknown.
   FOURCC_BPP_H264 = 0,
   FOURCC_BPP_IYUV = 12,
@@ -170,15 +208,15 @@ enum FourCCBpp {
   FOURCC_BPP_CM24 = 24,
 
   // Match any fourcc.
-  FOURCC_BPP_ANY  = 0,  // 0 means unknown.
+  FOURCC_BPP_ANY = 0,  // 0 means unknown.
 };
 
 // Converts fourcc aliases into canonical ones.
-LIBYUV_API uint32 CanonicalFourCC(uint32 fourcc);
+LIBYUV_API uint32_t CanonicalFourCC(uint32_t fourcc);
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_VIDEO_COMMON_H_  NOLINT
+#endif  // INCLUDE_LIBYUV_VIDEO_COMMON_H_
diff --git a/media/libyuv/infra/config/OWNERS b/media/libyuv/infra/config/OWNERS
new file mode 100644
index 0000000000..2c4f90a03c
--- /dev/null
+++ b/media/libyuv/infra/config/OWNERS
@@ -0,0 +1,3 @@
+fbarchard@chromium.org
+mbonadei@chromium.org
+jansson@google.com
diff --git a/media/libyuv/infra/config/PRESUBMIT.py b/media/libyuv/infra/config/PRESUBMIT.py
new file mode 100644
index 0000000000..f79e08ad6a
--- /dev/null
+++ b/media/libyuv/infra/config/PRESUBMIT.py
@@ -0,0 +1,13 @@
+# Copyright 2018 The PDFium Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+
+USE_PYTHON3 = True
+
+
+def CheckChangeOnUpload(input_api, output_api):
+  return input_api.canned_checks.CheckChangedLUCIConfigs(input_api, output_api)
+
+
+def CheckChangeOnCommit(input_api, output_api):
+  return input_api.canned_checks.CheckChangedLUCIConfigs(input_api, output_api)
diff --git a/media/libyuv/infra/config/README.md b/media/libyuv/infra/config/README.md
new file mode 100644
index 0000000000..e5e3b5f818
--- /dev/null
+++ b/media/libyuv/infra/config/README.md
@@ -0,0 +1,2 @@
+This folder contains libyuv project-wide configurations
+for chrome-infra services.
diff --git a/media/libyuv/infra/config/codereview.settings b/media/libyuv/infra/config/codereview.settings
new file mode 100644
index 0000000000..6d74227357
--- /dev/null
+++ b/media/libyuv/infra/config/codereview.settings
@@ -0,0 +1,6 @@
+# This file is used by gcl and git-cl to get repository specific information.
+CODE_REVIEW_SERVER: codereview.chromium.org
+PROJECT: libyuv
+GERRIT_HOST: True
+VIEW_VC: https://chromium.googlesource.com/libyuv/libyuv/+/
+
diff --git a/media/libyuv/infra/config/commit-queue.cfg b/media/libyuv/infra/config/commit-queue.cfg
new file mode 100644
index 0000000000..4a8d77f41d
--- /dev/null
+++ b/media/libyuv/infra/config/commit-queue.cfg
@@ -0,0 +1,143 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see Config message:
+#   https://luci-config.appspot.com/schemas/projects:commit-queue.cfg
+
+cq_status_host: "chromium-cq-status.appspot.com"
+submit_options {
+  max_burst: 4
+  burst_delay {
+    seconds: 480
+  }
+}
+config_groups {
+  name: "config"
+  gerrit {
+    url: "https://chromium-review.googlesource.com"
+    projects {
+      name: "libyuv/libyuv"
+      ref_regexp: "refs/heads/infra/config"
+    }
+  }
+  verifiers {
+    gerrit_cq_ability {
+      committer_list: "project-libyuv-committers"
+      dry_run_access_list: "project-libyuv-tryjob-access"
+    }
+    tryjob {
+      builders {
+        name: "libyuv/try/presubmit"
+      }
+      retry_config {
+        single_quota: 1
+        global_quota: 2
+        failure_weight: 1
+        transient_failure_weight: 1
+        timeout_weight: 2
+      }
+    }
+  }
+}
+config_groups {
+  name: "master"
+  gerrit {
+    url: "https://chromium-review.googlesource.com"
+    projects {
+      name: "libyuv/libyuv"
+      ref_regexp: "refs/heads/main"
+      ref_regexp: "refs/heads/master"
+    }
+  }
+  verifiers {
+    gerrit_cq_ability {
+      committer_list: "project-libyuv-committers"
+      dry_run_access_list: "project-libyuv-tryjob-access"
+    }
+    tryjob {
+      builders {
+        name: "libyuv/try/android"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/android_arm64"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/android_rel"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/android_x64"
+      }
+      builders {
+        name: "libyuv/try/android_x86"
+      }
+      builders {
+        name: "libyuv/try/ios_arm64"
+      }
+      builders {
+        name: "libyuv/try/ios_arm64_rel"
+      }
+      builders {
+        name: "libyuv/try/linux"
+      }
+      builders {
+        name: "libyuv/try/linux_asan"
+      }
+      builders {
+        name: "libyuv/try/linux_gcc"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/linux_msan"
+      }
+      builders {
+        name: "libyuv/try/linux_rel"
+      }
+      builders {
+        name: "libyuv/try/linux_tsan2"
+      }
+      builders {
+        name: "libyuv/try/linux_ubsan"
+      }
+      builders {
+        name: "libyuv/try/linux_ubsan_vptr"
+      }
+      builders {
+        name: "libyuv/try/mac"
+      }
+      builders {
+        name: "libyuv/try/mac_asan"
+      }
+      builders {
+        name: "libyuv/try/mac_rel"
+      }
+      builders {
+        name: "libyuv/try/win"
+      }
+      builders {
+        name: "libyuv/try/win_clang"
+      }
+      builders {
+        name: "libyuv/try/win_clang_rel"
+      }
+      builders {
+        name: "libyuv/try/win_rel"
+      }
+      builders {
+        name: "libyuv/try/win_x64_clang_rel"
+      }
+      builders {
+        name: "libyuv/try/win_x64_rel"
+      }
+      retry_config {
+        single_quota: 1
+        global_quota: 2
+        failure_weight: 1
+        transient_failure_weight: 1
+        timeout_weight: 2
+      }
+    }
+  }
+}
diff --git a/media/libyuv/infra/config/cr-buildbucket.cfg b/media/libyuv/infra/config/cr-buildbucket.cfg
new file mode 100644
index 0000000000..50ea625a94
--- /dev/null
+++ b/media/libyuv/infra/config/cr-buildbucket.cfg
@@ -0,0 +1,1963 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see BuildbucketCfg message:
+#   https://luci-config.appspot.com/schemas/projects:buildbucket.cfg
+
+buckets {
+  name: "ci"
+  acls {
+    role: WRITER
+    group: "project-libyuv-admins"
+  }
+  acls {
+    group: "all"
+  }
+  swarming {
+    builders {
+      name: "Android ARM64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Tester ARM32 Debug (Nexus 5X)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Tester ARM32 Release (Nexus 5X)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Tester ARM64 Debug (Nexus 5X)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android32 x86 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android64 x64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux Asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux MSan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux Tsan v2"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux UBSan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux UBSan vptr"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux32 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux32 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Mac Asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Mac64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Mac64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Debug (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Release (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Debug (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Release (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "iOS ARM64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "iOS ARM64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-trusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+  }
+}
+buckets {
+  name: "cron"
+  acls {
+    role: WRITER
+    group: "project-libyuv-admins"
+  }
+  acls {
+    group: "all"
+  }
+  swarming {
+    builders {
+      name: "DEPS Autoroller"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Linux"
+      dimensions: "pool:luci.webrtc.cron"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "recipe": "libyuv/roll_deps"'
+        '}'
+      execution_timeout_secs: 7200
+      build_numbers: YES
+      service_account: "libyuv-ci-autoroll-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+  }
+}
+buckets {
+  name: "try"
+  acls {
+    role: WRITER
+    group: "project-libyuv-admins"
+  }
+  acls {
+    group: "all"
+  }
+  acls {
+    role: SCHEDULER
+    group: "project-libyuv-tryjob-access"
+  }
+  acls {
+    role: SCHEDULER
+    group: "service-account-cq"
+  }
+  swarming {
+    builders {
+      name: "android"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_arm64"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_x64"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_x86"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "ios_arm64"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "ios_arm64_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_gcc"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_msan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_tsan2"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_ubsan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_ubsan_vptr"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "mac"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "mac_asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "mac_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "presubmit"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "run_presubmit",'
+        '  "repo_name": "libyuv",'
+        '  "runhooks": true'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_clang"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_clang_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_x64_clang_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_x64_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "$build/reclient": {'
+        '    "instance": "rbe-webrtc-untrusted",'
+        '    "metrics_project": "chromium-reclient-metrics"'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+  }
+}
diff --git a/media/libyuv/infra/config/luci-logdog.cfg b/media/libyuv/infra/config/luci-logdog.cfg
new file mode 100644
index 0000000000..adc75bef49
--- /dev/null
+++ b/media/libyuv/infra/config/luci-logdog.cfg
@@ -0,0 +1,9 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see ProjectConfig message:
+#   https://luci-config.appspot.com/schemas/projects:luci-logdog.cfg
+
+reader_auth_groups: "all"
+writer_auth_groups: "luci-logdog-chromium-writers"
+archive_gs_bucket: "chromium-luci-logdog"
diff --git a/media/libyuv/infra/config/luci-milo.cfg b/media/libyuv/infra/config/luci-milo.cfg
new file mode 100644
index 0000000000..baf786f21e
--- /dev/null
+++ b/media/libyuv/infra/config/luci-milo.cfg
@@ -0,0 +1,246 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see Project message:
+#   https://luci-config.appspot.com/schemas/projects:luci-milo.cfg
+
+consoles {
+  id: "main"
+  name: "libyuv Main Console"
+  repo_url: "https://chromium.googlesource.com/libyuv/libyuv"
+  refs: "regexp:refs/heads/main"
+  manifest_name: "REVISION"
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android ARM64 Debug"
+    category: "Android|Builder"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Debug"
+    category: "Android|Builder"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Release"
+    category: "Android|Builder"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android32 x86 Debug"
+    category: "Android|Builder|x86"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android64 x64 Debug"
+    category: "Android|Builder|x64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Tester ARM32 Debug (Nexus 5X)"
+    category: "Android|Tester|ARM 32"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Tester ARM32 Release (Nexus 5X)"
+    category: "Android|Tester|ARM 32"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Tester ARM64 Debug (Nexus 5X)"
+    category: "Android|Tester|ARM 64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux Asan"
+    category: "Linux"
+    short_name: "asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux MSan"
+    category: "Linux"
+    short_name: "msan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux Tsan v2"
+    category: "Linux"
+    short_name: "tsan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux UBSan"
+    category: "Linux|UBSan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux UBSan vptr"
+    category: "Linux|UBSan"
+    short_name: "vptr"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux32 Debug"
+    category: "Linux|32"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux32 Release"
+    category: "Linux|32"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux64 Debug"
+    category: "Linux|64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux64 Release"
+    category: "Linux|64"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Mac Asan"
+    category: "Mac"
+    short_name: "asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Mac64 Debug"
+    category: "Mac"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Mac64 Release"
+    category: "Mac"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Debug"
+    category: "Win|32|Debug"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Debug (Clang)"
+    category: "Win|32|Debug"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Release"
+    category: "Win|32|Release"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Release (Clang)"
+    category: "Win|32|Release"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Debug"
+    category: "Win|64|Debug"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Debug (Clang)"
+    category: "Win|64|Debug"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Release"
+    category: "Win|64|Release"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Release (Clang)"
+    category: "Win|64|Release"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/iOS ARM64 Debug"
+    category: "iOS|ARM64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/iOS ARM64 Release"
+    category: "iOS|ARM64"
+    short_name: "rel"
+  }
+  include_experimental_builds: true
+}
+consoles {
+  id: "cron"
+  name: "Cron"
+  builders {
+    name: "buildbucket/luci.libyuv.cron/DEPS Autoroller"
+  }
+  builder_view_only: true
+}
+consoles {
+  id: "try"
+  name: "libyuv Try Builders"
+  builders {
+    name: "buildbucket/luci.libyuv.try/android"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_arm64"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_x64"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_x86"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/ios_arm64"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/ios_arm64_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_gcc"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_msan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_tsan2"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_ubsan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_ubsan_vptr"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/mac"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/mac_asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/mac_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_clang"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_clang_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_x64_clang_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_x64_rel"
+  }
+  builder_view_only: true
+}
+logo_url: "https://storage.googleapis.com/chrome-infra-public/logo/libyuv-logo.png"
diff --git a/media/libyuv/infra/config/luci-scheduler.cfg b/media/libyuv/infra/config/luci-scheduler.cfg
new file mode 100644
index 0000000000..0ec5dd0e52
--- /dev/null
+++ b/media/libyuv/infra/config/luci-scheduler.cfg
@@ -0,0 +1,385 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see ProjectConfig message:
+#   https://luci-config.appspot.com/schemas/projects:luci-scheduler.cfg
+
+job {
+  id: "Android ARM64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android ARM64 Debug"
+  }
+}
+job {
+  id: "Android Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Debug"
+  }
+}
+job {
+  id: "Android Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Release"
+  }
+}
+job {
+  id: "Android Tester ARM32 Debug (Nexus 5X)"
+  realm: "ci"
+  acls {
+    role: TRIGGERER
+    granted_to: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Tester ARM32 Debug (Nexus 5X)"
+  }
+}
+job {
+  id: "Android Tester ARM32 Release (Nexus 5X)"
+  realm: "ci"
+  acls {
+    role: TRIGGERER
+    granted_to: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Tester ARM32 Release (Nexus 5X)"
+  }
+}
+job {
+  id: "Android Tester ARM64 Debug (Nexus 5X)"
+  realm: "ci"
+  acls {
+    role: TRIGGERER
+    granted_to: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Tester ARM64 Debug (Nexus 5X)"
+  }
+}
+job {
+  id: "Android32 x86 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android32 x86 Debug"
+  }
+}
+job {
+  id: "Android64 x64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android64 x64 Debug"
+  }
+}
+job {
+  id: "DEPS Autoroller"
+  realm: "cron"
+  schedule: "0 14 * * *"
+  acl_sets: "cron"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "cron"
+    builder: "DEPS Autoroller"
+  }
+}
+job {
+  id: "Linux Asan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux Asan"
+  }
+}
+job {
+  id: "Linux MSan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux MSan"
+  }
+}
+job {
+  id: "Linux Tsan v2"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux Tsan v2"
+  }
+}
+job {
+  id: "Linux UBSan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux UBSan"
+  }
+}
+job {
+  id: "Linux UBSan vptr"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux UBSan vptr"
+  }
+}
+job {
+  id: "Linux32 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux32 Debug"
+  }
+}
+job {
+  id: "Linux32 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux32 Release"
+  }
+}
+job {
+  id: "Linux64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux64 Debug"
+  }
+}
+job {
+  id: "Linux64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux64 Release"
+  }
+}
+job {
+  id: "Mac Asan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Mac Asan"
+  }
+}
+job {
+  id: "Mac64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Mac64 Debug"
+  }
+}
+job {
+  id: "Mac64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Mac64 Release"
+  }
+}
+job {
+  id: "Win32 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Debug"
+  }
+}
+job {
+  id: "Win32 Debug (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Debug (Clang)"
+  }
+}
+job {
+  id: "Win32 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Release"
+  }
+}
+job {
+  id: "Win32 Release (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Release (Clang)"
+  }
+}
+job {
+  id: "Win64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Debug"
+  }
+}
+job {
+  id: "Win64 Debug (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Debug (Clang)"
+  }
+}
+job {
+  id: "Win64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Release"
+  }
+}
+job {
+  id: "Win64 Release (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Release (Clang)"
+  }
+}
+job {
+  id: "iOS ARM64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "iOS ARM64 Debug"
+  }
+}
+job {
+  id: "iOS ARM64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "iOS ARM64 Release"
+  }
+}
+trigger {
+  id: "master-gitiles-trigger"
+  realm: "ci"
+  acl_sets: "ci"
+  triggers: "Android ARM64 Debug"
+  triggers: "Android Debug"
+  triggers: "Android Release"
+  triggers: "Android32 x86 Debug"
+  triggers: "Android64 x64 Debug"
+  triggers: "Linux Asan"
+  triggers: "Linux MSan"
+  triggers: "Linux Tsan v2"
+  triggers: "Linux UBSan"
+  triggers: "Linux UBSan vptr"
+  triggers: "Linux32 Debug"
+  triggers: "Linux32 Release"
+  triggers: "Linux64 Debug"
+  triggers: "Linux64 Release"
+  triggers: "Mac Asan"
+  triggers: "Mac64 Debug"
+  triggers: "Mac64 Release"
+  triggers: "Win32 Debug"
+  triggers: "Win32 Debug (Clang)"
+  triggers: "Win32 Release"
+  triggers: "Win32 Release (Clang)"
+  triggers: "Win64 Debug"
+  triggers: "Win64 Debug (Clang)"
+  triggers: "Win64 Release"
+  triggers: "Win64 Release (Clang)"
+  triggers: "iOS ARM64 Debug"
+  triggers: "iOS ARM64 Release"
+  gitiles {
+    repo: "https://chromium.googlesource.com/libyuv/libyuv"
+    refs: "regexp:refs/heads/main"
+  }
+}
+acl_sets {
+  name: "ci"
+  acls {
+    role: OWNER
+    granted_to: "group:project-libyuv-admins"
+  }
+  acls {
+    granted_to: "group:all"
+  }
+}
+acl_sets {
+  name: "cron"
+  acls {
+    role: OWNER
+    granted_to: "group:project-libyuv-admins"
+  }
+  acls {
+    granted_to: "group:all"
+  }
+}
diff --git a/media/libyuv/infra/config/main.star b/media/libyuv/infra/config/main.star
new file mode 100755
index 0000000000..b722b114b4
--- /dev/null
+++ b/media/libyuv/infra/config/main.star
@@ -0,0 +1,377 @@
+#!/usr/bin/env lucicfg
+# https://chromium.googlesource.com/infra/luci/luci-go/+/master/lucicfg/doc/
+
+"""LUCI project configuration for libyuv CQ and CI."""
+
+lucicfg.check_version("1.30.9")
+
+LIBYUV_GIT = "https://chromium.googlesource.com/libyuv/libyuv"
+LIBYUV_GERRIT = "https://chromium-review.googlesource.com/libyuv/libyuv"
+
+GOMA_BACKEND_RBE_PROD = {
+    "server_host": "goma.chromium.org",
+    "use_luci_auth": True,
+}
+
+GOMA_BACKEND_RBE_ATS_PROD = {
+    "server_host": "goma.chromium.org",
+    "use_luci_auth": True,
+    "enable_ats": True,
+}
+
+# Disable ATS on Windows CQ/try.
+GOMA_BACKEND_RBE_NO_ATS_PROD = {
+    "server_host": "goma.chromium.org",
+    "use_luci_auth": True,
+    "enable_ats": False,
+}
+
+RECLIENT_CI = {
+    "instance": "rbe-webrtc-trusted",
+    "metrics_project": "chromium-reclient-metrics",
+}
+
+RECLIENT_CQ = {
+    "instance": "rbe-webrtc-untrusted",
+    "metrics_project": "chromium-reclient-metrics",
+}
+
+# Use LUCI Scheduler BBv2 names and add Scheduler realms configs.
+lucicfg.enable_experiment("crbug.com/1182002")
+
+luci.builder.defaults.experiments.set(
+    {
+        "luci.recipes.use_python3": 100,
+    },
+)
+
+lucicfg.config(
+    lint_checks = ["default"],
+    config_dir = ".",
+    tracked_files = [
+        "commit-queue.cfg",
+        "cr-buildbucket.cfg",
+        "luci-logdog.cfg",
+        "luci-milo.cfg",
+        "luci-scheduler.cfg",
+        "project.cfg",
+        "realms.cfg",
+    ],
+)
+
+# Generates project.cfg
+
+luci.project(
+    name = "libyuv",
+    buildbucket = "cr-buildbucket.appspot.com",
+    logdog = "luci-logdog.appspot.com",
+    milo = "luci-milo.appspot.com",
+    notify = "luci-notify.appspot.com",
+    scheduler = "luci-scheduler.appspot.com",
+    swarming = "chromium-swarm.appspot.com",
+    acls = [
+        acl.entry(acl.PROJECT_CONFIGS_READER, groups = ["all"]),
+        acl.entry(acl.LOGDOG_READER, groups = ["all"]),
+        acl.entry(acl.LOGDOG_WRITER, groups = ["luci-logdog-chromium-writers"]),
+        acl.entry(acl.SCHEDULER_READER, groups = ["all"]),
+        acl.entry(acl.SCHEDULER_OWNER, groups = ["project-libyuv-admins"]),
+        acl.entry(acl.BUILDBUCKET_READER, groups = ["all"]),
+        acl.entry(acl.BUILDBUCKET_OWNER, groups = ["project-libyuv-admins"]),
+    ],
+    bindings = [
+        luci.binding(
+            roles = "role/swarming.taskTriggerer", # for LED tasks.
+            groups = "project-libyuv-admins",
+        ),
+        luci.binding(
+            roles = "role/configs.validator",
+            users = "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com",
+        ),
+    ],
+)
+
+# Generates luci-logdog.cfg
+
+luci.logdog(
+    gs_bucket = "chromium-luci-logdog",
+)
+
+# Generates luci-scheduler.cfg
+
+luci.gitiles_poller(
+    name = "master-gitiles-trigger",
+    bucket = "ci",
+    repo = LIBYUV_GIT,
+)
+
+# Generates luci-milo.cfg
+
+luci.milo(
+    logo = "https://storage.googleapis.com/chrome-infra-public/logo/libyuv-logo.png",
+)
+
+def libyuv_ci_view(name, category, short_name):
+    return luci.console_view_entry(
+        console_view = "main",
+        builder = name,
+        category = category,
+        short_name = short_name,
+    )
+
+def libyuv_try_view(name):
+    return luci.list_view_entry(
+        list_view = "try",
+        builder = name,
+    )
+
+luci.console_view(
+    name = "main",
+    title = "libyuv Main Console",
+    include_experimental_builds = True,
+    repo = LIBYUV_GIT,
+)
+
+luci.list_view(
+    name = "cron",
+    title = "Cron",
+    entries = ["DEPS Autoroller"],
+)
+
+luci.list_view(
+    name = "try",
+    title = "libyuv Try Builders",
+)
+
+# Generates commit-queue.cfg
+
+def libyuv_try_job_verifier(name, cq_group, experiment_percentage):
+    return luci.cq_tryjob_verifier(
+        builder = name,
+        cq_group = cq_group,
+        experiment_percentage = experiment_percentage,
+    )
+
+luci.cq(
+    status_host = "chromium-cq-status.appspot.com",
+    submit_max_burst = 4,
+    submit_burst_delay = 8 * time.minute,
+)
+
+luci.cq_group(
+    name = "master",
+    watch = [
+        cq.refset(
+            repo = LIBYUV_GERRIT,
+            refs = ["refs/heads/main", "refs/heads/master"],
+        ),
+    ],
+    acls = [
+        acl.entry(acl.CQ_COMMITTER, groups = ["project-libyuv-committers"]),
+        acl.entry(acl.CQ_DRY_RUNNER, groups = ["project-libyuv-tryjob-access"]),
+    ],
+    retry_config = cq.RETRY_ALL_FAILURES,
+    cancel_stale_tryjobs = True,
+)
+
+luci.cq_group(
+    name = "config",
+    watch = [
+        cq.refset(
+            repo = LIBYUV_GERRIT,
+            refs = ["refs/heads/infra/config"],
+        ),
+    ],
+    acls = [
+        acl.entry(acl.CQ_COMMITTER, groups = ["project-libyuv-committers"]),
+        acl.entry(acl.CQ_DRY_RUNNER, groups = ["project-libyuv-tryjob-access"]),
+    ],
+    retry_config = cq.RETRY_ALL_FAILURES,
+    cancel_stale_tryjobs = True,
+)
+
+# Generates cr-buildbucket.cfg
+
+luci.bucket(
+    name = "ci",
+)
+luci.bucket(
+    name = "try",
+    acls = [
+        acl.entry(acl.BUILDBUCKET_TRIGGERER, groups = [
+            "project-libyuv-tryjob-access",
+            "service-account-cq",
+        ]),
+    ],
+)
+luci.bucket(
+    name = "cron",
+)
+
+def get_os_dimensions(os):
+    if os == "android":
+        return {"device_type": "bullhead"}
+    if os == "ios" or os == "mac":
+        return {"os": "Mac-10.15", "cpu": "x86-64"}
+    elif os == "win":
+        return {"os": "Windows-10", "cores": "8", "cpu": "x86-64"}
+    elif os == "linux":
+        return {"os": "Ubuntu-18.04", "cores": "8", "cpu": "x86-64"}
+    return {}
+
+def get_os_properties(os, try_builder = False):
+    if os == "android":
+        return {"$build/goma": GOMA_BACKEND_RBE_PROD}
+    elif os in ("ios", "mac"):
+        return {"$build/goma": GOMA_BACKEND_RBE_PROD}
+    elif os == "win" and try_builder:
+        return {"$build/goma": GOMA_BACKEND_RBE_NO_ATS_PROD}
+    elif os == "win":
+        return {"$build/goma": GOMA_BACKEND_RBE_ATS_PROD}
+    elif os == "linux":
+        return {"$build/goma": GOMA_BACKEND_RBE_ATS_PROD}
+    return {}
+
+def libyuv_ci_builder(name, dimensions, properties, triggered_by):
+    return luci.builder(
+        name = name,
+        dimensions = dimensions,
+        properties = properties,
+        bucket = "ci",
+        service_account = "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com",
+        triggered_by = triggered_by,
+        swarming_tags = ["vpython:native-python-wrapper"],
+        execution_timeout = 180 * time.minute,
+        build_numbers = True,
+        executable = luci.recipe(
+            name = "libyuv/libyuv",
+            cipd_package = "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build",
+            use_python3 = True,
+        ),
+    )
+
+def libyuv_try_builder(name, dimensions, properties, recipe_name = "libyuv/libyuv"):
+    return luci.builder(
+        name = name,
+        dimensions = dimensions,
+        properties = properties,
+        bucket = "try",
+        service_account = "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com",
+        swarming_tags = ["vpython:native-python-wrapper"],
+        execution_timeout = 180 * time.minute,
+        build_numbers = True,
+        executable = luci.recipe(
+            name = recipe_name,
+            cipd_package = "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build",
+            use_python3 = True,
+        ),
+    )
+
+def ci_builder(name, os, category, short_name = None):
+    dimensions = get_os_dimensions(os)
+    properties = get_os_properties(os)
+    properties["$build/reclient"] = RECLIENT_CI
+
+    dimensions["pool"] = "luci.flex.ci"
+    properties["builder_group"] = "client.libyuv"
+
+    triggered_by = ["master-gitiles-trigger" if os != "android" else "Android Debug"]
+    libyuv_ci_view(name, category, short_name)
+    return libyuv_ci_builder(name, dimensions, properties, triggered_by)
+
+def try_builder(name, os, experiment_percentage = None):
+    dimensions = get_os_dimensions(os)
+    properties = get_os_properties(os, try_builder = True)
+    properties["$build/reclient"] = RECLIENT_CQ
+
+    dimensions["pool"] = "luci.flex.try"
+    properties["builder_group"] = "tryserver.libyuv"
+
+    if name == "presubmit":
+        recipe_name = "run_presubmit"
+        properties["repo_name"] = "libyuv"
+        properties["runhooks"] = True
+        libyuv_try_job_verifier(name, "config", experiment_percentage)
+        return libyuv_try_builder(name, dimensions, properties, recipe_name)
+
+    libyuv_try_job_verifier(name, "master", experiment_percentage)
+    libyuv_try_view(name)
+    return libyuv_try_builder(name, dimensions, properties)
+
+luci.builder(
+    name = "DEPS Autoroller",
+    bucket = "cron",
+    service_account = "libyuv-ci-autoroll-builder@chops-service-accounts.iam.gserviceaccount.com",
+    dimensions = {
+        "pool": "luci.webrtc.cron",
+        "os": "Linux",
+        "cpu": "x86-64",
+    },
+    swarming_tags = ["vpython:native-python-wrapper"],
+    execution_timeout = 120 * time.minute,
+    build_numbers = True,
+    schedule = "0 14 * * *",  # Every 2 hours.
+    executable = luci.recipe(
+        name = "libyuv/roll_deps",
+        cipd_package = "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build",
+        use_python3 = True,
+    ),
+)
+
+ci_builder("Android ARM64 Debug", "linux", "Android|Builder", "dbg")
+ci_builder("Android Debug", "linux", "Android|Builder", "dbg")
+ci_builder("Android Release", "linux", "Android|Builder", "rel")
+ci_builder("Android32 x86 Debug", "linux", "Android|Builder|x86", "dbg")
+ci_builder("Android64 x64 Debug", "linux", "Android|Builder|x64", "dbg")
+ci_builder("Android Tester ARM32 Debug (Nexus 5X)", "android", "Android|Tester|ARM 32", "dbg")
+ci_builder("Android Tester ARM32 Release (Nexus 5X)", "android", "Android|Tester|ARM 32", "rel")
+ci_builder("Android Tester ARM64 Debug (Nexus 5X)", "android", "Android|Tester|ARM 64", "dbg")
+ci_builder("Linux Asan", "linux", "Linux", "asan")
+ci_builder("Linux MSan", "linux", "Linux", "msan")
+ci_builder("Linux Tsan v2", "linux", "Linux", "tsan")
+ci_builder("Linux UBSan", "linux", "Linux|UBSan")
+ci_builder("Linux UBSan vptr", "linux", "Linux|UBSan", "vptr")
+ci_builder("Linux32 Debug", "linux", "Linux|32", "dbg")
+ci_builder("Linux32 Release", "linux", "Linux|32", "rel")
+ci_builder("Linux64 Debug", "linux", "Linux|64", "dbg")
+ci_builder("Linux64 Release", "linux", "Linux|64", "rel")
+ci_builder("Mac Asan", "mac", "Mac", "asan")
+ci_builder("Mac64 Debug", "mac", "Mac", "dbg")
+ci_builder("Mac64 Release", "mac", "Mac", "rel")
+ci_builder("Win32 Debug", "win", "Win|32|Debug")
+ci_builder("Win32 Debug (Clang)", "win", "Win|32|Debug", "clg")
+ci_builder("Win32 Release", "win", "Win|32|Release")
+ci_builder("Win32 Release (Clang)", "win", "Win|32|Release", "clg")
+ci_builder("Win64 Debug", "win", "Win|64|Debug", "clg")
+ci_builder("Win64 Debug (Clang)", "win", "Win|64|Debug", "clg")
+ci_builder("Win64 Release", "win", "Win|64|Release")
+ci_builder("Win64 Release (Clang)", "win", "Win|64|Release", "clg")
+ci_builder("iOS ARM64 Debug", "ios", "iOS|ARM64", "dbg")
+ci_builder("iOS ARM64 Release", "ios", "iOS|ARM64", "rel")
+
+# TODO(crbug.com/1242847): make this not experimental.
+try_builder("android", "android", experiment_percentage = 100)
+try_builder("android_arm64", "android", experiment_percentage = 100)
+try_builder("android_rel", "android", experiment_percentage = 100)
+
+try_builder("android_x64", "linux")
+try_builder("android_x86", "linux")
+try_builder("ios_arm64", "ios")
+try_builder("ios_arm64_rel", "ios")
+try_builder("linux", "linux")
+try_builder("linux_asan", "linux")
+try_builder("linux_gcc", "linux", experiment_percentage = 100)
+try_builder("linux_msan", "linux")
+try_builder("linux_rel", "linux")
+try_builder("linux_tsan2", "linux")
+try_builder("linux_ubsan", "linux")
+try_builder("linux_ubsan_vptr", "linux")
+try_builder("mac", "mac")
+try_builder("mac_asan", "mac")
+try_builder("mac_rel", "mac")
+try_builder("win", "win")
+try_builder("win_clang", "win")
+try_builder("win_clang_rel", "win")
+try_builder("win_rel", "win")
+try_builder("win_x64_clang_rel", "win")
+try_builder("win_x64_rel", "win")
+try_builder("presubmit", "linux")
diff --git a/media/libyuv/infra/config/project.cfg b/media/libyuv/infra/config/project.cfg
new file mode 100644
index 0000000000..52797c1c8a
--- /dev/null
+++ b/media/libyuv/infra/config/project.cfg
@@ -0,0 +1,15 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see ProjectCfg message:
+#   https://luci-config.appspot.com/schemas/projects:project.cfg
+
+name: "libyuv"
+access: "group:all"
+lucicfg {
+  version: "1.32.1"
+  package_dir: "."
+  config_dir: "."
+  entry_point: "main.star"
+  experiments: "crbug.com/1182002"
+}
diff --git a/media/libyuv/infra/config/realms.cfg b/media/libyuv/infra/config/realms.cfg
new file mode 100644
index 0000000000..16ffaac90f
--- /dev/null
+++ b/media/libyuv/infra/config/realms.cfg
@@ -0,0 +1,83 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see RealmsCfg message:
+#   https://luci-config.appspot.com/schemas/projects:realms.cfg
+
+realms {
+  name: "@root"
+  bindings {
+    role: "role/buildbucket.owner"
+    principals: "group:project-libyuv-admins"
+  }
+  bindings {
+    role: "role/buildbucket.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/configs.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/configs.validator"
+    principals: "user:libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  bindings {
+    role: "role/logdog.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/logdog.writer"
+    principals: "group:luci-logdog-chromium-writers"
+  }
+  bindings {
+    role: "role/scheduler.owner"
+    principals: "group:project-libyuv-admins"
+  }
+  bindings {
+    role: "role/scheduler.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/swarming.taskTriggerer"
+    principals: "group:project-libyuv-admins"
+  }
+}
+realms {
+  name: "ci"
+  bindings {
+    role: "role/buildbucket.builderServiceAccount"
+    principals: "user:libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  bindings {
+    role: "role/scheduler.triggerer"
+    principals: "user:libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+    conditions {
+      restrict {
+        attribute: "scheduler.job.name"
+        values: "Android Tester ARM32 Debug (Nexus 5X)"
+        values: "Android Tester ARM32 Release (Nexus 5X)"
+        values: "Android Tester ARM64 Debug (Nexus 5X)"
+      }
+    }
+  }
+}
+realms {
+  name: "cron"
+  bindings {
+    role: "role/buildbucket.builderServiceAccount"
+    principals: "user:libyuv-ci-autoroll-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+}
+realms {
+  name: "try"
+  bindings {
+    role: "role/buildbucket.builderServiceAccount"
+    principals: "user:libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  bindings {
+    role: "role/buildbucket.triggerer"
+    principals: "group:project-libyuv-tryjob-access"
+    principals: "group:service-account-cq"
+  }
+}
diff --git a/media/libyuv/libyuv.gni b/media/libyuv/libyuv.gni
new file mode 100644
index 0000000000..852f08ca9d
--- /dev/null
+++ b/media/libyuv/libyuv.gni
@@ -0,0 +1,23 @@
+# Copyright 2016 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+import("//build/config/arm.gni")
+import("//build/config/mips.gni")
+import("//build_overrides/build.gni")
+
+declare_args() {
+  libyuv_include_tests = !build_with_chromium
+  libyuv_disable_jpeg = false
+  libyuv_use_neon =
+      current_cpu == "arm64" ||
+      (current_cpu == "arm" && (arm_use_neon || arm_optionally_use_neon))
+  libyuv_use_msa =
+      (current_cpu == "mips64el" || current_cpu == "mipsel") && mips_use_msa
+  libyuv_use_mmi =
+      (current_cpu == "mips64el" || current_cpu == "mipsel") && mips_use_mmi
+}
diff --git a/media/libyuv/libyuv.gyp b/media/libyuv/libyuv.gyp
index 0f43f6c3f4..f73a1a4b74 100644
--- a/media/libyuv/libyuv.gyp
+++ b/media/libyuv/libyuv.gyp
@@ -18,23 +18,28 @@
   },
   'variables': {
     'use_system_libjpeg%': 0,
-    'libyuv_disable_jpeg%': 0,
+    # Can be enabled if your jpeg has GYP support.
+    'libyuv_disable_jpeg%': 1,
     # 'chromium_code' treats libyuv as internal and increases warning level.
     'chromium_code': 1,
     # clang compiler default variable usable by other apps that include libyuv.
     'clang%': 0,
     # Link-Time Optimizations.
     'use_lto%': 0,
-    'yuv_disable_asm%': 0,
-    'yuv_disable_avx2%': 0,
+    'mips_msa%': 0,  # Default to msa off.
     'build_neon': 0,
+    'build_msa': 0,
     'conditions': [
        ['(target_arch == "armv7" or target_arch == "armv7s" or \
        (target_arch == "arm" and arm_version >= 7) or target_arch == "arm64")\
-       and (arm_neon == 1 or arm_neon_optional == 1)',
-       {
+       and (arm_neon == 1 or arm_neon_optional == 1)', {
          'build_neon': 1,
        }],
+       ['(target_arch == "mipsel" or target_arch == "mips64el")\
+       and (mips_msa == 1)',
+       {
+         'build_msa': 1,
+       }],
     ],
   },
 
@@ -47,7 +52,7 @@
         'optimize': 'max',  # enable O2 and ltcg.
       },
       # Allows libyuv.a redistributable library without external dependencies.
-      # 'standalone_static_library': 1,
+      'standalone_static_library': 1,
       'conditions': [
        # Disable -Wunused-parameter
         ['clang == 1', {
@@ -65,11 +70,6 @@
             '-mfpu=vfpv3-d16',
             # '-mthumb',  # arm32 not thumb
           ],
-          'cflags_mozilla!': [
-            '-mfpu=vfp',
-            '-mfpu=vfpv3',
-            '-mfpu=vfpv3-d16',
-          ],
           'conditions': [
             # Disable LTO in libyuv_neon target due to gcc 4.9 compiler bug.
             ['clang == 0 and use_lto == 1', {
@@ -84,34 +84,15 @@
                 '-mfpu=neon',
                 # '-marm',  # arm32 not thumb
               ],
-              'cflags_mozilla': [
-                '-mfpu=neon',
-              ],
             }],
           ],
         }],
-        [ 'yuv_disable_asm != 0', {
-          'defines': [
-            # Enable the following 3 macros to turn off assembly for specified CPU.
-            'LIBYUV_DISABLE_X86',
-            'LIBYUV_DISABLE_NEON',
-            'LIBYUV_DISABLE_MIPS',
-          ],
-        }],
-        [ 'yuv_disable_avx2 == 1', {
+        ['build_msa != 0', {
           'defines': [
-            'LIBYUV_DISABLE_AVX2',
-          ]
-        }],
-        ['build_with_mozilla == 1', {
-          'defines': [
-            'HAVE_JPEG'
-          ],
-          'cflags_mozilla': [
-            '$(MOZ_JPEG_CFLAGS)',
+            'LIBYUV_MSA',
           ],
         }],
-        ['OS != "ios" and libyuv_disable_jpeg != 1 and build_with_mozilla != 1', {
+        ['OS != "ios" and libyuv_disable_jpeg != 1', {
           'defines': [
             'HAVE_JPEG'
           ],
@@ -140,9 +121,7 @@
         # Enable the following 3 macros to turn off assembly for specified CPU.
         # 'LIBYUV_DISABLE_X86',
         # 'LIBYUV_DISABLE_NEON',
-        # 'LIBYUV_DISABLE_MIPS',
-        # This disables AVX2 (Haswell) support, overriding compiler checks
-        # 'LIBYUV_DISABLE_AVX2',
+        # 'LIBYUV_DISABLE_DSPR2',
         # Enable the following macro to build libyuv as a shared library (dll).
         # 'LIBYUV_USING_SHARED_LIBRARY',
         # TODO(fbarchard): Make these into gyp defines.
@@ -180,3 +159,4 @@
 # tab-width:2
 # indent-tabs-mode:nil
 # End:
+# vim: set expandtab tabstop=2 shiftwidth=2:
diff --git a/media/libyuv/libyuv.gypi b/media/libyuv/libyuv.gypi
index 73fdec0a9f..48936aa7b0 100644
--- a/media/libyuv/libyuv.gypi
+++ b/media/libyuv/libyuv.gypi
@@ -18,6 +18,7 @@
       'include/libyuv/convert_from.h',
       'include/libyuv/convert_from_argb.h',
       'include/libyuv/cpu_id.h',
+      'include/libyuv/macros_msa.h',
       'include/libyuv/mjpeg_decoder.h',
       'include/libyuv/planar_functions.h',
       'include/libyuv/rotate.h',
@@ -26,7 +27,9 @@
       'include/libyuv/row.h',
       'include/libyuv/scale.h',
       'include/libyuv/scale_argb.h',
+      'include/libyuv/scale_rgb.h',
       'include/libyuv/scale_row.h',
+      'include/libyuv/scale_uv.h',
       'include/libyuv/version.h',
       'include/libyuv/video_common.h',
 
@@ -34,6 +37,7 @@
       'source/compare.cc',
       'source/compare_common.cc',
       'source/compare_gcc.cc',
+      'source/compare_msa.cc',
       'source/compare_neon.cc',
       'source/compare_neon64.cc',
       'source/compare_win.cc',
@@ -53,14 +57,14 @@
       'source/rotate_argb.cc',
       'source/rotate_common.cc',
       'source/rotate_gcc.cc',
-      'source/rotate_mips.cc',
+      'source/rotate_msa.cc',
       'source/rotate_neon.cc',
       'source/rotate_neon64.cc',
       'source/rotate_win.cc',
       'source/row_any.cc',
       'source/row_common.cc',
       'source/row_gcc.cc',
-      'source/row_mips.cc',
+      'source/row_msa.cc',
       'source/row_neon.cc',
       'source/row_neon64.cc',
       'source/row_win.cc',
@@ -69,9 +73,11 @@
       'source/scale_argb.cc',
       'source/scale_common.cc',
       'source/scale_gcc.cc',
-      'source/scale_mips.cc',
+      'source/scale_msa.cc',
       'source/scale_neon.cc',
       'source/scale_neon64.cc',
+      'source/scale_rgb.cc',
+      'source/scale_uv.cc',
       'source/scale_win.cc',
       'source/video_common.cc',
     ],
diff --git a/media/libyuv/libyuv_nacl.gyp b/media/libyuv/libyuv_nacl.gyp
deleted file mode 100644
index b8fe57ee3e..0000000000
--- a/media/libyuv/libyuv_nacl.gyp
+++ /dev/null
@@ -1,37 +0,0 @@
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-{
-  'includes': [
-    'libyuv.gypi',
-    '../../native_client/build/untrusted.gypi',
-  ],
-  'targets': [
-    {
-      'target_name': 'libyuv_nacl',
-      'type': 'none',
-      'variables': {
-        'nlib_target': 'libyuv_nacl.a',
-        'build_glibc': 0,
-        'build_newlib': 0,
-        'build_pnacl_newlib': 1,
-      },
-      'include_dirs': [
-        'include',
-      ],
-      'direct_dependent_settings': {
-        'include_dirs': [
-          'include',
-        ],
-      },
-      'sources': [
-        '<@(libyuv_sources)',
-      ],
-    },  # target libyuv_nacl
-  ]
-}
diff --git a/media/libyuv/libyuv_test.gyp b/media/libyuv/libyuv_test.gyp
deleted file mode 100644
index a060cc64af..0000000000
--- a/media/libyuv/libyuv_test.gyp
+++ /dev/null
@@ -1,229 +0,0 @@
-# Copyright 2011 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-{
-  'variables': {
-    'libyuv_disable_jpeg%': 0,
-  },
-  'targets': [
-    {
-      'target_name': 'libyuv_unittest',
-      'type': '<(gtest_target_type)',
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-        'testing/gtest.gyp:gtest',
-        'third_party/gflags/gflags.gyp:gflags',
-      ],
-      'direct_dependent_settings': {
-        'defines': [
-          'GTEST_RELATIVE_PATH',
-        ],
-      },
-      'export_dependent_settings': [
-        '<(DEPTH)/testing/gtest.gyp:gtest',
-      ],
-      'sources': [
-        # headers
-        'unit_test/unit_test.h',
-
-        # sources
-        'unit_test/basictypes_test.cc',
-        'unit_test/compare_test.cc',
-        'unit_test/color_test.cc',
-        'unit_test/convert_test.cc',
-        'unit_test/cpu_test.cc',
-        'unit_test/math_test.cc',
-        'unit_test/planar_test.cc',
-        'unit_test/rotate_argb_test.cc',
-        'unit_test/rotate_test.cc',
-        'unit_test/scale_argb_test.cc',
-        'unit_test/scale_test.cc',
-        'unit_test/unit_test.cc',
-        'unit_test/video_common_test.cc',
-      ],
-      'conditions': [
-        ['OS=="linux"', {
-          'cflags': [
-            '-fexceptions',
-          ],
-        }],
-        [ 'OS == "ios" and target_subarch == 64', {
-          'defines': [
-            'LIBYUV_DISABLE_NEON'
-          ],
-        }],
-        [ 'OS == "ios"', {
-          'xcode_settings': {
-            'DEBUGGING_SYMBOLS': 'YES',
-            'DEBUG_INFORMATION_FORMAT' : 'dwarf-with-dsym',
-            # Work around compile issue with isosim.mm, see
-            # https://code.google.com/p/libyuv/issues/detail?id=548 for details.
-            'WARNING_CFLAGS': [
-              '-Wno-sometimes-uninitialized',
-            ],
-          },
-          'cflags': [
-            '-Wno-sometimes-uninitialized',
-          ],
-        }],
-        [ 'OS != "ios" and libyuv_disable_jpeg != 1', {
-          'defines': [
-            'HAVE_JPEG',
-          ],
-        }],
-        ['OS=="android"', {
-          'dependencies': [
-            '<(DEPTH)/testing/android/native_test.gyp:native_test_native_code',
-          ],
-        }],
-        # TODO(YangZhang): These lines can be removed when high accuracy
-        # YUV to RGB to Neon is ported.
-        [ '(target_arch == "armv7" or target_arch == "armv7s" \
-          or (target_arch == "arm" and arm_version >= 7) \
-          or target_arch == "arm64") \
-          and (arm_neon == 1 or arm_neon_optional == 1)', {
-          'defines': [
-            'LIBYUV_NEON'
-          ],
-        }],
-      ], # conditions
-      'defines': [
-        # Enable the following 3 macros to turn off assembly for specified CPU.
-        # 'LIBYUV_DISABLE_X86',
-        # 'LIBYUV_DISABLE_NEON',
-        # 'LIBYUV_DISABLE_MIPS',
-        # Enable the following macro to build libyuv as a shared library (dll).
-        # 'LIBYUV_USING_SHARED_LIBRARY',
-      ],
-    },
-    {
-      'target_name': 'compare',
-      'type': 'executable',
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-      'sources': [
-        # sources
-        'util/compare.cc',
-      ],
-      'conditions': [
-        ['OS=="linux"', {
-          'cflags': [
-            '-fexceptions',
-          ],
-        }],
-      ], # conditions
-    },
-    {
-      'target_name': 'convert',
-      'type': 'executable',
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-      'sources': [
-        # sources
-        'util/convert.cc',
-      ],
-      'conditions': [
-        ['OS=="linux"', {
-          'cflags': [
-            '-fexceptions',
-          ],
-        }],
-      ], # conditions
-    },
-    # TODO(fbarchard): Enable SSE2 and OpenMP for better performance.
-    {
-      'target_name': 'psnr',
-      'type': 'executable',
-      'sources': [
-        # sources
-        'util/psnr_main.cc',
-        'util/psnr.cc',
-        'util/ssim.cc',
-      ],
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-      'conditions': [
-        [ 'OS == "ios" and target_subarch == 64', {
-          'defines': [
-            'LIBYUV_DISABLE_NEON'
-          ],
-        }],
-
-        [ 'OS != "ios" and libyuv_disable_jpeg != 1', {
-          'defines': [
-            'HAVE_JPEG',
-          ],
-        }],
-      ], # conditions
-    },
-
-    {
-      'target_name': 'cpuid',
-      'type': 'executable',
-      'sources': [
-        # sources
-        'util/cpuid.c',
-      ],
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-    },
-  ], # targets
-  'conditions': [
-    ['OS=="android"', {
-      'targets': [
-        {
-          # TODO(kjellander): Figure out what to change in build/apk_test.gypi
-          # to it can be used instead of the copied code below. Using it in its
-          # current version was not possible, since the target starts with 'lib',
-          # which somewhere confuses the variables.
-          'target_name': 'libyuv_unittest_apk',
-          'type': 'none',
-          'variables': {
-            # These are used to configure java_apk.gypi included below.
-            'test_type': 'gtest',
-            'apk_name': 'libyuv_unittest',
-            'test_suite_name': 'libyuv_unittest',
-            'intermediate_dir': '<(PRODUCT_DIR)/libyuv_unittest_apk',
-            'input_shlib_path': '<(SHARED_LIB_DIR)/<(SHARED_LIB_PREFIX)libyuv_unittest<(SHARED_LIB_SUFFIX)',
-            'final_apk_path': '<(intermediate_dir)/libyuv_unittest-debug.apk',
-            'java_in_dir': '<(DEPTH)/testing/android/native_test/java',
-            'test_runner_path': '<(DEPTH)/util/android/test_runner.py',
-            'native_lib_target': 'libyuv_unittest',
-            'gyp_managed_install': 0,
-          },
-          'includes': [
-            'build/android/test_runner.gypi',
-            'build/java_apk.gypi',
-           ],
-          'dependencies': [
-            '<(DEPTH)/base/base.gyp:base_java',
-            # TODO(kjellander): Figure out why base_build_config_gen is needed
-            # here. It really shouldn't since it's a dependency of base_java
-            # above, but there's always 0 tests run if it's missing.
-            '<(DEPTH)/base/base.gyp:base_build_config_gen',
-            '<(DEPTH)/build/android/pylib/device/commands/commands.gyp:chromium_commands',
-            '<(DEPTH)/build/android/pylib/remote/device/dummy/dummy.gyp:remote_device_dummy_apk',
-            '<(DEPTH)/testing/android/appurify_support.gyp:appurify_support_java',
-            '<(DEPTH)/testing/android/on_device_instrumentation.gyp:reporter_java',
-            '<(DEPTH)/tools/android/android_tools.gyp:android_tools',
-            'libyuv_unittest',
-          ],
-        },
-      ],
-    }],
-  ],
-}
-
-# Local Variables:
-# tab-width:2
-# indent-tabs-mode:nil
-# End:
diff --git a/media/libyuv/linux.mk b/media/libyuv/linux.mk
index ee5a3a7070..b541b47c16 100644
--- a/media/libyuv/linux.mk
+++ b/media/libyuv/linux.mk
@@ -13,13 +13,14 @@ LOCAL_OBJ_FILES := \
 	source/compare.o           \
 	source/compare_common.o    \
 	source/compare_gcc.o       \
-	source/compare_neon64.o    \
+	source/compare_msa.o       \
 	source/compare_neon.o      \
+	source/compare_neon64.o    \
 	source/compare_win.o       \
-	source/convert_argb.o      \
 	source/convert.o           \
-	source/convert_from_argb.o \
+	source/convert_argb.o      \
 	source/convert_from.o      \
+	source/convert_from_argb.o \
 	source/convert_jpeg.o      \
 	source/convert_to_argb.o   \
 	source/convert_to_i420.o   \
@@ -27,30 +28,32 @@ LOCAL_OBJ_FILES := \
 	source/mjpeg_decoder.o     \
 	source/mjpeg_validate.o    \
 	source/planar_functions.o  \
+	source/rotate.o            \
 	source/rotate_any.o        \
 	source/rotate_argb.o       \
-	source/rotate.o            \
 	source/rotate_common.o     \
 	source/rotate_gcc.o        \
-	source/rotate_mips.o       \
-	source/rotate_neon64.o     \
+	source/rotate_msa.o        \
 	source/rotate_neon.o       \
+	source/rotate_neon64.o     \
 	source/rotate_win.o        \
 	source/row_any.o           \
 	source/row_common.o        \
 	source/row_gcc.o           \
-	source/row_mips.o          \
-	source/row_neon64.o        \
+	source/row_msa.o           \
 	source/row_neon.o          \
+	source/row_neon64.o        \
 	source/row_win.o           \
+	source/scale.o             \
 	source/scale_any.o         \
 	source/scale_argb.o        \
-	source/scale.o             \
 	source/scale_common.o      \
 	source/scale_gcc.o         \
-	source/scale_mips.o        \
-	source/scale_neon64.o      \
+	source/scale_msa.o         \
 	source/scale_neon.o        \
+	source/scale_neon64.o      \
+	source/scale_rgb.o         \
+	source/scale_uv.o          \
 	source/scale_win.o         \
 	source/video_common.o
 
@@ -60,22 +63,32 @@ LOCAL_OBJ_FILES := \
 .c.o:
 	$(CC) -c $(CFLAGS) $*.c -o $*.o
 
-all: libyuv.a convert cpuid psnr
+all: libyuv.a i444tonv12_eg yuvconvert yuvconstants cpuid psnr
 
 libyuv.a: $(LOCAL_OBJ_FILES)
 	$(AR) $(ARFLAGS) $@ $(LOCAL_OBJ_FILES)
 
 # A C++ test utility that uses libyuv conversion.
-convert: util/convert.cc libyuv.a
-	$(CXX) $(CXXFLAGS) -Iutil/ -o $@ util/convert.cc libyuv.a
+yuvconvert: util/yuvconvert.cc libyuv.a
+	$(CXX) $(CXXFLAGS) -Iutil/ -o $@ util/yuvconvert.cc libyuv.a
+
+# A C test utility that generates yuvconstants for yuv to rgb.
+yuvconstants: util/yuvconstants.c libyuv.a
+	$(CXX) $(CXXFLAGS) -Iutil/ -lm -o $@ util/yuvconstants.c libyuv.a
 
 # A standalone test utility
 psnr: util/psnr.cc
 	$(CXX) $(CXXFLAGS) -Iutil/ -o $@ util/psnr.cc util/psnr_main.cc util/ssim.cc
 
+# A simple conversion example.
+i444tonv12_eg: util/i444tonv12_eg.cc libyuv.a
+	$(CXX) $(CXXFLAGS) -o $@ util/i444tonv12_eg.cc libyuv.a
+
 # A C test utility that uses libyuv conversion from C.
+# gcc 4.4 and older require -fno-exceptions to avoid link error on __gxx_personality_v0
+# CC=gcc-4.4 CXXFLAGS=-fno-exceptions CXX=g++-4.4 make -f linux.mk
 cpuid: util/cpuid.c libyuv.a
 	$(CC) $(CFLAGS) -o $@ util/cpuid.c libyuv.a
 
 clean:
-	/bin/rm -f source/*.o *.ii *.s libyuv.a convert cpuid psnr
+	/bin/rm -f source/*.o *.ii *.s libyuv.a i444tonv12_eg yuvconvert yuvconstants cpuid psnr
diff --git a/media/libyuv/moz.build b/media/libyuv/moz.build
index 1f910cc72f..b056785920 100644
--- a/media/libyuv/moz.build
+++ b/media/libyuv/moz.build
@@ -9,12 +9,16 @@ GYP_DIRS += ['.']
 
 libyuv_non_unified_sources = [
     'source/convert.cc',
+    'source/convert_argb.cc',
     'source/convert_from.cc',
+    'source/convert_from_argb.cc',
     'source/mjpeg_decoder.cc',
     'source/rotate_argb.cc',
     'source/row_common.cc',
     'source/scale.cc',
+    'source/scale_argb.cc',
     'source/scale_common.cc',
+    'source/scale_uv.cc',
 ]
 GYP_DIRS['.'].input = 'libyuv.gyp'
 GYP_DIRS['.'].variables = gyp_vars
diff --git a/media/libyuv/public.mk b/media/libyuv/public.mk
index 090d8cb659..1342307a56 100644
--- a/media/libyuv/public.mk
+++ b/media/libyuv/public.mk
@@ -1,13 +1,13 @@
-# This file contains all the common make variables which are useful for

-# anyone depending on this library.

-# Note that dependencies on NDK are not directly listed since NDK auto adds

-# them.

-

-LIBYUV_INCLUDES := $(LIBYUV_PATH)/include

-

-LIBYUV_C_FLAGS :=

-

-LIBYUV_CPP_FLAGS :=

-

-LIBYUV_LDLIBS :=

-LIBYUV_DEP_MODULES :=

+# This file contains all the common make variables which are useful for
+# anyone depending on this library.
+# Note that dependencies on NDK are not directly listed since NDK auto adds
+# them.
+
+LIBYUV_INCLUDES := $(LIBYUV_PATH)/include
+
+LIBYUV_C_FLAGS :=
+
+LIBYUV_CPP_FLAGS :=
+
+LIBYUV_LDLIBS :=
+LIBYUV_DEP_MODULES :=
diff --git a/media/libyuv/pylintrc b/media/libyuv/pylintrc
new file mode 100644
index 0000000000..b8bea33489
--- /dev/null
+++ b/media/libyuv/pylintrc
@@ -0,0 +1,17 @@
+[MESSAGES CONTROL]
+
+# Disable the message, report, category or checker with the given id(s).
+# TODO(kjellander): Reduce this list to as small as possible.
+disable=I0010,I0011,bad-continuation,broad-except,duplicate-code,eval-used,exec-used,fixme,invalid-name,missing-docstring,no-init,no-member,too-few-public-methods,too-many-ancestors,too-many-arguments,too-many-branches,too-many-function-args,too-many-instance-attributes,too-many-lines,too-many-locals,too-many-public-methods,too-many-return-statements,too-many-statements
+
+
+[REPORTS]
+
+# Don't write out full reports, just messages.
+reports=no
+
+
+[FORMAT]
+
+# We use two spaces for indents, instead of the usual four spaces or tab.
+indent-string='  '
diff --git a/media/libyuv/setup_links.py b/media/libyuv/setup_links.py
deleted file mode 100644
index b3c8c9a0af..0000000000
--- a/media/libyuv/setup_links.py
+++ /dev/null
@@ -1,476 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Setup links to a Chromium checkout for WebRTC.
-
-WebRTC standalone shares a lot of dependencies and build tools with Chromium.
-To do this, many of the paths of a Chromium checkout is emulated by creating
-symlinks to files and directories. This script handles the setup of symlinks to
-achieve this.
-
-It also handles cleanup of the legacy Subversion-based approach that was used
-before Chrome switched over their master repo from Subversion to Git.
-"""
-
-
-import ctypes
-import errno
-import logging
-import optparse
-import os
-import shelve
-import shutil
-import subprocess
-import sys
-import textwrap
-
-
-DIRECTORIES = [
-  'build',
-  'buildtools',
-  'mojo',  # TODO(kjellander): Remove, see webrtc:5629.
-  'native_client',
-  'net',
-  'testing',
-  'third_party/binutils',
-  'third_party/drmemory',
-  'third_party/instrumented_libraries',
-  'third_party/libjpeg',
-  'third_party/libjpeg_turbo',
-  'third_party/llvm-build',
-  'third_party/lss',
-  'third_party/yasm',
-  'third_party/WebKit',  # TODO(kjellander): Remove, see webrtc:5629.
-  'tools/clang',
-  'tools/gn',
-  'tools/gyp',
-  'tools/memory',
-  'tools/python',
-  'tools/swarming_client',
-  'tools/valgrind',
-  'tools/vim',
-  'tools/win',
-]
-
-from sync_chromium import get_target_os_list
-target_os = get_target_os_list()
-if 'ios' in target_os:
-  DIRECTORIES.append('third_party/class-dump')
-
-FILES = {
-  'tools/isolate_driver.py': None,
-  'third_party/BUILD.gn': None,
-}
-
-ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
-CHROMIUM_CHECKOUT = os.path.join('chromium', 'src')
-LINKS_DB = 'links'
-
-# Version management to make future upgrades/downgrades easier to support.
-SCHEMA_VERSION = 1
-
-
-def query_yes_no(question, default=False):
-  """Ask a yes/no question via raw_input() and return their answer.
-
-  Modified from http://stackoverflow.com/a/3041990.
-  """
-  prompt = " [%s/%%s]: "
-  prompt = prompt % ('Y' if default is True  else 'y')
-  prompt = prompt % ('N' if default is False else 'n')
-
-  if default is None:
-    default = 'INVALID'
-
-  while True:
-    sys.stdout.write(question + prompt)
-    choice = raw_input().lower()
-    if choice == '' and default != 'INVALID':
-      return default
-
-    if 'yes'.startswith(choice):
-      return True
-    elif 'no'.startswith(choice):
-      return False
-
-    print "Please respond with 'yes' or 'no' (or 'y' or 'n')."
-
-
-# Actions
-class Action(object):
-  def __init__(self, dangerous):
-    self.dangerous = dangerous
-
-  def announce(self, planning):
-    """Log a description of this action.
-
-    Args:
-      planning - True iff we're in the planning stage, False if we're in the
-                 doit stage.
-    """
-    pass
-
-  def doit(self, links_db):
-    """Execute the action, recording what we did to links_db, if necessary."""
-    pass
-
-
-class Remove(Action):
-  def __init__(self, path, dangerous):
-    super(Remove, self).__init__(dangerous)
-    self._priority = 0
-    self._path = path
-
-  def announce(self, planning):
-    log = logging.warn
-    filesystem_type = 'file'
-    if not self.dangerous:
-      log = logging.info
-      filesystem_type = 'link'
-    if planning:
-      log('Planning to remove %s: %s', filesystem_type, self._path)
-    else:
-      log('Removing %s: %s', filesystem_type, self._path)
-
-  def doit(self, _):
-    os.remove(self._path)
-
-
-class Rmtree(Action):
-  def __init__(self, path):
-    super(Rmtree, self).__init__(dangerous=True)
-    self._priority = 0
-    self._path = path
-
-  def announce(self, planning):
-    if planning:
-      logging.warn('Planning to remove directory: %s', self._path)
-    else:
-      logging.warn('Removing directory: %s', self._path)
-
-  def doit(self, _):
-    if sys.platform.startswith('win'):
-      # shutil.rmtree() doesn't work on Windows if any of the directories are
-      # read-only, which svn repositories are.
-      subprocess.check_call(['rd', '/q', '/s', self._path], shell=True)
-    else:
-      shutil.rmtree(self._path)
-
-
-class Makedirs(Action):
-  def __init__(self, path):
-    super(Makedirs, self).__init__(dangerous=False)
-    self._priority = 1
-    self._path = path
-
-  def doit(self, _):
-    try:
-      os.makedirs(self._path)
-    except OSError as e:
-      if e.errno != errno.EEXIST:
-        raise
-
-
-class Symlink(Action):
-  def __init__(self, source_path, link_path):
-    super(Symlink, self).__init__(dangerous=False)
-    self._priority = 2
-    self._source_path = source_path
-    self._link_path = link_path
-
-  def announce(self, planning):
-    if planning:
-      logging.info(
-          'Planning to create link from %s to %s', self._link_path,
-          self._source_path)
-    else:
-      logging.debug(
-          'Linking from %s to %s', self._link_path, self._source_path)
-
-  def doit(self, links_db):
-    # Files not in the root directory need relative path calculation.
-    # On Windows, use absolute paths instead since NTFS doesn't seem to support
-    # relative paths for symlinks.
-    if sys.platform.startswith('win'):
-      source_path = os.path.abspath(self._source_path)
-    else:
-      if os.path.dirname(self._link_path) != self._link_path:
-        source_path = os.path.relpath(self._source_path,
-                                      os.path.dirname(self._link_path))
-
-    os.symlink(source_path, os.path.abspath(self._link_path))
-    links_db[self._source_path] = self._link_path
-
-
-class LinkError(IOError):
-  """Failed to create a link."""
-  pass
-
-
-# Handles symlink creation on the different platforms.
-if sys.platform.startswith('win'):
-  def symlink(source_path, link_path):
-    flag = 1 if os.path.isdir(source_path) else 0
-    if not ctypes.windll.kernel32.CreateSymbolicLinkW(
-        unicode(link_path), unicode(source_path), flag):
-      raise OSError('Failed to create symlink to %s. Notice that only NTFS '
-                    'version 5.0 and up has all the needed APIs for '
-                    'creating symlinks.' % source_path)
-  os.symlink = symlink
-
-
-class WebRTCLinkSetup(object):
-  def __init__(self, links_db, force=False, dry_run=False, prompt=False):
-    self._force = force
-    self._dry_run = dry_run
-    self._prompt = prompt
-    self._links_db = links_db
-
-  def CreateLinks(self, on_bot):
-    logging.debug('CreateLinks')
-    # First, make a plan of action
-    actions = []
-
-    for source_path, link_path in FILES.iteritems():
-      actions += self._ActionForPath(
-          source_path, link_path, check_fn=os.path.isfile, check_msg='files')
-    for source_dir in DIRECTORIES:
-      actions += self._ActionForPath(
-          source_dir, None, check_fn=os.path.isdir,
-          check_msg='directories')
-
-    if not on_bot and self._force:
-      # When making the manual switch from legacy SVN checkouts to the new
-      # Git-based Chromium DEPS, the .gclient_entries file that contains cached
-      # URLs for all DEPS entries must be removed to avoid future sync problems.
-      entries_file = os.path.join(os.path.dirname(ROOT_DIR), '.gclient_entries')
-      if os.path.exists(entries_file):
-        actions.append(Remove(entries_file, dangerous=True))
-
-    actions.sort()
-
-    if self._dry_run:
-      for action in actions:
-        action.announce(planning=True)
-      logging.info('Not doing anything because dry-run was specified.')
-      sys.exit(0)
-
-    if any(a.dangerous for a in actions):
-      logging.warn('Dangerous actions:')
-      for action in (a for a in actions if a.dangerous):
-        action.announce(planning=True)
-      print
-
-      if not self._force:
-        logging.error(textwrap.dedent("""\
-        @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
-                              A C T I O N     R E Q I R E D
-        @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
-
-        Because chromium/src is transitioning to Git (from SVN), we needed to
-        change the way that the WebRTC standalone checkout works. Instead of
-        individually syncing subdirectories of Chromium in SVN, we're now
-        syncing Chromium (and all of its DEPS, as defined by its own DEPS file),
-        into the `chromium/src` directory.
-
-        As such, all Chromium directories which are currently pulled by DEPS are
-        now replaced with a symlink into the full Chromium checkout.
-
-        To avoid disrupting developers, we've chosen to not delete your
-        directories forcibly, in case you have some work in progress in one of
-        them :).
-
-        ACTION REQUIRED:
-        Before running `gclient sync|runhooks` again, you must run:
-        %s%s --force
-
-        Which will replace all directories which now must be symlinks, after
-        prompting with a summary of the work-to-be-done.
-        """), 'python ' if sys.platform.startswith('win') else '', sys.argv[0])
-        sys.exit(1)
-      elif self._prompt:
-        if not query_yes_no('Would you like to perform the above plan?'):
-          sys.exit(1)
-
-    for action in actions:
-      action.announce(planning=False)
-      action.doit(self._links_db)
-
-    if not on_bot and self._force:
-      logging.info('Completed!\n\nNow run `gclient sync|runhooks` again to '
-                   'let the remaining hooks (that probably were interrupted) '
-                   'execute.')
-
-  def CleanupLinks(self):
-    logging.debug('CleanupLinks')
-    for source, link_path  in self._links_db.iteritems():
-      if source == 'SCHEMA_VERSION':
-        continue
-      if os.path.islink(link_path) or sys.platform.startswith('win'):
-        # os.path.islink() always returns false on Windows
-        # See http://bugs.python.org/issue13143.
-        logging.debug('Removing link to %s at %s', source, link_path)
-        if not self._dry_run:
-          if os.path.exists(link_path):
-            if sys.platform.startswith('win') and os.path.isdir(link_path):
-              subprocess.check_call(['rmdir', '/q', '/s', link_path],
-                                    shell=True)
-            else:
-              os.remove(link_path)
-          del self._links_db[source]
-
-  @staticmethod
-  def _ActionForPath(source_path, link_path=None, check_fn=None,
-                     check_msg=None):
-    """Create zero or more Actions to link to a file or directory.
-
-    This will be a symlink on POSIX platforms. On Windows this requires
-    that NTFS is version 5.0 or higher (Vista or newer).
-
-    Args:
-      source_path: Path relative to the Chromium checkout root.
-        For readability, the path may contain slashes, which will
-        automatically be converted to the right path delimiter on Windows.
-      link_path: The location for the link to create. If omitted it will be the
-        same path as source_path.
-      check_fn: A function returning true if the type of filesystem object is
-        correct for the attempted call. Otherwise an error message with
-        check_msg will be printed.
-      check_msg: String used to inform the user of an invalid attempt to create
-        a file.
-    Returns:
-      A list of Action objects.
-    """
-    def fix_separators(path):
-      if sys.platform.startswith('win'):
-        return path.replace(os.altsep, os.sep)
-      else:
-        return path
-
-    assert check_fn
-    assert check_msg
-    link_path = link_path or source_path
-    link_path = fix_separators(link_path)
-
-    source_path = fix_separators(source_path)
-    source_path = os.path.join(CHROMIUM_CHECKOUT, source_path)
-    if os.path.exists(source_path) and not check_fn:
-      raise LinkError('_LinkChromiumPath can only be used to link to %s: '
-                      'Tried to link to: %s' % (check_msg, source_path))
-
-    if not os.path.exists(source_path):
-      logging.debug('Silently ignoring missing source: %s. This is to avoid '
-                    'errors on platform-specific dependencies.', source_path)
-      return []
-
-    actions = []
-
-    if os.path.exists(link_path) or os.path.islink(link_path):
-      if os.path.islink(link_path):
-        actions.append(Remove(link_path, dangerous=False))
-      elif os.path.isfile(link_path):
-        actions.append(Remove(link_path, dangerous=True))
-      elif os.path.isdir(link_path):
-        actions.append(Rmtree(link_path))
-      else:
-        raise LinkError('Don\'t know how to plan: %s' % link_path)
-
-    # Create parent directories to the target link if needed.
-    target_parent_dirs = os.path.dirname(link_path)
-    if (target_parent_dirs and
-        target_parent_dirs != link_path and
-        not os.path.exists(target_parent_dirs)):
-      actions.append(Makedirs(target_parent_dirs))
-
-    actions.append(Symlink(source_path, link_path))
-
-    return actions
-
-def _initialize_database(filename):
-  links_database = shelve.open(filename)
-
-  # Wipe the database if this version of the script ends up looking at a
-  # newer (future) version of the links db, just to be sure.
-  version = links_database.get('SCHEMA_VERSION')
-  if version and version != SCHEMA_VERSION:
-    logging.info('Found database with schema version %s while this script only '
-                 'supports %s. Wiping previous database contents.', version,
-                 SCHEMA_VERSION)
-    links_database.clear()
-  links_database['SCHEMA_VERSION'] = SCHEMA_VERSION
-  return links_database
-
-
-def main():
-  on_bot = os.environ.get('CHROME_HEADLESS') == '1'
-
-  parser = optparse.OptionParser()
-  parser.add_option('-d', '--dry-run', action='store_true', default=False,
-                    help='Print what would be done, but don\'t perform any '
-                         'operations. This will automatically set logging to '
-                         'verbose.')
-  parser.add_option('-c', '--clean-only', action='store_true', default=False,
-                    help='Only clean previously created links, don\'t create '
-                         'new ones. This will automatically set logging to '
-                         'verbose.')
-  parser.add_option('-f', '--force', action='store_true', default=on_bot,
-                    help='Force link creation. CAUTION: This deletes existing '
-                         'folders and files in the locations where links are '
-                         'about to be created.')
-  parser.add_option('-n', '--no-prompt', action='store_false', dest='prompt',
-                    default=(not on_bot),
-                    help='Prompt if we\'re planning to do a dangerous action')
-  parser.add_option('-v', '--verbose', action='store_const',
-                    const=logging.DEBUG, default=logging.INFO,
-                    help='Print verbose output for debugging.')
-  options, _ = parser.parse_args()
-
-  if options.dry_run or options.force or options.clean_only:
-    options.verbose = logging.DEBUG
-  logging.basicConfig(format='%(message)s', level=options.verbose)
-
-  # Work from the root directory of the checkout.
-  script_dir = os.path.dirname(os.path.abspath(__file__))
-  os.chdir(script_dir)
-
-  if sys.platform.startswith('win'):
-    def is_admin():
-      try:
-        return os.getuid() == 0
-      except AttributeError:
-        return ctypes.windll.shell32.IsUserAnAdmin() != 0
-    if not is_admin():
-      logging.error('On Windows, you now need to have administrator '
-                    'privileges for the shell running %s (or '
-                    '`gclient sync|runhooks`).\nPlease start another command '
-                    'prompt as Administrator and try again.', sys.argv[0])
-      return 1
-
-  if not os.path.exists(CHROMIUM_CHECKOUT):
-    logging.error('Cannot find a Chromium checkout at %s. Did you run "gclient '
-                  'sync" before running this script?', CHROMIUM_CHECKOUT)
-    return 2
-
-  links_database = _initialize_database(LINKS_DB)
-  try:
-    symlink_creator = WebRTCLinkSetup(links_database, options.force,
-                                      options.dry_run, options.prompt)
-    symlink_creator.CleanupLinks()
-    if not options.clean_only:
-      symlink_creator.CreateLinks(on_bot)
-  except LinkError as e:
-    print >> sys.stderr, e.message
-    return 3
-  finally:
-    links_database.close()
-  return 0
-
-
-if __name__ == '__main__':
-  sys.exit(main())
diff --git a/media/libyuv/source/compare.cc b/media/libyuv/source/compare.cc
index e3846bdfdd..50a736bdda 100644
--- a/media/libyuv/source/compare.cc
+++ b/media/libyuv/source/compare.cc
@@ -29,10 +29,10 @@ extern "C" {
 
 // hash seed of 5381 recommended.
 LIBYUV_API
-uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
+uint32_t HashDjb2(const uint8_t* src, uint64_t count, uint32_t seed) {
   const int kBlockSize = 1 << 15;  // 32768;
   int remainder;
-  uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) =
+  uint32_t (*HashDjb2_SSE)(const uint8_t* src, int count, uint32_t seed) =
       HashDjb2_C;
 #if defined(HAS_HASHDJB2_SSE41)
   if (TestCpuFlag(kCpuHasSSE41)) {
@@ -45,37 +45,37 @@ uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
   }
 #endif
 
-  while (count >= (uint64)(kBlockSize)) {
+  while (count >= (uint64_t)kBlockSize) {
     seed = HashDjb2_SSE(src, kBlockSize, seed);
     src += kBlockSize;
     count -= kBlockSize;
   }
-  remainder = (int)(count) & ~15;
+  remainder = (int)count & ~15;
   if (remainder) {
     seed = HashDjb2_SSE(src, remainder, seed);
     src += remainder;
     count -= remainder;
   }
-  remainder = (int)(count) & 15;
+  remainder = (int)count & 15;
   if (remainder) {
     seed = HashDjb2_C(src, remainder, seed);
   }
   return seed;
 }
 
-static uint32 ARGBDetectRow_C(const uint8* argb, int width) {
+static uint32_t ARGBDetectRow_C(const uint8_t* argb, int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
     if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
       return FOURCC_BGRA;
     }
-    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+    if (argb[3] != 255) {  // Fourth byte is not Alpha of 255, so not BGRA.
       return FOURCC_ARGB;
     }
     if (argb[4] != 255) {  // Second pixel first byte is not Alpha of 255.
       return FOURCC_BGRA;
     }
-    if (argb[7] != 255) {  // Second pixel 4th byte is not Alpha of 255.
+    if (argb[7] != 255) {  // Second pixel fourth byte is not Alpha of 255.
       return FOURCC_ARGB;
     }
     argb += 8;
@@ -94,8 +94,11 @@ static uint32 ARGBDetectRow_C(const uint8* argb, int width) {
 // Scan an opaque argb image and return fourcc based on alpha offset.
 // Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
 LIBYUV_API
-uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) {
-  uint32 fourcc = 0;
+uint32_t ARGBDetect(const uint8_t* argb,
+                    int stride_argb,
+                    int width,
+                    int height) {
+  uint32_t fourcc = 0;
   int h;
 
   // Coalesce rows.
@@ -111,19 +114,81 @@ uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) {
   return fourcc;
 }
 
+// NEON version accumulates in 16 bit shorts which overflow at 65536 bytes.
+// So actual maximum is 1 less loop, which is 64436 - 32 bytes.
+
+LIBYUV_API
+uint64_t ComputeHammingDistance(const uint8_t* src_a,
+                                const uint8_t* src_b,
+                                int count) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  const int kSimdSize = 64;
+  // SIMD for multiple of 64, and C for remainder
+  int remainder = count & (kBlockSize - 1) & ~(kSimdSize - 1);
+  uint64_t diff = 0;
+  int i;
+  uint32_t (*HammingDistance)(const uint8_t* src_a, const uint8_t* src_b,
+                              int count) = HammingDistance_C;
+#if defined(HAS_HAMMINGDISTANCE_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    HammingDistance = HammingDistance_NEON;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    HammingDistance = HammingDistance_SSSE3;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_SSE42)
+  if (TestCpuFlag(kCpuHasSSE42)) {
+    HammingDistance = HammingDistance_SSE42;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HammingDistance = HammingDistance_AVX2;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    HammingDistance = HammingDistance_MSA;
+  }
+#endif
+
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+ : diff)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    diff += HammingDistance(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    diff += HammingDistance(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & (kSimdSize - 1);
+  if (remainder) {
+    diff += HammingDistance_C(src_a, src_b, remainder);
+  }
+  return diff;
+}
+
 // TODO(fbarchard): Refactor into row function.
 LIBYUV_API
-uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
-                             int count) {
+uint64_t ComputeSumSquareError(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count) {
   // SumSquareError returns values 0 to 65535 for each squared difference.
-  // Up to 65536 of those can be summed and remain within a uint32.
-  // After each block of 65536 pixels, accumulate into a uint64.
+  // Up to 65536 of those can be summed and remain within a uint32_t.
+  // After each block of 65536 pixels, accumulate into a uint64_t.
   const int kBlockSize = 65536;
   int remainder = count & (kBlockSize - 1) & ~31;
-  uint64 sse = 0;
+  uint64_t sse = 0;
   int i;
-  uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
-      SumSquareError_C;
+  uint32_t (*SumSquareError)(const uint8_t* src_a, const uint8_t* src_b,
+                             int count) = SumSquareError_C;
 #if defined(HAS_SUMSQUAREERROR_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     SumSquareError = SumSquareError_NEON;
@@ -141,8 +206,13 @@ uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
     SumSquareError = SumSquareError_AVX2;
   }
 #endif
+#if defined(HAS_SUMSQUAREERROR_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SumSquareError = SumSquareError_MSA;
+  }
+#endif
 #ifdef _OPENMP
-#pragma omp parallel for reduction(+: sse)
+#pragma omp parallel for reduction(+ : sse)
 #endif
   for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
     sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
@@ -162,14 +232,16 @@ uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
 }
 
 LIBYUV_API
-uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
-                                  const uint8* src_b, int stride_b,
-                                  int width, int height) {
-  uint64 sse = 0;
+uint64_t ComputeSumSquareErrorPlane(const uint8_t* src_a,
+                                    int stride_a,
+                                    const uint8_t* src_b,
+                                    int stride_b,
+                                    int width,
+                                    int height) {
+  uint64_t sse = 0;
   int h;
   // Coalesce rows.
-  if (stride_a == width &&
-      stride_b == width) {
+  if (stride_a == width && stride_b == width) {
     width *= height;
     height = 1;
     stride_a = stride_b = 0;
@@ -183,66 +255,76 @@ uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
 }
 
 LIBYUV_API
-double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
+double SumSquareErrorToPsnr(uint64_t sse, uint64_t count) {
   double psnr;
   if (sse > 0) {
-    double mse = (double)(count) / (double)(sse);
+    double mse = (double)count / (double)sse;
     psnr = 10.0 * log10(255.0 * 255.0 * mse);
   } else {
-    psnr = kMaxPsnr;      // Limit to prevent divide by 0
+    psnr = kMaxPsnr;  // Limit to prevent divide by 0
   }
 
-  if (psnr > kMaxPsnr)
+  if (psnr > kMaxPsnr) {
     psnr = kMaxPsnr;
+  }
 
   return psnr;
 }
 
 LIBYUV_API
-double CalcFramePsnr(const uint8* src_a, int stride_a,
-                     const uint8* src_b, int stride_b,
-                     int width, int height) {
-  const uint64 samples = width * height;
-  const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
-                                                src_b, stride_b,
-                                                width, height);
+double CalcFramePsnr(const uint8_t* src_a,
+                     int stride_a,
+                     const uint8_t* src_b,
+                     int stride_b,
+                     int width,
+                     int height) {
+  const uint64_t samples = (uint64_t)width * (uint64_t)height;
+  const uint64_t sse = ComputeSumSquareErrorPlane(src_a, stride_a, src_b,
+                                                  stride_b, width, height);
   return SumSquareErrorToPsnr(sse, samples);
 }
 
 LIBYUV_API
-double I420Psnr(const uint8* src_y_a, int stride_y_a,
-                const uint8* src_u_a, int stride_u_a,
-                const uint8* src_v_a, int stride_v_a,
-                const uint8* src_y_b, int stride_y_b,
-                const uint8* src_u_b, int stride_u_b,
-                const uint8* src_v_b, int stride_v_b,
-                int width, int height) {
-  const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
-                                                  src_y_b, stride_y_b,
-                                                  width, height);
+double I420Psnr(const uint8_t* src_y_a,
+                int stride_y_a,
+                const uint8_t* src_u_a,
+                int stride_u_a,
+                const uint8_t* src_v_a,
+                int stride_v_a,
+                const uint8_t* src_y_b,
+                int stride_y_b,
+                const uint8_t* src_u_b,
+                int stride_u_b,
+                const uint8_t* src_v_b,
+                int stride_v_b,
+                int width,
+                int height) {
+  const uint64_t sse_y = ComputeSumSquareErrorPlane(
+      src_y_a, stride_y_a, src_y_b, stride_y_b, width, height);
   const int width_uv = (width + 1) >> 1;
   const int height_uv = (height + 1) >> 1;
-  const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
-                                                  src_u_b, stride_u_b,
-                                                  width_uv, height_uv);
-  const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
-                                                  src_v_b, stride_v_b,
-                                                  width_uv, height_uv);
-  const uint64 samples = width * height + 2 * (width_uv * height_uv);
-  const uint64 sse = sse_y + sse_u + sse_v;
+  const uint64_t sse_u = ComputeSumSquareErrorPlane(
+      src_u_a, stride_u_a, src_u_b, stride_u_b, width_uv, height_uv);
+  const uint64_t sse_v = ComputeSumSquareErrorPlane(
+      src_v_a, stride_v_a, src_v_b, stride_v_b, width_uv, height_uv);
+  const uint64_t samples = (uint64_t)width * (uint64_t)height +
+                           2 * ((uint64_t)width_uv * (uint64_t)height_uv);
+  const uint64_t sse = sse_y + sse_u + sse_v;
   return SumSquareErrorToPsnr(sse, samples);
 }
 
-static const int64 cc1 =  26634;  // (64^2*(.01*255)^2
-static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
+static const int64_t cc1 = 26634;   // (64^2*(.01*255)^2
+static const int64_t cc2 = 239708;  // (64^2*(.03*255)^2
 
-static double Ssim8x8_C(const uint8* src_a, int stride_a,
-                        const uint8* src_b, int stride_b) {
-  int64 sum_a = 0;
-  int64 sum_b = 0;
-  int64 sum_sq_a = 0;
-  int64 sum_sq_b = 0;
-  int64 sum_axb = 0;
+static double Ssim8x8_C(const uint8_t* src_a,
+                        int stride_a,
+                        const uint8_t* src_b,
+                        int stride_b) {
+  int64_t sum_a = 0;
+  int64_t sum_b = 0;
+  int64_t sum_sq_a = 0;
+  int64_t sum_sq_b = 0;
+  int64_t sum_axb = 0;
 
   int i;
   for (i = 0; i < 8; ++i) {
@@ -260,27 +342,27 @@ static double Ssim8x8_C(const uint8* src_a, int stride_a,
   }
 
   {
-    const int64 count = 64;
+    const int64_t count = 64;
     // scale the constants by number of pixels
-    const int64 c1 = (cc1 * count * count) >> 12;
-    const int64 c2 = (cc2 * count * count) >> 12;
+    const int64_t c1 = (cc1 * count * count) >> 12;
+    const int64_t c2 = (cc2 * count * count) >> 12;
 
-    const int64 sum_a_x_sum_b = sum_a * sum_b;
+    const int64_t sum_a_x_sum_b = sum_a * sum_b;
 
-    const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
-                         (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
+    const int64_t ssim_n = (2 * sum_a_x_sum_b + c1) *
+                           (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
 
-    const int64 sum_a_sq = sum_a*sum_a;
-    const int64 sum_b_sq = sum_b*sum_b;
+    const int64_t sum_a_sq = sum_a * sum_a;
+    const int64_t sum_b_sq = sum_b * sum_b;
 
-    const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
-                         (count * sum_sq_a - sum_a_sq +
-                          count * sum_sq_b - sum_b_sq + c2);
+    const int64_t ssim_d =
+        (sum_a_sq + sum_b_sq + c1) *
+        (count * sum_sq_a - sum_a_sq + count * sum_sq_b - sum_b_sq + c2);
 
-    if (ssim_d == 0.0) {
+    if (ssim_d == 0) {
       return DBL_MAX;
     }
-    return ssim_n * 1.0 / ssim_d;
+    return (double)ssim_n / (double)ssim_d;
   }
 }
 
@@ -288,13 +370,16 @@ static double Ssim8x8_C(const uint8* src_a, int stride_a,
 // on the 4x4 pixel grid. Such arrangement allows the windows to overlap
 // block boundaries to penalize blocking artifacts.
 LIBYUV_API
-double CalcFrameSsim(const uint8* src_a, int stride_a,
-                     const uint8* src_b, int stride_b,
-                     int width, int height) {
+double CalcFrameSsim(const uint8_t* src_a,
+                     int stride_a,
+                     const uint8_t* src_b,
+                     int stride_b,
+                     int width,
+                     int height) {
   int samples = 0;
   double ssim_total = 0;
-  double (*Ssim8x8)(const uint8* src_a, int stride_a,
-                    const uint8* src_b, int stride_b) = Ssim8x8_C;
+  double (*Ssim8x8)(const uint8_t* src_a, int stride_a, const uint8_t* src_b,
+                    int stride_b) = Ssim8x8_C;
 
   // sample point start with each 4x4 location
   int i;
@@ -314,22 +399,27 @@ double CalcFrameSsim(const uint8* src_a, int stride_a,
 }
 
 LIBYUV_API
-double I420Ssim(const uint8* src_y_a, int stride_y_a,
-                const uint8* src_u_a, int stride_u_a,
-                const uint8* src_v_a, int stride_v_a,
-                const uint8* src_y_b, int stride_y_b,
-                const uint8* src_u_b, int stride_u_b,
-                const uint8* src_v_b, int stride_v_b,
-                int width, int height) {
-  const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
-                                      src_y_b, stride_y_b, width, height);
+double I420Ssim(const uint8_t* src_y_a,
+                int stride_y_a,
+                const uint8_t* src_u_a,
+                int stride_u_a,
+                const uint8_t* src_v_a,
+                int stride_v_a,
+                const uint8_t* src_y_b,
+                int stride_y_b,
+                const uint8_t* src_u_b,
+                int stride_u_b,
+                const uint8_t* src_v_b,
+                int stride_v_b,
+                int width,
+                int height) {
+  const double ssim_y =
+      CalcFrameSsim(src_y_a, stride_y_a, src_y_b, stride_y_b, width, height);
   const int width_uv = (width + 1) >> 1;
   const int height_uv = (height + 1) >> 1;
-  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
-                                      src_u_b, stride_u_b,
+  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a, src_u_b, stride_u_b,
                                       width_uv, height_uv);
-  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
-                                      src_v_b, stride_v_b,
+  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a, src_v_b, stride_v_b,
                                       width_uv, height_uv);
   return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
 }
diff --git a/media/libyuv/source/compare_common.cc b/media/libyuv/source/compare_common.cc
index 42fc589354..d1cab8d2b4 100644
--- a/media/libyuv/source/compare_common.cc
+++ b/media/libyuv/source/compare_common.cc
@@ -17,20 +17,50 @@ namespace libyuv {
 extern "C" {
 #endif
 
-uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) {
-  uint32 sse = 0u;
+// Hakmem method for hamming distance.
+uint32_t HammingDistance_C(const uint8_t* src_a,
+                           const uint8_t* src_b,
+                           int count) {
+  uint32_t diff = 0u;
+
+  int i;
+  for (i = 0; i < count - 3; i += 4) {
+    uint32_t x = *((const uint32_t*)src_a) ^ *((const uint32_t*)src_b);
+    uint32_t u = x - ((x >> 1) & 0x55555555);
+    u = ((u >> 2) & 0x33333333) + (u & 0x33333333);
+    diff += ((((u + (u >> 4)) & 0x0f0f0f0f) * 0x01010101) >> 24);
+    src_a += 4;
+    src_b += 4;
+  }
+
+  for (; i < count; ++i) {
+    uint32_t x = *src_a ^ *src_b;
+    uint32_t u = x - ((x >> 1) & 0x55);
+    u = ((u >> 2) & 0x33) + (u & 0x33);
+    diff += (u + (u >> 4)) & 0x0f;
+    src_a += 1;
+    src_b += 1;
+  }
+
+  return diff;
+}
+
+uint32_t SumSquareError_C(const uint8_t* src_a,
+                          const uint8_t* src_b,
+                          int count) {
+  uint32_t sse = 0u;
   int i;
   for (i = 0; i < count; ++i) {
     int diff = src_a[i] - src_b[i];
-    sse += (uint32)(diff * diff);
+    sse += (uint32_t)(diff * diff);
   }
   return sse;
 }
 
 // hash seed of 5381 recommended.
 // Internal C version of HashDjb2 with int sized count for efficiency.
-uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) {
-  uint32 hash = seed;
+uint32_t HashDjb2_C(const uint8_t* src, int count, uint32_t seed) {
+  uint32_t hash = seed;
   int i;
   for (i = 0; i < count; ++i) {
     hash += (hash << 5) + src[i];
diff --git a/media/libyuv/source/compare_gcc.cc b/media/libyuv/source/compare_gcc.cc
index 1b83edb166..33cbe25ded 100644
--- a/media/libyuv/source/compare_gcc.cc
+++ b/media/libyuv/source/compare_gcc.cc
@@ -19,127 +19,336 @@ extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
-
-uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
-  uint32 sse;
-  asm volatile (
-    "pxor      %%xmm0,%%xmm0                   \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x10, 0) ",%0          \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x10, 1) ",%1          \n"
-    "movdqa    %%xmm1,%%xmm3                   \n"
-    "psubusb   %%xmm2,%%xmm1                   \n"
-    "psubusb   %%xmm3,%%xmm2                   \n"
-    "por       %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "punpckhbw %%xmm5,%%xmm2                   \n"
-    "pmaddwd   %%xmm1,%%xmm1                   \n"
-    "pmaddwd   %%xmm2,%%xmm2                   \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "paddd     %%xmm2,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-
-    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "movd      %%xmm0,%3                       \n"
-
-  : "+r"(src_a),      // %0
-    "+r"(src_b),      // %1
-    "+r"(count),      // %2
-    "=g"(sse)         // %3
-  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(__x86_64__)
+uint32_t HammingDistance_SSE42(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count) {
+  uint64_t diff = 0u;
+
+  asm volatile(
+      "xor         %3,%3                         \n"
+      "xor         %%r8,%%r8                     \n"
+      "xor         %%r9,%%r9                     \n"
+      "xor         %%r10,%%r10                   \n"
+
+      // Process 32 bytes per loop.
+      LABELALIGN
+      "1:                                        \n"
+      "mov         (%0),%%rcx                    \n"
+      "mov         0x8(%0),%%rdx                 \n"
+      "xor         (%1),%%rcx                    \n"
+      "xor         0x8(%1),%%rdx                 \n"
+      "popcnt      %%rcx,%%rcx                   \n"
+      "popcnt      %%rdx,%%rdx                   \n"
+      "mov         0x10(%0),%%rsi                \n"
+      "mov         0x18(%0),%%rdi                \n"
+      "xor         0x10(%1),%%rsi                \n"
+      "xor         0x18(%1),%%rdi                \n"
+      "popcnt      %%rsi,%%rsi                   \n"
+      "popcnt      %%rdi,%%rdi                   \n"
+      "add         $0x20,%0                      \n"
+      "add         $0x20,%1                      \n"
+      "add         %%rcx,%3                      \n"
+      "add         %%rdx,%%r8                    \n"
+      "add         %%rsi,%%r9                    \n"
+      "add         %%rdi,%%r10                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+
+      "add         %%r8, %3                      \n"
+      "add         %%r9, %3                      \n"
+      "add         %%r10, %3                     \n"
+      : "+r"(src_a),  // %0
+        "+r"(src_b),  // %1
+        "+r"(count),  // %2
+        "=r"(diff)    // %3
+      :
+      : "memory", "cc", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10");
+
+  return (uint32_t)(diff);
+}
+#else
+uint32_t HammingDistance_SSE42(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count) {
+  uint32_t diff = 0u;
+
+  asm volatile(
+      // Process 16 bytes per loop.
+      LABELALIGN
+      "1:                                        \n"
+      "mov         (%0),%%ecx                    \n"
+      "mov         0x4(%0),%%edx                 \n"
+      "xor         (%1),%%ecx                    \n"
+      "xor         0x4(%1),%%edx                 \n"
+      "popcnt      %%ecx,%%ecx                   \n"
+      "add         %%ecx,%3                      \n"
+      "popcnt      %%edx,%%edx                   \n"
+      "add         %%edx,%3                      \n"
+      "mov         0x8(%0),%%ecx                 \n"
+      "mov         0xc(%0),%%edx                 \n"
+      "xor         0x8(%1),%%ecx                 \n"
+      "xor         0xc(%1),%%edx                 \n"
+      "popcnt      %%ecx,%%ecx                   \n"
+      "add         %%ecx,%3                      \n"
+      "popcnt      %%edx,%%edx                   \n"
+      "add         %%edx,%3                      \n"
+      "add         $0x10,%0                      \n"
+      "add         $0x10,%1                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_a),  // %0
+        "+r"(src_b),  // %1
+        "+r"(count),  // %2
+        "+r"(diff)    // %3
+      :
+      : "memory", "cc", "ecx", "edx");
+
+  return diff;
+}
+#endif
+
+static const vec8 kNibbleMask = {15, 15, 15, 15, 15, 15, 15, 15,
+                                 15, 15, 15, 15, 15, 15, 15, 15};
+static const vec8 kBitCount = {0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4};
+
+uint32_t HammingDistance_SSSE3(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count) {
+  uint32_t diff = 0u;
+
+  asm volatile(
+      "movdqa      %4,%%xmm2                     \n"
+      "movdqa      %5,%%xmm3                     \n"
+      "pxor        %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "sub         %0,%1                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqa      (%0),%%xmm4                   \n"
+      "movdqa      0x10(%0), %%xmm5              \n"
+      "pxor        (%0,%1), %%xmm4               \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "pand        %%xmm2,%%xmm6                 \n"
+      "psrlw       $0x4,%%xmm4                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "pshufb      %%xmm6,%%xmm7                 \n"
+      "pand        %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm6                 \n"
+      "pshufb      %%xmm4,%%xmm6                 \n"
+      "paddb       %%xmm7,%%xmm6                 \n"
+      "pxor        0x10(%0,%1),%%xmm5            \n"
+      "add         $0x20,%0                      \n"
+      "movdqa      %%xmm5,%%xmm4                 \n"
+      "pand        %%xmm2,%%xmm5                 \n"
+      "psrlw       $0x4,%%xmm4                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "pshufb      %%xmm5,%%xmm7                 \n"
+      "pand        %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "pshufb      %%xmm4,%%xmm5                 \n"
+      "paddb       %%xmm7,%%xmm5                 \n"
+      "paddb       %%xmm5,%%xmm6                 \n"
+      "psadbw      %%xmm1,%%xmm6                 \n"
+      "paddd       %%xmm6,%%xmm0                 \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+
+      "pshufd      $0xaa,%%xmm0,%%xmm1           \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "movd        %%xmm0, %3                    \n"
+      : "+r"(src_a),       // %0
+        "+r"(src_b),       // %1
+        "+r"(count),       // %2
+        "=r"(diff)         // %3
+      : "m"(kNibbleMask),  // %4
+        "m"(kBitCount)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+
+  return diff;
+}
+
+#ifdef HAS_HAMMINGDISTANCE_AVX2
+uint32_t HammingDistance_AVX2(const uint8_t* src_a,
+                              const uint8_t* src_b,
+                              int count) {
+  uint32_t diff = 0u;
+
+  asm volatile(
+      "vbroadcastf128 %4,%%ymm2                  \n"
+      "vbroadcastf128 %5,%%ymm3                  \n"
+      "vpxor       %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpxor       %%ymm1,%%ymm1,%%ymm1          \n"
+      "sub         %0,%1                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqa     (%0),%%ymm4                   \n"
+      "vmovdqa     0x20(%0), %%ymm5              \n"
+      "vpxor       (%0,%1), %%ymm4, %%ymm4       \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm6          \n"
+      "vpsrlw      $0x4,%%ymm4,%%ymm4            \n"
+      "vpshufb     %%ymm6,%%ymm3,%%ymm6          \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm4          \n"
+      "vpshufb     %%ymm4,%%ymm3,%%ymm4          \n"
+      "vpaddb      %%ymm4,%%ymm6,%%ymm6          \n"
+      "vpxor       0x20(%0,%1),%%ymm5,%%ymm4     \n"
+      "add         $0x40,%0                      \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm5          \n"
+      "vpsrlw      $0x4,%%ymm4,%%ymm4            \n"
+      "vpshufb     %%ymm5,%%ymm3,%%ymm5          \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm4          \n"
+      "vpshufb     %%ymm4,%%ymm3,%%ymm4          \n"
+      "vpaddb      %%ymm5,%%ymm4,%%ymm4          \n"
+      "vpaddb      %%ymm6,%%ymm4,%%ymm4          \n"
+      "vpsadbw     %%ymm1,%%ymm4,%%ymm4          \n"
+      "vpaddd      %%ymm0,%%ymm4,%%ymm0          \n"
+      "sub         $0x40,%2                      \n"
+      "jg          1b                            \n"
+
+      "vpermq      $0xb1,%%ymm0,%%ymm1           \n"
+      "vpaddd      %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xaa,%%ymm0,%%ymm1           \n"
+      "vpaddd      %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovd       %%xmm0, %3                    \n"
+      "vzeroupper                                \n"
+      : "+r"(src_a),       // %0
+        "+r"(src_b),       // %1
+        "+r"(count),       // %2
+        "=r"(diff)         // %3
+      : "m"(kNibbleMask),  // %4
+        "m"(kBitCount)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+
+  return diff;
+}
+#endif  // HAS_HAMMINGDISTANCE_AVX2
+
+uint32_t SumSquareError_SSE2(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count) {
+  uint32_t sse;
+  asm volatile(
+      "pxor        %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm2                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "psubusb     %%xmm2,%%xmm1                 \n"
+      "psubusb     %%xmm3,%%xmm2                 \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpckhbw   %%xmm5,%%xmm2                 \n"
+      "pmaddwd     %%xmm1,%%xmm1                 \n"
+      "pmaddwd     %%xmm2,%%xmm2                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "paddd       %%xmm2,%%xmm0                 \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+
+      "pshufd      $0xee,%%xmm0,%%xmm1           \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "pshufd      $0x1,%%xmm0,%%xmm1            \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "movd        %%xmm0,%3                     \n"
+
+      : "+r"(src_a),  // %0
+        "+r"(src_b),  // %1
+        "+r"(count),  // %2
+        "=g"(sse)     // %3
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
   return sse;
 }
 
-static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
-static uvec32 kHashMul0 = {
-  0x0c3525e1,  // 33 ^ 15
-  0xa3476dc1,  // 33 ^ 14
-  0x3b4039a1,  // 33 ^ 13
-  0x4f5f0981,  // 33 ^ 12
+static const uvec32 kHash16x33 = {0x92d9e201, 0, 0, 0};  // 33 ^ 16
+static const uvec32 kHashMul0 = {
+    0x0c3525e1,  // 33 ^ 15
+    0xa3476dc1,  // 33 ^ 14
+    0x3b4039a1,  // 33 ^ 13
+    0x4f5f0981,  // 33 ^ 12
 };
-static uvec32 kHashMul1 = {
-  0x30f35d61,  // 33 ^ 11
-  0x855cb541,  // 33 ^ 10
-  0x040a9121,  // 33 ^ 9
-  0x747c7101,  // 33 ^ 8
+static const uvec32 kHashMul1 = {
+    0x30f35d61,  // 33 ^ 11
+    0x855cb541,  // 33 ^ 10
+    0x040a9121,  // 33 ^ 9
+    0x747c7101,  // 33 ^ 8
 };
-static uvec32 kHashMul2 = {
-  0xec41d4e1,  // 33 ^ 7
-  0x4cfa3cc1,  // 33 ^ 6
-  0x025528a1,  // 33 ^ 5
-  0x00121881,  // 33 ^ 4
+static const uvec32 kHashMul2 = {
+    0xec41d4e1,  // 33 ^ 7
+    0x4cfa3cc1,  // 33 ^ 6
+    0x025528a1,  // 33 ^ 5
+    0x00121881,  // 33 ^ 4
 };
-static uvec32 kHashMul3 = {
-  0x00008c61,  // 33 ^ 3
-  0x00000441,  // 33 ^ 2
-  0x00000021,  // 33 ^ 1
-  0x00000001,  // 33 ^ 0
+static const uvec32 kHashMul3 = {
+    0x00008c61,  // 33 ^ 3
+    0x00000441,  // 33 ^ 2
+    0x00000021,  // 33 ^ 1
+    0x00000001,  // 33 ^ 0
 };
 
-uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
-  uint32 hash;
-  asm volatile (
-    "movd      %2,%%xmm0                       \n"
-    "pxor      %%xmm7,%%xmm7                   \n"
-    "movdqa    %4,%%xmm6                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x10, 0) ",%0          \n"
-    "pmulld    %%xmm6,%%xmm0                   \n"
-    "movdqa    %5,%%xmm5                       \n"
-    "movdqa    %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm2,%%xmm3                   \n"
-    "punpcklwd %%xmm7,%%xmm3                   \n"
-    "pmulld    %%xmm5,%%xmm3                   \n"
-    "movdqa    %6,%%xmm5                       \n"
-    "movdqa    %%xmm2,%%xmm4                   \n"
-    "punpckhwd %%xmm7,%%xmm4                   \n"
-    "pmulld    %%xmm5,%%xmm4                   \n"
-    "movdqa    %7,%%xmm5                       \n"
-    "punpckhbw %%xmm7,%%xmm1                   \n"
-    "movdqa    %%xmm1,%%xmm2                   \n"
-    "punpcklwd %%xmm7,%%xmm2                   \n"
-    "pmulld    %%xmm5,%%xmm2                   \n"
-    "movdqa    %8,%%xmm5                       \n"
-    "punpckhwd %%xmm7,%%xmm1                   \n"
-    "pmulld    %%xmm5,%%xmm1                   \n"
-    "paddd     %%xmm4,%%xmm3                   \n"
-    "paddd     %%xmm2,%%xmm1                   \n"
-    "paddd     %%xmm3,%%xmm1                   \n"
-    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
-    "paddd     %%xmm2,%%xmm1                   \n"
-    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
-    "paddd     %%xmm2,%%xmm1                   \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "sub       $0x10,%1                        \n"
-    "jg        1b                              \n"
-    "movd      %%xmm0,%3                       \n"
-  : "+r"(src),        // %0
-    "+r"(count),      // %1
-    "+rm"(seed),      // %2
-    "=g"(hash)        // %3
-  : "m"(kHash16x33),  // %4
-    "m"(kHashMul0),   // %5
-    "m"(kHashMul1),   // %6
-    "m"(kHashMul2),   // %7
-    "m"(kHashMul3)    // %8
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+uint32_t HashDjb2_SSE41(const uint8_t* src, int count, uint32_t seed) {
+  uint32_t hash;
+  asm volatile(
+      "movd        %2,%%xmm0                     \n"
+      "pxor        %%xmm7,%%xmm7                 \n"
+      "movdqa      %4,%%xmm6                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "pmulld      %%xmm6,%%xmm0                 \n"
+      "movdqa      %5,%%xmm5                     \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm7,%%xmm3                 \n"
+      "pmulld      %%xmm5,%%xmm3                 \n"
+      "movdqa      %6,%%xmm5                     \n"
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "punpckhwd   %%xmm7,%%xmm4                 \n"
+      "pmulld      %%xmm5,%%xmm4                 \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "punpckhbw   %%xmm7,%%xmm1                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklwd   %%xmm7,%%xmm2                 \n"
+      "pmulld      %%xmm5,%%xmm2                 \n"
+      "movdqa      %8,%%xmm5                     \n"
+      "punpckhwd   %%xmm7,%%xmm1                 \n"
+      "pmulld      %%xmm5,%%xmm1                 \n"
+      "paddd       %%xmm4,%%xmm3                 \n"
+      "paddd       %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm3,%%xmm1                 \n"
+      "pshufd      $0xe,%%xmm1,%%xmm2            \n"
+      "paddd       %%xmm2,%%xmm1                 \n"
+      "pshufd      $0x1,%%xmm1,%%xmm2            \n"
+      "paddd       %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "sub         $0x10,%1                      \n"
+      "jg          1b                            \n"
+      "movd        %%xmm0,%3                     \n"
+      : "+r"(src),        // %0
+        "+r"(count),      // %1
+        "+rm"(seed),      // %2
+        "=g"(hash)        // %3
+      : "m"(kHash16x33),  // %4
+        "m"(kHashMul0),   // %5
+        "m"(kHashMul1),   // %6
+        "m"(kHashMul2),   // %7
+        "m"(kHashMul3)    // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
   return hash;
 }
 #endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
@@ -148,4 +357,3 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
 }  // extern "C"
 }  // namespace libyuv
 #endif
-
diff --git a/media/libyuv/source/compare_mmi.cc b/media/libyuv/source/compare_mmi.cc
new file mode 100644
index 0000000000..7640d9468c
--- /dev/null
+++ b/media/libyuv/source/compare_mmi.cc
@@ -0,0 +1,123 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Mips MMI.
+#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+// Hakmem method for hamming distance.
+uint32_t HammingDistance_MMI(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count) {
+  uint32_t diff = 0u;
+
+  uint64_t temp = 0, temp1 = 0, ta = 0, tb = 0;
+  uint64_t c1 = 0x5555555555555555;
+  uint64_t c2 = 0x3333333333333333;
+  uint64_t c3 = 0x0f0f0f0f0f0f0f0f;
+  uint32_t c4 = 0x01010101;
+  uint64_t s1 = 1, s2 = 2, s3 = 4;
+  __asm__ volatile(
+      "1:	\n\t"
+      "ldc1   %[ta],    0(%[src_a])          \n\t"
+      "ldc1   %[tb],    0(%[src_b])          \n\t"
+      "xor    %[temp],  %[ta],      %[tb]    \n\t"
+      "psrlw  %[temp1], %[temp],    %[s1]    \n\t"  // temp1=x>>1
+      "and    %[temp1], %[temp1],   %[c1]    \n\t"  // temp1&=c1
+      "psubw  %[temp1], %[temp],    %[temp1] \n\t"  // x-temp1
+      "and    %[temp],  %[temp1],   %[c2]    \n\t"  // t = (u&c2)
+      "psrlw  %[temp1], %[temp1],   %[s2]    \n\t"  // u>>2
+      "and    %[temp1], %[temp1],   %[c2]    \n\t"  // u>>2 & c2
+      "paddw  %[temp1], %[temp1],   %[temp]  \n\t"  // t1 = t1+t
+      "psrlw  %[temp],  %[temp1],   %[s3]    \n\t"  // u>>4
+      "paddw  %[temp1], %[temp1],   %[temp]  \n\t"  // u+(u>>4)
+      "and    %[temp1], %[temp1],   %[c3]    \n\t"  //&c3
+      "dmfc1  $t0,      %[temp1]             \n\t"
+      "dsrl32 $t0,      $t0,        0        \n\t "
+      "mul    $t0,      $t0,        %[c4]    \n\t"
+      "dsrl   $t0,      $t0,        24       \n\t"
+      "dadd   %[diff],  %[diff],    $t0      \n\t"
+      "dmfc1  $t0,      %[temp1]             \n\t"
+      "mul    $t0,      $t0,        %[c4]    \n\t"
+      "dsrl   $t0,      $t0,        24       \n\t"
+      "dadd   %[diff],  %[diff],    $t0      \n\t"
+      "daddiu %[src_a], %[src_a],   8        \n\t"
+      "daddiu %[src_b], %[src_b],   8        \n\t"
+      "addiu  %[count], %[count],  -8        \n\t"
+      "bgtz   %[count], 1b \n\t"
+      "nop                            \n\t"
+      : [diff] "+r"(diff), [src_a] "+r"(src_a), [src_b] "+r"(src_b),
+        [count] "+r"(count), [ta] "+f"(ta), [tb] "+f"(tb), [temp] "+f"(temp),
+        [temp1] "+f"(temp1)
+      : [c1] "f"(c1), [c2] "f"(c2), [c3] "f"(c3), [c4] "r"(c4), [s1] "f"(s1),
+        [s2] "f"(s2), [s3] "f"(s3)
+      : "memory");
+  return diff;
+}
+
+uint32_t SumSquareError_MMI(const uint8_t* src_a,
+                            const uint8_t* src_b,
+                            int count) {
+  uint32_t sse = 0u;
+  uint32_t sse_hi = 0u, sse_lo = 0u;
+
+  uint64_t src1, src2;
+  uint64_t diff, diff_hi, diff_lo;
+  uint64_t sse_sum, sse_tmp;
+
+  const uint64_t mask = 0x0ULL;
+
+  __asm__ volatile(
+      "xor        %[sse_sum],      %[sse_sum],        %[sse_sum]    \n\t"
+
+      "1:                                                           \n\t"
+      "ldc1       %[src1],         0x00(%[src_a])                   \n\t"
+      "ldc1       %[src2],         0x00(%[src_b])                   \n\t"
+      "pasubub    %[diff],         %[src1],           %[src2]       \n\t"
+      "punpcklbh  %[diff_lo],      %[diff],           %[mask]       \n\t"
+      "punpckhbh  %[diff_hi],      %[diff],           %[mask]       \n\t"
+      "pmaddhw    %[sse_tmp],      %[diff_lo],        %[diff_lo]    \n\t"
+      "paddw      %[sse_sum],      %[sse_sum],        %[sse_tmp]    \n\t"
+      "pmaddhw    %[sse_tmp],      %[diff_hi],        %[diff_hi]    \n\t"
+      "paddw      %[sse_sum],      %[sse_sum],        %[sse_tmp]    \n\t"
+
+      "daddiu     %[src_a],        %[src_a],          0x08          \n\t"
+      "daddiu     %[src_b],        %[src_b],          0x08          \n\t"
+      "daddiu     %[count],        %[count],         -0x08          \n\t"
+      "bnez       %[count],        1b                               \n\t"
+
+      "mfc1       %[sse_lo],       %[sse_sum]                       \n\t"
+      "mfhc1      %[sse_hi],       %[sse_sum]                       \n\t"
+      "daddu      %[sse],          %[sse_hi],         %[sse_lo]     \n\t"
+      : [sse] "+&r"(sse), [diff] "=&f"(diff), [src1] "=&f"(src1),
+        [src2] "=&f"(src2), [diff_lo] "=&f"(diff_lo), [diff_hi] "=&f"(diff_hi),
+        [sse_sum] "=&f"(sse_sum), [sse_tmp] "=&f"(sse_tmp),
+        [sse_hi] "+&r"(sse_hi), [sse_lo] "+&r"(sse_lo)
+      : [src_a] "r"(src_a), [src_b] "r"(src_b), [count] "r"(count),
+        [mask] "f"(mask)
+      : "memory");
+
+  return sse;
+}
+
+#endif  // !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/compare_msa.cc b/media/libyuv/source/compare_msa.cc
new file mode 100644
index 0000000000..0b807d37be
--- /dev/null
+++ b/media/libyuv/source/compare_msa.cc
@@ -0,0 +1,97 @@
+/*
+ *  Copyright 2017 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+#include "libyuv/row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+uint32_t HammingDistance_MSA(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count) {
+  uint32_t diff = 0u;
+  int i;
+  v16u8 src0, src1, src2, src3;
+  v2i64 vec0 = {0}, vec1 = {0};
+
+  for (i = 0; i < count; i += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_a, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_a, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_b, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_b, 16);
+    src0 ^= src2;
+    src1 ^= src3;
+    vec0 += __msa_pcnt_d((v2i64)src0);
+    vec1 += __msa_pcnt_d((v2i64)src1);
+    src_a += 32;
+    src_b += 32;
+  }
+
+  vec0 += vec1;
+  diff = (uint32_t)__msa_copy_u_w((v4i32)vec0, 0);
+  diff += (uint32_t)__msa_copy_u_w((v4i32)vec0, 2);
+  return diff;
+}
+
+uint32_t SumSquareError_MSA(const uint8_t* src_a,
+                            const uint8_t* src_b,
+                            int count) {
+  uint32_t sse = 0u;
+  int i;
+  v16u8 src0, src1, src2, src3;
+  v8i16 vec0, vec1, vec2, vec3;
+  v4i32 reg0 = {0}, reg1 = {0}, reg2 = {0}, reg3 = {0};
+  v2i64 tmp0;
+
+  for (i = 0; i < count; i += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_a, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_a, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_b, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_b, 16);
+    vec0 = (v8i16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8i16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8i16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8i16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = __msa_hsub_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hsub_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hsub_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hsub_u_h((v16u8)vec3, (v16u8)vec3);
+    reg0 = __msa_dpadd_s_w(reg0, vec0, vec0);
+    reg1 = __msa_dpadd_s_w(reg1, vec1, vec1);
+    reg2 = __msa_dpadd_s_w(reg2, vec2, vec2);
+    reg3 = __msa_dpadd_s_w(reg3, vec3, vec3);
+    src_a += 32;
+    src_b += 32;
+  }
+
+  reg0 += reg1;
+  reg2 += reg3;
+  reg0 += reg2;
+  tmp0 = __msa_hadd_s_d(reg0, reg0);
+  sse = (uint32_t)__msa_copy_u_w((v4i32)tmp0, 0);
+  sse += (uint32_t)__msa_copy_u_w((v4i32)tmp0, 2);
+  return sse;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
diff --git a/media/libyuv/source/compare_neon.cc b/media/libyuv/source/compare_neon.cc
index 49aa3b4eef..afdd601216 100644
--- a/media/libyuv/source/compare_neon.cc
+++ b/media/libyuv/source/compare_neon.cc
@@ -21,40 +21,70 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
     !defined(__aarch64__)
 
-uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
-  volatile uint32 sse;
-  asm volatile (
-    "vmov.u8    q8, #0                         \n"
-    "vmov.u8    q10, #0                        \n"
-    "vmov.u8    q9, #0                         \n"
-    "vmov.u8    q11, #0                        \n"
+// 256 bits at a time
+// uses short accumulator which restricts count to 131 KB
+uint32_t HammingDistance_NEON(const uint8_t* src_a,
+                              const uint8_t* src_b,
+                              int count) {
+  uint32_t diff;
 
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"
-    MEMACCESS(1)
-    "vld1.8     {q1}, [%1]!                    \n"
-    "subs       %2, %2, #16                    \n"
-    "vsubl.u8   q2, d0, d2                     \n"
-    "vsubl.u8   q3, d1, d3                     \n"
-    "vmlal.s16  q8, d4, d4                     \n"
-    "vmlal.s16  q9, d6, d6                     \n"
-    "vmlal.s16  q10, d5, d5                    \n"
-    "vmlal.s16  q11, d7, d7                    \n"
-    "bgt        1b                             \n"
+  asm volatile(
+      "vmov.u16    q4, #0                        \n"  // accumulator
 
-    "vadd.u32   q8, q8, q9                     \n"
-    "vadd.u32   q10, q10, q11                  \n"
-    "vadd.u32   q11, q8, q10                   \n"
-    "vpaddl.u32 q1, q11                        \n"
-    "vadd.u64   d0, d2, d3                     \n"
-    "vmov.32    %3, d0[0]                      \n"
-    : "+r"(src_a),
-      "+r"(src_b),
-      "+r"(count),
-      "=r"(sse)
-    :
-    : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+      "1:                                        \n"
+      "vld1.8      {q0, q1}, [%0]!               \n"
+      "vld1.8      {q2, q3}, [%1]!               \n"
+      "veor.32     q0, q0, q2                    \n"
+      "veor.32     q1, q1, q3                    \n"
+      "vcnt.i8     q0, q0                        \n"
+      "vcnt.i8     q1, q1                        \n"
+      "subs        %2, %2, #32                   \n"
+      "vadd.u8     q0, q0, q1                    \n"  // 16 byte counts
+      "vpadal.u8   q4, q0                        \n"  // 8 shorts
+      "bgt         1b                            \n"
+
+      "vpaddl.u16  q0, q4                        \n"  // 4 ints
+      "vpadd.u32   d0, d0, d1                    \n"
+      "vpadd.u32   d0, d0, d0                    \n"
+      "vmov.32     %3, d0[0]                     \n"
+
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff)
+      :
+      : "cc", "q0", "q1", "q2", "q3", "q4");
+  return diff;
+}
+
+uint32_t SumSquareError_NEON(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count) {
+  uint32_t sse;
+  asm volatile(
+      "vmov.u8     q8, #0                        \n"
+      "vmov.u8     q10, #0                       \n"
+      "vmov.u8     q9, #0                        \n"
+      "vmov.u8     q11, #0                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"
+      "vld1.8      {q1}, [%1]!                   \n"
+      "subs        %2, %2, #16                   \n"
+      "vsubl.u8    q2, d0, d2                    \n"
+      "vsubl.u8    q3, d1, d3                    \n"
+      "vmlal.s16   q8, d4, d4                    \n"
+      "vmlal.s16   q9, d6, d6                    \n"
+      "vmlal.s16   q10, d5, d5                   \n"
+      "vmlal.s16   q11, d7, d7                   \n"
+      "bgt         1b                            \n"
+
+      "vadd.u32    q8, q8, q9                    \n"
+      "vadd.u32    q10, q10, q11                 \n"
+      "vadd.u32    q11, q8, q10                  \n"
+      "vpaddl.u32  q1, q11                       \n"
+      "vadd.u64    d0, d2, d3                    \n"
+      "vmov.32     %3, d0[0]                     \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
   return sse;
 }
 
diff --git a/media/libyuv/source/compare_neon64.cc b/media/libyuv/source/compare_neon64.cc
index f9c7df98c8..70fb9b9143 100644
--- a/media/libyuv/source/compare_neon64.cc
+++ b/media/libyuv/source/compare_neon64.cc
@@ -20,39 +20,69 @@ extern "C" {
 
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
-uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
-  volatile uint32 sse;
-  asm volatile (
-    "eor        v16.16b, v16.16b, v16.16b      \n"
-    "eor        v18.16b, v18.16b, v18.16b      \n"
-    "eor        v17.16b, v17.16b, v17.16b      \n"
-    "eor        v19.16b, v19.16b, v19.16b      \n"
+// 256 bits at a time
+// uses short accumulator which restricts count to 131 KB
+uint32_t HammingDistance_NEON(const uint8_t* src_a,
+                              const uint8_t* src_b,
+                              int count) {
+  uint32_t diff;
+  asm volatile(
+      "movi        v4.8h, #0                     \n"
 
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"
-    MEMACCESS(1)
-    "ld1        {v1.16b}, [%1], #16            \n"
-    "subs       %w2, %w2, #16                  \n"
-    "usubl      v2.8h, v0.8b, v1.8b            \n"
-    "usubl2     v3.8h, v0.16b, v1.16b          \n"
-    "smlal      v16.4s, v2.4h, v2.4h           \n"
-    "smlal      v17.4s, v3.4h, v3.4h           \n"
-    "smlal2     v18.4s, v2.8h, v2.8h           \n"
-    "smlal2     v19.4s, v3.8h, v3.8h           \n"
-    "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld1         {v0.16b, v1.16b}, [%0], #32   \n"
+      "ld1         {v2.16b, v3.16b}, [%1], #32   \n"
+      "eor         v0.16b, v0.16b, v2.16b        \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "eor         v1.16b, v1.16b, v3.16b        \n"
+      "cnt         v0.16b, v0.16b                \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "cnt         v1.16b, v1.16b                \n"
+      "subs        %w2, %w2, #32                 \n"
+      "add         v0.16b, v0.16b, v1.16b        \n"
+      "uadalp      v4.8h, v0.16b                 \n"
+      "b.gt        1b                            \n"
 
-    "add        v16.4s, v16.4s, v17.4s         \n"
-    "add        v18.4s, v18.4s, v19.4s         \n"
-    "add        v19.4s, v16.4s, v18.4s         \n"
-    "addv       s0, v19.4s                     \n"
-    "fmov       %w3, s0                        \n"
-    : "+r"(src_a),
-      "+r"(src_b),
-      "+r"(count),
-      "=r"(sse)
-    :
-    : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
+      "uaddlv      s4, v4.8h                     \n"
+      "fmov        %w3, s4                       \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff)
+      :
+      : "cc", "v0", "v1", "v2", "v3", "v4");
+  return diff;
+}
+
+uint32_t SumSquareError_NEON(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count) {
+  uint32_t sse;
+  asm volatile(
+      "eor         v16.16b, v16.16b, v16.16b     \n"
+      "eor         v18.16b, v18.16b, v18.16b     \n"
+      "eor         v17.16b, v17.16b, v17.16b     \n"
+      "eor         v19.16b, v19.16b, v19.16b     \n"
+
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"
+      "ld1         {v1.16b}, [%1], #16           \n"
+      "subs        %w2, %w2, #16                 \n"
+      "usubl       v2.8h, v0.8b, v1.8b           \n"
+      "usubl2      v3.8h, v0.16b, v1.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "smlal       v16.4s, v2.4h, v2.4h          \n"
+      "smlal       v17.4s, v3.4h, v3.4h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "smlal2      v18.4s, v2.8h, v2.8h          \n"
+      "smlal2      v19.4s, v3.8h, v3.8h          \n"
+      "b.gt        1b                            \n"
+
+      "add         v16.4s, v16.4s, v17.4s        \n"
+      "add         v18.4s, v18.4s, v19.4s        \n"
+      "add         v19.4s, v16.4s, v18.4s        \n"
+      "addv        s0, v19.4s                    \n"
+      "fmov        %w3, s0                       \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
+      :
+      : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
   return sse;
 }
 
diff --git a/media/libyuv/source/compare_win.cc b/media/libyuv/source/compare_win.cc
index dc86fe25b1..9bb27f1dd1 100644
--- a/media/libyuv/source/compare_win.cc
+++ b/media/libyuv/source/compare_win.cc
@@ -13,20 +13,40 @@
 #include "libyuv/compare_row.h"
 #include "libyuv/row.h"
 
+#if defined(_MSC_VER)
+#include <intrin.h>  // For __popcnt
+#endif
+
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for 32 bit Visual C x86 and clangcl
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+// This module is for 32 bit Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(_M_IX86)
+
+uint32_t HammingDistance_SSE42(const uint8_t* src_a,
+                               const uint8_t* src_b,
+                               int count) {
+  uint32_t diff = 0u;
+
+  int i;
+  for (i = 0; i < count - 3; i += 4) {
+    uint32_t x = *((uint32_t*)src_a) ^ *((uint32_t*)src_b);  // NOLINT
+    src_a += 4;
+    src_b += 4;
+    diff += __popcnt(x);
+  }
+  return diff;
+}
 
-__declspec(naked)
-uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+__declspec(naked) uint32_t
+    SumSquareError_SSE2(const uint8_t* src_a, const uint8_t* src_b, int count) {
   __asm {
-    mov        eax, [esp + 4]    // src_a
-    mov        edx, [esp + 8]    // src_b
-    mov        ecx, [esp + 12]   // count
+    mov        eax, [esp + 4]  // src_a
+    mov        edx, [esp + 8]  // src_b
+    mov        ecx, [esp + 12]  // count
     pxor       xmm0, xmm0
     pxor       xmm5, xmm5
 
@@ -58,16 +78,15 @@ uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
   }
 }
 
-// Visual C 2012 required for AVX2.
-#if _MSC_VER >= 1700
+#ifdef HAS_SUMSQUAREERROR_AVX2
 // C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
-#pragma warning(disable: 4752)
-__declspec(naked)
-uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+#pragma warning(disable : 4752)
+__declspec(naked) uint32_t
+    SumSquareError_AVX2(const uint8_t* src_a, const uint8_t* src_b, int count) {
   __asm {
-    mov        eax, [esp + 4]    // src_a
-    mov        edx, [esp + 8]    // src_b
-    mov        ecx, [esp + 12]   // count
+    mov        eax, [esp + 4]  // src_a
+    mov        edx, [esp + 8]  // src_b
+    mov        ecx, [esp + 12]  // count
     vpxor      ymm0, ymm0, ymm0  // sum
     vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
     sub        edx, eax
@@ -99,67 +118,67 @@ uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
     ret
   }
 }
-#endif  // _MSC_VER >= 1700
+#endif  // HAS_SUMSQUAREERROR_AVX2
 
-uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+uvec32 kHash16x33 = {0x92d9e201, 0, 0, 0};  // 33 ^ 16
 uvec32 kHashMul0 = {
-  0x0c3525e1,  // 33 ^ 15
-  0xa3476dc1,  // 33 ^ 14
-  0x3b4039a1,  // 33 ^ 13
-  0x4f5f0981,  // 33 ^ 12
+    0x0c3525e1,  // 33 ^ 15
+    0xa3476dc1,  // 33 ^ 14
+    0x3b4039a1,  // 33 ^ 13
+    0x4f5f0981,  // 33 ^ 12
 };
 uvec32 kHashMul1 = {
-  0x30f35d61,  // 33 ^ 11
-  0x855cb541,  // 33 ^ 10
-  0x040a9121,  // 33 ^ 9
-  0x747c7101,  // 33 ^ 8
+    0x30f35d61,  // 33 ^ 11
+    0x855cb541,  // 33 ^ 10
+    0x040a9121,  // 33 ^ 9
+    0x747c7101,  // 33 ^ 8
 };
 uvec32 kHashMul2 = {
-  0xec41d4e1,  // 33 ^ 7
-  0x4cfa3cc1,  // 33 ^ 6
-  0x025528a1,  // 33 ^ 5
-  0x00121881,  // 33 ^ 4
+    0xec41d4e1,  // 33 ^ 7
+    0x4cfa3cc1,  // 33 ^ 6
+    0x025528a1,  // 33 ^ 5
+    0x00121881,  // 33 ^ 4
 };
 uvec32 kHashMul3 = {
-  0x00008c61,  // 33 ^ 3
-  0x00000441,  // 33 ^ 2
-  0x00000021,  // 33 ^ 1
-  0x00000001,  // 33 ^ 0
+    0x00008c61,  // 33 ^ 3
+    0x00000441,  // 33 ^ 2
+    0x00000021,  // 33 ^ 1
+    0x00000001,  // 33 ^ 0
 };
 
-__declspec(naked)
-uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+__declspec(naked) uint32_t
+    HashDjb2_SSE41(const uint8_t* src, int count, uint32_t seed) {
   __asm {
-    mov        eax, [esp + 4]    // src
-    mov        ecx, [esp + 8]    // count
+    mov        eax, [esp + 4]  // src
+    mov        ecx, [esp + 8]  // count
     movd       xmm0, [esp + 12]  // seed
 
-    pxor       xmm7, xmm7        // constant 0 for unpck
+    pxor       xmm7, xmm7  // constant 0 for unpck
     movdqa     xmm6, xmmword ptr kHash16x33
 
   wloop:
-    movdqu     xmm1, [eax]       // src[0-15]
+    movdqu     xmm1, [eax]  // src[0-15]
     lea        eax, [eax + 16]
-    pmulld     xmm0, xmm6        // hash *= 33 ^ 16
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
     movdqa     xmm5, xmmword ptr kHashMul0
     movdqa     xmm2, xmm1
-    punpcklbw  xmm2, xmm7        // src[0-7]
+    punpcklbw  xmm2, xmm7  // src[0-7]
     movdqa     xmm3, xmm2
-    punpcklwd  xmm3, xmm7        // src[0-3]
+    punpcklwd  xmm3, xmm7  // src[0-3]
     pmulld     xmm3, xmm5
     movdqa     xmm5, xmmword ptr kHashMul1
     movdqa     xmm4, xmm2
-    punpckhwd  xmm4, xmm7        // src[4-7]
+    punpckhwd  xmm4, xmm7  // src[4-7]
     pmulld     xmm4, xmm5
     movdqa     xmm5, xmmword ptr kHashMul2
-    punpckhbw  xmm1, xmm7        // src[8-15]
+    punpckhbw  xmm1, xmm7  // src[8-15]
     movdqa     xmm2, xmm1
-    punpcklwd  xmm2, xmm7        // src[8-11]
+    punpcklwd  xmm2, xmm7  // src[8-11]
     pmulld     xmm2, xmm5
     movdqa     xmm5, xmmword ptr kHashMul3
-    punpckhwd  xmm1, xmm7        // src[12-15]
+    punpckhwd  xmm1, xmm7  // src[12-15]
     pmulld     xmm1, xmm5
-    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm3, xmm4  // add 16 results
     paddd      xmm1, xmm2
     paddd      xmm1, xmm3
 
@@ -171,18 +190,18 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     sub        ecx, 16
     jg         wloop
 
-    movd       eax, xmm0         // return hash
+    movd       eax, xmm0  // return hash
     ret
   }
 }
 
 // Visual C 2012 required for AVX2.
-#if _MSC_VER >= 1700
-__declspec(naked)
-uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+#ifdef HAS_HASHDJB2_AVX2
+__declspec(naked) uint32_t
+    HashDjb2_AVX2(const uint8_t* src, int count, uint32_t seed) {
   __asm {
-    mov        eax, [esp + 4]    // src
-    mov        ecx, [esp + 8]    // count
+    mov        eax, [esp + 4]  // src
+    mov        ecx, [esp + 8]  // count
     vmovd      xmm0, [esp + 12]  // seed
 
   wloop:
@@ -196,7 +215,7 @@ uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
     vpmulld    xmm2, xmm2, xmmword ptr kHashMul2
     lea        eax, [eax + 16]
     vpmulld    xmm1, xmm1, xmmword ptr kHashMul3
-    vpaddd     xmm3, xmm3, xmm4        // add 16 results
+    vpaddd     xmm3, xmm3, xmm4  // add 16 results
     vpaddd     xmm1, xmm1, xmm2
     vpaddd     xmm1, xmm1, xmm3
     vpshufd    xmm2, xmm1, 0x0e  // upper 2 dwords
@@ -207,12 +226,12 @@ uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
     sub        ecx, 16
     jg         wloop
 
-    vmovd      eax, xmm0         // return hash
+    vmovd      eax, xmm0  // return hash
     vzeroupper
     ret
   }
 }
-#endif  // _MSC_VER >= 1700
+#endif  // HAS_HASHDJB2_AVX2
 
 #endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
 
diff --git a/media/libyuv/source/convert.cc b/media/libyuv/source/convert.cc
index e332bc505c..37b7091b10 100644
--- a/media/libyuv/source/convert.cc
+++ b/media/libyuv/source/convert.cc
@@ -14,64 +14,82 @@
 #include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
-#include "libyuv/scale.h"  // For ScalePlane()
 #include "libyuv/row.h"
+#include "libyuv/scale.h"      // For ScalePlane()
+#include "libyuv/scale_row.h"  // For FixedDiv
+#include "libyuv/scale_uv.h"   // For UVScale()
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
+// Subsample amount uses a shift.
+//   v is value
+//   a is amount to add to round up
+//   s is shift to subsample down
 #define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
 static __inline int Abs(int v) {
   return v >= 0 ? v : -v;
 }
 
 // Any I4xx To I420 format with mirroring.
-static int I4xxToI420(const uint8* src_y, int src_stride_y,
-                      const uint8* src_u, int src_stride_u,
-                      const uint8* src_v, int src_stride_v,
-                      uint8* dst_y, int dst_stride_y,
-                      uint8* dst_u, int dst_stride_u,
-                      uint8* dst_v, int dst_stride_v,
-                      int src_y_width, int src_y_height,
-                      int src_uv_width, int src_uv_height) {
+static int I4xxToI420(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_y,
+                      int dst_stride_y,
+                      uint8_t* dst_u,
+                      int dst_stride_u,
+                      uint8_t* dst_v,
+                      int dst_stride_v,
+                      int src_y_width,
+                      int src_y_height,
+                      int src_uv_width,
+                      int src_uv_height) {
   const int dst_y_width = Abs(src_y_width);
   const int dst_y_height = Abs(src_y_height);
   const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
   const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
-  if (src_y_width == 0 || src_y_height == 0 ||
-      src_uv_width == 0 || src_uv_height == 0) {
+  if (src_uv_width <= 0 || src_uv_height == 0) {
     return -1;
   }
-  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
-             dst_y, dst_stride_y, dst_y_width, dst_y_height,
-             kFilterBilinear);
-  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
-             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
-             kFilterBilinear);
-  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
-             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
-             kFilterBilinear);
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, src_y_width, src_y_height, dst_y,
+               dst_stride_y, dst_y_width, dst_y_height, kFilterBilinear);
+  }
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height, dst_u,
+             dst_stride_u, dst_uv_width, dst_uv_height, kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height, dst_v,
+             dst_stride_v, dst_uv_width, dst_uv_height, kFilterBilinear);
   return 0;
 }
 
-// Copy I420 with optional flipping
+// Copy I420 with optional flipping.
 // TODO(fbarchard): Use Scale plane which supports mirroring, but ensure
 // is does row coalescing.
 LIBYUV_API
-int I420Copy(const uint8* src_y, int src_stride_y,
-             const uint8* src_u, int src_stride_u,
-             const uint8* src_v, int src_stride_v,
-             uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int width, int height) {
+int I420Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_u,
+             int src_stride_u,
+             const uint8_t* src_v,
+             int src_stride_v,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_u || !src_v ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -95,304 +113,1305 @@ int I420Copy(const uint8* src_y, int src_stride_y,
   return 0;
 }
 
-// 422 chroma is 1/2 width, 1x height
-// 420 chroma is 1/2 width, 1/2 height
+// Copy I010 with optional flipping.
 LIBYUV_API
-int I422ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  const int src_uv_width = SUBSAMPLE(width, 1, 1);
-  return I4xxToI420(src_y, src_stride_y,
-                    src_u, src_stride_u,
-                    src_v, src_stride_v,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height,
-                    src_uv_width, height);
+int I010Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
 }
 
-// 444 chroma is 1x width, 1x height
-// 420 chroma is 1/2 width, 1/2 height
+static int Planar16bitTo8bit(const uint16_t* src_y,
+                             int src_stride_y,
+                             const uint16_t* src_u,
+                             int src_stride_u,
+                             const uint16_t* src_v,
+                             int src_stride_v,
+                             uint8_t* dst_y,
+                             int dst_stride_y,
+                             uint8_t* dst_u,
+                             int dst_stride_u,
+                             uint8_t* dst_v,
+                             int dst_stride_v,
+                             int width,
+                             int height,
+                             int subsample_x,
+                             int subsample_y,
+                             int depth) {
+  int uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  int uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  int scale = 1 << (24 - depth);
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    uv_height = -uv_height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (uv_height - 1) * src_stride_u;
+    src_v = src_v + (uv_height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert16To8Plane(src_y, src_stride_y, dst_y, dst_stride_y, scale, width,
+                    height);
+  // Convert UV planes.
+  Convert16To8Plane(src_u, src_stride_u, dst_u, dst_stride_u, scale, uv_width,
+                    uv_height);
+  Convert16To8Plane(src_v, src_stride_v, dst_v, dst_stride_v, scale, uv_width,
+                    uv_height);
+  return 0;
+}
+
+// Convert 10 bit YUV to 8 bit.
 LIBYUV_API
-int I444ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  return I4xxToI420(src_y, src_stride_y,
-                    src_u, src_stride_u,
-                    src_v, src_stride_v,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height,
-                    width, height);
+int I010ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           1, 10);
+}
+
+LIBYUV_API
+int I210ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  const int depth = 10;
+  const int scale = 1 << (24 - depth);
+
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  {
+    const int uv_width = SUBSAMPLE(width, 1, 1);
+    const int uv_height = SUBSAMPLE(height, 1, 1);
+    const int dy = FixedDiv(height, uv_height);
+
+    Convert16To8Plane(src_y, src_stride_y, dst_y, dst_stride_y, scale, width,
+                      height);
+    ScalePlaneVertical_16To8(height, uv_width, uv_height, src_stride_u,
+                             dst_stride_u, src_u, dst_u, 0, 32768, dy,
+                             /*bpp=*/1, scale, kFilterBilinear);
+    ScalePlaneVertical_16To8(height, uv_width, uv_height, src_stride_v,
+                             dst_stride_v, src_v, dst_v, 0, 32768, dy,
+                             /*bpp=*/1, scale, kFilterBilinear);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I210ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           0, 10);
+}
+
+LIBYUV_API
+int I410ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  const int depth = 10;
+  const int scale = 1 << (24 - depth);
+
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  {
+    const int uv_width = SUBSAMPLE(width, 1, 1);
+    const int uv_height = SUBSAMPLE(height, 1, 1);
+
+    Convert16To8Plane(src_y, src_stride_y, dst_y, dst_stride_y, scale, width,
+                      height);
+    ScalePlaneDown2_16To8(width, height, uv_width, uv_height, src_stride_u,
+                          dst_stride_u, src_u, dst_u, scale, kFilterBilinear);
+    ScalePlaneDown2_16To8(width, height, uv_width, uv_height, src_stride_v,
+                          dst_stride_v, src_v, dst_v, scale, kFilterBilinear);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I410ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 0,
+                           0, 10);
+}
+
+LIBYUV_API
+int I012ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           1, 12);
+}
+
+LIBYUV_API
+int I212ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           0, 12);
+}
+
+LIBYUV_API
+int I412ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 0,
+                           0, 12);
+}
+
+// Any Ix10 To I010 format with mirroring.
+static int Ix10ToI010(const uint16_t* src_y,
+                      int src_stride_y,
+                      const uint16_t* src_u,
+                      int src_stride_u,
+                      const uint16_t* src_v,
+                      int src_stride_v,
+                      uint16_t* dst_y,
+                      int dst_stride_y,
+                      uint16_t* dst_u,
+                      int dst_stride_u,
+                      uint16_t* dst_v,
+                      int dst_stride_v,
+                      int width,
+                      int height,
+                      int subsample_x,
+                      int subsample_y) {
+  const int dst_y_width = Abs(width);
+  const int dst_y_height = Abs(height);
+  const int src_uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  const int src_uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
+  const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+  if (dst_y) {
+    ScalePlane_12(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  dst_y_width, dst_y_height, kFilterBilinear);
+  }
+  ScalePlane_12(src_u, src_stride_u, src_uv_width, src_uv_height, dst_u,
+                dst_stride_u, dst_uv_width, dst_uv_height, kFilterBilinear);
+  ScalePlane_12(src_v, src_stride_v, src_uv_width, src_uv_height, dst_v,
+                dst_stride_v, dst_uv_width, dst_uv_height, kFilterBilinear);
+  return 0;
+}
+
+LIBYUV_API
+int I410ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Ix10ToI010(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, 0, 0);
+}
+
+LIBYUV_API
+int I210ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Ix10ToI010(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, 1, 0);
+}
+
+// Any I[420]1[02] to P[420]1[02] format with mirroring.
+static int IxxxToPxxx(const uint16_t* src_y,
+                      int src_stride_y,
+                      const uint16_t* src_u,
+                      int src_stride_u,
+                      const uint16_t* src_v,
+                      int src_stride_v,
+                      uint16_t* dst_y,
+                      int dst_stride_y,
+                      uint16_t* dst_uv,
+                      int dst_stride_uv,
+                      int width,
+                      int height,
+                      int subsample_x,
+                      int subsample_y,
+                      int depth) {
+  const int uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  const int uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  ConvertToMSBPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height,
+                       depth);
+  MergeUVPlane_16(src_u, src_stride_u, src_v, src_stride_v, dst_uv,
+                  dst_stride_uv, uv_width, uv_height, depth);
+  return 0;
+}
+
+LIBYUV_API
+int I010ToP010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 1, 10);
 }
 
-// 411 chroma is 1/4 width, 1x height
+LIBYUV_API
+int I210ToP210(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 0, 10);
+}
+
+LIBYUV_API
+int I012ToP012(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 1, 12);
+}
+
+LIBYUV_API
+int I212ToP212(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 0, 12);
+}
+
+// 422 chroma is 1/2 width, 1x height
 // 420 chroma is 1/2 width, 1/2 height
 LIBYUV_API
-int I411ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  const int src_uv_width = SUBSAMPLE(width, 3, 2);
-  return I4xxToI420(src_y, src_stride_y,
-                    src_u, src_stride_u,
-                    src_v, src_stride_v,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height,
-                    src_uv_width, height);
+int I422ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  const int src_uv_width = SUBSAMPLE(width, 1, 1);
+  return I4xxToI420(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, src_uv_width, height);
 }
 
-// I400 is greyscale typically used in MJPG
 LIBYUV_API
-int I400ToI420(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int I422ToI210(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int halfwidth = (width + 1) >> 1;
-  int halfheight = (height + 1) >> 1;
-  if (!src_y || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert8To16Plane(src_y, src_stride_y, dst_y, dst_stride_y, 1024, width,
+                    height);
+  // Convert UV planes.
+  Convert8To16Plane(src_u, src_stride_u, dst_u, dst_stride_u, 1024, halfwidth,
+                    height);
+  Convert8To16Plane(src_v, src_stride_v, dst_v, dst_stride_v, 1024, halfwidth,
+                    height);
+  return 0;
+}
+
+// TODO(fbarchard): Implement row conversion.
+LIBYUV_API
+int I422ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
     halfheight = (height + 1) >> 1;
     src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
     src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
   }
-  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-  SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128);
-  SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128);
+
+  // Allocate u and v buffers
+  align_buffer_64(plane_u, halfwidth * halfheight * 2);
+  uint8_t* plane_v = plane_u + halfwidth * halfheight;
+
+  I422ToI420(src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+             dst_y, dst_stride_y, plane_u, halfwidth, plane_v, halfwidth, width,
+             height);
+  MergeUVPlane(plane_v, halfwidth, plane_u, halfwidth, dst_vu, dst_stride_vu,
+               halfwidth, halfheight);
+  free_aligned_buffer_64(plane_u);
   return 0;
 }
 
-static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
-                       uint8* dst, int dst_stride,
-                       int width, int height) {
-  int y;
-  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
-#if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+LIBYUV_API
+int MM21ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (!src_uv || !dst_uv || width <= 0) {
+    return -1;
   }
-#endif
-#if defined(HAS_COPYROW_AVX)
-  if (TestCpuFlag(kCpuHasAVX)) {
-    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+
+  int sign = height < 0 ? -1 : 1;
+
+  if (dst_y) {
+    DetilePlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height, 32);
   }
-#endif
-#if defined(HAS_COPYROW_ERMS)
-  if (TestCpuFlag(kCpuHasERMS)) {
-    CopyRow = CopyRow_ERMS;
+  DetilePlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, (width + 1) & ~1,
+              (height + sign) / 2, 16);
+
+  return 0;
+}
+
+LIBYUV_API
+int MM21ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int sign = height < 0 ? -1 : 1;
+
+  if (!src_uv || !dst_u || !dst_v || width <= 0) {
+    return -1;
   }
-#endif
-#if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+
+  if (dst_y) {
+    DetilePlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height, 32);
   }
-#endif
-#if defined(HAS_COPYROW_MIPS)
-  if (TestCpuFlag(kCpuHasMIPS)) {
-    CopyRow = CopyRow_MIPS;
+  DetileSplitUVPlane(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                     dst_stride_v, (width + 1) & ~1, (height + sign) / 2, 16);
+
+  return 0;
+}
+
+LIBYUV_API
+int MM21ToYUY2(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height) {
+  if (!src_y || !src_uv || !dst_yuy2 || width <= 0) {
+    return -1;
   }
-#endif
 
-  // Copy plane
-  for (y = 0; y < height - 1; y += 2) {
-    CopyRow(src, dst, width);
-    CopyRow(src + src_stride_0, dst + dst_stride, width);
-    src += src_stride_0 + src_stride_1;
-    dst += dst_stride * 2;
+  DetileToYUY2(src_y, src_stride_y, src_uv, src_stride_uv, dst_yuy2,
+               dst_stride_yuy2, width, height, 32);
+
+  return 0;
+}
+
+// Convert MT2T into P010. See tinyurl.com/mtk-10bit-video-format for format
+// documentation.
+// TODO(greenjustin): Add an MT2T to I420 conversion.
+LIBYUV_API
+int MT2TToP010(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || !height || !src_uv || !dst_uv) {
+    return -1;
   }
-  if (height & 1) {
-    CopyRow(src, dst, width);
+
+  {
+    int u_width = (width + 1) / 2;
+    int uv_width = 2 * u_width;
+    int y = 0;
+    int uv_height = uv_height = (height + 1) / 2;
+    const int tile_width = 16;
+    const int y_tile_height = 32;
+    const int uv_tile_height = 16;
+    int padded_width = (width + tile_width - 1) & ~(tile_width - 1);
+    int y_tile_row_size = padded_width * y_tile_height * 10 / 8;
+    int uv_tile_row_size = padded_width * uv_tile_height * 10 / 8;
+    size_t row_buf_size = padded_width * y_tile_height * sizeof(uint16_t);
+    void (*UnpackMT2T)(const uint8_t* src, uint16_t* dst, size_t size) =
+        UnpackMT2T_C;
+    align_buffer_64(row_buf, row_buf_size);
+
+#if defined(HAS_UNPACKMT2T_NEON)
+    if (TestCpuFlag(kCpuHasNEON)) {
+      UnpackMT2T = UnpackMT2T_NEON;
+    }
+#endif
+    // Negative height means invert the image.
+    if (height < 0) {
+      height = -height;
+      uv_height = (height + 1) / 2;
+      if (dst_y) {
+        dst_y = dst_y + (height - 1) * dst_stride_y;
+        dst_stride_y = -dst_stride_y;
+      }
+      dst_uv = dst_uv + (uv_height - 1) * dst_stride_uv;
+      dst_stride_uv = -dst_stride_uv;
+    }
+
+    // Unpack and detile Y in rows of tiles
+    if (src_y && dst_y) {
+      for (y = 0; y < (height & ~(y_tile_height - 1)); y += y_tile_height) {
+        UnpackMT2T(src_y, (uint16_t*)row_buf, y_tile_row_size);
+        DetilePlane_16((uint16_t*)row_buf, padded_width, dst_y, dst_stride_y,
+                       width, y_tile_height, y_tile_height);
+        src_y += src_stride_y * y_tile_height;
+        dst_y += dst_stride_y * y_tile_height;
+      }
+      if (height & (y_tile_height - 1)) {
+        UnpackMT2T(src_y, (uint16_t*)row_buf, y_tile_row_size);
+        DetilePlane_16((uint16_t*)row_buf, padded_width, dst_y, dst_stride_y,
+                       width, height & (y_tile_height - 1), y_tile_height);
+      }
+    }
+
+    // Unpack and detile UV plane
+    for (y = 0; y < (uv_height & ~(uv_tile_height - 1)); y += uv_tile_height) {
+      UnpackMT2T(src_uv, (uint16_t*)row_buf, uv_tile_row_size);
+      DetilePlane_16((uint16_t*)row_buf, padded_width, dst_uv, dst_stride_uv,
+                     uv_width, uv_tile_height, uv_tile_height);
+      src_uv += src_stride_uv * uv_tile_height;
+      dst_uv += dst_stride_uv * uv_tile_height;
+    }
+    if (uv_height & (uv_tile_height - 1)) {
+      UnpackMT2T(src_uv, (uint16_t*)row_buf, uv_tile_row_size);
+      DetilePlane_16((uint16_t*)row_buf, padded_width, dst_uv, dst_stride_uv,
+                     uv_width, uv_height & (uv_tile_height - 1),
+                     uv_tile_height);
+    }
+    free_aligned_buffer_64(row_buf);
   }
+  return 0;
 }
 
-// Support converting from FOURCC_M420
-// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for
-// easy conversion to I420.
-// M420 format description:
-// M420 is row biplanar 420: 2 rows of Y and 1 row of UV.
-// Chroma is half width / half height. (420)
-// src_stride_m420 is row planar. Normally this will be the width in pixels.
-//   The UV plane is half width, but 2 values, so src_stride_m420 applies to
-//   this as well as the two Y planes.
-static int X420ToI420(const uint8* src_y,
-                      int src_stride_y0, int src_stride_y1,
-                      const uint8* src_uv, int src_stride_uv,
-                      uint8* dst_y, int dst_stride_y,
-                      uint8* dst_u, int dst_stride_u,
-                      uint8* dst_v, int dst_stride_v,
-                      int width, int height) {
+#ifdef I422TONV21_ROW_VERSION
+// Unittest fails for this version.
+// 422 chroma is 1/2 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+// Swap src_u and src_v to implement I422ToNV12
+LIBYUV_API
+int I422ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
   int y;
+  void (*MergeUVRow)(const uint8_t* src_u, const uint8_t* src_v,
+                     uint8_t* dst_uv, int width) = MergeUVRow_C;
+  void (*InterpolateRow)(uint8_t* dst_ptr, const uint8_t* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) = SplitUVRow_C;
-  if (!src_y || !src_uv ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_u || !src_v || !dst_vu || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     halfheight = (height + 1) >> 1;
-    dst_y = dst_y + (height - 1) * dst_stride_y;
-    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
-    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
-    dst_stride_y = -dst_stride_y;
-    dst_stride_u = -dst_stride_u;
-    dst_stride_v = -dst_stride_v;
-  }
-  // Coalesce rows.
-  if (src_stride_y0 == width &&
-      src_stride_y1 == width &&
-      dst_stride_y == width) {
-    width *= height;
-    height = 1;
-    src_stride_y0 = src_stride_y1 = dst_stride_y = 0;
-  }
-  // Coalesce rows.
-  if (src_stride_uv == halfwidth * 2 &&
-      dst_stride_u == halfwidth &&
-      dst_stride_v == halfwidth) {
-    halfwidth *= halfheight;
-    halfheight = 1;
-    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
   }
-#if defined(HAS_SPLITUVROW_SSE2)
+#if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    SplitUVRow = SplitUVRow_Any_SSE2;
+    MergeUVRow = MergeUVRow_Any_SSE2;
     if (IS_ALIGNED(halfwidth, 16)) {
-      SplitUVRow = SplitUVRow_SSE2;
+      MergeUVRow = MergeUVRow_SSE2;
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_AVX2)
+#if defined(HAS_MERGEUVROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    SplitUVRow = SplitUVRow_Any_AVX2;
+    MergeUVRow = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow = MergeUVRow_Any_AVX512BW;
     if (IS_ALIGNED(halfwidth, 32)) {
-      SplitUVRow = SplitUVRow_AVX2;
+      MergeUVRow = MergeUVRow_AVX512BW;
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_NEON)
+#if defined(HAS_MERGEUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    SplitUVRow = SplitUVRow_Any_NEON;
+    MergeUVRow = MergeUVRow_Any_NEON;
     if (IS_ALIGNED(halfwidth, 16)) {
-      SplitUVRow = SplitUVRow_NEON;
+      MergeUVRow = MergeUVRow_NEON;
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
-      IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
-      IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
-    SplitUVRow = SplitUVRow_Any_DSPR2;
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow = MergeUVRow_Any_MSA;
     if (IS_ALIGNED(halfwidth, 16)) {
-      SplitUVRow = SplitUVRow_DSPR2;
+      MergeUVRow = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow = MergeUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
 
   if (dst_y) {
-    if (src_stride_y0 == src_stride_y1) {
-      CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height);
-    } else {
-      CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y,
-                 width, height);
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, halfwidth, height);
+  }
+  {
+    // Allocate 2 rows of vu.
+    int awidth = halfwidth * 2;
+    align_buffer_64(row_vu_0, awidth * 2);
+    uint8_t* row_vu_1 = row_vu_0 + awidth;
+
+    for (y = 0; y < height - 1; y += 2) {
+      MergeUVRow(src_v, src_u, row_vu_0, halfwidth);
+      MergeUVRow(src_v + src_stride_v, src_u + src_stride_u, row_vu_1,
+                 halfwidth);
+      InterpolateRow(dst_vu, row_vu_0, awidth, awidth, 128);
+      src_u += src_stride_u * 2;
+      src_v += src_stride_v * 2;
+      dst_vu += dst_stride_vu;
+    }
+    if (height & 1) {
+      MergeUVRow(src_v, src_u, dst_vu, halfwidth);
     }
+    free_aligned_buffer_64(row_vu_0);
   }
+  return 0;
+}
+#endif  // I422TONV21_ROW_VERSION
 
-  for (y = 0; y < halfheight; ++y) {
-    // Copy a row of UV.
-    SplitUVRow(src_uv, dst_u, dst_v, halfwidth);
-    dst_u += dst_stride_u;
-    dst_v += dst_stride_v;
-    src_uv += src_stride_uv;
+// 444 chroma is 1x width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I444ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return I4xxToI420(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, width, height);
+}
+
+LIBYUV_API
+int I444ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (!src_y || !src_u || !src_v || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
   }
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  HalfMergeUVPlane(src_u, src_stride_u, src_v, src_stride_v, dst_uv,
+                   dst_stride_uv, width, height);
   return 0;
 }
 
-// Convert NV12 to I420.
 LIBYUV_API
-int NV12ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_uv, int src_stride_uv,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  return X420ToI420(src_y, src_stride_y, src_stride_y,
-                    src_uv, src_stride_uv,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
+int I444ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  return I444ToNV12(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                    src_stride_u, dst_y, dst_stride_y, dst_vu, dst_stride_vu,
                     width, height);
 }
 
+// I400 is greyscale typically used in MJPG
+LIBYUV_API
+int I400ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128);
+  SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128);
+  return 0;
+}
+
+// I400 is greyscale typically used in MJPG
+LIBYUV_API
+int I400ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!dst_vu || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  SetPlane(dst_vu, dst_stride_vu, halfwidth * 2, halfheight, 128);
+  return 0;
+}
+
+// Convert NV12 to I420.
+// TODO(fbarchard): Consider inverting destination. Faster on ARM with prfm.
+LIBYUV_API
+int NV12ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_uv || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == halfwidth * 2 && dst_stride_u == halfwidth &&
+      dst_stride_v == halfwidth) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+
+  // Split UV plane - NV12 / NV21
+  SplitUVPlane(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v, dst_stride_v,
+               halfwidth, halfheight);
+
+  return 0;
+}
+
 // Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
 LIBYUV_API
-int NV21ToI420(const uint8* src_y, int src_stride_y,
-               const uint8* src_vu, int src_stride_vu,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  return X420ToI420(src_y, src_stride_y, src_stride_y,
-                    src_vu, src_stride_vu,
-                    dst_y, dst_stride_y,
-                    dst_v, dst_stride_v,
-                    dst_u, dst_stride_u,
+int NV21ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return NV12ToI420(src_y, src_stride_y, src_vu, src_stride_vu, dst_y,
+                    dst_stride_y, dst_v, dst_stride_v, dst_u, dst_stride_u,
                     width, height);
 }
 
-// Convert M420 to I420.
 LIBYUV_API
-int M420ToI420(const uint8* src_m420, int src_stride_m420,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2,
-                    src_m420 + src_stride_m420 * 2, src_stride_m420 * 3,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height);
+int NV12ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+               Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1),
+          SUBSAMPLE(height, 1, 1), dst_uv, dst_stride_uv, Abs(width),
+          Abs(height), kFilterBilinear);
+  return 0;
+}
+
+LIBYUV_API
+int NV16ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+               Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1), height, dst_uv,
+          dst_stride_uv, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
+}
+
+// Any P[420]1[02] to I[420]1[02] format with mirroring.
+static int PxxxToIxxx(const uint16_t* src_y,
+                      int src_stride_y,
+                      const uint16_t* src_uv,
+                      int src_stride_uv,
+                      uint16_t* dst_y,
+                      int dst_stride_y,
+                      uint16_t* dst_u,
+                      int dst_stride_u,
+                      uint16_t* dst_v,
+                      int dst_stride_v,
+                      int width,
+                      int height,
+                      int subsample_x,
+                      int subsample_y,
+                      int depth) {
+  const int uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  const int uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+  ConvertToLSBPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height,
+                       depth);
+  SplitUVPlane_16(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                  dst_stride_v, uv_width, uv_height, depth);
+  return 0;
+}
+
+LIBYUV_API
+int P010ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return PxxxToIxxx(src_y, src_stride_y, src_uv, src_stride_uv, dst_y,
+                    dst_stride_y, dst_u, dst_stride_u, dst_v, dst_stride_v,
+                    width, height, 1, 1, 10);
+}
+
+LIBYUV_API
+int P012ToI012(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return PxxxToIxxx(src_y, src_stride_y, src_uv, src_stride_uv, dst_y,
+                    dst_stride_y, dst_u, dst_stride_u, dst_v, dst_stride_v,
+                    width, height, 1, 1, 12);
+}
+
+LIBYUV_API
+int P010ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_16(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale_16(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1),
+             SUBSAMPLE(height, 1, 1), dst_uv, dst_stride_uv, Abs(width),
+             Abs(height), kFilterBilinear);
+  return 0;
+}
+
+LIBYUV_API
+int P210ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_16(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale_16(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1), height, dst_uv,
+             dst_stride_uv, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
 }
 
 // Convert YUY2 to I420.
 LIBYUV_API
-int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int YUY2ToI420(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*YUY2ToUVRow)(const uint8* src_yuy2, int src_stride_yuy2,
-      uint8* dst_u, uint8* dst_v, int width) = YUY2ToUVRow_C;
-  void (*YUY2ToYRow)(const uint8* src_yuy2,
-      uint8* dst_y, int width) = YUY2ToYRow_C;
+  void (*YUY2ToUVRow)(const uint8_t* src_yuy2, int src_stride_yuy2,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      YUY2ToUVRow_C;
+  void (*YUY2ToYRow)(const uint8_t* src_yuy2, uint8_t* dst_y, int width) =
+      YUY2ToYRow_C;
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -429,6 +1448,26 @@ int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
     }
   }
 #endif
+#if defined(HAS_YUY2TOYROW_MSA) && defined(HAS_YUY2TOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    YUY2ToYRow = YUY2ToYRow_Any_MSA;
+    YUY2ToUVRow = YUY2ToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_MSA;
+      YUY2ToUVRow = YUY2ToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_LASX) && defined(HAS_YUY2TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    YUY2ToYRow = YUY2ToYRow_Any_LASX;
+    YUY2ToUVRow = YUY2ToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_LASX;
+      YUY2ToUVRow = YUY2ToUVRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height - 1; y += 2) {
     YUY2ToUVRow(src_yuy2, src_stride_yuy2, dst_u, dst_v, width);
@@ -448,16 +1487,22 @@ int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
 
 // Convert UYVY to I420.
 LIBYUV_API
-int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int UYVYToI420(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*UYVYToUVRow)(const uint8* src_uyvy, int src_stride_uyvy,
-      uint8* dst_u, uint8* dst_v, int width) = UYVYToUVRow_C;
-  void (*UYVYToYRow)(const uint8* src_uyvy,
-      uint8* dst_y, int width) = UYVYToYRow_C;
+  void (*UYVYToUVRow)(const uint8_t* src_uyvy, int src_stride_uyvy,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      UYVYToUVRow_C;
+  void (*UYVYToYRow)(const uint8_t* src_uyvy, uint8_t* dst_y, int width) =
+      UYVYToYRow_C;
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -494,6 +1539,26 @@ int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
     }
   }
 #endif
+#if defined(HAS_UYVYTOYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    UYVYToYRow = UYVYToYRow_Any_MSA;
+    UYVYToUVRow = UYVYToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_MSA;
+      UYVYToUVRow = UYVYToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    UYVYToYRow = UYVYToYRow_Any_LASX;
+    UYVYToUVRow = UYVYToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_LASX;
+      UYVYToUVRow = UYVYToUVRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height - 1; y += 2) {
     UYVYToUVRow(src_uyvy, src_stride_uyvy, dst_u, dst_v, width);
@@ -511,21 +1576,163 @@ int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
   return 0;
 }
 
+// Convert AYUV to NV12.
+LIBYUV_API
+int AYUVToNV12(const uint8_t* src_ayuv,
+               int src_stride_ayuv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  int y;
+  void (*AYUVToUVRow)(const uint8_t* src_ayuv, int src_stride_ayuv,
+                      uint8_t* dst_uv, int width) = AYUVToUVRow_C;
+  void (*AYUVToYRow)(const uint8_t* src_ayuv, uint8_t* dst_y, int width) =
+      AYUVToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ayuv = src_ayuv + (height - 1) * src_stride_ayuv;
+    src_stride_ayuv = -src_stride_ayuv;
+  }
+// place holders for future intel code
+#if defined(HAS_AYUVTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    AYUVToUVRow = AYUVToUVRow_Any_SSE2;
+    AYUVToYRow = AYUVToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      AYUVToUVRow = AYUVToUVRow_SSE2;
+      AYUVToYRow = AYUVToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_AYUVTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AYUVToUVRow = AYUVToUVRow_Any_AVX2;
+    AYUVToYRow = AYUVToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      AYUVToUVRow = AYUVToUVRow_AVX2;
+      AYUVToYRow = AYUVToYRow_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_AYUVTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AYUVToYRow = AYUVToYRow_Any_NEON;
+    AYUVToUVRow = AYUVToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      AYUVToYRow = AYUVToYRow_NEON;
+      AYUVToUVRow = AYUVToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    AYUVToUVRow(src_ayuv, src_stride_ayuv, dst_uv, width);
+    AYUVToYRow(src_ayuv, dst_y, width);
+    AYUVToYRow(src_ayuv + src_stride_ayuv, dst_y + dst_stride_y, width);
+    src_ayuv += src_stride_ayuv * 2;
+    dst_y += dst_stride_y * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    AYUVToUVRow(src_ayuv, 0, dst_uv, width);
+    AYUVToYRow(src_ayuv, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert AYUV to NV21.
+LIBYUV_API
+int AYUVToNV21(const uint8_t* src_ayuv,
+               int src_stride_ayuv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int y;
+  void (*AYUVToVURow)(const uint8_t* src_ayuv, int src_stride_ayuv,
+                      uint8_t* dst_vu, int width) = AYUVToVURow_C;
+  void (*AYUVToYRow)(const uint8_t* src_ayuv, uint8_t* dst_y, int width) =
+      AYUVToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ayuv = src_ayuv + (height - 1) * src_stride_ayuv;
+    src_stride_ayuv = -src_stride_ayuv;
+  }
+// place holders for future intel code
+#if defined(HAS_AYUVTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    AYUVToVURow = AYUVToVURow_Any_SSE2;
+    AYUVToYRow = AYUVToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      AYUVToVURow = AYUVToVURow_SSE2;
+      AYUVToYRow = AYUVToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_AYUVTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AYUVToVURow = AYUVToVURow_Any_AVX2;
+    AYUVToYRow = AYUVToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      AYUVToVURow = AYUVToVURow_AVX2;
+      AYUVToYRow = AYUVToYRow_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_AYUVTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AYUVToYRow = AYUVToYRow_Any_NEON;
+    AYUVToVURow = AYUVToVURow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      AYUVToYRow = AYUVToYRow_NEON;
+      AYUVToVURow = AYUVToVURow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    AYUVToVURow(src_ayuv, src_stride_ayuv, dst_vu, width);
+    AYUVToYRow(src_ayuv, dst_y, width);
+    AYUVToYRow(src_ayuv + src_stride_ayuv, dst_y + dst_stride_y, width);
+    src_ayuv += src_stride_ayuv * 2;
+    dst_y += dst_stride_y * 2;
+    dst_vu += dst_stride_vu;
+  }
+  if (height & 1) {
+    AYUVToVURow(src_ayuv, 0, dst_vu, width);
+    AYUVToYRow(src_ayuv, dst_y, width);
+  }
+  return 0;
+}
+
 // Convert ARGB to I420.
 LIBYUV_API
-int ARGBToI420(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ARGBToI420(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  if (!src_argb ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -534,39 +1741,81 @@ int ARGBToI420(const uint8* src_argb, int src_stride_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_NEON;
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_NEON;
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
@@ -589,19 +1838,23 @@ int ARGBToI420(const uint8* src_argb, int src_stride_argb,
 
 // Convert BGRA to I420.
 LIBYUV_API
-int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int BGRAToI420(const uint8_t* src_bgra,
+               int src_stride_bgra,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*BGRAToUVRow)(const uint8* src_bgra0, int src_stride_bgra,
-      uint8* dst_u, uint8* dst_v, int width) = BGRAToUVRow_C;
-  void (*BGRAToYRow)(const uint8* src_bgra, uint8* dst_y, int width) =
+  void (*BGRAToUVRow)(const uint8_t* src_bgra0, int src_stride_bgra,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      BGRAToUVRow_C;
+  void (*BGRAToYRow)(const uint8_t* src_bgra, uint8_t* dst_y, int width) =
       BGRAToYRow_C;
-  if (!src_bgra ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_bgra || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -610,31 +1863,83 @@ int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
     src_bgra = src_bgra + (height - 1) * src_stride_bgra;
     src_stride_bgra = -src_stride_bgra;
   }
-#if defined(HAS_BGRATOYROW_SSSE3) && defined(HAS_BGRATOUVROW_SSSE3)
+#if defined(HAS_BGRATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    BGRAToYRow = BGRAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToYRow = BGRAToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    BGRAToUVRow = BGRAToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToUVRow = BGRAToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
     BGRAToYRow = BGRAToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      BGRAToUVRow = BGRAToUVRow_SSSE3;
       BGRAToYRow = BGRAToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_BGRATOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    BGRAToYRow = BGRAToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      BGRAToYRow = BGRAToYRow_NEON;
+#if defined(HAS_BGRATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToUVRow = BGRAToUVRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_BGRATOUVROW_NEON)
-    if (TestCpuFlag(kCpuHasNEON)) {
-      BGRAToUVRow = BGRAToUVRow_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        BGRAToUVRow = BGRAToUVRow_NEON;
-      }
+#if defined(HAS_BGRATOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    BGRAToYRow = BGRAToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      BGRAToYRow = BGRAToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    BGRAToUVRow = BGRAToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      BGRAToUVRow = BGRAToUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOYROW_MSA) && defined(HAS_BGRATOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    BGRAToYRow = BGRAToYRow_Any_MSA;
+    BGRAToUVRow = BGRAToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToYRow = BGRAToYRow_MSA;
     }
+    if (IS_ALIGNED(width, 32)) {
+      BGRAToUVRow = BGRAToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOYROW_LSX) && defined(HAS_BGRATOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    BGRAToYRow = BGRAToYRow_Any_LSX;
+    BGRAToUVRow = BGRAToUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToYRow = BGRAToYRow_LSX;
+      BGRAToUVRow = BGRAToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    BGRAToYRow = BGRAToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      BGRAToYRow = BGRAToYRow_LASX;
+    }
+  }
 #endif
 
   for (y = 0; y < height - 1; y += 2) {
@@ -655,19 +1960,23 @@ int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
 
 // Convert ABGR to I420.
 LIBYUV_API
-int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ABGRToI420(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*ABGRToUVRow)(const uint8* src_abgr0, int src_stride_abgr,
-      uint8* dst_u, uint8* dst_v, int width) = ABGRToUVRow_C;
-  void (*ABGRToYRow)(const uint8* src_abgr, uint8* dst_y, int width) =
+  void (*ABGRToUVRow)(const uint8_t* src_abgr0, int src_stride_abgr,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8_t* src_abgr, uint8_t* dst_y, int width) =
       ABGRToYRow_C;
-  if (!src_abgr ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_abgr || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -676,20 +1985,42 @@ int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
     src_abgr = src_abgr + (height - 1) * src_stride_abgr;
     src_stride_abgr = -src_stride_abgr;
   }
-#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
+#if defined(HAS_ABGRTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
     ABGRToYRow = ABGRToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ABGRToUVRow = ABGRToUVRow_SSSE3;
       ABGRToYRow = ABGRToYRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYRow = ABGRToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ABGRTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ABGRToYRow = ABGRToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ABGRToYRow = ABGRToYRow_NEON;
     }
   }
@@ -702,6 +2033,34 @@ int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
     }
   }
 #endif
+#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYRow = ABGRToYRow_Any_MSA;
+    ABGRToUVRow = ABGRToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_MSA;
+      ABGRToUVRow = ABGRToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LSX) && defined(HAS_ABGRTOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYRow = ABGRToYRow_Any_LSX;
+    ABGRToUVRow = ABGRToUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_LSX;
+      ABGRToUVRow = ABGRToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYRow = ABGRToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height - 1; y += 2) {
     ABGRToUVRow(src_abgr, src_stride_abgr, dst_u, dst_v, width);
@@ -721,19 +2080,23 @@ int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
 
 // Convert RGBA to I420.
 LIBYUV_API
-int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int RGBAToI420(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*RGBAToUVRow)(const uint8* src_rgba0, int src_stride_rgba,
-      uint8* dst_u, uint8* dst_v, int width) = RGBAToUVRow_C;
-  void (*RGBAToYRow)(const uint8* src_rgba, uint8* dst_y, int width) =
+  void (*RGBAToUVRow)(const uint8_t* src_rgba0, int src_stride_rgba,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RGBAToUVRow_C;
+  void (*RGBAToYRow)(const uint8_t* src_rgba, uint8_t* dst_y, int width) =
       RGBAToYRow_C;
-  if (!src_rgba ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_rgba || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -742,20 +2105,26 @@ int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
     src_rgba = src_rgba + (height - 1) * src_stride_rgba;
     src_stride_rgba = -src_stride_rgba;
   }
-#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
+#if defined(HAS_RGBATOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
     RGBAToYRow = RGBAToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      RGBAToUVRow = RGBAToUVRow_SSSE3;
       RGBAToYRow = RGBAToYRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_RGBATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_RGBATOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGBAToYRow = RGBAToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RGBAToYRow = RGBAToYRow_NEON;
     }
   }
@@ -768,6 +2137,34 @@ int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
     }
   }
 #endif
+#if defined(HAS_RGBATOYROW_MSA) && defined(HAS_RGBATOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGBAToYRow = RGBAToYRow_Any_MSA;
+    RGBAToUVRow = RGBAToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYRow = RGBAToYRow_MSA;
+      RGBAToUVRow = RGBAToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYROW_LSX) && defined(HAS_RGBATOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGBAToYRow = RGBAToYRow_Any_LSX;
+    RGBAToUVRow = RGBAToUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYRow = RGBAToYRow_LSX;
+      RGBAToUVRow = RGBAToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYROW_LASX)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToYRow = RGBAToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGBAToYRow = RGBAToYRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height - 1; y += 2) {
     RGBAToUVRow(src_rgba, src_stride_rgba, dst_u, dst_v, width);
@@ -785,29 +2182,41 @@ int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
   return 0;
 }
 
+// Enabled if 1 pass is available
+#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
+     defined(HAS_RGB24TOYROW_LSX))
+#define HAS_RGB24TOYROW
+#endif
+
 // Convert RGB24 to I420.
 LIBYUV_API
-int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
-                uint8* dst_y, int dst_stride_y,
-                uint8* dst_u, int dst_stride_u,
-                uint8* dst_v, int dst_stride_v,
-                int width, int height) {
+int RGB24ToI420(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                uint8_t* dst_u,
+                int dst_stride_u,
+                uint8_t* dst_v,
+                int dst_stride_v,
+                int width,
+                int height) {
   int y;
-#if defined(HAS_RGB24TOYROW_NEON)
-  void (*RGB24ToUVRow)(const uint8* src_rgb24, int src_stride_rgb24,
-      uint8* dst_u, uint8* dst_v, int width) = RGB24ToUVRow_C;
-  void (*RGB24ToYRow)(const uint8* src_rgb24, uint8* dst_y, int width) =
+#if defined(HAS_RGB24TOYROW)
+  void (*RGB24ToUVRow)(const uint8_t* src_rgb24, int src_stride_rgb24,
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RGB24ToUVRow_C;
+  void (*RGB24ToYRow)(const uint8_t* src_rgb24, uint8_t* dst_y, int width) =
       RGB24ToYRow_C;
 #else
-  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
+  void (*RGB24ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
       RGB24ToARGBRow_C;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
 #endif
-  if (!src_rgb24 || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_rgb24 || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -817,20 +2226,53 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
     src_stride_rgb24 = -src_stride_rgb24;
   }
 
+#if defined(HAS_RGB24TOYROW)
+
 // Neon version does direct RGB24 to YUV.
-#if defined(HAS_RGB24TOYROW_NEON)
+#if defined(HAS_RGB24TOYROW_NEON) && defined(HAS_RGB24TOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
     RGB24ToYRow = RGB24ToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RGB24ToYRow = RGB24ToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RGB24ToUVRow = RGB24ToUVRow_NEON;
-      }
+      RGB24ToUVRow = RGB24ToUVRow_NEON;
     }
   }
+#endif
+#if defined(HAS_RGB24TOYROW_MSA) && defined(HAS_RGB24TOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_MSA;
+    RGB24ToYRow = RGB24ToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYRow = RGB24ToYRow_MSA;
+      RGB24ToUVRow = RGB24ToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYROW_LSX) && defined(HAS_RGB24TOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_LSX;
+    RGB24ToYRow = RGB24ToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYRow = RGB24ToYRow_LSX;
+      RGB24ToUVRow = RGB24ToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYROW_LASX) && defined(HAS_RGB24TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_LASX;
+    RGB24ToYRow = RGB24ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYRow = RGB24ToYRow_LASX;
+      RGB24ToUVRow = RGB24ToUVRow_LASX;
+    }
+  }
+#endif
+
 // Other platforms do intermediate conversion from RGB24 to ARGB.
-#else
+#else  // HAS_RGB24TOYROW
+
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
@@ -839,43 +2281,58 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RGB24TOYROW
+
   {
+#if !defined(HAS_RGB24TOYROW)
     // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if defined(HAS_RGB24TOYROW_NEON)
+#if defined(HAS_RGB24TOYROW)
       RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
       RGB24ToYRow(src_rgb24, dst_y, width);
       RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
 #else
       RGB24ToARGBRow(src_rgb24, row, width);
-      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
-      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + row_size, width);
+      ARGBToUVRow(row, row_size, dst_u, dst_v, width);
       ARGBToYRow(row, dst_y, width);
-      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      ARGBToYRow(row + row_size, dst_y + dst_stride_y, width);
 #endif
       src_rgb24 += src_stride_rgb24 * 2;
       dst_y += dst_stride_y * 2;
@@ -883,7 +2340,7 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if defined(HAS_RGB24TOYROW_NEON)
+#if defined(HAS_RGB24TOYROW)
       RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
       RGB24ToYRow(src_rgb24, dst_y, width);
 #else
@@ -892,36 +2349,218 @@ int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !defined(HAS_RGB24TOYROW_NEON)
+#if !defined(HAS_RGB24TOYROW)
     free_aligned_buffer_64(row);
+#endif
   }
+  return 0;
+}
+#undef HAS_RGB24TOYROW
+
+// Enabled if 1 pass is available
+#if defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA)
+#define HAS_RGB24TOYJROW
 #endif
+
+// Convert RGB24 to J420.
+LIBYUV_API
+int RGB24ToJ420(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                uint8_t* dst_u,
+                int dst_stride_u,
+                uint8_t* dst_v,
+                int dst_stride_v,
+                int width,
+                int height) {
+  int y;
+#if defined(HAS_RGB24TOYJROW)
+  void (*RGB24ToUVJRow)(const uint8_t* src_rgb24, int src_stride_rgb24,
+                        uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RGB24ToUVJRow_C;
+  void (*RGB24ToYJRow)(const uint8_t* src_rgb24, uint8_t* dst_y, int width) =
+      RGB24ToYJRow_C;
+#else
+  void (*RGB24ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
+      RGB24ToARGBRow_C;
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
+      ARGBToYJRow_C;
+#endif
+  if (!src_rgb24 || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+
+#if defined(HAS_RGB24TOYJROW)
+
+// Neon version does direct RGB24 to YUV.
+#if defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToUVJRow = RGB24ToUVJRow_Any_NEON;
+    RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_NEON;
+      RGB24ToUVJRow = RGB24ToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_MSA) && defined(HAS_RGB24TOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGB24ToUVJRow = RGB24ToUVJRow_Any_MSA;
+    RGB24ToYJRow = RGB24ToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_MSA;
+      RGB24ToUVJRow = RGB24ToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYJRow = RGB24ToYJRow_LASX;
+    }
+  }
+#endif
+
+// Other platforms do intermediate conversion from RGB24 to ARGB.
+#else  // HAS_RGB24TOYJROW
+
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RGB24TOYJROW
+
+  {
+#if !defined(HAS_RGB24TOYJROW)
+    // Allocate 2 rows of ARGB.
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB24TOYJROW)
+      RGB24ToUVJRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
+      RGB24ToYJRow(src_rgb24, dst_y, width);
+      RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
+#else
+      RGB24ToARGBRow(src_rgb24, row, width);
+      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + row_size, width);
+      ARGBToUVJRow(row, row_size, dst_u, dst_v, width);
+      ARGBToYJRow(row, dst_y, width);
+      ARGBToYJRow(row + row_size, dst_y + dst_stride_y, width);
+#endif
+      src_rgb24 += src_stride_rgb24 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RGB24TOYJROW)
+      RGB24ToUVJRow(src_rgb24, 0, dst_u, dst_v, width);
+      RGB24ToYJRow(src_rgb24, dst_y, width);
+#else
+      RGB24ToARGBRow(src_rgb24, row, width);
+      ARGBToUVJRow(row, 0, dst_u, dst_v, width);
+      ARGBToYJRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RGB24TOYJROW)
+    free_aligned_buffer_64(row);
+#endif
+  }
   return 0;
 }
+#undef HAS_RGB24TOYJROW
+
+// Enabled if 1 pass is available
+#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
+     defined(HAS_RAWTOYROW_LSX))
+#define HAS_RAWTOYROW
+#endif
 
 // Convert RAW to I420.
 LIBYUV_API
-int RAWToI420(const uint8* src_raw, int src_stride_raw,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int width, int height) {
+int RAWToI420(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height) {
   int y;
-#if defined(HAS_RAWTOYROW_NEON)
-  void (*RAWToUVRow)(const uint8* src_raw, int src_stride_raw,
-      uint8* dst_u, uint8* dst_v, int width) = RAWToUVRow_C;
-  void (*RAWToYRow)(const uint8* src_raw, uint8* dst_y, int width) =
+#if defined(HAS_RAWTOYROW)
+  void (*RAWToUVRow)(const uint8_t* src_raw, int src_stride_raw, uint8_t* dst_u,
+                     uint8_t* dst_v, int width) = RAWToUVRow_C;
+  void (*RAWToYRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
       RAWToYRow_C;
 #else
-  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
+  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
       RAWToARGBRow_C;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
 #endif
-  if (!src_raw || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_raw || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -931,20 +2570,53 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
     src_stride_raw = -src_stride_raw;
   }
 
+#if defined(HAS_RAWTOYROW)
+
 // Neon version does direct RAW to YUV.
-#if defined(HAS_RAWTOYROW_NEON)
+#if defined(HAS_RAWTOYROW_NEON) && defined(HAS_RAWTOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RAWToUVRow = RAWToUVRow_Any_NEON;
     RAWToYRow = RAWToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RAWToYRow = RAWToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RAWToUVRow = RAWToUVRow_NEON;
-      }
+      RAWToUVRow = RAWToUVRow_NEON;
     }
   }
+#endif
+#if defined(HAS_RAWTOYROW_MSA) && defined(HAS_RAWTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToUVRow = RAWToUVRow_Any_MSA;
+    RAWToYRow = RAWToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYRow = RAWToYRow_MSA;
+      RAWToUVRow = RAWToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYROW_LSX) && defined(HAS_RAWTOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToUVRow = RAWToUVRow_Any_LSX;
+    RAWToYRow = RAWToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYRow = RAWToYRow_LSX;
+      RAWToUVRow = RAWToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYROW_LASX) && defined(HAS_RAWTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToUVRow = RAWToUVRow_Any_LASX;
+    RAWToYRow = RAWToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYRow = RAWToYRow_LASX;
+      RAWToUVRow = RAWToUVRow_LASX;
+    }
+  }
+#endif
+
 // Other platforms do intermediate conversion from RAW to ARGB.
-#else
+#else  // HAS_RAWTOYROW
+
 #if defined(HAS_RAWTOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
@@ -953,43 +2625,58 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RAWTOYROW
+
   {
+#if !defined(HAS_RAWTOYROW)
     // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if defined(HAS_RAWTOYROW_NEON)
+#if defined(HAS_RAWTOYROW)
       RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
       RAWToYRow(src_raw, dst_y, width);
       RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
 #else
       RAWToARGBRow(src_raw, row, width);
-      RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
-      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + row_size, width);
+      ARGBToUVRow(row, row_size, dst_u, dst_v, width);
       ARGBToYRow(row, dst_y, width);
-      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      ARGBToYRow(row + row_size, dst_y + dst_stride_y, width);
 #endif
       src_raw += src_stride_raw * 2;
       dst_y += dst_stride_y * 2;
@@ -997,7 +2684,7 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if defined(HAS_RAWTOYROW_NEON)
+#if defined(HAS_RAWTOYROW)
       RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
       RAWToYRow(src_raw, dst_y, width);
 #else
@@ -1006,36 +2693,214 @@ int RAWToI420(const uint8* src_raw, int src_stride_raw,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !defined(HAS_RAWTOYROW_NEON)
+#if !defined(HAS_RAWTOYROW)
     free_aligned_buffer_64(row);
+#endif
+  }
+  return 0;
+}
+#undef HAS_RAWTOYROW
+
+// Enabled if 1 pass is available
+#if defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA)
+#define HAS_RAWTOYJROW
+#endif
+
+// Convert RAW to J420.
+LIBYUV_API
+int RAWToJ420(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height) {
+  int y;
+#if defined(HAS_RAWTOYJROW)
+  void (*RAWToUVJRow)(const uint8_t* src_raw, int src_stride_raw,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RAWToUVJRow_C;
+  void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
+      RAWToYJRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
+      ARGBToYJRow_C;
+#endif
+  if (!src_raw || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+#if defined(HAS_RAWTOYJROW)
+
+// Neon version does direct RAW to YUV.
+#if defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVJRow = RAWToUVJRow_Any_NEON;
+    RAWToYJRow = RAWToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_NEON;
+      RAWToUVJRow = RAWToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_MSA) && defined(HAS_RAWTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToUVJRow = RAWToUVJRow_Any_MSA;
+    RAWToYJRow = RAWToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_MSA;
+      RAWToUVJRow = RAWToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToYJRow = RAWToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToYJRow = RAWToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYJRow = RAWToYJRow_LASX;
+    }
+  }
+#endif
+
+// Other platforms do intermediate conversion from RAW to ARGB.
+#else  // HAS_RAWTOYJROW
+
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
   }
 #endif
+#endif  // HAS_RAWTOYJROW
+
+  {
+#if !defined(HAS_RAWTOYJROW)
+    // Allocate 2 rows of ARGB.
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, src_stride_raw, dst_u, dst_v, width);
+      RAWToYJRow(src_raw, dst_y, width);
+      RAWToYJRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + row_size, width);
+      ARGBToUVJRow(row, row_size, dst_u, dst_v, width);
+      ARGBToYJRow(row, dst_y, width);
+      ARGBToYJRow(row + row_size, dst_y + dst_stride_y, width);
+#endif
+      src_raw += src_stride_raw * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, 0, dst_u, dst_v, width);
+      RAWToYJRow(src_raw, dst_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      ARGBToUVJRow(row, 0, dst_u, dst_v, width);
+      ARGBToYJRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RAWTOYJROW)
+    free_aligned_buffer_64(row);
+#endif
+  }
   return 0;
 }
+#undef HAS_RAWTOYJROW
 
 // Convert RGB565 to I420.
 LIBYUV_API
-int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
-                 uint8* dst_y, int dst_stride_y,
-                 uint8* dst_u, int dst_stride_u,
-                 uint8* dst_v, int dst_stride_v,
-                 int width, int height) {
+int RGB565ToI420(const uint8_t* src_rgb565,
+                 int src_stride_rgb565,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 uint8_t* dst_u,
+                 int dst_stride_u,
+                 uint8_t* dst_v,
+                 int dst_stride_v,
+                 int width,
+                 int height) {
   int y;
-#if defined(HAS_RGB565TOYROW_NEON)
-  void (*RGB565ToUVRow)(const uint8* src_rgb565, int src_stride_rgb565,
-      uint8* dst_u, uint8* dst_v, int width) = RGB565ToUVRow_C;
-  void (*RGB565ToYRow)(const uint8* src_rgb565, uint8* dst_y, int width) =
+#if (defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
+     defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
+  void (*RGB565ToUVRow)(const uint8_t* src_rgb565, int src_stride_rgb565,
+                        uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RGB565ToUVRow_C;
+  void (*RGB565ToYRow)(const uint8_t* src_rgb565, uint8_t* dst_y, int width) =
       RGB565ToYRow_C;
 #else
-  void (*RGB565ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
-      RGB565ToARGBRow_C;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*RGB565ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb,
+                          int width) = RGB565ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
 #endif
-  if (!src_rgb565 || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_rgb565 || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1057,6 +2922,39 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
       }
     }
   }
+// MSA version does direct RGB565 to YUV.
+#elif (defined(HAS_RGB565TOYROW_MSA) || defined(HAS_RGB565TOYROW_LSX) || \
+       defined(HAS_RGB565TOYROW_LASX))
+#if defined(HAS_RGB565TOYROW_MSA) && defined(HAS_RGB565TOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_MSA;
+    RGB565ToYRow = RGB565ToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToYRow = RGB565ToYRow_MSA;
+      RGB565ToUVRow = RGB565ToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOYROW_LSX) && defined(HAS_RGB565TOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_LSX;
+    RGB565ToYRow = RGB565ToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToYRow = RGB565ToYRow_LSX;
+      RGB565ToUVRow = RGB565ToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOYROW_LASX) && defined(HAS_RGB565TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_LASX;
+    RGB565ToYRow = RGB565ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB565ToYRow = RGB565ToYRow_LASX;
+      RGB565ToUVRow = RGB565ToUVRow_LASX;
+    }
+  }
+#endif
 // Other platforms do intermediate conversion from RGB565 to ARGB.
 #else
 #if defined(HAS_RGB565TOARGBROW_SSE2)
@@ -1075,43 +2973,58 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#endif
   {
+#if !(defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
+      defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
     // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 #endif
-
     for (y = 0; y < height - 1; y += 2) {
-#if defined(HAS_RGB565TOYROW_NEON)
+#if (defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
+     defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
       RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width);
       RGB565ToYRow(src_rgb565, dst_y, width);
       RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width);
 #else
       RGB565ToARGBRow(src_rgb565, row, width);
-      RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + kRowSize, width);
-      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + row_size, width);
+      ARGBToUVRow(row, row_size, dst_u, dst_v, width);
       ARGBToYRow(row, dst_y, width);
-      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      ARGBToYRow(row + row_size, dst_y + dst_stride_y, width);
 #endif
       src_rgb565 += src_stride_rgb565 * 2;
       dst_y += dst_stride_y * 2;
@@ -1119,7 +3032,8 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if defined(HAS_RGB565TOYROW_NEON)
+#if (defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
+     defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
       RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width);
       RGB565ToYRow(src_rgb565, dst_y, width);
 #else
@@ -1128,36 +3042,45 @@ int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !defined(HAS_RGB565TOYROW_NEON)
+#if !(defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
+      defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
     free_aligned_buffer_64(row);
-  }
 #endif
+  }
   return 0;
 }
 
 // Convert ARGB1555 to I420.
 LIBYUV_API
-int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
-                   uint8* dst_y, int dst_stride_y,
-                   uint8* dst_u, int dst_stride_u,
-                   uint8* dst_v, int dst_stride_v,
-                   int width, int height) {
+int ARGB1555ToI420(const uint8_t* src_argb1555,
+                   int src_stride_argb1555,
+                   uint8_t* dst_y,
+                   int dst_stride_y,
+                   uint8_t* dst_u,
+                   int dst_stride_u,
+                   uint8_t* dst_v,
+                   int dst_stride_v,
+                   int width,
+                   int height) {
   int y;
-#if defined(HAS_ARGB1555TOYROW_NEON)
-  void (*ARGB1555ToUVRow)(const uint8* src_argb1555, int src_stride_argb1555,
-      uint8* dst_u, uint8* dst_v, int width) = ARGB1555ToUVRow_C;
-  void (*ARGB1555ToYRow)(const uint8* src_argb1555, uint8* dst_y, int width) =
-      ARGB1555ToYRow_C;
+#if (defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
+     defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
+  void (*ARGB1555ToUVRow)(const uint8_t* src_argb1555, int src_stride_argb1555,
+                          uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGB1555ToUVRow_C;
+  void (*ARGB1555ToYRow)(const uint8_t* src_argb1555, uint8_t* dst_y,
+                         int width) = ARGB1555ToYRow_C;
 #else
-  void (*ARGB1555ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
-      ARGB1555ToARGBRow_C;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGB1555ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb,
+                            int width) = ARGB1555ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
 #endif
-  if (!src_argb1555 || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_argb1555 || !dst_y || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1179,6 +3102,39 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
       }
     }
   }
+// MSA version does direct ARGB1555 to YUV.
+#elif (defined(HAS_ARGB1555TOYROW_MSA) || defined(HAS_ARGB1555TOYROW_LSX) || \
+       defined(HAS_ARGB1555TOYROW_LASX))
+#if defined(HAS_ARGB1555TOYROW_MSA) && defined(HAS_ARGB1555TOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_MSA;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_MSA;
+      ARGB1555ToUVRow = ARGB1555ToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOYROW_LSX) && defined(HAS_ARGB1555TOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_LSX;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_LSX;
+      ARGB1555ToUVRow = ARGB1555ToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOYROW_LASX) && defined(HAS_ARGB1555TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_LASX;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_LASX;
+      ARGB1555ToUVRow = ARGB1555ToUVRow_LASX;
+    }
+  }
+#endif
 // Other platforms do intermediate conversion from ARGB1555 to ARGB.
 #else
 #if defined(HAS_ARGB1555TOARGBROW_SSE2)
@@ -1197,45 +3153,61 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#endif
   {
+#if !(defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
+      defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
     // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if defined(HAS_ARGB1555TOYROW_NEON)
+#if (defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
+     defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
       ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width);
       ARGB1555ToYRow(src_argb1555, dst_y, width);
       ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y,
                      width);
 #else
       ARGB1555ToARGBRow(src_argb1555, row, width);
-      ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + kRowSize,
+      ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + row_size,
                         width);
-      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToUVRow(row, row_size, dst_u, dst_v, width);
       ARGBToYRow(row, dst_y, width);
-      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      ARGBToYRow(row + row_size, dst_y + dst_stride_y, width);
 #endif
       src_argb1555 += src_stride_argb1555 * 2;
       dst_y += dst_stride_y * 2;
@@ -1243,7 +3215,8 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if defined(HAS_ARGB1555TOYROW_NEON)
+#if (defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
+     defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
       ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width);
       ARGB1555ToYRow(src_argb1555, dst_y, width);
 #else
@@ -1252,36 +3225,44 @@ int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !defined(HAS_ARGB1555TOYROW_NEON)
+#if !(defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
+      defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
     free_aligned_buffer_64(row);
-  }
 #endif
+  }
   return 0;
 }
 
 // Convert ARGB4444 to I420.
 LIBYUV_API
-int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
-                   uint8* dst_y, int dst_stride_y,
-                   uint8* dst_u, int dst_stride_u,
-                   uint8* dst_v, int dst_stride_v,
-                   int width, int height) {
+int ARGB4444ToI420(const uint8_t* src_argb4444,
+                   int src_stride_argb4444,
+                   uint8_t* dst_y,
+                   int dst_stride_y,
+                   uint8_t* dst_u,
+                   int dst_stride_u,
+                   uint8_t* dst_v,
+                   int dst_stride_v,
+                   int width,
+                   int height) {
   int y;
 #if defined(HAS_ARGB4444TOYROW_NEON)
-  void (*ARGB4444ToUVRow)(const uint8* src_argb4444, int src_stride_argb4444,
-      uint8* dst_u, uint8* dst_v, int width) = ARGB4444ToUVRow_C;
-  void (*ARGB4444ToYRow)(const uint8* src_argb4444, uint8* dst_y, int width) =
-      ARGB4444ToYRow_C;
+  void (*ARGB4444ToUVRow)(const uint8_t* src_argb4444, int src_stride_argb4444,
+                          uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGB4444ToUVRow_C;
+  void (*ARGB4444ToYRow)(const uint8_t* src_argb4444, uint8_t* dst_y,
+                         int width) = ARGB4444ToYRow_C;
 #else
-  void (*ARGB4444ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
-      ARGB4444ToARGBRow_C;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGB4444ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb,
+                            int width) = ARGB4444ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
 #endif
-  if (!src_argb4444 || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_argb4444 || !dst_y || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1321,30 +3302,99 @@ int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGB4444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+      if (IS_ALIGNED(width, 32)) {
+        ARGBToUVRow = ARGBToUVRow_MSA;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
+    }
+  }
+#endif
+#endif
+
   {
+#if !(defined(HAS_ARGB4444TOYROW_NEON))
     // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
@@ -1355,11 +3405,11 @@ int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
                      width);
 #else
       ARGB4444ToARGBRow(src_argb4444, row, width);
-      ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + kRowSize,
+      ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + row_size,
                         width);
-      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToUVRow(row, row_size, dst_u, dst_v, width);
       ARGBToYRow(row, dst_y, width);
-      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      ARGBToYRow(row + row_size, dst_y + dst_stride_y, width);
 #endif
       src_argb4444 += src_stride_argb4444 * 2;
       dst_y += dst_stride_y * 2;
@@ -1376,13 +3426,202 @@ int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !defined(HAS_ARGB4444TOYROW_NEON)
+#if !(defined(HAS_ARGB4444TOYROW_NEON))
     free_aligned_buffer_64(row);
+#endif
+  }
+  return 0;
+}
+
+// Convert RGB24 to J400.
+LIBYUV_API
+int RGB24ToJ400(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_yj,
+                int dst_stride_yj,
+                int width,
+                int height) {
+  int y;
+  void (*RGB24ToYJRow)(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) =
+      RGB24ToYJRow_C;
+  if (!src_rgb24 || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_yj = 0;
+  }
+#if defined(HAS_RGB24TOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYJRow = RGB24ToYJRow_AVX2;
+    }
   }
 #endif
+#if defined(HAS_RGB24TOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYJRow = RGB24ToYJRow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB24ToYJRow(src_rgb24, dst_yj, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_yj += dst_stride_yj;
+  }
   return 0;
 }
 
+// Convert RAW to J400.
+LIBYUV_API
+int RAWToJ400(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_yj,
+              int dst_stride_yj,
+              int width,
+              int height) {
+  int y;
+  void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_yj, int width) =
+      RAWToYJRow_C;
+  if (!src_raw || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_yj = 0;
+  }
+
+#if defined(HAS_RAWTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToYJRow = RAWToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RAWToYJRow = RAWToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYJRow = RAWToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToYJRow = RAWToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToYJRow = RAWToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToYJRow = RAWToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToYJRow = RAWToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYJRow = RAWToYJRow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RAWToYJRow(src_raw, dst_yj, width);
+    src_raw += src_stride_raw;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+// Convert Android420 to I420.
+LIBYUV_API
+int Android420ToI420(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     int src_pixel_stride_uv,
+                     uint8_t* dst_y,
+                     int dst_stride_y,
+                     uint8_t* dst_u,
+                     int dst_stride_u,
+                     uint8_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height) {
+  return Android420ToI420Rotate(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                                src_stride_v, src_pixel_stride_uv, dst_y,
+                                dst_stride_y, dst_u, dst_stride_u, dst_v,
+                                dst_stride_v, width, height, kRotate0);
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/media/libyuv/source/convert_argb.cc b/media/libyuv/source/convert_argb.cc
index fb9582d627..64425c5967 100644
--- a/media/libyuv/source/convert_argb.cc
+++ b/media/libyuv/source/convert_argb.cc
@@ -10,6 +10,9 @@
 
 #include "libyuv/convert_argb.h"
 
+#include <assert.h>
+
+#include "libyuv/convert_from_argb.h"
 #include "libyuv/cpu_id.h"
 #ifdef HAVE_JPEG
 #include "libyuv/mjpeg_decoder.h"
@@ -17,6 +20,7 @@
 #include "libyuv/planar_functions.h"  // For CopyPlane and ARGBShuffle.
 #include "libyuv/rotate_argb.h"
 #include "libyuv/row.h"
+#include "libyuv/scale_row.h"  // For ScaleRowUp2_Linear and ScaleRowUp2_Bilinear
 #include "libyuv/video_common.h"
 
 #ifdef __cplusplus
@@ -26,11 +30,13 @@ extern "C" {
 
 // Copy ARGB with optional flipping
 LIBYUV_API
-int ARGBCopy(const uint8* src_argb, int src_stride_argb,
-             uint8* dst_argb, int dst_stride_argb,
-             int width, int height) {
-  if (!src_argb || !dst_argb ||
-      width <= 0 || height == 0) {
+int ARGBCopy(const uint8_t* src_argb,
+             int src_stride_argb,
+             uint8_t* dst_argb,
+             int dst_stride_argb,
+             int width,
+             int height) {
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -40,27 +46,31 @@ int ARGBCopy(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
 
-  CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
-            width * 4, height);
+  CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width * 4,
+            height);
   return 0;
 }
 
-// Convert I422 to ARGB with matrix
-static int I420ToARGBMatrix(const uint8* src_y, int src_stride_y,
-                            const uint8* src_u, int src_stride_u,
-                            const uint8* src_v, int src_stride_v,
-                            uint8* dst_argb, int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width, int height) {
+// Convert I420 to ARGB with matrix.
+LIBYUV_API
+int I420ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
-  void (*I422ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I422ToARGBRow_C;
-  if (!src_y || !src_u || !src_v || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -85,6 +95,15 @@ static int I420ToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
@@ -93,13 +112,20 @@ static int I420ToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
-#if defined(HAS_I422TOARGBROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
-    I422ToARGBRow = I422ToARGBRow_DSPR2;
+#if defined(HAS_I422TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGBRow = I422ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
   }
 #endif
 
@@ -117,111 +143,168 @@ static int I420ToARGBMatrix(const uint8* src_y, int src_stride_y,
 
 // Convert I420 to ARGB.
 LIBYUV_API
-int I420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I420ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvI601Constants,
-                          width, height);
+int I420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
 }
 
 // Convert I420 to ABGR.
 LIBYUV_API
-int I420ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I420ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int I420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuI601Constants,  // Use Yvu matrix
                           width, height);
 }
 
 // Convert J420 to ARGB.
 LIBYUV_API
-int J420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I420ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvJPEGConstants,
-                          width, height);
+int J420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvJPEGConstants, width, height);
 }
 
 // Convert J420 to ABGR.
 LIBYUV_API
-int J420ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I420ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int J420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuJPEGConstants,  // Use Yvu matrix
                           width, height);
 }
 
 // Convert H420 to ARGB.
 LIBYUV_API
-int H420ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I420ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvH709Constants,
-                          width, height);
+int H420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvH709Constants, width, height);
 }
 
 // Convert H420 to ABGR.
 LIBYUV_API
-int H420ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I420ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int H420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuH709Constants,  // Use Yvu matrix
                           width, height);
 }
 
-// Convert I422 to ARGB with matrix
-static int I422ToARGBMatrix(const uint8* src_y, int src_stride_y,
-                            const uint8* src_u, int src_stride_u,
-                            const uint8* src_v, int src_stride_v,
-                            uint8* dst_argb, int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width, int height) {
+// Convert U420 to ARGB.
+LIBYUV_API
+int U420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U420 to ABGR.
+LIBYUV_API
+int U420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I422 to ARGB with matrix.
+LIBYUV_API
+int I422ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
-  void (*I422ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I422ToARGBRow_C;
-  if (!src_y || !src_u || !src_v ||
-      !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -231,10 +314,8 @@ static int I422ToARGBMatrix(const uint8* src_y, int src_stride_y,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      src_stride_u * 2 == width &&
-      src_stride_v * 2 == width &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_y == width && src_stride_u * 2 == width &&
+      src_stride_v * 2 == width && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
@@ -255,6 +336,15 @@ static int I422ToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
@@ -263,13 +353,20 @@ static int I422ToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
-#if defined(HAS_I422TOARGBROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
-    I422ToARGBRow = I422ToARGBRow_DSPR2;
+#if defined(HAS_I422TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGBRow = I422ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
   }
 #endif
 
@@ -285,111 +382,168 @@ static int I422ToARGBMatrix(const uint8* src_y, int src_stride_y,
 
 // Convert I422 to ARGB.
 LIBYUV_API
-int I422ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I422ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvI601Constants,
-                          width, height);
+int I422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
 }
 
 // Convert I422 to ABGR.
 LIBYUV_API
-int I422ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I422ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int I422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuI601Constants,  // Use Yvu matrix
                           width, height);
 }
 
 // Convert J422 to ARGB.
 LIBYUV_API
-int J422ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I422ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvJPEGConstants,
-                          width, height);
+int J422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvJPEGConstants, width, height);
 }
 
 // Convert J422 to ABGR.
 LIBYUV_API
-int J422ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I422ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int J422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuJPEGConstants,  // Use Yvu matrix
                           width, height);
 }
 
 // Convert H422 to ARGB.
 LIBYUV_API
-int H422ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I422ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvH709Constants,
-                          width, height);
+int H422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvH709Constants, width, height);
 }
 
 // Convert H422 to ABGR.
 LIBYUV_API
-int H422ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I422ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int H422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuH709Constants,  // Use Yvu matrix
                           width, height);
 }
 
-// Convert I444 to ARGB with matrix
-static int I444ToARGBMatrix(const uint8* src_y, int src_stride_y,
-                            const uint8* src_u, int src_stride_u,
-                            const uint8* src_v, int src_stride_v,
-                            uint8* dst_argb, int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width, int height) {
+// Convert U422 to ARGB.
+LIBYUV_API
+int U422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U422 to ABGR.
+LIBYUV_API
+int U422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I444 to ARGB with matrix.
+LIBYUV_API
+int I444ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
-  void (*I444ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I444ToARGBRow_C;
-  if (!src_y || !src_u || !src_v ||
-      !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I444ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -399,9 +553,7 @@ static int I444ToARGBMatrix(const uint8* src_y, int src_stride_y,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      src_stride_u == width &&
-      src_stride_v == width &&
+  if (src_stride_y == width && src_stride_u == width && src_stride_v == width &&
       dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
@@ -431,6 +583,22 @@ static int I444ToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToARGBRow = I444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I444ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
@@ -444,110 +612,1126 @@ static int I444ToARGBMatrix(const uint8* src_y, int src_stride_y,
 
 // Convert I444 to ARGB.
 LIBYUV_API
-int I444ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I444ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvI601Constants,
-                          width, height);
+int I444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
 }
 
 // Convert I444 to ABGR.
 LIBYUV_API
-int I444ToABGR(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_abgr, int dst_stride_abgr,
-               int width, int height) {
-  return I444ToARGBMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_abgr, dst_stride_abgr,
+int I444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuI601Constants,  // Use Yvu matrix
                           width, height);
 }
 
 // Convert J444 to ARGB.
 LIBYUV_API
-int J444ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return I444ToARGBMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_argb, dst_stride_argb,
-                          &kYuvJPEGConstants,
+int J444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvJPEGConstants, width, height);
+}
+
+// Convert J444 to ABGR.
+LIBYUV_API
+int J444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuJPEGConstants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert H444 to ARGB.
+LIBYUV_API
+int H444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert H444 to ABGR.
+LIBYUV_API
+int H444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuH709Constants,  // Use Yvu matrix
                           width, height);
 }
 
-// Convert I411 to ARGB.
+// Convert U444 to ARGB.
 LIBYUV_API
-int I411ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int U444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U444 to ABGR.
+LIBYUV_API
+int U444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I444 to RGB24 with matrix.
+LIBYUV_API
+int I444ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
   int y;
-  void (*I411ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I411ToARGBRow_C;
-  if (!src_y || !src_u || !src_v ||
-      !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*I444ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                         const uint8_t* v_buf, uint8_t* rgb_buf,
+                         const struct YuvConstants* yuvconstants, int width) =
+      I444ToRGB24Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
-    dst_stride_argb = -dst_stride_argb;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      src_stride_u * 4 == width &&
-      src_stride_v * 4 == width &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_y == width && src_stride_u == width && src_stride_v == width &&
+      dst_stride_rgb24 == width * 3) {
     width *= height;
     height = 1;
-    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_rgb24 = 0;
   }
-#if defined(HAS_I411TOARGBROW_SSSE3)
+#if defined(HAS_I444TORGB24ROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    I411ToARGBRow = I411ToARGBRow_Any_SSSE3;
+    I444ToRGB24Row = I444ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToRGB24Row = I444ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToRGB24Row = I444ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
-      I411ToARGBRow = I411ToARGBRow_SSSE3;
+      I444ToRGB24Row = I444ToRGB24Row_NEON;
     }
   }
 #endif
-#if defined(HAS_I411TOARGBROW_AVX2)
+#if defined(HAS_I444TORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToRGB24Row = I444ToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToRGB24Row = I444ToRGB24Row_LSX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444ToRGB24Row(src_y, src_u, src_v, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I444 to RGB24.
+LIBYUV_API
+int I444ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I444ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvI601Constants, width, height);
+}
+
+// Convert I444 to RAW.
+LIBYUV_API
+int I444ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I444ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuI601Constants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert 10 bit YUV to ARGB with matrix.
+// TODO(fbarchard): Consider passing scale multiplier to I210ToARGB to
+// multiply 10 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I210ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I210ToAR30Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210ToAR30Row = I210ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210ToAR30Row = I210ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210TOAR30ROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    I411ToARGBRow = I411ToARGBRow_Any_AVX2;
+    I210ToAR30Row = I210ToAR30Row_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
-      I411ToARGBRow = I411ToARGBRow_AVX2;
+      I210ToAR30Row = I210ToAR30Row_AVX2;
     }
   }
 #endif
-#if defined(HAS_I411TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I411ToARGBRow = I411ToARGBRow_Any_NEON;
+  for (y = 0; y < height; ++y) {
+    I210ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I010 to AR30.
+LIBYUV_API
+int I010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert H010 to AR30.
+LIBYUV_API
+int H010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert U010 to AR30.
+LIBYUV_API
+int U010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert I010 to AB30.
+LIBYUV_API
+int I010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuI601Constants, width, height);
+}
+
+// Convert H010 to AB30.
+LIBYUV_API
+int H010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuH709Constants, width, height);
+}
+
+// Convert U010 to AB30.
+LIBYUV_API
+int U010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert 12 bit YUV to ARGB with matrix.
+// TODO(fbarchard): Consider passing scale multiplier to I212ToARGB to
+// multiply 12 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I012ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I212ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I212ToAR30Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I212TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I212ToAR30Row = I212ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I212ToAR30Row = I212ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I212TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I212ToAR30Row = I212ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I212ToAR30Row = I212ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I212ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert 10 bit YUV to ARGB with matrix.
+// TODO(fbarchard): Consider passing scale multiplier to I210ToARGB to
+// multiply 10 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I210ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I210ToAR30Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210ToAR30Row = I210ToAR30Row_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I411ToARGBRow = I411ToARGBRow_NEON;
+      I210ToAR30Row = I210ToAR30Row_SSSE3;
     }
   }
 #endif
+#if defined(HAS_I210TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToAR30Row = I210ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToAR30Row = I210ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I210ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I210 to AR30.
+LIBYUV_API
+int I210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert H210 to AR30.
+LIBYUV_API
+int H210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert U210 to AR30.
+LIBYUV_API
+int U210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert I210 to AB30.
+LIBYUV_API
+int I210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuI601Constants, width, height);
+}
+
+// Convert H210 to AB30.
+LIBYUV_API
+int H210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuH709Constants, width, height);
+}
+
+// Convert U210 to AB30.
+LIBYUV_API
+int U210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYuv2020Constants, width, height);
+}
+
+LIBYUV_API
+int I410ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I410ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToAR30Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToAR30Row = I410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToAR30Row = I410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToAR30Row = I410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToAR30Row = I410ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I410ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert 10 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I210ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I210ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210ToARGBRow = I210ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210ToARGBRow = I210ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToARGBRow = I210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToARGBRow = I210ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I210ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I010 to ARGB.
+LIBYUV_API
+int I010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I010 to ABGR.
+LIBYUV_API
+int I010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert H010 to ARGB.
+LIBYUV_API
+int H010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert H010 to ABGR.
+LIBYUV_API
+int H010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuH709Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert U010 to ARGB.
+LIBYUV_API
+int U010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U010 to ABGR.
+LIBYUV_API
+int U010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert 12 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I012ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I212ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I212ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I212TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I212ToARGBRow = I212ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I212ToARGBRow = I212ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I212TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I212ToARGBRow = I212ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I212ToARGBRow = I212ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I212ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert 10 bit 422 YUV to ARGB with matrix.
+LIBYUV_API
+int I210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I210ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I210ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210ToARGBRow = I210ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210ToARGBRow = I210ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToARGBRow = I210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToARGBRow = I210ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I210ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
 
+// Convert I210 to ARGB.
+LIBYUV_API
+int I210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I210 to ABGR.
+LIBYUV_API
+int I210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert H210 to ARGB.
+LIBYUV_API
+int H210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert H210 to ABGR.
+LIBYUV_API
+int H210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuH709Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert U210 to ARGB.
+LIBYUV_API
+int U210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U210 to ABGR.
+LIBYUV_API
+int U210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+LIBYUV_API
+int I410ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I410ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToARGBRow = I410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToARGBRow = I410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToARGBRow = I410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToARGBRow = I410ToARGBRow_AVX2;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
-    I411ToARGBRow(src_y, src_u, src_v, dst_argb, &kYuvI601Constants, width);
+    I410ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;
     src_y += src_stride_y;
     src_u += src_stride_u;
@@ -556,26 +1740,233 @@ int I411ToARGB(const uint8* src_y, int src_stride_y,
   return 0;
 }
 
-// Convert I420 with Alpha to preattenuated ARGB.
-static int I420AlphaToARGBMatrix(const uint8* src_y, int src_stride_y,
-                                 const uint8* src_u, int src_stride_u,
-                                 const uint8* src_v, int src_stride_v,
-                                 const uint8* src_a, int src_stride_a,
-                                 uint8* dst_argb, int dst_stride_argb,
-                                 const struct YuvConstants* yuvconstants,
-                                 int width, int height, int attenuate) {
+LIBYUV_API
+int P010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
-  void (*I422AlphaToARGBRow)(const uint8* y_buf,
-                             const uint8* u_buf,
-                             const uint8* v_buf,
-                             const uint8* a_buf,
-                             uint8* dst_argb,
+  void (*P210ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToARGBRow = P210ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToARGBRow = P210ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToARGBRow = P210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToARGBRow = P210ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToARGBRow = P210ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToARGBRow = P210ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToARGBRow = P210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToARGBRow = P210ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToAR30Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToAR30Row = P210ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToAR30Row = P210ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToAR30Row = P210ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToAR30Row = P210ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToAR30Row(src_y, src_uv, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToAR30Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToAR30Row = P210ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToAR30Row = P210ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToAR30Row = P210ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToAR30Row = P210ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToAR30Row(src_y, src_uv, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+// Convert I420 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I420AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I422AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width) = I422AlphaToARGBRow_C;
-  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
                            int width) = ARGBAttenuateRow_C;
-  if (!src_y || !src_u || !src_v || !dst_argb ||
-      width <= 0 || height == 0) {
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -608,13 +1999,20 @@ static int I420AlphaToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
-#if defined(HAS_I422ALPHATOARGBROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
-    I422AlphaToARGBRow = I422AlphaToARGBRow_DSPR2;
+#if defined(HAS_I422ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_LASX;
+    }
   }
 #endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
@@ -641,6 +2039,14 @@ static int I420AlphaToARGBMatrix(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -659,51 +2065,687 @@ static int I420AlphaToARGBMatrix(const uint8* src_y, int src_stride_y,
   return 0;
 }
 
+// Convert I422 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I422AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I422AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I422AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I422ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I444 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I444AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I444AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I444AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
 // Convert I420 with Alpha to ARGB.
 LIBYUV_API
-int I420AlphaToARGB(const uint8* src_y, int src_stride_y,
-                    const uint8* src_u, int src_stride_u,
-                    const uint8* src_v, int src_stride_v,
-                    const uint8* src_a, int src_stride_a,
-                    uint8* dst_argb, int dst_stride_argb,
-                    int width, int height, int attenuate) {
-  return I420AlphaToARGBMatrix(src_y, src_stride_y,
-                               src_u, src_stride_u,
-                               src_v, src_stride_v,
-                               src_a, src_stride_a,
-                               dst_argb, dst_stride_argb,
-                               &kYuvI601Constants,
-                               width, height, attenuate);
+int I420AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I420AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, src_a, src_stride_a, dst_argb,
+                               dst_stride_argb, &kYuvI601Constants, width,
+                               height, attenuate);
 }
 
 // Convert I420 with Alpha to ABGR.
 LIBYUV_API
-int I420AlphaToABGR(const uint8* src_y, int src_stride_y,
-                    const uint8* src_u, int src_stride_u,
-                    const uint8* src_v, int src_stride_v,
-                    const uint8* src_a, int src_stride_a,
-                    uint8* dst_abgr, int dst_stride_abgr,
-                    int width, int height, int attenuate) {
-  return I420AlphaToARGBMatrix(src_y, src_stride_y,
-                               src_v, src_stride_v,  // Swap U and V
-                               src_u, src_stride_u,
-                               src_a, src_stride_a,
-                               dst_abgr, dst_stride_abgr,
-                               &kYvuI601Constants,  // Use Yvu matrix
-                               width, height, attenuate);
+int I420AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I420AlphaToARGBMatrix(
+      src_y, src_stride_y, src_v, src_stride_v,  // Swap U and V
+      src_u, src_stride_u, src_a, src_stride_a, dst_abgr, dst_stride_abgr,
+      &kYvuI601Constants,  // Use Yvu matrix
+      width, height, attenuate);
 }
 
-// Convert I400 to ARGB.
+// Convert I422 with Alpha to ARGB.
+LIBYUV_API
+int I422AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I422AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, src_a, src_stride_a, dst_argb,
+                               dst_stride_argb, &kYuvI601Constants, width,
+                               height, attenuate);
+}
+
+// Convert I422 with Alpha to ABGR.
+LIBYUV_API
+int I422AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I422AlphaToARGBMatrix(
+      src_y, src_stride_y, src_v, src_stride_v,  // Swap U and V
+      src_u, src_stride_u, src_a, src_stride_a, dst_abgr, dst_stride_abgr,
+      &kYvuI601Constants,  // Use Yvu matrix
+      width, height, attenuate);
+}
+
+// Convert I444 with Alpha to ARGB.
+LIBYUV_API
+int I444AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I444AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, src_a, src_stride_a, dst_argb,
+                               dst_stride_argb, &kYuvI601Constants, width,
+                               height, attenuate);
+}
+
+// Convert I444 with Alpha to ABGR.
+LIBYUV_API
+int I444AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I444AlphaToARGBMatrix(
+      src_y, src_stride_y, src_v, src_stride_v,  // Swap U and V
+      src_u, src_stride_u, src_a, src_stride_a, dst_abgr, dst_stride_abgr,
+      &kYvuI601Constants,  // Use Yvu matrix
+      width, height, attenuate);
+}
+
+// Convert I010 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I010AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I210AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I210AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I210AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I210 with Alpha to preattenuated ARGB with matrix.
 LIBYUV_API
-int I400ToARGB(const uint8* src_y, int src_stride_y,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int I210AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
   int y;
-  void (*I400ToARGBRow)(const uint8* y_buf,
-                     uint8* rgb_buf,
-                     int width) = I400ToARGBRow_C;
-  if (!src_y || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*I210AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I210AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I210AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I410 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I410AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I410AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I410AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I410AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB with matrix.
+LIBYUV_API
+int I400ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I400ToARGBRow)(const uint8_t* y_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I400ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -713,8 +2755,7 @@ int I400ToARGB(const uint8* src_y, int src_stride_y,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_y == width && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_y = dst_stride_argb = 0;
@@ -743,25 +2784,55 @@ int I400ToARGB(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I400TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I400ToARGBRow = I400ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I400TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I400ToARGBRow = I400ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
-    I400ToARGBRow(src_y, dst_argb, width);
+    I400ToARGBRow(src_y, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;
     src_y += src_stride_y;
   }
   return 0;
 }
 
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I400ToARGBMatrix(src_y, src_stride_y, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
+}
+
 // Convert J400 to ARGB.
 LIBYUV_API
-int J400ToARGB(const uint8* src_y, int src_stride_y,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int J400ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
   int y;
-  void (*J400ToARGBRow)(const uint8* src_y, uint8* dst_argb, int width) =
+  void (*J400ToARGBRow)(const uint8_t* src_y, uint8_t* dst_argb, int width) =
       J400ToARGBRow_C;
-  if (!src_y || !dst_argb ||
-      width <= 0 || height == 0) {
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -771,8 +2842,7 @@ int J400ToARGB(const uint8* src_y, int src_stride_y,
     src_stride_y = -src_stride_y;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_y == width && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_y = dst_stride_argb = 0;
@@ -801,6 +2871,22 @@ int J400ToARGB(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_J400TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    J400ToARGBRow = J400ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      J400ToARGBRow = J400ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_J400TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    J400ToARGBRow = J400ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      J400ToARGBRow = J400ToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     J400ToARGBRow(src_y, dst_argb, width);
     src_y += src_stride_y;
@@ -810,85 +2896,105 @@ int J400ToARGB(const uint8* src_y, int src_stride_y,
 }
 
 // Shuffle table for converting BGRA to ARGB.
-static uvec8 kShuffleMaskBGRAToARGB = {
-  3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u
-};
+static const uvec8 kShuffleMaskBGRAToARGB = {
+    3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u};
 
 // Shuffle table for converting ABGR to ARGB.
-static uvec8 kShuffleMaskABGRToARGB = {
-  2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u
-};
+static const uvec8 kShuffleMaskABGRToARGB = {
+    2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u};
 
 // Shuffle table for converting RGBA to ARGB.
-static uvec8 kShuffleMaskRGBAToARGB = {
-  1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u
-};
+static const uvec8 kShuffleMaskRGBAToARGB = {
+    1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u};
+
+// Shuffle table for converting AR64 to AB64.
+static const uvec8 kShuffleMaskAR64ToAB64 = {
+    4u, 5u, 2u, 3u, 0u, 1u, 6u, 7u, 12u, 13u, 10u, 11u, 8u, 9u, 14u, 15u};
 
 // Convert BGRA to ARGB.
 LIBYUV_API
-int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return ARGBShuffle(src_bgra, src_stride_bgra,
-                     dst_argb, dst_stride_argb,
-                     (const uint8*)(&kShuffleMaskBGRAToARGB),
-                     width, height);
+int BGRAToARGB(const uint8_t* src_bgra,
+               int src_stride_bgra,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra, dst_argb, dst_stride_argb,
+                     (const uint8_t*)&kShuffleMaskBGRAToARGB, width, height);
 }
 
 // Convert ARGB to BGRA (same as BGRAToARGB).
 LIBYUV_API
-int ARGBToBGRA(const uint8* src_bgra, int src_stride_bgra,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return ARGBShuffle(src_bgra, src_stride_bgra,
-                     dst_argb, dst_stride_argb,
-                     (const uint8*)(&kShuffleMaskBGRAToARGB),
-                     width, height);
+int ARGBToBGRA(const uint8_t* src_bgra,
+               int src_stride_bgra,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra, dst_argb, dst_stride_argb,
+                     (const uint8_t*)&kShuffleMaskBGRAToARGB, width, height);
 }
 
 // Convert ABGR to ARGB.
 LIBYUV_API
-int ABGRToARGB(const uint8* src_abgr, int src_stride_abgr,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return ARGBShuffle(src_abgr, src_stride_abgr,
-                     dst_argb, dst_stride_argb,
-                     (const uint8*)(&kShuffleMaskABGRToARGB),
-                     width, height);
+int ABGRToARGB(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr, dst_argb, dst_stride_argb,
+                     (const uint8_t*)&kShuffleMaskABGRToARGB, width, height);
 }
 
 // Convert ARGB to ABGR to (same as ABGRToARGB).
 LIBYUV_API
-int ARGBToABGR(const uint8* src_abgr, int src_stride_abgr,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return ARGBShuffle(src_abgr, src_stride_abgr,
-                     dst_argb, dst_stride_argb,
-                     (const uint8*)(&kShuffleMaskABGRToARGB),
-                     width, height);
+int ARGBToABGR(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr, dst_argb, dst_stride_argb,
+                     (const uint8_t*)&kShuffleMaskABGRToARGB, width, height);
 }
 
 // Convert RGBA to ARGB.
 LIBYUV_API
-int RGBAToARGB(const uint8* src_rgba, int src_stride_rgba,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
-  return ARGBShuffle(src_rgba, src_stride_rgba,
-                     dst_argb, dst_stride_argb,
-                     (const uint8*)(&kShuffleMaskRGBAToARGB),
-                     width, height);
+int RGBAToARGB(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return ARGBShuffle(src_rgba, src_stride_rgba, dst_argb, dst_stride_argb,
+                     (const uint8_t*)&kShuffleMaskRGBAToARGB, width, height);
+}
+
+// Convert AR64 To AB64.
+LIBYUV_API
+int AR64ToAB64(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height) {
+  return AR64Shuffle(src_ar64, src_stride_ar64, dst_ab64, dst_stride_ab64,
+                     (const uint8_t*)&kShuffleMaskAR64ToAB64, width, height);
 }
 
 // Convert RGB24 to ARGB.
 LIBYUV_API
-int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
-                uint8* dst_argb, int dst_stride_argb,
-                int width, int height) {
+int RGB24ToARGB(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_argb,
+                int dst_stride_argb,
+                int width,
+                int height) {
   int y;
-  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
+  void (*RGB24ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
       RGB24ToARGBRow_C;
-  if (!src_rgb24 || !dst_argb ||
-      width <= 0 || height == 0) {
+  if (!src_rgb24 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -898,8 +3004,7 @@ int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
     src_stride_rgb24 = -src_stride_rgb24;
   }
   // Coalesce rows.
-  if (src_stride_rgb24 == width * 3 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_rgb24 == width * 3 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_rgb24 = dst_stride_argb = 0;
@@ -920,6 +3025,30 @@ int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
     }
   }
 #endif
+#if defined(HAS_RGB24TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGB24ToARGBRow(src_rgb24, dst_argb, width);
@@ -931,14 +3060,16 @@ int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
 
 // Convert RAW to ARGB.
 LIBYUV_API
-int RAWToARGB(const uint8* src_raw, int src_stride_raw,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height) {
+int RAWToARGB(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height) {
   int y;
-  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int width) =
+  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
       RAWToARGBRow_C;
-  if (!src_raw || !dst_argb ||
-      width <= 0 || height == 0) {
+  if (!src_raw || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -948,8 +3079,7 @@ int RAWToARGB(const uint8* src_raw, int src_stride_raw,
     src_stride_raw = -src_stride_raw;
   }
   // Coalesce rows.
-  if (src_stride_raw == width * 3 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_raw == width * 3 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_raw = dst_stride_argb = 0;
@@ -970,6 +3100,30 @@ int RAWToARGB(const uint8* src_raw, int src_stride_raw,
     }
   }
 #endif
+#if defined(HAS_RAWTOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToARGBRow = RAWToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToARGBRow = RAWToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToARGBRow = RAWToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToARGBRow = RAWToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RAWToARGBRow(src_raw, dst_argb, width);
@@ -979,16 +3133,69 @@ int RAWToARGB(const uint8* src_raw, int src_stride_raw,
   return 0;
 }
 
+// Convert RAW to RGBA.
+LIBYUV_API
+int RAWToRGBA(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_rgba,
+              int dst_stride_rgba,
+              int width,
+              int height) {
+  int y;
+  void (*RAWToRGBARow)(const uint8_t* src_rgb, uint8_t* dst_rgba, int width) =
+      RAWToRGBARow_C;
+  if (!src_raw || !dst_rgba || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 && dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_rgba = 0;
+  }
+#if defined(HAS_RAWTORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToRGBARow = RAWToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToRGBARow = RAWToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RAWTORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToRGBARow = RAWToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToRGBARow = RAWToRGBARow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RAWToRGBARow(src_raw, dst_rgba, width);
+    src_raw += src_stride_raw;
+    dst_rgba += dst_stride_rgba;
+  }
+  return 0;
+}
+
 // Convert RGB565 to ARGB.
 LIBYUV_API
-int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
-                 uint8* dst_argb, int dst_stride_argb,
-                 int width, int height) {
+int RGB565ToARGB(const uint8_t* src_rgb565,
+                 int src_stride_rgb565,
+                 uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int width,
+                 int height) {
   int y;
-  void (*RGB565ToARGBRow)(const uint8* src_rgb565, uint8* dst_argb, int width) =
-      RGB565ToARGBRow_C;
-  if (!src_rgb565 || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*RGB565ToARGBRow)(const uint8_t* src_rgb565, uint8_t* dst_argb,
+                          int width) = RGB565ToARGBRow_C;
+  if (!src_rgb565 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -998,8 +3205,7 @@ int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
     src_stride_rgb565 = -src_stride_rgb565;
   }
   // Coalesce rows.
-  if (src_stride_rgb565 == width * 2 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_rgb565 == width * 2 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_rgb565 = dst_stride_argb = 0;
@@ -1028,6 +3234,30 @@ int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
     }
   }
 #endif
+#if defined(HAS_RGB565TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGB565ToARGBRow(src_rgb565, dst_argb, width);
@@ -1039,14 +3269,16 @@ int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
 
 // Convert ARGB1555 to ARGB.
 LIBYUV_API
-int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
-                   uint8* dst_argb, int dst_stride_argb,
-                   int width, int height) {
+int ARGB1555ToARGB(const uint8_t* src_argb1555,
+                   int src_stride_argb1555,
+                   uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   int width,
+                   int height) {
   int y;
-  void (*ARGB1555ToARGBRow)(const uint8* src_argb1555, uint8* dst_argb,
-      int width) = ARGB1555ToARGBRow_C;
-  if (!src_argb1555 || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*ARGB1555ToARGBRow)(const uint8_t* src_argb1555, uint8_t* dst_argb,
+                            int width) = ARGB1555ToARGBRow_C;
+  if (!src_argb1555 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1056,8 +3288,7 @@ int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
     src_stride_argb1555 = -src_stride_argb1555;
   }
   // Coalesce rows.
-  if (src_stride_argb1555 == width * 2 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb1555 == width * 2 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb1555 = dst_stride_argb = 0;
@@ -1086,6 +3317,30 @@ int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
     }
   }
 #endif
+#if defined(HAS_ARGB1555TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
@@ -1097,14 +3352,16 @@ int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
 
 // Convert ARGB4444 to ARGB.
 LIBYUV_API
-int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
-                   uint8* dst_argb, int dst_stride_argb,
-                   int width, int height) {
+int ARGB4444ToARGB(const uint8_t* src_argb4444,
+                   int src_stride_argb4444,
+                   uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   int width,
+                   int height) {
   int y;
-  void (*ARGB4444ToARGBRow)(const uint8* src_argb4444, uint8* dst_argb,
-      int width) = ARGB4444ToARGBRow_C;
-  if (!src_argb4444 || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*ARGB4444ToARGBRow)(const uint8_t* src_argb4444, uint8_t* dst_argb,
+                            int width) = ARGB4444ToARGBRow_C;
+  if (!src_argb4444 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1114,8 +3371,7 @@ int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
     src_stride_argb4444 = -src_stride_argb4444;
   }
   // Coalesce rows.
-  if (src_stride_argb4444 == width * 2 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb4444 == width * 2 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb4444 = dst_stride_argb = 0;
@@ -1144,6 +3400,30 @@ int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
     }
   }
 #endif
+#if defined(HAS_ARGB4444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
@@ -1153,20 +3433,237 @@ int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
   return 0;
 }
 
-// Convert NV12 to ARGB.
+// Convert AR30 to ARGB.
+LIBYUV_API
+int AR30ToARGB(const uint8_t* src_ar30,
+               int src_stride_ar30,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  int y;
+  if (!src_ar30 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar30 = src_ar30 + (height - 1) * src_stride_ar30;
+    src_stride_ar30 = -src_stride_ar30;
+  }
+  // Coalesce rows.
+  if (src_stride_ar30 == width * 4 && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar30 = dst_stride_argb = 0;
+  }
+  for (y = 0; y < height; ++y) {
+    AR30ToARGBRow_C(src_ar30, dst_argb, width);
+    src_ar30 += src_stride_ar30;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert AR30 to ABGR.
+LIBYUV_API
+int AR30ToABGR(const uint8_t* src_ar30,
+               int src_stride_ar30,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  int y;
+  if (!src_ar30 || !dst_abgr || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar30 = src_ar30 + (height - 1) * src_stride_ar30;
+    src_stride_ar30 = -src_stride_ar30;
+  }
+  // Coalesce rows.
+  if (src_stride_ar30 == width * 4 && dst_stride_abgr == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar30 = dst_stride_abgr = 0;
+  }
+  for (y = 0; y < height; ++y) {
+    AR30ToABGRRow_C(src_ar30, dst_abgr, width);
+    src_ar30 += src_stride_ar30;
+    dst_abgr += dst_stride_abgr;
+  }
+  return 0;
+}
+
+// Convert AR30 to AB30.
+LIBYUV_API
+int AR30ToAB30(const uint8_t* src_ar30,
+               int src_stride_ar30,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  int y;
+  if (!src_ar30 || !dst_ab30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar30 = src_ar30 + (height - 1) * src_stride_ar30;
+    src_stride_ar30 = -src_stride_ar30;
+  }
+  // Coalesce rows.
+  if (src_stride_ar30 == width * 4 && dst_stride_ab30 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar30 = dst_stride_ab30 = 0;
+  }
+  for (y = 0; y < height; ++y) {
+    AR30ToAB30Row_C(src_ar30, dst_ab30, width);
+    src_ar30 += src_stride_ar30;
+    dst_ab30 += dst_stride_ab30;
+  }
+  return 0;
+}
+
+// Convert AR64 to ARGB.
+LIBYUV_API
+int AR64ToARGB(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  int y;
+  void (*AR64ToARGBRow)(const uint16_t* src_ar64, uint8_t* dst_argb,
+                        int width) = AR64ToARGBRow_C;
+  if (!src_ar64 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar64 = src_ar64 + (height - 1) * src_stride_ar64;
+    src_stride_ar64 = -src_stride_ar64;
+  }
+  // Coalesce rows.
+  if (src_stride_ar64 == width * 4 && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar64 = dst_stride_argb = 0;
+  }
+#if defined(HAS_AR64TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    AR64ToARGBRow = AR64ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      AR64ToARGBRow = AR64ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_AR64TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AR64ToARGBRow = AR64ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      AR64ToARGBRow = AR64ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_AR64TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AR64ToARGBRow = AR64ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      AR64ToARGBRow = AR64ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    AR64ToARGBRow(src_ar64, dst_argb, width);
+    src_ar64 += src_stride_ar64;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert AB64 to ARGB.
 LIBYUV_API
-int NV12ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_uv, int src_stride_uv,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int AB64ToARGB(const uint16_t* src_ab64,
+               int src_stride_ab64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
   int y;
-  void (*NV12ToARGBRow)(const uint8* y_buf,
-                        const uint8* uv_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = NV12ToARGBRow_C;
-  if (!src_y || !src_uv || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*AB64ToARGBRow)(const uint16_t* src_ar64, uint8_t* dst_argb,
+                        int width) = AB64ToARGBRow_C;
+  if (!src_ab64 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ab64 = src_ab64 + (height - 1) * src_stride_ab64;
+    src_stride_ab64 = -src_stride_ab64;
+  }
+  // Coalesce rows.
+  if (src_stride_ab64 == width * 4 && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ab64 = dst_stride_argb = 0;
+  }
+#if defined(HAS_AB64TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    AB64ToARGBRow = AB64ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      AB64ToARGBRow = AB64ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_AB64TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AB64ToARGBRow = AB64ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      AB64ToARGBRow = AB64ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_AB64TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AB64ToARGBRow = AB64ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      AB64ToARGBRow = AB64ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    AB64ToARGBRow(src_ab64, dst_argb, width);
+    src_ab64 += src_stride_ab64;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert NV12 to ARGB with matrix.
+LIBYUV_API
+int NV12ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*NV12ToARGBRow)(
+      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = NV12ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1199,9 +3696,33 @@ int NV12ToARGB(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_NV12TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
-    NV12ToARGBRow(src_y, src_uv, dst_argb, &kYuvI601Constants, width);
+    NV12ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;
     src_y += src_stride_y;
     if (y & 1) {
@@ -1211,20 +3732,23 @@ int NV12ToARGB(const uint8* src_y, int src_stride_y,
   return 0;
 }
 
-// Convert NV21 to ARGB.
+// Convert NV21 to ARGB with matrix.
 LIBYUV_API
-int NV21ToARGB(const uint8* src_y, int src_stride_y,
-               const uint8* src_uv, int src_stride_uv,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int NV21ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_vu,
+                     int src_stride_vu,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
-  void (*NV21ToARGBRow)(const uint8* y_buf,
-                        const uint8* uv_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = NV21ToARGBRow_C;
-  if (!src_y || !src_uv || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*NV21ToARGBRow)(
+      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = NV21ToARGBRow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_vu || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1257,92 +3781,350 @@ int NV21ToARGB(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_NV21TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToARGBRow = NV21ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
-    NV21ToARGBRow(src_y, src_uv, dst_argb, &kYuvI601Constants, width);
+    NV21ToARGBRow(src_y, src_vu, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;
     src_y += src_stride_y;
     if (y & 1) {
-      src_uv += src_stride_uv;
+      src_vu += src_stride_vu;
     }
   }
   return 0;
 }
 
-// Convert M420 to ARGB.
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return NV12ToARGBMatrix(src_y, src_stride_y, src_uv, src_stride_uv, dst_argb,
+                          dst_stride_argb, &kYuvI601Constants, width, height);
+}
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return NV21ToARGBMatrix(src_y, src_stride_y, src_vu, src_stride_vu, dst_argb,
+                          dst_stride_argb, &kYuvI601Constants, width, height);
+}
+
+// Convert NV12 to ABGR.
+// To output ABGR instead of ARGB swap the UV and use a mirrored yuv matrix.
+// To swap the UV use NV12 instead of NV21.LIBYUV_API
 LIBYUV_API
-int M420ToARGB(const uint8* src_m420, int src_stride_m420,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int NV12ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return NV21ToARGBMatrix(src_y, src_stride_y, src_uv, src_stride_uv, dst_abgr,
+                          dst_stride_abgr, &kYvuI601Constants, width, height);
+}
+
+// Convert NV21 to ABGR.
+LIBYUV_API
+int NV21ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return NV12ToARGBMatrix(src_y, src_stride_y, src_vu, src_stride_vu, dst_abgr,
+                          dst_stride_abgr, &kYvuI601Constants, width, height);
+}
+
+// TODO(fbarchard): Consider SSSE3 2 step conversion.
+// Convert NV12 to RGB24 with matrix.
+LIBYUV_API
+int NV12ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_uv,
+                      int src_stride_uv,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
   int y;
-  void (*NV12ToARGBRow)(const uint8* y_buf,
-                        const uint8* uv_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = NV12ToARGBRow_C;
-  if (!src_m420 || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*NV12ToRGB24Row)(
+      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = NV12ToRGB24Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_rgb24 || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
-    dst_stride_argb = -dst_stride_argb;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
   }
-#if defined(HAS_NV12TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+#if defined(HAS_NV12TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToRGB24Row = NV12ToRGB24Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+      NV12ToRGB24Row = NV12ToRGB24Row_NEON;
     }
   }
 #endif
-#if defined(HAS_NV12TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+#if defined(HAS_NV12TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToRGB24Row = NV12ToRGB24Row_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+      NV12ToRGB24Row = NV12ToRGB24Row_SSSE3;
     }
   }
 #endif
-#if defined(HAS_NV12TOARGBROW_NEON)
+#if defined(HAS_NV12TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToRGB24Row = NV12ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      NV12ToRGB24Row = NV12ToRGB24Row_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB24Row(src_y, src_uv, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to RGB24 with matrix.
+LIBYUV_API
+int NV21ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_vu,
+                      int src_stride_vu,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
+  int y;
+  void (*NV21ToRGB24Row)(
+      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = NV21ToRGB24Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_vu || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_NV21TORGB24ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    NV21ToRGB24Row = NV21ToRGB24Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_NEON;
+      NV21ToRGB24Row = NV21ToRGB24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToRGB24Row = NV21ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToRGB24Row = NV21ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToRGB24Row = NV21ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      NV21ToRGB24Row = NV21ToRGB24Row_AVX2;
     }
   }
 #endif
 
-  for (y = 0; y < height - 1; y += 2) {
-    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb,
-                  &kYuvI601Constants, width);
-    NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2,
-                  dst_argb + dst_stride_argb, &kYuvI601Constants, width);
-    dst_argb += dst_stride_argb * 2;
-    src_m420 += src_stride_m420 * 3;
+  for (y = 0; y < height; ++y) {
+    NV21ToRGB24Row(src_y, src_vu, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB24.
+LIBYUV_API
+int NV12ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_uv,
+                int src_stride_uv,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return NV12ToRGB24Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                           dst_rgb24, dst_stride_rgb24, &kYuvI601Constants,
+                           width, height);
+}
+
+// Convert NV21 to RGB24.
+LIBYUV_API
+int NV21ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_vu,
+                int src_stride_vu,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return NV21ToRGB24Matrix(src_y, src_stride_y, src_vu, src_stride_vu,
+                           dst_rgb24, dst_stride_rgb24, &kYuvI601Constants,
+                           width, height);
+}
+
+// Convert NV12 to RAW.
+LIBYUV_API
+int NV12ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_uv,
+              int src_stride_uv,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return NV21ToRGB24Matrix(src_y, src_stride_y, src_uv, src_stride_uv, dst_raw,
+                           dst_stride_raw, &kYvuI601Constants, width, height);
+}
+
+// Convert NV21 to RAW.
+LIBYUV_API
+int NV21ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_vu,
+              int src_stride_vu,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return NV12ToRGB24Matrix(src_y, src_stride_y, src_vu, src_stride_vu, dst_raw,
+                           dst_stride_raw, &kYvuI601Constants, width, height);
+}
+
+// Convert NV21 to YUV24
+int NV21ToYUV24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_vu,
+                int src_stride_vu,
+                uint8_t* dst_yuv24,
+                int dst_stride_yuv24,
+                int width,
+                int height) {
+  int y;
+  void (*NV21ToYUV24Row)(const uint8_t* src_y, const uint8_t* src_vu,
+                         uint8_t* dst_yuv24, int width) = NV21ToYUV24Row_C;
+  if (!src_y || !src_vu || !dst_yuv24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuv24 = dst_yuv24 + (height - 1) * dst_stride_yuv24;
+    dst_stride_yuv24 = -dst_stride_yuv24;
+  }
+#if defined(HAS_NV21TOYUV24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV21ToYUV24Row = NV21ToYUV24Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToYUV24Row = NV21ToYUV24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOYUV24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToYUV24Row = NV21ToYUV24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToYUV24Row = NV21ToYUV24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOYUV24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToYUV24Row = NV21ToYUV24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      NV21ToYUV24Row = NV21ToYUV24Row_AVX2;
+    }
   }
-  if (height & 1) {
-    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb,
-                  &kYuvI601Constants, width);
+#endif
+  for (y = 0; y < height; ++y) {
+    NV21ToYUV24Row(src_y, src_vu, dst_yuv24, width);
+    dst_yuv24 += dst_stride_yuv24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
   }
   return 0;
 }
 
 // Convert YUY2 to ARGB.
 LIBYUV_API
-int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int YUY2ToARGB(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
   int y;
-  void (*YUY2ToARGBRow)(const uint8* src_yuy2,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) =
+  void (*YUY2ToARGBRow)(const uint8_t* src_yuy2, uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants, int width) =
       YUY2ToARGBRow_C;
-  if (!src_yuy2 || !dst_argb ||
-      width <= 0 || height == 0) {
+  if (!src_yuy2 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1352,8 +4134,7 @@ int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
     src_stride_yuy2 = -src_stride_yuy2;
   }
   // Coalesce rows.
-  if (src_stride_yuy2 == width * 2 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_yuy2 == width * 2 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_yuy2 = dst_stride_argb = 0;
@@ -1382,6 +4163,22 @@ int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
     }
   }
 #endif
+#if defined(HAS_YUY2TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     YUY2ToARGBRow(src_yuy2, dst_argb, &kYuvI601Constants, width);
     src_yuy2 += src_stride_yuy2;
@@ -1392,17 +4189,17 @@ int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
 
 // Convert UYVY to ARGB.
 LIBYUV_API
-int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int UYVYToARGB(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
   int y;
-  void (*UYVYToARGBRow)(const uint8* src_uyvy,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) =
+  void (*UYVYToARGBRow)(const uint8_t* src_uyvy, uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants, int width) =
       UYVYToARGBRow_C;
-  if (!src_uyvy || !dst_argb ||
-      width <= 0 || height == 0) {
+  if (!src_uyvy || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1412,8 +4209,7 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
     src_stride_uyvy = -src_stride_uyvy;
   }
   // Coalesce rows.
-  if (src_stride_uyvy == width * 2 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_uyvy == width * 2 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_uyvy = dst_stride_argb = 0;
@@ -1442,6 +4238,22 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
     }
   }
 #endif
+#if defined(HAS_UYVYTOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     UYVYToARGBRow(src_uyvy, dst_argb, &kYuvI601Constants, width);
     src_uyvy += src_stride_uyvy;
@@ -1449,6 +4261,3726 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
   }
   return 0;
 }
+static void WeavePixels(const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        int src_pixel_stride_uv,
+                        uint8_t* dst_uv,
+                        int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_uv[0] = *src_u;
+    dst_uv[1] = *src_v;
+    dst_uv += 2;
+    src_u += src_pixel_stride_uv;
+    src_v += src_pixel_stride_uv;
+  }
+}
+
+// Convert Android420 to ARGB with matrix.
+LIBYUV_API
+int Android420ToARGBMatrix(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height) {
+  int y;
+  uint8_t* dst_uv;
+  const ptrdiff_t vu_off = src_v - src_u;
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+
+  // I420
+  if (src_pixel_stride_uv == 1) {
+    return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_argb, dst_stride_argb,
+                            yuvconstants, width, height);
+    // NV21
+  }
+  if (src_pixel_stride_uv == 2 && vu_off == -1 &&
+      src_stride_u == src_stride_v) {
+    return NV21ToARGBMatrix(src_y, src_stride_y, src_v, src_stride_v, dst_argb,
+                            dst_stride_argb, yuvconstants, width, height);
+    // NV12
+  }
+  if (src_pixel_stride_uv == 2 && vu_off == 1 && src_stride_u == src_stride_v) {
+    return NV12ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, dst_argb,
+                            dst_stride_argb, yuvconstants, width, height);
+  }
+
+  // General case fallback creates NV12
+  align_buffer_64(plane_uv, halfwidth * 2 * halfheight);
+  dst_uv = plane_uv;
+  for (y = 0; y < halfheight; ++y) {
+    WeavePixels(src_u, src_v, src_pixel_stride_uv, dst_uv, halfwidth);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += halfwidth * 2;
+  }
+  NV12ToARGBMatrix(src_y, src_stride_y, plane_uv, halfwidth * 2, dst_argb,
+                   dst_stride_argb, yuvconstants, width, height);
+  free_aligned_buffer_64(plane_uv);
+  return 0;
+}
+
+// Convert Android420 to ARGB.
+LIBYUV_API
+int Android420ToARGB(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     int src_pixel_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     int width,
+                     int height) {
+  return Android420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                                src_stride_v, src_pixel_stride_uv, dst_argb,
+                                dst_stride_argb, &kYuvI601Constants, width,
+                                height);
+}
+
+// Convert Android420 to ABGR.
+LIBYUV_API
+int Android420ToABGR(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     int src_pixel_stride_uv,
+                     uint8_t* dst_abgr,
+                     int dst_stride_abgr,
+                     int width,
+                     int height) {
+  return Android420ToARGBMatrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                                src_stride_u, src_pixel_stride_uv, dst_abgr,
+                                dst_stride_abgr, &kYvuI601Constants, width,
+                                height);
+}
+
+// Convert I422 to RGBA with matrix.
+LIBYUV_API
+int I422ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGBARow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgba || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGBARow = I422ToRGBARow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGBARow = I422ToRGBARow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGBARow = I422ToRGBARow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height) {
+  return I422ToRGBAMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_rgba, dst_stride_rgba,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height) {
+  return I422ToRGBAMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_bgra, dst_stride_bgra,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert NV12 to RGB565 with matrix.
+LIBYUV_API
+int NV12ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_uv,
+                       int src_stride_uv,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height) {
+  int y;
+  void (*NV12ToRGB565Row)(
+      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = NV12ToRGB565Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_LSX;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, yuvconstants, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return NV12ToRGB565Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                            dst_rgb565, dst_stride_rgb565, &kYuvI601Constants,
+                            width, height);
+}
+
+// Convert I422 to RGBA with matrix.
+LIBYUV_API
+int I420ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGBARow_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgba || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGBARow = I422ToRGBARow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGBARow = I422ToRGBARow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGBARow = I422ToRGBARow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGBA.
+LIBYUV_API
+int I420ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height) {
+  return I420ToRGBAMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_rgba, dst_stride_rgba,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I420 to BGRA.
+LIBYUV_API
+int I420ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height) {
+  return I420ToRGBAMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_bgra, dst_stride_bgra,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I420 to RGB24 with matrix.
+LIBYUV_API
+int I420ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
+  int y;
+  void (*I422ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                         const uint8_t* v_buf, uint8_t* rgb_buf,
+                         const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB24Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB24Row = I422ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB24Row = I422ToRGB24Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB24.
+LIBYUV_API
+int I420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvI601Constants, width, height);
+}
+
+// Convert I420 to RAW.
+LIBYUV_API
+int I420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuI601Constants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert J420 to RGB24.
+LIBYUV_API
+int J420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvJPEGConstants, width, height);
+}
+
+// Convert J420 to RAW.
+LIBYUV_API
+int J420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuJPEGConstants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert H420 to RGB24.
+LIBYUV_API
+int H420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvH709Constants, width, height);
+}
+
+// Convert H420 to RAW.
+LIBYUV_API
+int H420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuH709Constants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert I422 to RGB24 with matrix.
+LIBYUV_API
+int I422ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
+  int y;
+  void (*I422ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                         const uint8_t* v_buf, uint8_t* rgb_buf,
+                         const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB24Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB24Row = I422ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB24Row = I422ToRGB24Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGB24.
+LIBYUV_API
+int I422ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I422ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvI601Constants, width, height);
+}
+
+// Convert I422 to RAW.
+LIBYUV_API
+int I422ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I422ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuI601Constants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert I420 to ARGB1555.
+LIBYUV_API
+int I420ToARGB1555(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb1555,
+                   int dst_stride_argb1555,
+                   int width,
+                   int height) {
+  int y;
+  void (*I422ToARGB1555Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                            const uint8_t* v_buf, uint8_t* rgb_buf,
+                            const struct YuvConstants* yuvconstants,
+                            int width) = I422ToARGB1555Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb1555 || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
+    dst_stride_argb1555 = -dst_stride_argb1555;
+  }
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, &kYuvI601Constants,
+                      width);
+    dst_argb1555 += dst_stride_argb1555;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ARGB4444.
+LIBYUV_API
+int I420ToARGB4444(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb4444,
+                   int dst_stride_argb4444,
+                   int width,
+                   int height) {
+  int y;
+  void (*I422ToARGB4444Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                            const uint8_t* v_buf, uint8_t* rgb_buf,
+                            const struct YuvConstants* yuvconstants,
+                            int width) = I422ToARGB4444Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb4444 || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
+    dst_stride_argb4444 = -dst_stride_argb4444;
+  }
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, &kYuvI601Constants,
+                      width);
+    dst_argb4444 += dst_stride_argb4444;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565 with specified color matrix.
+LIBYUV_API
+int I420ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                          const uint8_t* v_buf, uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB565Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB565Row = I422ToRGB565Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, yuvconstants, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565.
+LIBYUV_API
+int I420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvI601Constants, width, height);
+}
+
+// Convert J420 to RGB565.
+LIBYUV_API
+int J420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvJPEGConstants, width, height);
+}
+
+// Convert H420 to RGB565.
+LIBYUV_API
+int H420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvH709Constants, width, height);
+}
+
+// Convert I422 to RGB565 with specified color matrix.
+LIBYUV_API
+int I422ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                          const uint8_t* v_buf, uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB565Row_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB565Row = I422ToRGB565Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, yuvconstants, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGB565.
+LIBYUV_API
+int I422ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I422ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvI601Constants, width, height);
+}
+
+// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
+static const uint8_t kDither565_4x4[16] = {
+    0, 4, 1, 5, 6, 2, 7, 3, 1, 5, 0, 4, 7, 3, 6, 2,
+};
+
+// Convert I420 to RGB565 with dithering.
+LIBYUV_API
+int I420ToRGB565Dither(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const uint8_t* dither4x4,
+                       int width,
+                       int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToARGBRow_C;
+  void (*ARGBToRGB565DitherRow)(const uint8_t* src_argb, uint8_t* dst_rgb,
+                                const uint32_t dither4, int width) =
+      ARGBToRGB565DitherRow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+  if (!dither4x4) {
+    dither4x4 = kDither565_4x4;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGBRow = I422ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_LASX;
+    }
+  }
+#endif
+  {
+    // Allocate a row of argb.
+    align_buffer_64(row_argb, width * 4);
+    for (y = 0; y < height; ++y) {
+      I422ToARGBRow(src_y, src_u, src_v, row_argb, &kYuvI601Constants, width);
+      ARGBToRGB565DitherRow(row_argb, dst_rgb565,
+                            *(const uint32_t*)(dither4x4 + ((y & 3) << 2)),
+                            width);
+      dst_rgb565 += dst_stride_rgb565;
+      src_y += src_stride_y;
+      if (y & 1) {
+        src_u += src_stride_u;
+        src_v += src_stride_v;
+      }
+    }
+    free_aligned_buffer_64(row_argb);
+  }
+  return 0;
+}
+
+// Convert I420 to AR30 with matrix.
+LIBYUV_API
+int I420ToAR30Matrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I422ToAR30Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToAR30Row_C;
+
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+
+#if defined(HAS_I422TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToAR30Row = I422ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToAR30Row = I422ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToAR30Row = I422ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToAR30Row = I422ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to AR30.
+LIBYUV_API
+int I420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert H420 to AR30.
+LIBYUV_API
+int H420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYvuH709Constants, width, height);
+}
+
+// Convert I420 to AB30.
+LIBYUV_API
+int I420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuI601Constants, width, height);
+}
+
+// Convert H420 to AB30.
+LIBYUV_API
+int H420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuH709Constants, width, height);
+}
+
+static int I420ToARGBMatrixBilinear(const uint8_t* src_y,
+                                    int src_stride_y,
+                                    const uint8_t* src_u,
+                                    int src_stride_u,
+                                    const uint8_t* src_v,
+                                    int src_stride_v,
+                                    uint8_t* dst_argb,
+                                    int dst_stride_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I444ToARGBRow_C;
+  void (*Scale2RowUp_Bilinear)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                               uint8_t* dst_ptr, ptrdiff_t dst_stride,
+                               int dst_width) = ScaleRowUp2_Bilinear_Any_C;
+  void (*ScaleRowUp2_Linear)(const uint8_t* src_ptr, uint8_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToARGBRow = I444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToARGBRow = I444ToARGBRow_LASX;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_SSE2;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_SSSE3;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_AVX2;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_NEON;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 4);
+  uint8_t* temp_u_1 = row;
+  uint8_t* temp_u_2 = row + row_size;
+  uint8_t* temp_v_1 = row + row_size * 2;
+  uint8_t* temp_v_2 = row + row_size * 3;
+
+  ScaleRowUp2_Linear(src_u, temp_u_1, width);
+  ScaleRowUp2_Linear(src_v, temp_v_1, width);
+  I444ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear(src_u, src_stride_u, temp_u_1, row_size, width);
+    Scale2RowUp_Bilinear(src_v, src_stride_v, temp_v_1, row_size, width);
+    I444ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    I444ToARGBRow(src_y, temp_u_2, temp_v_2, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    ScaleRowUp2_Linear(src_u, temp_u_1, width);
+    ScaleRowUp2_Linear(src_v, temp_v_1, width);
+    I444ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I422ToARGBMatrixLinear(const uint8_t* src_y,
+                                  int src_stride_y,
+                                  const uint8_t* src_u,
+                                  int src_stride_u,
+                                  const uint8_t* src_v,
+                                  int src_stride_v,
+                                  uint8_t* dst_argb,
+                                  int dst_stride_argb,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I444ToARGBRow_C;
+  void (*ScaleRowUp2_Linear)(const uint8_t* src_ptr, uint8_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToARGBRow = I444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToARGBRow = I444ToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 2);
+  uint8_t* temp_u = row;
+  uint8_t* temp_v = row + row_size;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear(src_u, temp_u, width);
+    ScaleRowUp2_Linear(src_v, temp_v, width);
+    I444ToARGBRow(src_y, temp_u, temp_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I420ToRGB24MatrixBilinear(const uint8_t* src_y,
+                                     int src_stride_y,
+                                     const uint8_t* src_u,
+                                     int src_stride_u,
+                                     const uint8_t* src_v,
+                                     int src_stride_v,
+                                     uint8_t* dst_rgb24,
+                                     int dst_stride_rgb24,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width,
+                                     int height) {
+  int y;
+  void (*I444ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                         const uint8_t* v_buf, uint8_t* rgb_buf,
+                         const struct YuvConstants* yuvconstants, int width) =
+      I444ToRGB24Row_C;
+  void (*Scale2RowUp_Bilinear)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                               uint8_t* dst_ptr, ptrdiff_t dst_stride,
+                               int dst_width) = ScaleRowUp2_Bilinear_Any_C;
+  void (*ScaleRowUp2_Linear)(const uint8_t* src_ptr, uint8_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I444TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToRGB24Row = I444ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToRGB24Row = I444ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToRGB24Row = I444ToRGB24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToRGB24Row = I444ToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToRGB24Row = I444ToRGB24Row_LASX;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_SSE2;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_SSSE3;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_AVX2;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_NEON;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 4);
+  uint8_t* temp_u_1 = row;
+  uint8_t* temp_u_2 = row + row_size;
+  uint8_t* temp_v_1 = row + row_size * 2;
+  uint8_t* temp_v_2 = row + row_size * 3;
+
+  ScaleRowUp2_Linear(src_u, temp_u_1, width);
+  ScaleRowUp2_Linear(src_v, temp_v_1, width);
+  I444ToRGB24Row(src_y, temp_u_1, temp_v_1, dst_rgb24, yuvconstants, width);
+  dst_rgb24 += dst_stride_rgb24;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear(src_u, src_stride_u, temp_u_1, row_size, width);
+    Scale2RowUp_Bilinear(src_v, src_stride_v, temp_v_1, row_size, width);
+    I444ToRGB24Row(src_y, temp_u_1, temp_v_1, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    I444ToRGB24Row(src_y, temp_u_2, temp_v_2, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    ScaleRowUp2_Linear(src_u, temp_u_1, width);
+    ScaleRowUp2_Linear(src_v, temp_v_1, width);
+    I444ToRGB24Row(src_y, temp_u_1, temp_v_1, dst_rgb24, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I010ToAR30MatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_u,
+                                    int src_stride_u,
+                                    const uint16_t* src_v,
+                                    int src_stride_v,
+                                    uint8_t* dst_ar30,
+                                    int dst_stride_ar30,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*I410ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToAR30Row_C;
+  void (*Scale2RowUp_Bilinear_12)(
+      const uint16_t* src_ptr, ptrdiff_t src_stride, uint16_t* dst_ptr,
+      ptrdiff_t dst_stride, int dst_width) = ScaleRowUp2_Bilinear_16_Any_C;
+  void (*ScaleRowUp2_Linear_12)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                                int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToAR30Row = I410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToAR30Row = I410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToAR30Row = I410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToAR30Row = I410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_AVX2;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_NEON;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 4 * sizeof(uint16_t));
+  uint16_t* temp_u_1 = (uint16_t*)(row);
+  uint16_t* temp_u_2 = (uint16_t*)(row) + row_size;
+  uint16_t* temp_v_1 = (uint16_t*)(row) + row_size * 2;
+  uint16_t* temp_v_2 = (uint16_t*)(row) + row_size * 3;
+
+  ScaleRowUp2_Linear_12(src_u, temp_u_1, width);
+  ScaleRowUp2_Linear_12(src_v, temp_v_1, width);
+  I410ToAR30Row(src_y, temp_u_1, temp_v_1, dst_ar30, yuvconstants, width);
+  dst_ar30 += dst_stride_ar30;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear_12(src_u, src_stride_u, temp_u_1, row_size, width);
+    Scale2RowUp_Bilinear_12(src_v, src_stride_v, temp_v_1, row_size, width);
+    I410ToAR30Row(src_y, temp_u_1, temp_v_1, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    I410ToAR30Row(src_y, temp_u_2, temp_v_2, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    ScaleRowUp2_Linear_12(src_u, temp_u_1, width);
+    ScaleRowUp2_Linear_12(src_v, temp_v_1, width);
+    I410ToAR30Row(src_y, temp_u_1, temp_v_1, dst_ar30, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+
+  return 0;
+}
+
+static int I210ToAR30MatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_u,
+                                  int src_stride_u,
+                                  const uint16_t* src_v,
+                                  int src_stride_v,
+                                  uint8_t* dst_ar30,
+                                  int dst_stride_ar30,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*I410ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToAR30Row_C;
+  void (*ScaleRowUp2_Linear_12)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                                int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToAR30Row = I410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToAR30Row = I410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToAR30Row = I410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToAR30Row = I410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 2 * sizeof(uint16_t));
+  uint16_t* temp_u = (uint16_t*)(row);
+  uint16_t* temp_v = (uint16_t*)(row) + row_size;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear_12(src_u, temp_u, width);
+    ScaleRowUp2_Linear_12(src_v, temp_v, width);
+    I410ToAR30Row(src_y, temp_u, temp_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I010ToARGBMatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_u,
+                                    int src_stride_u,
+                                    const uint16_t* src_v,
+                                    int src_stride_v,
+                                    uint8_t* dst_argb,
+                                    int dst_stride_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*I410ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToARGBRow_C;
+  void (*Scale2RowUp_Bilinear_12)(
+      const uint16_t* src_ptr, ptrdiff_t src_stride, uint16_t* dst_ptr,
+      ptrdiff_t dst_stride, int dst_width) = ScaleRowUp2_Bilinear_16_Any_C;
+  void (*ScaleRowUp2_Linear_12)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                                int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToARGBRow = I410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToARGBRow = I410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToARGBRow = I410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToARGBRow = I410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_AVX2;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_NEON;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 4 * sizeof(uint16_t));
+  uint16_t* temp_u_1 = (uint16_t*)(row);
+  uint16_t* temp_u_2 = (uint16_t*)(row) + row_size;
+  uint16_t* temp_v_1 = (uint16_t*)(row) + row_size * 2;
+  uint16_t* temp_v_2 = (uint16_t*)(row) + row_size * 3;
+
+  ScaleRowUp2_Linear_12(src_u, temp_u_1, width);
+  ScaleRowUp2_Linear_12(src_v, temp_v_1, width);
+  I410ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear_12(src_u, src_stride_u, temp_u_1, row_size, width);
+    Scale2RowUp_Bilinear_12(src_v, src_stride_v, temp_v_1, row_size, width);
+    I410ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    I410ToARGBRow(src_y, temp_u_2, temp_v_2, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    ScaleRowUp2_Linear_12(src_u, temp_u_1, width);
+    ScaleRowUp2_Linear_12(src_v, temp_v_1, width);
+    I410ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I210ToARGBMatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_u,
+                                  int src_stride_u,
+                                  const uint16_t* src_v,
+                                  int src_stride_v,
+                                  uint8_t* dst_argb,
+                                  int dst_stride_argb,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*I410ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToARGBRow_C;
+  void (*ScaleRowUp2_Linear_12)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                                int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToARGBRow = I410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToARGBRow = I410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToARGBRow = I410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToARGBRow = I410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 2 * sizeof(uint16_t));
+  uint16_t* temp_u = (uint16_t*)(row);
+  uint16_t* temp_v = (uint16_t*)(row) + row_size;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear_12(src_u, temp_u, width);
+    ScaleRowUp2_Linear_12(src_v, temp_v, width);
+    I410ToARGBRow(src_y, temp_u, temp_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I420AlphaToARGBMatrixBilinear(
+    const uint8_t* src_y,
+    int src_stride_y,
+    const uint8_t* src_u,
+    int src_stride_u,
+    const uint8_t* src_v,
+    int src_stride_v,
+    const uint8_t* src_a,
+    int src_stride_a,
+    uint8_t* dst_argb,
+    int dst_stride_argb,
+    const struct YuvConstants* yuvconstants,
+    int width,
+    int height,
+    int attenuate) {
+  int y;
+  void (*I444AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I444AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*Scale2RowUp_Bilinear)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                               uint8_t* dst_ptr, ptrdiff_t dst_stride,
+                               int dst_width) = ScaleRowUp2_Bilinear_Any_C;
+  void (*ScaleRowUp2_Linear)(const uint8_t* src_ptr, uint8_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_SSE2;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_SSSE3;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_AVX2;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear = ScaleRowUp2_Bilinear_Any_NEON;
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 4);
+  uint8_t* temp_u_1 = row;
+  uint8_t* temp_u_2 = row + row_size;
+  uint8_t* temp_v_1 = row + row_size * 2;
+  uint8_t* temp_v_2 = row + row_size * 3;
+
+  ScaleRowUp2_Linear(src_u, temp_u_1, width);
+  ScaleRowUp2_Linear(src_v, temp_v_1, width);
+  I444AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                     width);
+  if (attenuate) {
+    ARGBAttenuateRow(dst_argb, dst_argb, width);
+  }
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+  src_a += src_stride_a;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear(src_u, src_stride_u, temp_u_1, row_size, width);
+    Scale2RowUp_Bilinear(src_v, src_stride_v, temp_v_1, row_size, width);
+    I444AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_a += src_stride_a;
+    I444AlphaToARGBRow(src_y, temp_u_2, temp_v_2, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    src_a += src_stride_a;
+  }
+
+  if (!(height & 1)) {
+    ScaleRowUp2_Linear(src_u, temp_u_1, width);
+    ScaleRowUp2_Linear(src_v, temp_v_1, width);
+    I444AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I422AlphaToARGBMatrixLinear(const uint8_t* src_y,
+                                       int src_stride_y,
+                                       const uint8_t* src_u,
+                                       int src_stride_u,
+                                       const uint8_t* src_v,
+                                       int src_stride_v,
+                                       const uint8_t* src_a,
+                                       int src_stride_a,
+                                       uint8_t* dst_argb,
+                                       int dst_stride_argb,
+                                       const struct YuvConstants* yuvconstants,
+                                       int width,
+                                       int height,
+                                       int attenuate) {
+  int y;
+  void (*I444AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I444AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*ScaleRowUp2_Linear)(const uint8_t* src_ptr, uint8_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 2);
+  uint8_t* temp_u = row;
+  uint8_t* temp_v = row + row_size;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear(src_u, temp_u, width);
+    ScaleRowUp2_Linear(src_v, temp_v, width);
+    I444AlphaToARGBRow(src_y, temp_u, temp_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I010AlphaToARGBMatrixBilinear(
+    const uint16_t* src_y,
+    int src_stride_y,
+    const uint16_t* src_u,
+    int src_stride_u,
+    const uint16_t* src_v,
+    int src_stride_v,
+    const uint16_t* src_a,
+    int src_stride_a,
+    uint8_t* dst_argb,
+    int dst_stride_argb,
+    const struct YuvConstants* yuvconstants,
+    int width,
+    int height,
+    int attenuate) {
+  int y;
+  void (*I410AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I410AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*Scale2RowUp_Bilinear_12)(
+      const uint16_t* src_ptr, ptrdiff_t src_stride, uint16_t* dst_ptr,
+      ptrdiff_t dst_stride, int dst_width) = ScaleRowUp2_Bilinear_16_Any_C;
+  void (*ScaleRowUp2_Linear_12)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                                int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_AVX2;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_BILINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear_12 = ScaleRowUp2_Bilinear_12_Any_NEON;
+    ScaleRowUp2_Linear_12 = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 4 * sizeof(uint16_t));
+  uint16_t* temp_u_1 = (uint16_t*)(row);
+  uint16_t* temp_u_2 = (uint16_t*)(row) + row_size;
+  uint16_t* temp_v_1 = (uint16_t*)(row) + row_size * 2;
+  uint16_t* temp_v_2 = (uint16_t*)(row) + row_size * 3;
+
+  ScaleRowUp2_Linear_12(src_u, temp_u_1, width);
+  ScaleRowUp2_Linear_12(src_v, temp_v_1, width);
+  I410AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                     width);
+  if (attenuate) {
+    ARGBAttenuateRow(dst_argb, dst_argb, width);
+  }
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+  src_a += src_stride_a;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear_12(src_u, src_stride_u, temp_u_1, row_size, width);
+    Scale2RowUp_Bilinear_12(src_v, src_stride_v, temp_v_1, row_size, width);
+    I410AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_a += src_stride_a;
+    I410AlphaToARGBRow(src_y, temp_u_2, temp_v_2, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_a += src_stride_a;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    ScaleRowUp2_Linear_12(src_u, temp_u_1, width);
+    ScaleRowUp2_Linear_12(src_v, temp_v_1, width);
+    I410AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I210AlphaToARGBMatrixLinear(const uint16_t* src_y,
+                                       int src_stride_y,
+                                       const uint16_t* src_u,
+                                       int src_stride_u,
+                                       const uint16_t* src_v,
+                                       int src_stride_v,
+                                       const uint16_t* src_a,
+                                       int src_stride_a,
+                                       uint8_t* dst_argb,
+                                       int dst_stride_argb,
+                                       const struct YuvConstants* yuvconstants,
+                                       int width,
+                                       int height,
+                                       int attenuate) {
+  int y;
+  void (*I410AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I410AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*ScaleRowUp2_Linear)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 2 * sizeof(uint16_t));
+  uint16_t* temp_u = (uint16_t*)(row);
+  uint16_t* temp_v = (uint16_t*)(row) + row_size;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear(src_u, temp_u, width);
+    ScaleRowUp2_Linear(src_v, temp_v, width);
+    I410AlphaToARGBRow(src_y, temp_u, temp_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P010ToARGBMatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_uv,
+                                    int src_stride_uv,
+                                    uint8_t* dst_argb,
+                                    int dst_stride_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*P410ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToARGBRow_C;
+  void (*Scale2RowUp_Bilinear_16)(
+      const uint16_t* src_ptr, ptrdiff_t src_stride, uint16_t* dst_ptr,
+      ptrdiff_t dst_stride, int dst_width) = ScaleUVRowUp2_Bilinear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToARGBRow = P410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToARGBRow = P410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToARGBRow = P410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToARGBRow = P410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    Scale2RowUp_Bilinear_16 = ScaleUVRowUp2_Bilinear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear_16 = ScaleUVRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear_16 = ScaleUVRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (2 * width + 31) & ~31;
+  align_buffer_64(row, row_size * 2 * sizeof(uint16_t));
+  uint16_t* temp_uv_1 = (uint16_t*)(row);
+  uint16_t* temp_uv_2 = (uint16_t*)(row) + row_size;
+
+  Scale2RowUp_Bilinear_16(src_uv, 0, temp_uv_1, row_size, width);
+  P410ToARGBRow(src_y, temp_uv_1, dst_argb, yuvconstants, width);
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear_16(src_uv, src_stride_uv, temp_uv_1, row_size, width);
+    P410ToARGBRow(src_y, temp_uv_1, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    P410ToARGBRow(src_y, temp_uv_2, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp_Bilinear_16(src_uv, 0, temp_uv_1, row_size, width);
+    P410ToARGBRow(src_y, temp_uv_1, dst_argb, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P210ToARGBMatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_uv,
+                                  int src_stride_uv,
+                                  uint8_t* dst_argb,
+                                  int dst_stride_argb,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*P410ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToARGBRow_C;
+  void (*ScaleRowUp2_Linear)(const uint16_t* src_uv, uint16_t* dst_uv,
+                             int dst_width) = ScaleUVRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToARGBRow = P410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToARGBRow = P410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToARGBRow = P410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToARGBRow = P410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    ScaleRowUp2_Linear = ScaleUVRowUp2_Linear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear = ScaleUVRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear = ScaleUVRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  const int row_size = (2 * width + 31) & ~31;
+  align_buffer_64(row, row_size * sizeof(uint16_t));
+  uint16_t* temp_uv = (uint16_t*)(row);
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear(src_uv, temp_uv, width);
+    P410ToARGBRow(src_y, temp_uv, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P010ToAR30MatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_uv,
+                                    int src_stride_uv,
+                                    uint8_t* dst_ar30,
+                                    int dst_stride_ar30,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*P410ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToAR30Row_C;
+  void (*Scale2RowUp_Bilinear_16)(
+      const uint16_t* src_ptr, ptrdiff_t src_stride, uint16_t* dst_ptr,
+      ptrdiff_t dst_stride, int dst_width) = ScaleUVRowUp2_Bilinear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToAR30Row = P410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToAR30Row = P410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToAR30Row = P410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToAR30Row = P410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    Scale2RowUp_Bilinear_16 = ScaleUVRowUp2_Bilinear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp_Bilinear_16 = ScaleUVRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp_Bilinear_16 = ScaleUVRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (2 * width + 31) & ~31;
+  align_buffer_64(row, row_size * 2 * sizeof(uint16_t));
+  uint16_t* temp_uv_1 = (uint16_t*)(row);
+  uint16_t* temp_uv_2 = (uint16_t*)(row) + row_size;
+
+  Scale2RowUp_Bilinear_16(src_uv, 0, temp_uv_1, row_size, width);
+  P410ToAR30Row(src_y, temp_uv_1, dst_ar30, yuvconstants, width);
+  dst_ar30 += dst_stride_ar30;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp_Bilinear_16(src_uv, src_stride_uv, temp_uv_1, row_size, width);
+    P410ToAR30Row(src_y, temp_uv_1, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    P410ToAR30Row(src_y, temp_uv_2, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp_Bilinear_16(src_uv, 0, temp_uv_1, row_size, width);
+    P410ToAR30Row(src_y, temp_uv_1, dst_ar30, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P210ToAR30MatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_uv,
+                                  int src_stride_uv,
+                                  uint8_t* dst_ar30,
+                                  int dst_stride_ar30,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*P410ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToAR30Row_C;
+  void (*ScaleRowUp2_Linear)(const uint16_t* src_uv, uint16_t* dst_uv,
+                             int dst_width) = ScaleUVRowUp2_Linear_16_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToAR30Row = P410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToAR30Row = P410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToAR30Row = P410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToAR30Row = P410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    ScaleRowUp2_Linear = ScaleUVRowUp2_Linear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear = ScaleUVRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear = ScaleUVRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  const int row_size = (2 * width + 31) & ~31;
+  align_buffer_64(row, row_size * sizeof(uint16_t));
+  uint16_t* temp_uv = (uint16_t*)(row);
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear(src_uv, temp_uv, width);
+    P410ToAR30Row(src_y, temp_uv, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I422ToRGB24MatrixLinear(const uint8_t* src_y,
+                                   int src_stride_y,
+                                   const uint8_t* src_u,
+                                   int src_stride_u,
+                                   const uint8_t* src_v,
+                                   int src_stride_v,
+                                   uint8_t* dst_rgb24,
+                                   int dst_stride_rgb24,
+                                   const struct YuvConstants* yuvconstants,
+                                   int width,
+                                   int height) {
+  int y;
+  void (*I444ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                         const uint8_t* v_buf, uint8_t* rgb_buf,
+                         const struct YuvConstants* yuvconstants, int width) =
+      I444ToRGB24Row_C;
+  void (*ScaleRowUp2_Linear)(const uint8_t* src_ptr, uint8_t* dst_ptr,
+                             int dst_width) = ScaleRowUp2_Linear_Any_C;
+  assert(yuvconstants);
+  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I444TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToRGB24Row = I444ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToRGB24Row = I444ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToRGB24Row = I444ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToRGB24Row = I444ToRGB24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int row_size = (width + 31) & ~31;
+  align_buffer_64(row, row_size * 2);
+  uint8_t* temp_u = row;
+  uint8_t* temp_v = row + row_size;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp2_Linear(src_u, temp_u, width);
+    ScaleRowUp2_Linear(src_v, temp_v, width);
+    I444ToRGB24Row(src_y, temp_u, temp_v, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+LIBYUV_API
+int I422ToRGB24MatrixFilter(const uint8_t* src_y,
+                            int src_stride_y,
+                            const uint8_t* src_u,
+                            int src_stride_u,
+                            const uint8_t* src_v,
+                            int src_stride_v,
+                            uint8_t* dst_rgb24,
+                            int dst_stride_rgb24,
+                            const struct YuvConstants* yuvconstants,
+                            int width,
+                            int height,
+                            enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I422ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, dst_rgb24, dst_stride_rgb24,
+                               yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I422ToRGB24MatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_rgb24, dst_stride_rgb24, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I420ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+      return I420ToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+    case kFilterLinear:
+      // Actually we can do this, but probably there's no usage.
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I422ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I422ToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I420ToRGB24MatrixFilter(const uint8_t* src_y,
+                            int src_stride_y,
+                            const uint8_t* src_u,
+                            int src_stride_u,
+                            const uint8_t* src_v,
+                            int src_stride_v,
+                            uint8_t* dst_rgb24,
+                            int dst_stride_rgb24,
+                            const struct YuvConstants* yuvconstants,
+                            int width,
+                            int height,
+                            enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, dst_rgb24, dst_stride_rgb24,
+                               yuvconstants, width, height);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return I420ToRGB24MatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_rgb24, dst_stride_rgb24, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I010ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_ar30, dst_stride_ar30,
+                              yuvconstants, width, height);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return I010ToAR30MatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_ar30, dst_stride_ar30, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_ar30, dst_stride_ar30,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I210ToAR30MatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_ar30, dst_stride_ar30, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I010ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return I010ToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I210ToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I420AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I420AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return I420AlphaToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I422AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I422AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I422AlphaToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I010AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I010AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return I010AlphaToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I210AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I210AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I210AlphaToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+  }
+
+  return -1;
+}
+
+// TODO(fb): Verify this function works correctly.  P010 is like NV12 but 10 bit
+// UV is biplanar.
+LIBYUV_API
+int P010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P010ToARGBMatrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_argb, dst_stride_argb, yuvconstants, width,
+                              height);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return P010ToARGBMatrixBilinear(src_y, src_stride_y, src_uv,
+                                      src_stride_uv, dst_argb, dst_stride_argb,
+                                      yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P210ToARGBMatrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_argb, dst_stride_argb, yuvconstants, width,
+                              height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return P210ToARGBMatrixLinear(src_y, src_stride_y, src_uv, src_stride_uv,
+                                    dst_argb, dst_stride_argb, yuvconstants,
+                                    width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P010ToAR30Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_ar30, dst_stride_ar30, yuvconstants, width,
+                              height);
+    case kFilterLinear:  // TODO(fb): Implement Linear using Bilinear stride 0
+    case kFilterBilinear:
+    case kFilterBox:
+      return P010ToAR30MatrixBilinear(src_y, src_stride_y, src_uv,
+                                      src_stride_uv, dst_ar30, dst_stride_ar30,
+                                      yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P210ToAR30Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_ar30, dst_stride_ar30, yuvconstants, width,
+                              height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return P210ToAR30MatrixLinear(src_y, src_stride_y, src_uv, src_stride_uv,
+                                    dst_ar30, dst_stride_ar30, yuvconstants,
+                                    width, height);
+  }
+
+  return -1;
+}
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/media/libyuv/source/convert_from.cc b/media/libyuv/source/convert_from.cc
index 46abdebcd6..8bd07e4ce2 100644
--- a/media/libyuv/source/convert_from.cc
+++ b/media/libyuv/source/convert_from.cc
@@ -15,9 +15,9 @@
 #include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
+#include "libyuv/row.h"
 #include "libyuv/scale.h"  // For ScalePlane()
 #include "libyuv/video_common.h"
-#include "libyuv/row.h"
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -30,107 +30,288 @@ static __inline int Abs(int v) {
 }
 
 // I420 To any I4xx YUV format with mirroring.
-static int I420ToI4xx(const uint8* src_y, int src_stride_y,
-                      const uint8* src_u, int src_stride_u,
-                      const uint8* src_v, int src_stride_v,
-                      uint8* dst_y, int dst_stride_y,
-                      uint8* dst_u, int dst_stride_u,
-                      uint8* dst_v, int dst_stride_v,
-                      int src_y_width, int src_y_height,
-                      int dst_uv_width, int dst_uv_height) {
+// TODO(fbarchard): Consider kFilterNone for Y, or CopyPlane
+
+static int I420ToI4xx(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_y,
+                      int dst_stride_y,
+                      uint8_t* dst_u,
+                      int dst_stride_u,
+                      uint8_t* dst_v,
+                      int dst_stride_v,
+                      int src_y_width,
+                      int src_y_height,
+                      int dst_uv_width,
+                      int dst_uv_height) {
   const int dst_y_width = Abs(src_y_width);
   const int dst_y_height = Abs(src_y_height);
   const int src_uv_width = SUBSAMPLE(src_y_width, 1, 1);
   const int src_uv_height = SUBSAMPLE(src_y_height, 1, 1);
-  if (src_y_width == 0 || src_y_height == 0 ||
-      dst_uv_width <= 0 || dst_uv_height <= 0) {
+  if (src_y_width == 0 || src_y_height == 0 || dst_uv_width <= 0 ||
+      dst_uv_height <= 0) {
     return -1;
   }
-  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
-             dst_y, dst_stride_y, dst_y_width, dst_y_height,
-             kFilterBilinear);
-  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
-             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
-             kFilterBilinear);
-  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
-             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
-             kFilterBilinear);
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, src_y_width, src_y_height, dst_y,
+               dst_stride_y, dst_y_width, dst_y_height, kFilterBilinear);
+  }
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height, dst_u,
+             dst_stride_u, dst_uv_width, dst_uv_height, kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height, dst_v,
+             dst_stride_v, dst_uv_width, dst_uv_height, kFilterBilinear);
+  return 0;
+}
+
+// Convert 8 bit YUV to 10 bit.
+LIBYUV_API
+int I420ToI010(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert8To16Plane(src_y, src_stride_y, dst_y, dst_stride_y, 1024, width,
+                    height);
+  // Convert UV planes.
+  Convert8To16Plane(src_u, src_stride_u, dst_u, dst_stride_u, 1024, halfwidth,
+                    halfheight);
+  Convert8To16Plane(src_v, src_stride_v, dst_v, dst_stride_v, 1024, halfwidth,
+                    halfheight);
+  return 0;
+}
+
+// Convert 8 bit YUV to 12 bit.
+LIBYUV_API
+int I420ToI012(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert8To16Plane(src_y, src_stride_y, dst_y, dst_stride_y, 4096, width,
+                    height);
+  // Convert UV planes.
+  Convert8To16Plane(src_u, src_stride_u, dst_u, dst_stride_u, 4096, halfwidth,
+                    halfheight);
+  Convert8To16Plane(src_v, src_stride_v, dst_v, dst_stride_v, 4096, halfwidth,
+                    halfheight);
   return 0;
 }
 
 // 420 chroma is 1/2 width, 1/2 height
 // 422 chroma is 1/2 width, 1x height
 LIBYUV_API
-int I420ToI422(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int I420ToI422(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   const int dst_uv_width = (Abs(width) + 1) >> 1;
   const int dst_uv_height = Abs(height);
-  return I420ToI4xx(src_y, src_stride_y,
-                    src_u, src_stride_u,
-                    src_v, src_stride_v,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height,
-                    dst_uv_width, dst_uv_height);
+  return I420ToI4xx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, dst_uv_width,
+                    dst_uv_height);
 }
 
 // 420 chroma is 1/2 width, 1/2 height
 // 444 chroma is 1x width, 1x height
 LIBYUV_API
-int I420ToI444(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int I420ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   const int dst_uv_width = Abs(width);
   const int dst_uv_height = Abs(height);
-  return I420ToI4xx(src_y, src_stride_y,
-                    src_u, src_stride_u,
-                    src_v, src_stride_v,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height,
-                    dst_uv_width, dst_uv_height);
+  return I420ToI4xx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, dst_uv_width,
+                    dst_uv_height);
 }
 
-// 420 chroma is 1/2 width, 1/2 height
-// 411 chroma is 1/4 width, 1x height
+// 420 chroma to 444 chroma, 10/12 bit version
 LIBYUV_API
-int I420ToI411(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  const int dst_uv_width = (Abs(width) + 3) >> 2;
-  const int dst_uv_height = Abs(height);
-  return I420ToI4xx(src_y, src_stride_y,
-                    src_u, src_stride_u,
-                    src_v, src_stride_v,
-                    dst_y, dst_stride_y,
-                    dst_u, dst_stride_u,
-                    dst_v, dst_stride_v,
-                    width, height,
-                    dst_uv_width, dst_uv_height);
+int I010ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  if (width == 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_12(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  ScalePlane_12(src_u, src_stride_u, SUBSAMPLE(width, 1, 1),
+                SUBSAMPLE(height, 1, 1), dst_u, dst_stride_u, Abs(width),
+                Abs(height), kFilterBilinear);
+  ScalePlane_12(src_v, src_stride_v, SUBSAMPLE(width, 1, 1),
+                SUBSAMPLE(height, 1, 1), dst_v, dst_stride_v, Abs(width),
+                Abs(height), kFilterBilinear);
+  return 0;
+}
+
+// 422 chroma to 444 chroma, 10/12 bit version
+LIBYUV_API
+int I210ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  if (width == 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_12(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  ScalePlane_12(src_u, src_stride_u, SUBSAMPLE(width, 1, 1), height, dst_u,
+                dst_stride_u, Abs(width), Abs(height), kFilterBilinear);
+  ScalePlane_12(src_v, src_stride_v, SUBSAMPLE(width, 1, 1), height, dst_v,
+                dst_stride_v, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
+}
+
+// 422 chroma is 1/2 width, 1x height
+// 444 chroma is 1x width, 1x height
+LIBYUV_API
+int I422ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  if (width == 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+               Abs(width), Abs(height), kFilterBilinear);
+  }
+  ScalePlane(src_u, src_stride_u, SUBSAMPLE(width, 1, 1), height, dst_u,
+             dst_stride_u, Abs(width), Abs(height), kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, SUBSAMPLE(width, 1, 1), height, dst_v,
+             dst_stride_v, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
 }
 
 // Copy to I400. Source can be I420,422,444,400,NV12,NV21
 LIBYUV_API
-int I400Copy(const uint8* src_y, int src_stride_y,
-             uint8* dst_y, int dst_stride_y,
-             int width, int height) {
-  if (!src_y || !dst_y ||
-      width <= 0 || height == 0) {
+int I400Copy(const uint8_t* src_y,
+             int src_stride_y,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             int width,
+             int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -144,17 +325,21 @@ int I400Copy(const uint8* src_y, int src_stride_y,
 }
 
 LIBYUV_API
-int I422ToYUY2(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_yuy2, int dst_stride_yuy2,
-               int width, int height) {
+int I422ToYUY2(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height) {
   int y;
-  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
-                        const uint8* src_v, uint8* dst_yuy2, int width) =
+  void (*I422ToYUY2Row)(const uint8_t* src_y, const uint8_t* src_u,
+                        const uint8_t* src_v, uint8_t* dst_yuy2, int width) =
       I422ToYUY2Row_C;
-  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
-      width <= 0 || height == 0) {
+  if (!src_y || !src_u || !src_v || !dst_yuy2 || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -164,10 +349,8 @@ int I422ToYUY2(const uint8* src_y, int src_stride_y,
     dst_stride_yuy2 = -dst_stride_yuy2;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      src_stride_u * 2 == width &&
-      src_stride_v * 2 == width &&
-      dst_stride_yuy2 == width * 2) {
+  if (src_stride_y == width && src_stride_u * 2 == width &&
+      src_stride_v * 2 == width && dst_stride_yuy2 == width * 2) {
     width *= height;
     height = 1;
     src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0;
@@ -180,6 +363,14 @@ int I422ToYUY2(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
@@ -200,17 +391,21 @@ int I422ToYUY2(const uint8* src_y, int src_stride_y,
 }
 
 LIBYUV_API
-int I420ToYUY2(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_yuy2, int dst_stride_yuy2,
-               int width, int height) {
+int I420ToYUY2(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height) {
   int y;
-  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
-                        const uint8* src_v, uint8* dst_yuy2, int width) =
+  void (*I422ToYUY2Row)(const uint8_t* src_y, const uint8_t* src_u,
+                        const uint8_t* src_v, uint8_t* dst_yuy2, int width) =
       I422ToYUY2Row_C;
-  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
-      width <= 0 || height == 0) {
+  if (!src_y || !src_u || !src_v || !dst_yuy2 || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -227,6 +422,14 @@ int I420ToYUY2(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
@@ -235,6 +438,22 @@ int I420ToYUY2(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height - 1; y += 2) {
     I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
@@ -252,17 +471,21 @@ int I420ToYUY2(const uint8* src_y, int src_stride_y,
 }
 
 LIBYUV_API
-int I422ToUYVY(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_uyvy, int dst_stride_uyvy,
-               int width, int height) {
+int I422ToUYVY(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_uyvy,
+               int dst_stride_uyvy,
+               int width,
+               int height) {
   int y;
-  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
-                        const uint8* src_v, uint8* dst_uyvy, int width) =
+  void (*I422ToUYVYRow)(const uint8_t* src_y, const uint8_t* src_u,
+                        const uint8_t* src_v, uint8_t* dst_uyvy, int width) =
       I422ToUYVYRow_C;
-  if (!src_y || !src_u || !src_v || !dst_uyvy ||
-      width <= 0 || height == 0) {
+  if (!src_y || !src_u || !src_v || !dst_uyvy || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -272,10 +495,8 @@ int I422ToUYVY(const uint8* src_y, int src_stride_y,
     dst_stride_uyvy = -dst_stride_uyvy;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      src_stride_u * 2 == width &&
-      src_stride_v * 2 == width &&
-      dst_stride_uyvy == width * 2) {
+  if (src_stride_y == width && src_stride_u * 2 == width &&
+      src_stride_v * 2 == width && dst_stride_uyvy == width * 2) {
     width *= height;
     height = 1;
     src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0;
@@ -288,6 +509,14 @@ int I422ToUYVY(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
@@ -296,6 +525,22 @@ int I422ToUYVY(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
@@ -308,17 +553,21 @@ int I422ToUYVY(const uint8* src_y, int src_stride_y,
 }
 
 LIBYUV_API
-int I420ToUYVY(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_uyvy, int dst_stride_uyvy,
-               int width, int height) {
+int I420ToUYVY(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_uyvy,
+               int dst_stride_uyvy,
+               int width,
+               int height) {
   int y;
-  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
-                        const uint8* src_v, uint8* dst_uyvy, int width) =
+  void (*I422ToUYVYRow)(const uint8_t* src_y, const uint8_t* src_u,
+                        const uint8_t* src_v, uint8_t* dst_uyvy, int width) =
       I422ToUYVYRow_C;
-  if (!src_y || !src_u || !src_v || !dst_uyvy ||
-      width <= 0 || height == 0) {
+  if (!src_y || !src_u || !src_v || !dst_uyvy || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -335,6 +584,14 @@ int I420ToUYVY(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
@@ -343,6 +600,22 @@ int I420ToUYVY(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height - 1; y += 2) {
     I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
@@ -360,799 +633,218 @@ int I420ToUYVY(const uint8* src_y, int src_stride_y,
 }
 
 LIBYUV_API
-int I420ToNV12(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height) {
-  int y;
-  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
-      int width) = MergeUVRow_C;
-  // Coalesce rows.
-  int halfwidth = (width + 1) >> 1;
-  int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_u || !src_v || !dst_y || !dst_uv ||
-      width <= 0 || height == 0) {
+int I420ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) / 2;
+  int halfheight = (height + 1) / 2;
+  if (!src_y || !src_u || !src_v || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     halfheight = (height + 1) >> 1;
-    dst_y = dst_y + (height - 1) * dst_stride_y;
-    dst_uv = dst_uv + (halfheight - 1) * dst_stride_uv;
-    dst_stride_y = -dst_stride_y;
-    dst_stride_uv = -dst_stride_uv;
-  }
-  if (src_stride_y == width &&
-      dst_stride_y == width) {
-    width *= height;
-    height = 1;
-    src_stride_y = dst_stride_y = 0;
-  }
-  // Coalesce rows.
-  if (src_stride_u == halfwidth &&
-      src_stride_v == halfwidth &&
-      dst_stride_uv == halfwidth * 2) {
-    halfwidth *= halfheight;
-    halfheight = 1;
-    src_stride_u = src_stride_v = dst_stride_uv = 0;
-  }
-#if defined(HAS_MERGEUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    MergeUVRow_ = MergeUVRow_Any_SSE2;
-    if (IS_ALIGNED(halfwidth, 16)) {
-      MergeUVRow_ = MergeUVRow_SSE2;
-    }
-  }
-#endif
-#if defined(HAS_MERGEUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    MergeUVRow_ = MergeUVRow_Any_AVX2;
-    if (IS_ALIGNED(halfwidth, 32)) {
-      MergeUVRow_ = MergeUVRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_MERGEUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    MergeUVRow_ = MergeUVRow_Any_NEON;
-    if (IS_ALIGNED(halfwidth, 16)) {
-      MergeUVRow_ = MergeUVRow_NEON;
-    }
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
   }
-#endif
-
-  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-  for (y = 0; y < halfheight; ++y) {
-    // Merge a row of U and V into a row of UV.
-    MergeUVRow_(src_u, src_v, dst_uv, halfwidth);
-    src_u += src_stride_u;
-    src_v += src_stride_v;
-    dst_uv += dst_stride_uv;
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
   }
+  MergeUVPlane(src_u, src_stride_u, src_v, src_stride_v, dst_uv, dst_stride_uv,
+               halfwidth, halfheight);
   return 0;
 }
 
 LIBYUV_API
-int I420ToNV21(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_vu, int dst_stride_vu,
-               int width, int height) {
-  return I420ToNV12(src_y, src_stride_y,
-                    src_v, src_stride_v,
-                    src_u, src_stride_u,
-                    dst_y, dst_stride_y,
-                    dst_vu, dst_stride_vu,
+int I420ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  return I420ToNV12(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                    src_stride_u, dst_y, dst_stride_y, dst_vu, dst_stride_vu,
                     width, height);
 }
 
-// Convert I422 to RGBA with matrix
-static int I420ToRGBAMatrix(const uint8* src_y, int src_stride_y,
-                            const uint8* src_u, int src_stride_u,
-                            const uint8* src_v, int src_stride_v,
-                            uint8* dst_rgba, int dst_stride_rgba,
-                            const struct YuvConstants* yuvconstants,
-                            int width, int height) {
-  int y;
-  void (*I422ToRGBARow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I422ToRGBARow_C;
-  if (!src_y || !src_u || !src_v || !dst_rgba ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
-    dst_stride_rgba = -dst_stride_rgba;
-  }
-#if defined(HAS_I422TORGBAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGBARow = I422ToRGBARow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGBARow = I422ToRGBARow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
-      IS_ALIGNED(dst_rgba, 4) && IS_ALIGNED(dst_stride_rgba, 4)) {
-    I422ToRGBARow = I422ToRGBARow_DSPR2;
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
-    dst_rgba += dst_stride_rgba;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGBA.
-LIBYUV_API
-int I420ToRGBA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_rgba, int dst_stride_rgba,
-               int width, int height) {
-  return I420ToRGBAMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_rgba, dst_stride_rgba,
-                          &kYuvI601Constants,
-                          width, height);
-}
-
-// Convert I420 to BGRA.
-LIBYUV_API
-int I420ToBGRA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_bgra, int dst_stride_bgra,
-               int width, int height) {
-  return I420ToRGBAMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_bgra, dst_stride_bgra,
-                          &kYvuI601Constants,  // Use Yvu matrix
-                          width, height);
-}
-
-// Convert I420 to RGB24 with matrix
-static int I420ToRGB24Matrix(const uint8* src_y, int src_stride_y,
-                             const uint8* src_u, int src_stride_u,
-                             const uint8* src_v, int src_stride_v,
-                             uint8* dst_rgb24, int dst_stride_rgb24,
-                             const struct YuvConstants* yuvconstants,
-                             int width, int height) {
-  int y;
-  void (*I422ToRGB24Row)(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* rgb_buf,
-                         const struct YuvConstants* yuvconstants,
-                         int width) = I422ToRGB24Row_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb24 ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
-    dst_stride_rgb24 = -dst_stride_rgb24;
-  }
-#if defined(HAS_I422TORGB24ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB24ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGB24Row = I422ToRGB24Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB24ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB24Row = I422ToRGB24Row_NEON;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, yuvconstants, width);
-    dst_rgb24 += dst_stride_rgb24;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGB24.
-LIBYUV_API
-int I420ToRGB24(const uint8* src_y, int src_stride_y,
-                const uint8* src_u, int src_stride_u,
-                const uint8* src_v, int src_stride_v,
-                uint8* dst_rgb24, int dst_stride_rgb24,
-                int width, int height) {
-  return I420ToRGB24Matrix(src_y, src_stride_y,
-                           src_u, src_stride_u,
-                           src_v, src_stride_v,
-                           dst_rgb24, dst_stride_rgb24,
-                           &kYuvI601Constants,
-                           width, height);
-}
-
-// Convert I420 to RAW.
-LIBYUV_API
-int I420ToRAW(const uint8* src_y, int src_stride_y,
-              const uint8* src_u, int src_stride_u,
-              const uint8* src_v, int src_stride_v,
-              uint8* dst_raw, int dst_stride_raw,
-              int width, int height) {
-  return I420ToRGB24Matrix(src_y, src_stride_y,
-                           src_v, src_stride_v,  // Swap U and V
-                           src_u, src_stride_u,
-                           dst_raw, dst_stride_raw,
-                           &kYvuI601Constants,  // Use Yvu matrix
-                           width, height);
-}
-
-// Convert I420 to ARGB1555.
-LIBYUV_API
-int I420ToARGB1555(const uint8* src_y, int src_stride_y,
-                   const uint8* src_u, int src_stride_u,
-                   const uint8* src_v, int src_stride_v,
-                   uint8* dst_argb1555, int dst_stride_argb1555,
-                   int width, int height) {
-  int y;
-  void (*I422ToARGB1555Row)(const uint8* y_buf,
-                            const uint8* u_buf,
-                            const uint8* v_buf,
-                            uint8* rgb_buf,
-                            const struct YuvConstants* yuvconstants,
-                            int width) = I422ToARGB1555Row_C;
-  if (!src_y || !src_u || !src_v || !dst_argb1555 ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
-    dst_stride_argb1555 = -dst_stride_argb1555;
-  }
-#if defined(HAS_I422TOARGB1555ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB1555ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB1555ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, &kYuvI601Constants,
-                      width);
-    dst_argb1555 += dst_stride_argb1555;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-
-// Convert I420 to ARGB4444.
-LIBYUV_API
-int I420ToARGB4444(const uint8* src_y, int src_stride_y,
-                   const uint8* src_u, int src_stride_u,
-                   const uint8* src_v, int src_stride_v,
-                   uint8* dst_argb4444, int dst_stride_argb4444,
-                   int width, int height) {
-  int y;
-  void (*I422ToARGB4444Row)(const uint8* y_buf,
-                            const uint8* u_buf,
-                            const uint8* v_buf,
-                            uint8* rgb_buf,
-                            const struct YuvConstants* yuvconstants,
-                            int width) = I422ToARGB4444Row_C;
-  if (!src_y || !src_u || !src_v || !dst_argb4444 ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
-    dst_stride_argb4444 = -dst_stride_argb4444;
-  }
-#if defined(HAS_I422TOARGB4444ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB4444ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB4444ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, &kYuvI601Constants,
-                      width);
-    dst_argb4444 += dst_stride_argb4444;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGB565.
-LIBYUV_API
-int I420ToRGB565(const uint8* src_y, int src_stride_y,
-                 const uint8* src_u, int src_stride_u,
-                 const uint8* src_v, int src_stride_v,
-                 uint8* dst_rgb565, int dst_stride_rgb565,
-                 int width, int height) {
-  int y;
-  void (*I422ToRGB565Row)(const uint8* y_buf,
-                          const uint8* u_buf,
-                          const uint8* v_buf,
-                          uint8* rgb_buf,
-                          const struct YuvConstants* yuvconstants,
-                          int width) = I422ToRGB565Row_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-#if defined(HAS_I422TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGB565Row = I422ToRGB565Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_NEON;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, &kYuvI601Constants, width);
-    dst_rgb565 += dst_stride_rgb565;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
-static const uint8 kDither565_4x4[16] = {
-  0, 4, 1, 5,
-  6, 2, 7, 3,
-  1, 5, 0, 4,
-  7, 3, 6, 2,
-};
-
-// Convert I420 to RGB565 with dithering.
-LIBYUV_API
-int I420ToRGB565Dither(const uint8* src_y, int src_stride_y,
-                       const uint8* src_u, int src_stride_u,
-                       const uint8* src_v, int src_stride_v,
-                       uint8* dst_rgb565, int dst_stride_rgb565,
-                       const uint8* dither4x4, int width, int height) {
-  int y;
-  void (*I422ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I422ToARGBRow_C;
-  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
-      const uint32 dither4, int width) = ARGBToRGB565DitherRow_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-  if (!dither4x4) {
-    dither4x4 = kDither565_4x4;
-  }
-#if defined(HAS_I422TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGBRow = I422ToARGBRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToARGBRow = I422ToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) {
-    I422ToARGBRow = I422ToARGBRow_DSPR2;
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
-    }
-  }
-#endif
-  {
-    // Allocate a row of argb.
-    align_buffer_64(row_argb, width * 4);
-    for (y = 0; y < height; ++y) {
-      I422ToARGBRow(src_y, src_u, src_v, row_argb, &kYuvI601Constants, width);
-      ARGBToRGB565DitherRow(row_argb, dst_rgb565,
-                            *(uint32*)(dither4x4 + ((y & 3) << 2)), width);
-      dst_rgb565 += dst_stride_rgb565;
-      src_y += src_stride_y;
-      if (y & 1) {
-        src_u += src_stride_u;
-        src_v += src_stride_v;
-      }
-    }
-    free_aligned_buffer_64(row_argb);
-  }
-  return 0;
-}
-
 // Convert I420 to specified format
 LIBYUV_API
-int ConvertFromI420(const uint8* y, int y_stride,
-                    const uint8* u, int u_stride,
-                    const uint8* v, int v_stride,
-                    uint8* dst_sample, int dst_sample_stride,
-                    int width, int height,
-                    uint32 fourcc) {
-  uint32 format = CanonicalFourCC(fourcc);
+int ConvertFromI420(const uint8_t* y,
+                    int y_stride,
+                    const uint8_t* u,
+                    int u_stride,
+                    const uint8_t* v,
+                    int v_stride,
+                    uint8_t* dst_sample,
+                    int dst_sample_stride,
+                    int width,
+                    int height,
+                    uint32_t fourcc) {
+  uint32_t format = CanonicalFourCC(fourcc);
   int r = 0;
-  if (!y || !u|| !v || !dst_sample ||
-      width <= 0 || height == 0) {
+  if (!y || !u || !v || !dst_sample || width <= 0 || height == 0) {
     return -1;
   }
   switch (format) {
     // Single plane formats
     case FOURCC_YUY2:
-      r = I420ToYUY2(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width * 2,
-                     width, height);
+      r = I420ToYUY2(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2, width,
+                     height);
       break;
     case FOURCC_UYVY:
-      r = I420ToUYVY(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width * 2,
-                     width, height);
+      r = I420ToUYVY(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2, width,
+                     height);
       break;
     case FOURCC_RGBP:
-      r = I420ToRGB565(y, y_stride,
-                       u, u_stride,
-                       v, v_stride,
-                       dst_sample,
-                       dst_sample_stride ? dst_sample_stride : width * 2,
-                       width, height);
+      r = I420ToRGB565(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                       dst_sample_stride ? dst_sample_stride : width * 2, width,
+                       height);
       break;
     case FOURCC_RGBO:
-      r = I420ToARGB1555(y, y_stride,
-                         u, u_stride,
-                         v, v_stride,
-                         dst_sample,
+      r = I420ToARGB1555(y, y_stride, u, u_stride, v, v_stride, dst_sample,
                          dst_sample_stride ? dst_sample_stride : width * 2,
                          width, height);
       break;
     case FOURCC_R444:
-      r = I420ToARGB4444(y, y_stride,
-                         u, u_stride,
-                         v, v_stride,
-                         dst_sample,
+      r = I420ToARGB4444(y, y_stride, u, u_stride, v, v_stride, dst_sample,
                          dst_sample_stride ? dst_sample_stride : width * 2,
                          width, height);
       break;
     case FOURCC_24BG:
-      r = I420ToRGB24(y, y_stride,
-                      u, u_stride,
-                      v, v_stride,
-                      dst_sample,
-                      dst_sample_stride ? dst_sample_stride : width * 3,
-                      width, height);
+      r = I420ToRGB24(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                      dst_sample_stride ? dst_sample_stride : width * 3, width,
+                      height);
       break;
     case FOURCC_RAW:
-      r = I420ToRAW(y, y_stride,
-                    u, u_stride,
-                    v, v_stride,
-                    dst_sample,
-                    dst_sample_stride ? dst_sample_stride : width * 3,
-                    width, height);
+      r = I420ToRAW(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                    dst_sample_stride ? dst_sample_stride : width * 3, width,
+                    height);
       break;
     case FOURCC_ARGB:
-      r = I420ToARGB(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width * 4,
-                     width, height);
+      r = I420ToARGB(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4, width,
+                     height);
       break;
     case FOURCC_BGRA:
-      r = I420ToBGRA(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width * 4,
-                     width, height);
+      r = I420ToBGRA(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4, width,
+                     height);
       break;
     case FOURCC_ABGR:
-      r = I420ToABGR(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width * 4,
-                     width, height);
+      r = I420ToABGR(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4, width,
+                     height);
       break;
     case FOURCC_RGBA:
-      r = I420ToRGBA(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width * 4,
-                     width, height);
+      r = I420ToRGBA(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4, width,
+                     height);
+      break;
+    case FOURCC_AR30:
+      r = I420ToAR30(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4, width,
+                     height);
       break;
     case FOURCC_I400:
-      r = I400Copy(y, y_stride,
-                   dst_sample,
-                   dst_sample_stride ? dst_sample_stride : width,
-                   width, height);
+      r = I400Copy(y, y_stride, dst_sample,
+                   dst_sample_stride ? dst_sample_stride : width, width,
+                   height);
       break;
     case FOURCC_NV12: {
-      uint8* dst_uv = dst_sample + width * height;
-      r = I420ToNV12(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width,
-                     dst_uv,
-                     dst_sample_stride ? dst_sample_stride : width,
-                     width, height);
+      int dst_y_stride = dst_sample_stride ? dst_sample_stride : width;
+      uint8_t* dst_uv = dst_sample + dst_y_stride * height;
+      r = I420ToNV12(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width, dst_uv,
+                     dst_sample_stride ? dst_sample_stride : width, width,
+                     height);
       break;
     }
     case FOURCC_NV21: {
-      uint8* dst_vu = dst_sample + width * height;
-      r = I420ToNV21(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample,
-                     dst_sample_stride ? dst_sample_stride : width,
-                     dst_vu,
-                     dst_sample_stride ? dst_sample_stride : width,
-                     width, height);
+      int dst_y_stride = dst_sample_stride ? dst_sample_stride : width;
+      uint8_t* dst_vu = dst_sample + dst_y_stride * height;
+      r = I420ToNV21(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width, dst_vu,
+                     dst_sample_stride ? dst_sample_stride : width, width,
+                     height);
       break;
     }
-    // TODO(fbarchard): Add M420.
     // Triplanar formats
-    // TODO(fbarchard): halfstride instead of halfwidth
     case FOURCC_I420:
     case FOURCC_YV12: {
-      int halfwidth = (width + 1) / 2;
+      dst_sample_stride = dst_sample_stride ? dst_sample_stride : width;
+      int halfstride = (dst_sample_stride + 1) / 2;
       int halfheight = (height + 1) / 2;
-      uint8* dst_u;
-      uint8* dst_v;
+      uint8_t* dst_u;
+      uint8_t* dst_v;
       if (format == FOURCC_YV12) {
-        dst_v = dst_sample + width * height;
-        dst_u = dst_v + halfwidth * halfheight;
+        dst_v = dst_sample + dst_sample_stride * height;
+        dst_u = dst_v + halfstride * halfheight;
       } else {
-        dst_u = dst_sample + width * height;
-        dst_v = dst_u + halfwidth * halfheight;
+        dst_u = dst_sample + dst_sample_stride * height;
+        dst_v = dst_u + halfstride * halfheight;
       }
-      r = I420Copy(y, y_stride,
-                   u, u_stride,
-                   v, v_stride,
-                   dst_sample, width,
-                   dst_u, halfwidth,
-                   dst_v, halfwidth,
+      r = I420Copy(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                   dst_sample_stride, dst_u, halfstride, dst_v, halfstride,
                    width, height);
       break;
     }
     case FOURCC_I422:
     case FOURCC_YV16: {
-      int halfwidth = (width + 1) / 2;
-      uint8* dst_u;
-      uint8* dst_v;
+      dst_sample_stride = dst_sample_stride ? dst_sample_stride : width;
+      int halfstride = (dst_sample_stride + 1) / 2;
+      uint8_t* dst_u;
+      uint8_t* dst_v;
       if (format == FOURCC_YV16) {
-        dst_v = dst_sample + width * height;
-        dst_u = dst_v + halfwidth * height;
+        dst_v = dst_sample + dst_sample_stride * height;
+        dst_u = dst_v + halfstride * height;
       } else {
-        dst_u = dst_sample + width * height;
-        dst_v = dst_u + halfwidth * height;
+        dst_u = dst_sample + dst_sample_stride * height;
+        dst_v = dst_u + halfstride * height;
       }
-      r = I420ToI422(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample, width,
-                     dst_u, halfwidth,
-                     dst_v, halfwidth,
+      r = I420ToI422(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride, dst_u, halfstride, dst_v, halfstride,
                      width, height);
       break;
     }
     case FOURCC_I444:
     case FOURCC_YV24: {
-      uint8* dst_u;
-      uint8* dst_v;
+      dst_sample_stride = dst_sample_stride ? dst_sample_stride : width;
+      uint8_t* dst_u;
+      uint8_t* dst_v;
       if (format == FOURCC_YV24) {
-        dst_v = dst_sample + width * height;
-        dst_u = dst_v + width * height;
+        dst_v = dst_sample + dst_sample_stride * height;
+        dst_u = dst_v + dst_sample_stride * height;
       } else {
-        dst_u = dst_sample + width * height;
-        dst_v = dst_u + width * height;
+        dst_u = dst_sample + dst_sample_stride * height;
+        dst_v = dst_u + dst_sample_stride * height;
       }
-      r = I420ToI444(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample, width,
-                     dst_u, width,
-                     dst_v, width,
-                     width, height);
+      r = I420ToI444(y, y_stride, u, u_stride, v, v_stride, dst_sample,
+                     dst_sample_stride, dst_u, dst_sample_stride, dst_v,
+                     dst_sample_stride, width, height);
       break;
     }
-    case FOURCC_I411: {
-      int quarterwidth = (width + 3) / 4;
-      uint8* dst_u = dst_sample + width * height;
-      uint8* dst_v = dst_u + quarterwidth * height;
-      r = I420ToI411(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     dst_sample, width,
-                     dst_u, quarterwidth,
-                     dst_v, quarterwidth,
-                     width, height);
-      break;
-    }
-
     // Formats not supported - MJPG, biplanar, some rgb formats.
     default:
       return -1;  // unknown fourcc - return failure code.
diff --git a/media/libyuv/source/convert_from_argb.cc b/media/libyuv/source/convert_from_argb.cc
index 2a8682b7eb..55516cbd8d 100644
--- a/media/libyuv/source/convert_from_argb.cc
+++ b/media/libyuv/source/convert_from_argb.cc
@@ -22,16 +22,21 @@ extern "C" {
 
 // ARGB little endian (bgra in memory) to I444
 LIBYUV_API
-int ARGBToI444(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ARGBToI444(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  void (*ARGBToUV444Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-      int width) = ARGBToUV444Row_C;
+  void (*ARGBToUV444Row)(const uint8_t* src_argb, uint8_t* dst_u,
+                         uint8_t* dst_v, int width) = ARGBToUV444Row_C;
   if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
@@ -41,20 +46,18 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_y == width &&
-      dst_stride_u == width &&
-      dst_stride_v == width) {
+  if (src_stride_argb == width * 4 && dst_stride_y == width &&
+      dst_stride_u == width && dst_stride_v == width) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
 #if defined(HAS_ARGBTOUV444ROW_SSSE3)
-    if (TestCpuFlag(kCpuHasSSSE3)) {
-      ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUV444Row = ARGBToUV444Row_SSSE3;
-      }
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV444Row = ARGBToUV444Row_SSSE3;
+    }
   }
 #endif
 #if defined(HAS_ARGBTOUV444ROW_NEON)
@@ -65,6 +68,22 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOUV444ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV444Row = ARGBToUV444Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV444ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUV444Row = ARGBToUV444Row_LASX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
@@ -84,11 +103,35 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToUV444Row(src_argb, dst_u, dst_v, width);
@@ -103,19 +146,23 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
 
 // ARGB little endian (bgra in memory) to I422
 LIBYUV_API
-int ARGBToI422(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ARGBToI422(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  if (!src_argb ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -125,38 +172,48 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_y == width &&
-      dst_stride_u * 2 == width &&
-      dst_stride_v * 2 == width) {
+  if (src_stride_argb == width * 4 && dst_stride_y == width &&
+      dst_stride_u * 2 == width && dst_stride_v * 2 == width) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -169,6 +226,36 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToUVRow(src_argb, 0, dst_u, dst_v, width);
@@ -181,35 +268,49 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
-// ARGB little endian (bgra in memory) to I411
 LIBYUV_API
-int ARGBToI411(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ARGBToNV12(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUV411Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-      int width) = ARGBToUV411Row_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+  void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v,
+                      uint8_t* dst_uv, int width) = MergeUVRow_C;
+  if (!src_argb || !dst_y || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
+  // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-  // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_y == width &&
-      dst_stride_u * 4 == width &&
-      dst_stride_v * 4 == width) {
-    width *= height;
-    height = 1;
-    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
   }
+#endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
@@ -218,6 +319,14 @@ int ARGBToI411(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYRow = ARGBToYRow_Any_AVX2;
@@ -226,50 +335,136 @@ int ARGBToI411(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_NEON;
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUV411ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUV411Row = ARGBToUV411Row_NEON;
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
     }
   }
 #endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX512BW;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX512BW;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8_t* row_v = row_u + ((halfwidth + 31) & ~31);
 
-  for (y = 0; y < height; ++y) {
-    ARGBToUV411Row(src_argb, dst_u, dst_v, width);
-    ARGBToYRow(src_argb, dst_y, width);
-    src_argb += src_stride_argb;
-    dst_y += dst_stride_y;
-    dst_u += dst_stride_u;
-    dst_v += dst_stride_v;
+    for (y = 0; y < height - 1; y += 2) {
+      ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+      ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+      src_argb += src_stride_argb * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
   }
   return 0;
 }
 
+// Same as NV12 but U and V swapped.
 LIBYUV_API
-int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height) {
+int ARGBToNV21(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
   int y;
   int halfwidth = (width + 1) >> 1;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb0, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
-                      int width) = MergeUVRow_C;
-  if (!src_argb ||
-      !dst_y || !dst_uv ||
-      width <= 0 || height == 0) {
+  void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v,
+                      uint8_t* dst_vu, int width) = MergeUVRow_C;
+  if (!src_argb || !dst_y || !dst_vu || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -278,30 +473,42 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -314,6 +521,36 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
@@ -325,11 +562,19 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_MERGEUVROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     MergeUVRow_ = MergeUVRow_Any_AVX2;
-    if (IS_ALIGNED(halfwidth, 32)) {
+    if (IS_ALIGNED(halfwidth, 16)) {
       MergeUVRow_ = MergeUVRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX512BW;
+    if (IS_ALIGNED(halfwidth, 64)) {
+      MergeUVRow_ = MergeUVRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_MERGEUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     MergeUVRow_ = MergeUVRow_Any_NEON;
@@ -338,23 +583,39 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
   {
     // Allocate a rows of uv.
     align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
-    uint8* row_v = row_u + ((halfwidth + 31) & ~31);
+    uint8_t* row_v = row_u + ((halfwidth + 31) & ~31);
 
     for (y = 0; y < height - 1; y += 2) {
       ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
-      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
       ARGBToYRow(src_argb, dst_y, width);
       ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
       src_argb += src_stride_argb * 2;
       dst_y += dst_stride_y * 2;
-      dst_uv += dst_stride_uv;
+      dst_vu += dst_stride_vu;
     }
     if (height & 1) {
       ARGBToUVRow(src_argb, 0, row_u, row_v, width);
-      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
       ARGBToYRow(src_argb, dst_y, width);
     }
     free_aligned_buffer_64(row_u);
@@ -362,64 +623,106 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
-// Same as NV12 but U and V swapped.
 LIBYUV_API
-int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height) {
+int ABGRToNV12(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
   int y;
   int halfwidth = (width + 1) >> 1;
-  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
-      ARGBToYRow_C;
-  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
-                      int width) = MergeUVRow_C;
-  if (!src_argb ||
-      !dst_y || !dst_uv ||
-      width <= 0 || height == 0) {
+  void (*ABGRToUVRow)(const uint8_t* src_abgr0, int src_stride_abgr,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8_t* src_abgr, uint8_t* dst_y, int width) =
+      ABGRToYRow_C;
+  void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v,
+                      uint8_t* dst_uv, int width) = MergeUVRow_C;
+  if (!src_abgr || !dst_y || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    src_argb = src_argb + (height - 1) * src_stride_argb;
-    src_stride_argb = -src_stride_argb;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ABGRTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
-    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
-      ARGBToYRow = ARGBToYRow_SSSE3;
+      ABGRToYRow = ABGRToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ABGRTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
-    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    ABGRToYRow = ABGRToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
-      ARGBToYRow = ARGBToYRow_AVX2;
+      ABGRToYRow = ABGRToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_NEON)
+#if defined(HAS_ABGRTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_NEON;
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_NEON;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVROW_NEON)
+#if defined(HAS_ABGRTOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    ABGRToUVRow = ABGRToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_NEON;
+      ABGRToUVRow = ABGRToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYRow = ABGRToYRow_Any_MSA;
+    ABGRToUVRow = ABGRToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYRow = ABGRToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYRow = ABGRToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_LASX;
     }
   }
 #endif
@@ -434,11 +737,19 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_MERGEUVROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     MergeUVRow_ = MergeUVRow_Any_AVX2;
-    if (IS_ALIGNED(halfwidth, 32)) {
+    if (IS_ALIGNED(halfwidth, 16)) {
       MergeUVRow_ = MergeUVRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX512BW;
+    if (IS_ALIGNED(halfwidth, 64)) {
+      MergeUVRow_ = MergeUVRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_MERGEUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     MergeUVRow_ = MergeUVRow_Any_NEON;
@@ -447,24 +758,216 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
   {
     // Allocate a rows of uv.
     align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
-    uint8* row_v = row_u + ((halfwidth + 31) & ~31);
+    uint8_t* row_v = row_u + ((halfwidth + 31) & ~31);
 
     for (y = 0; y < height - 1; y += 2) {
-      ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
-      MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
-      ARGBToYRow(src_argb, dst_y, width);
-      ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
-      src_argb += src_stride_argb * 2;
+      ABGRToUVRow(src_abgr, src_stride_abgr, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ABGRToYRow(src_abgr, dst_y, width);
+      ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+      src_abgr += src_stride_abgr * 2;
       dst_y += dst_stride_y * 2;
       dst_uv += dst_stride_uv;
     }
     if (height & 1) {
-      ARGBToUVRow(src_argb, 0, row_u, row_v, width);
-      MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
-      ARGBToYRow(src_argb, dst_y, width);
+      ABGRToUVRow(src_abgr, 0, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ABGRToYRow(src_abgr, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+
+// Same as NV12 but U and V swapped.
+LIBYUV_API
+int ABGRToNV21(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ABGRToUVRow)(const uint8_t* src_abgr0, int src_stride_abgr,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8_t* src_abgr, uint8_t* dst_y, int width) =
+      ABGRToYRow_C;
+  void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v,
+                      uint8_t* dst_vu, int width) = MergeUVRow_C;
+  if (!src_abgr || !dst_y || !dst_vu || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+#if defined(HAS_ABGRTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYRow = ABGRToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVRow = ABGRToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYRow = ABGRToYRow_Any_MSA;
+    ABGRToUVRow = ABGRToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYRow = ABGRToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYRow = ABGRToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX512BW;
+    if (IS_ALIGNED(halfwidth, 64)) {
+      MergeUVRow_ = MergeUVRow_AVX512BW;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8_t* row_v = row_u + ((halfwidth + 31) & ~31);
+
+    for (y = 0; y < height - 1; y += 2) {
+      ABGRToUVRow(src_abgr, src_stride_abgr, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
+      ABGRToYRow(src_abgr, dst_y, width);
+      ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+      src_abgr += src_stride_abgr * 2;
+      dst_y += dst_stride_y * 2;
+      dst_vu += dst_stride_vu;
+    }
+    if (height & 1) {
+      ABGRToUVRow(src_abgr, 0, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
+      ABGRToYRow(src_abgr, dst_y, width);
     }
     free_aligned_buffer_64(row_u);
   }
@@ -473,19 +976,23 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB to YUY2.
 LIBYUV_API
-int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yuy2, int dst_stride_yuy2,
-               int width, int height) {
+int ARGBToYUY2(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yuy2,
+               int dst_stride_yuy2,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUVRow)(const uint8* src_argb, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
-      const uint8* src_v, uint8* dst_yuy2, int width) = I422ToYUY2Row_C;
+  void (*I422ToYUY2Row)(const uint8_t* src_y, const uint8_t* src_u,
+                        const uint8_t* src_v, uint8_t* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
 
-  if (!src_argb || !dst_yuy2 ||
-      width <= 0 || height == 0) {
+  if (!src_argb || !dst_yuy2 || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -495,36 +1002,47 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
     dst_stride_yuy2 = -dst_stride_yuy2;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_yuy2 == width * 2) {
+  if (src_stride_argb == width * 4 && dst_stride_yuy2 == width * 2) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_yuy2 = 0;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -537,6 +1055,36 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
@@ -545,6 +1093,14 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
@@ -553,12 +1109,28 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_LASX;
+    }
+  }
+#endif
 
   {
     // Allocate a rows of yuv.
     align_buffer_64(row_y, ((width + 63) & ~63) * 2);
-    uint8* row_u = row_y + ((width + 63) & ~63);
-    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+    uint8_t* row_u = row_y + ((width + 63) & ~63);
+    uint8_t* row_v = row_u + ((width + 63) & ~63) / 2;
 
     for (y = 0; y < height; ++y) {
       ARGBToUVRow(src_argb, 0, row_u, row_v, width);
@@ -575,19 +1147,23 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB to UYVY.
 LIBYUV_API
-int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_uyvy, int dst_stride_uyvy,
-               int width, int height) {
+int ARGBToUYVY(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_uyvy,
+               int dst_stride_uyvy,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUVRow)(const uint8* src_argb, int src_stride_argb,
-      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToUVRow)(const uint8_t* src_argb, int src_stride_argb,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
-  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
-      const uint8* src_v, uint8* dst_uyvy, int width) = I422ToUYVYRow_C;
+  void (*I422ToUYVYRow)(const uint8_t* src_y, const uint8_t* src_u,
+                        const uint8_t* src_v, uint8_t* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
 
-  if (!src_argb || !dst_uyvy ||
-      width <= 0 || height == 0) {
+  if (!src_argb || !dst_uyvy || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -597,36 +1173,47 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
     dst_stride_uyvy = -dst_stride_uyvy;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_uyvy == width * 2) {
+  if (src_stride_argb == width * 4 && dst_stride_uyvy == width * 2) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_uyvy = 0;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -639,6 +1226,36 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
@@ -647,6 +1264,14 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
@@ -655,12 +1280,28 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_LASX;
+    }
+  }
+#endif
 
   {
     // Allocate a rows of yuv.
     align_buffer_64(row_y, ((width + 63) & ~63) * 2);
-    uint8* row_u = row_y + ((width + 63) & ~63);
-    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+    uint8_t* row_u = row_y + ((width + 63) & ~63);
+    uint8_t* row_v = row_u + ((width + 63) & ~63) / 2;
 
     for (y = 0; y < height; ++y) {
       ARGBToUVRow(src_argb, 0, row_u, row_v, width);
@@ -677,11 +1318,14 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB to I400.
 LIBYUV_API
-int ARGBToI400(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height) {
+int ARGBToI400(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int width) =
+  void (*ARGBToYRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
       ARGBToYRow_C;
   if (!src_argb || !dst_y || width <= 0 || height == 0) {
     return -1;
@@ -692,8 +1336,7 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_y == width) {
+  if (src_stride_argb == width * 4 && dst_stride_y == width) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_y = 0;
@@ -717,11 +1360,35 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToYRow(src_argb, dst_y, width);
@@ -732,28 +1399,31 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb,
 }
 
 // Shuffle table for converting ARGB to RGBA.
-static uvec8 kShuffleMaskARGBToRGBA = {
-  3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u
-};
+static const uvec8 kShuffleMaskARGBToRGBA = {
+    3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u};
 
 // Convert ARGB to RGBA.
 LIBYUV_API
-int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_rgba, int dst_stride_rgba,
-               int width, int height) {
-  return ARGBShuffle(src_argb, src_stride_argb,
-                     dst_rgba, dst_stride_rgba,
-                     (const uint8*)(&kShuffleMaskARGBToRGBA),
-                     width, height);
+int ARGBToRGBA(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height) {
+  return ARGBShuffle(src_argb, src_stride_argb, dst_rgba, dst_stride_rgba,
+                     (const uint8_t*)(&kShuffleMaskARGBToRGBA), width, height);
 }
 
 // Convert ARGB To RGB24.
 LIBYUV_API
-int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
-                uint8* dst_rgb24, int dst_stride_rgb24,
-                int width, int height) {
+int ARGBToRGB24(const uint8_t* src_argb,
+                int src_stride_argb,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
   int y;
-  void (*ARGBToRGB24Row)(const uint8* src_argb, uint8* dst_rgb, int width) =
+  void (*ARGBToRGB24Row)(const uint8_t* src_argb, uint8_t* dst_rgb, int width) =
       ARGBToRGB24Row_C;
   if (!src_argb || !dst_rgb24 || width <= 0 || height == 0) {
     return -1;
@@ -764,8 +1434,7 @@ int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_rgb24 == width * 3) {
+  if (src_stride_argb == width * 4 && dst_stride_rgb24 == width * 3) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_rgb24 = 0;
@@ -778,14 +1447,46 @@ int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB24ROW_AVX512VBMI)
+  if (TestCpuFlag(kCpuHasAVX512VBMI)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_AVX512VBMI;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_AVX512VBMI;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTORGB24ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToRGB24Row = ARGBToRGB24Row_NEON;
     }
   }
 #endif
+#if defined(HAS_ARGBTORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB24ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToRGB24Row(src_argb, dst_rgb24, width);
@@ -797,11 +1498,14 @@ int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB To RAW.
 LIBYUV_API
-int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
-              uint8* dst_raw, int dst_stride_raw,
-              int width, int height) {
+int ARGBToRAW(const uint8_t* src_argb,
+              int src_stride_argb,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
   int y;
-  void (*ARGBToRAWRow)(const uint8* src_argb, uint8* dst_rgb, int width) =
+  void (*ARGBToRAWRow)(const uint8_t* src_argb, uint8_t* dst_rgb, int width) =
       ARGBToRAWRow_C;
   if (!src_argb || !dst_raw || width <= 0 || height == 0) {
     return -1;
@@ -812,8 +1516,7 @@ int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_raw == width * 3) {
+  if (src_stride_argb == width * 4 && dst_stride_raw == width * 3) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_raw = 0;
@@ -826,6 +1529,14 @@ int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORAWROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRAWRow = ARGBToRAWRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTORAWROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
@@ -834,6 +1545,22 @@ int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORAWROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRAWRow = ARGBToRAWRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORAWROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRAWRow = ARGBToRAWRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToRAWRow(src_argb, dst_raw, width);
@@ -844,21 +1571,23 @@ int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
 }
 
 // Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
-static const uint8 kDither565_4x4[16] = {
-  0, 4, 1, 5,
-  6, 2, 7, 3,
-  1, 5, 0, 4,
-  7, 3, 6, 2,
+static const uint8_t kDither565_4x4[16] = {
+    0, 4, 1, 5, 6, 2, 7, 3, 1, 5, 0, 4, 7, 3, 6, 2,
 };
 
 // Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes).
 LIBYUV_API
-int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_rgb565, int dst_stride_rgb565,
-                       const uint8* dither4x4, int width, int height) {
+int ARGBToRGB565Dither(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const uint8_t* dither4x4,
+                       int width,
+                       int height) {
   int y;
-  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
-      const uint32 dither4, int width) = ARGBToRGB565DitherRow_C;
+  void (*ARGBToRGB565DitherRow)(const uint8_t* src_argb, uint8_t* dst_rgb,
+                                const uint32_t dither4, int width) =
+      ARGBToRGB565DitherRow_C;
   if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
     return -1;
   }
@@ -894,9 +1623,27 @@ int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORGB565DITHERROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_LASX;
+    }
+  }
+#endif
+
   for (y = 0; y < height; ++y) {
     ARGBToRGB565DitherRow(src_argb, dst_rgb565,
-                          *(uint32*)(dither4x4 + ((y & 3) << 2)), width);
+                          *(const uint32_t*)(dither4x4 + ((y & 3) << 2)),
+                          width);
     src_argb += src_stride_argb;
     dst_rgb565 += dst_stride_rgb565;
   }
@@ -906,12 +1653,15 @@ int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
 // Convert ARGB To RGB565.
 // TODO(fbarchard): Consider using dither function low level with zeros.
 LIBYUV_API
-int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
-                 uint8* dst_rgb565, int dst_stride_rgb565,
-                 int width, int height) {
+int ARGBToRGB565(const uint8_t* src_argb,
+                 int src_stride_argb,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
   int y;
-  void (*ARGBToRGB565Row)(const uint8* src_argb, uint8* dst_rgb, int width) =
-      ARGBToRGB565Row_C;
+  void (*ARGBToRGB565Row)(const uint8_t* src_argb, uint8_t* dst_rgb,
+                          int width) = ARGBToRGB565Row_C;
   if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
     return -1;
   }
@@ -921,8 +1671,7 @@ int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_rgb565 == width * 2) {
+  if (src_stride_argb == width * 4 && dst_stride_rgb565 == width * 2) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_rgb565 = 0;
@@ -951,6 +1700,22 @@ int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToRGB565Row(src_argb, dst_rgb565, width);
@@ -962,12 +1727,15 @@ int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB To ARGB1555.
 LIBYUV_API
-int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_argb1555, int dst_stride_argb1555,
-                   int width, int height) {
+int ARGBToARGB1555(const uint8_t* src_argb,
+                   int src_stride_argb,
+                   uint8_t* dst_argb1555,
+                   int dst_stride_argb1555,
+                   int width,
+                   int height) {
   int y;
-  void (*ARGBToARGB1555Row)(const uint8* src_argb, uint8* dst_rgb, int width) =
-      ARGBToARGB1555Row_C;
+  void (*ARGBToARGB1555Row)(const uint8_t* src_argb, uint8_t* dst_rgb,
+                            int width) = ARGBToARGB1555Row_C;
   if (!src_argb || !dst_argb1555 || width <= 0 || height == 0) {
     return -1;
   }
@@ -977,8 +1745,7 @@ int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb1555 == width * 2) {
+  if (src_stride_argb == width * 4 && dst_stride_argb1555 == width * 2) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb1555 = 0;
@@ -1007,6 +1774,22 @@ int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOARGB1555ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToARGB1555Row(src_argb, dst_argb1555, width);
@@ -1018,12 +1801,15 @@ int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB To ARGB4444.
 LIBYUV_API
-int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_argb4444, int dst_stride_argb4444,
-                   int width, int height) {
+int ARGBToARGB4444(const uint8_t* src_argb,
+                   int src_stride_argb,
+                   uint8_t* dst_argb4444,
+                   int dst_stride_argb4444,
+                   int width,
+                   int height) {
   int y;
-  void (*ARGBToARGB4444Row)(const uint8* src_argb, uint8* dst_rgb, int width) =
-      ARGBToARGB4444Row_C;
+  void (*ARGBToARGB4444Row)(const uint8_t* src_argb, uint8_t* dst_rgb,
+                            int width) = ARGBToARGB4444Row_C;
   if (!src_argb || !dst_argb4444 || width <= 0 || height == 0) {
     return -1;
   }
@@ -1033,8 +1819,7 @@ int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb4444 == width * 2) {
+  if (src_stride_argb == width * 4 && dst_stride_argb4444 == width * 2) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb4444 = 0;
@@ -1063,6 +1848,22 @@ int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOARGB4444ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToARGB4444Row(src_argb, dst_argb4444, width);
@@ -1072,21 +1873,123 @@ int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// Convert ABGR To AR30.
+LIBYUV_API
+int ABGRToAR30(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  int y;
+  void (*ABGRToAR30Row)(const uint8_t* src_abgr, uint8_t* dst_rgb, int width) =
+      ABGRToAR30Row_C;
+  if (!src_abgr || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_abgr == width * 4 && dst_stride_ar30 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_abgr = dst_stride_ar30 = 0;
+  }
+#if defined(HAS_ABGRTOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToAR30Row = ABGRToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ABGRToAR30Row = ABGRToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToAR30Row = ABGRToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ABGRToAR30Row = ABGRToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ABGRToAR30Row(src_abgr, dst_ar30, width);
+    src_abgr += src_stride_abgr;
+    dst_ar30 += dst_stride_ar30;
+  }
+  return 0;
+}
+
+// Convert ARGB To AR30.
+LIBYUV_API
+int ARGBToAR30(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  int y;
+  void (*ARGBToAR30Row)(const uint8_t* src_argb, uint8_t* dst_rgb, int width) =
+      ARGBToAR30Row_C;
+  if (!src_argb || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 && dst_stride_ar30 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_ar30 = 0;
+  }
+#if defined(HAS_ARGBTOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToAR30Row = ARGBToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToAR30Row = ARGBToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToAR30Row = ARGBToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAR30Row = ARGBToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBToAR30Row(src_argb, dst_ar30, width);
+    src_argb += src_stride_argb;
+    dst_ar30 += dst_stride_ar30;
+  }
+  return 0;
+}
+
 // Convert ARGB to J420. (JPeg full range I420).
 LIBYUV_API
-int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yj, int dst_stride_yj,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ARGBToJ420(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C;
-  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int width) =
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_uj, uint8_t* dst_vj, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_yj, int width) =
       ARGBToYJRow_C;
-  if (!src_argb ||
-      !dst_yj || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_argb || !dst_yj || !dst_uj || !dst_vj || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1095,16 +1998,38 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
       ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
@@ -1113,34 +2038,58 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_NEON;
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
+#if defined(HAS_ARGBTOYJROW_MSA) && defined(HAS_ARGBTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYJRow = ARGBToYJRow_Any_MSA;
+    ARGBToUVJRow = ARGBToUVJRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_NEON;
+      ARGBToYJRow = ARGBToYJRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_LSX;
+      ARGBToUVJRow = ARGBToUVJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LASX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_LASX;
+      ARGBToUVJRow = ARGBToUVJRow_LASX;
     }
   }
 #endif
 
   for (y = 0; y < height - 1; y += 2) {
-    ARGBToUVJRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToUVJRow(src_argb, src_stride_argb, dst_uj, dst_vj, width);
     ARGBToYJRow(src_argb, dst_yj, width);
     ARGBToYJRow(src_argb + src_stride_argb, dst_yj + dst_stride_yj, width);
     src_argb += src_stride_argb * 2;
     dst_yj += dst_stride_yj * 2;
-    dst_u += dst_stride_u;
-    dst_v += dst_stride_v;
+    dst_uj += dst_stride_uj;
+    dst_vj += dst_stride_vj;
   }
   if (height & 1) {
-    ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToUVJRow(src_argb, 0, dst_uj, dst_vj, width);
     ARGBToYJRow(src_argb, dst_yj, width);
   }
   return 0;
@@ -1148,19 +2097,23 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB to J422. (JPeg full range I422).
 LIBYUV_API
-int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yj, int dst_stride_yj,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int ARGBToJ422(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C;
-  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int width) =
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_uj, uint8_t* dst_vj, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_yj, int width) =
       ARGBToYJRow_C;
-  if (!src_argb ||
-      !dst_yj || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+  if (!src_argb || !dst_yj || !dst_uj || !dst_vj || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1170,24 +2123,28 @@ int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_yj == width &&
-      dst_stride_u * 2 == width &&
-      dst_stride_v * 2 == width) {
+  if (src_stride_argb == width * 4 && dst_stride_yj == width &&
+      dst_stride_uj * 2 == width && dst_stride_vj * 2 == width) {
     width *= height;
     height = 1;
-    src_stride_argb = dst_stride_yj = dst_stride_u = dst_stride_v = 0;
+    src_stride_argb = dst_stride_yj = dst_stride_uj = dst_stride_vj = 0;
   }
-#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+#if defined(HAS_ARGBTOYJROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
       ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
@@ -1196,10 +2153,18 @@ int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
@@ -1212,25 +2177,60 @@ int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYJROW_MSA) && defined(HAS_ARGBTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYJRow = ARGBToYJRow_Any_MSA;
+    ARGBToUVJRow = ARGBToUVJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_LSX;
+      ARGBToUVJRow = ARGBToUVJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LASX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_LASX;
+      ARGBToUVJRow = ARGBToUVJRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
-    ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToUVJRow(src_argb, 0, dst_uj, dst_vj, width);
     ARGBToYJRow(src_argb, dst_yj, width);
     src_argb += src_stride_argb;
     dst_yj += dst_stride_yj;
-    dst_u += dst_stride_u;
-    dst_v += dst_stride_v;
+    dst_uj += dst_stride_uj;
+    dst_vj += dst_stride_vj;
   }
   return 0;
 }
 
 // Convert ARGB to J400.
 LIBYUV_API
-int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_yj, int dst_stride_yj,
-               int width, int height) {
+int ARGBToJ400(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height) {
   int y;
-  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int width) =
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_yj, int width) =
       ARGBToYJRow_C;
   if (!src_argb || !dst_yj || width <= 0 || height == 0) {
     return -1;
@@ -1241,8 +2241,7 @@ int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_yj == width) {
+  if (src_stride_argb == width * 4 && dst_stride_yj == width) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_yj = 0;
@@ -1266,11 +2265,19 @@ int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
 #endif
+#if defined(HAS_ARGBTOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYJRow = ARGBToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToYJRow(src_argb, dst_yj, width);
@@ -1280,6 +2287,753 @@ int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// Convert RGBA to J400.
+LIBYUV_API
+int RGBAToJ400(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height) {
+  int y;
+  void (*RGBAToYJRow)(const uint8_t* src_rgba, uint8_t* dst_yj, int width) =
+      RGBAToYJRow_C;
+  if (!src_rgba || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_rgba = src_rgba + (height - 1) * src_stride_rgba;
+    src_stride_rgba = -src_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_rgba == width * 4 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_rgba = dst_stride_yj = 0;
+  }
+#if defined(HAS_RGBATOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToYJRow = RGBAToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGBAToYJRow = RGBAToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RGBAToYJRow = RGBAToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToYJRow = RGBAToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGBAToYJRow = RGBAToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGBAToYJRow = RGBAToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGBAToYJRow = RGBAToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGBAToYJRow = RGBAToYJRow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGBAToYJRow(src_rgba, dst_yj, width);
+    src_rgba += src_stride_rgba;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+// Convert ABGR to J420. (JPeg full range I420).
+LIBYUV_API
+int ABGRToJ420(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height) {
+  int y;
+  void (*ABGRToUVJRow)(const uint8_t* src_abgr0, int src_stride_abgr,
+                       uint8_t* dst_uj, uint8_t* dst_vj, int width) =
+      ABGRToUVJRow_C;
+  void (*ABGRToYJRow)(const uint8_t* src_abgr, uint8_t* dst_yj, int width) =
+      ABGRToYJRow_C;
+  if (!src_abgr || !dst_yj || !dst_uj || !dst_vj || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+#if defined(HAS_ABGRTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToYJRow = ABGRToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVJRow = ABGRToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVJRow = ABGRToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYJRow = ABGRToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVJRow = ABGRToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVJRow = ABGRToUVJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYJRow = ABGRToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVJRow = ABGRToUVJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVJRow = ABGRToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_MSA) && defined(HAS_ABGRTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYJRow = ABGRToYJRow_Any_MSA;
+    ABGRToUVJRow = ABGRToUVJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_MSA;
+      ABGRToUVJRow = ABGRToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ABGRToUVJRow(src_abgr, src_stride_abgr, dst_uj, dst_vj, width);
+    ABGRToYJRow(src_abgr, dst_yj, width);
+    ABGRToYJRow(src_abgr + src_stride_abgr, dst_yj + dst_stride_yj, width);
+    src_abgr += src_stride_abgr * 2;
+    dst_yj += dst_stride_yj * 2;
+    dst_uj += dst_stride_uj;
+    dst_vj += dst_stride_vj;
+  }
+  if (height & 1) {
+    ABGRToUVJRow(src_abgr, 0, dst_uj, dst_vj, width);
+    ABGRToYJRow(src_abgr, dst_yj, width);
+  }
+  return 0;
+}
+
+// Convert ABGR to J422. (JPeg full range I422).
+LIBYUV_API
+int ABGRToJ422(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               uint8_t* dst_uj,
+               int dst_stride_uj,
+               uint8_t* dst_vj,
+               int dst_stride_vj,
+               int width,
+               int height) {
+  int y;
+  void (*ABGRToUVJRow)(const uint8_t* src_abgr0, int src_stride_abgr,
+                       uint8_t* dst_uj, uint8_t* dst_vj, int width) =
+      ABGRToUVJRow_C;
+  void (*ABGRToYJRow)(const uint8_t* src_abgr, uint8_t* dst_yj, int width) =
+      ABGRToYJRow_C;
+  if (!src_abgr || !dst_yj || !dst_uj || !dst_vj || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_abgr == width * 4 && dst_stride_yj == width &&
+      dst_stride_uj * 2 == width && dst_stride_vj * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_abgr = dst_stride_yj = dst_stride_uj = dst_stride_vj = 0;
+  }
+#if defined(HAS_ABGRTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToYJRow = ABGRToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVJRow = ABGRToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVJRow = ABGRToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYJRow = ABGRToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVJRow = ABGRToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVJRow = ABGRToUVJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYJRow = ABGRToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVJRow = ABGRToUVJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVJRow = ABGRToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_MSA) && defined(HAS_ABGRTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYJRow = ABGRToYJRow_Any_MSA;
+    ABGRToUVJRow = ABGRToUVJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVJRow = ABGRToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ABGRToUVJRow(src_abgr, 0, dst_uj, dst_vj, width);
+    ABGRToYJRow(src_abgr, dst_yj, width);
+    src_abgr += src_stride_abgr;
+    dst_yj += dst_stride_yj;
+    dst_uj += dst_stride_uj;
+    dst_vj += dst_stride_vj;
+  }
+  return 0;
+}
+
+// Convert ABGR to J400.
+LIBYUV_API
+int ABGRToJ400(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height) {
+  int y;
+  void (*ABGRToYJRow)(const uint8_t* src_abgr, uint8_t* dst_yj, int width) =
+      ABGRToYJRow_C;
+  if (!src_abgr || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_abgr == width * 4 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_abgr = dst_stride_yj = 0;
+  }
+#if defined(HAS_ABGRTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToYJRow = ABGRToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYJRow = ABGRToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYJRow = ABGRToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYJRow = ABGRToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ABGRToYJRow(src_abgr, dst_yj, width);
+    src_abgr += src_stride_abgr;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+// Convert ARGB to AR64.
+LIBYUV_API
+int ARGBToAR64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ar64,
+               int dst_stride_ar64,
+               int width,
+               int height) {
+  int y;
+  void (*ARGBToAR64Row)(const uint8_t* src_argb, uint16_t* dst_ar64,
+                        int width) = ARGBToAR64Row_C;
+  if (!src_argb || !dst_ar64 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 && dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_ar64 = 0;
+  }
+#if defined(HAS_ARGBTOAR64ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToAR64Row = ARGBToAR64Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToAR64Row = ARGBToAR64Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAR64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToAR64Row = ARGBToAR64Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAR64Row = ARGBToAR64Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAR64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToAR64Row = ARGBToAR64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAR64Row = ARGBToAR64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToAR64Row(src_argb, dst_ar64, width);
+    src_argb += src_stride_argb;
+    dst_ar64 += dst_stride_ar64;
+  }
+  return 0;
+}
+
+// Convert ARGB to AB64.
+LIBYUV_API
+int ARGBToAB64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height) {
+  int y;
+  void (*ARGBToAB64Row)(const uint8_t* src_argb, uint16_t* dst_ar64,
+                        int width) = ARGBToAB64Row_C;
+  if (!src_argb || !dst_ab64 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 && dst_stride_ab64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_ab64 = 0;
+  }
+#if defined(HAS_ARGBTOAB64ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToAB64Row = ARGBToAB64Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToAB64Row = ARGBToAB64Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAB64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToAB64Row = ARGBToAB64Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAB64Row = ARGBToAB64Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAB64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToAB64Row = ARGBToAB64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAB64Row = ARGBToAB64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToAB64Row(src_argb, dst_ab64, width);
+    src_argb += src_stride_argb;
+    dst_ab64 += dst_stride_ab64;
+  }
+  return 0;
+}
+
+// Enabled if 1 pass is available
+#if defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA)
+#define HAS_RAWTOYJROW
+#endif
+
+// RAW to JNV21 full range NV21
+LIBYUV_API
+int RAWToJNV21(const uint8_t* src_raw,
+               int src_stride_raw,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+#if defined(HAS_RAWTOYJROW)
+  void (*RAWToUVJRow)(const uint8_t* src_raw, int src_stride_raw,
+                      uint8_t* dst_uj, uint8_t* dst_vj, int width) =
+      RAWToUVJRow_C;
+  void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
+      RAWToYJRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_uj, uint8_t* dst_vj, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
+      ARGBToYJRow_C;
+#endif
+  void (*MergeUVRow_)(const uint8_t* src_uj, const uint8_t* src_vj,
+                      uint8_t* dst_vu, int width) = MergeUVRow_C;
+  if (!src_raw || !dst_y || !dst_vu || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+#if defined(HAS_RAWTOYJROW)
+
+// Neon version does direct RAW to YUV.
+#if defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVJRow = RAWToUVJRow_Any_NEON;
+    RAWToYJRow = RAWToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_NEON;
+      RAWToUVJRow = RAWToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_MSA) && defined(HAS_RAWTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToUVJRow = RAWToUVJRow_Any_MSA;
+    RAWToYJRow = RAWToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_MSA;
+      RAWToUVJRow = RAWToUVJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToYJRow = RAWToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToYJRow = RAWToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYJRow = RAWToYJRow_LASX;
+    }
+  }
+#endif
+
+// Other platforms do intermediate conversion from RAW to ARGB.
+#else  // HAS_RAWTOYJROW
+
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RAWTOYJROW
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX512BW;
+    if (IS_ALIGNED(halfwidth, 64)) {
+      MergeUVRow_ = MergeUVRow_AVX512BW;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
+  {
+    // Allocate a row of uv.
+    align_buffer_64(row_uj, ((halfwidth + 31) & ~31) * 2);
+    uint8_t* row_vj = row_uj + ((halfwidth + 31) & ~31);
+#if !defined(HAS_RAWTOYJROW)
+    // Allocate 2 rows of ARGB.
+    const int row_size = (width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, src_stride_raw, row_uj, row_vj, width);
+      MergeUVRow_(row_vj, row_uj, dst_vu, halfwidth);
+      RAWToYJRow(src_raw, dst_y, width);
+      RAWToYJRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + row_size, width);
+      ARGBToUVJRow(row, row_size, row_uj, row_vj, width);
+      MergeUVRow_(row_vj, row_uj, dst_vu, halfwidth);
+      ARGBToYJRow(row, dst_y, width);
+      ARGBToYJRow(row + row_size, dst_y + dst_stride_y, width);
+#endif
+      src_raw += src_stride_raw * 2;
+      dst_y += dst_stride_y * 2;
+      dst_vu += dst_stride_vu;
+    }
+    if (height & 1) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, 0, row_uj, row_vj, width);
+      MergeUVRow_(row_vj, row_uj, dst_vu, halfwidth);
+      RAWToYJRow(src_raw, dst_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      ARGBToUVJRow(row, 0, row_uj, row_vj, width);
+      MergeUVRow_(row_vj, row_uj, dst_vu, halfwidth);
+      ARGBToYJRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RAWTOYJROW)
+    free_aligned_buffer_64(row);
+#endif
+    free_aligned_buffer_64(row_uj);
+  }
+  return 0;
+}
+#undef HAS_RAWTOYJROW
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/media/libyuv/source/convert_jpeg.cc b/media/libyuv/source/convert_jpeg.cc
index 90f550a26a..d7556ee91b 100644
--- a/media/libyuv/source/convert_jpeg.cc
+++ b/media/libyuv/source/convert_jpeg.cc
@@ -22,28 +22,24 @@ extern "C" {
 
 #ifdef HAVE_JPEG
 struct I420Buffers {
-  uint8* y;
+  uint8_t* y;
   int y_stride;
-  uint8* u;
+  uint8_t* u;
   int u_stride;
-  uint8* v;
+  uint8_t* v;
   int v_stride;
   int w;
   int h;
 };
 
 static void JpegCopyI420(void* opaque,
-                         const uint8* const* data,
+                         const uint8_t* const* data,
                          const int* strides,
                          int rows) {
   I420Buffers* dest = (I420Buffers*)(opaque);
-  I420Copy(data[0], strides[0],
-           data[1], strides[1],
-           data[2], strides[2],
-           dest->y, dest->y_stride,
-           dest->u, dest->u_stride,
-           dest->v, dest->v_stride,
-           dest->w, rows);
+  I420Copy(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+           dest->y, dest->y_stride, dest->u, dest->u_stride, dest->v,
+           dest->v_stride, dest->w, rows);
   dest->y += rows * dest->y_stride;
   dest->u += ((rows + 1) >> 1) * dest->u_stride;
   dest->v += ((rows + 1) >> 1) * dest->v_stride;
@@ -51,17 +47,13 @@ static void JpegCopyI420(void* opaque,
 }
 
 static void JpegI422ToI420(void* opaque,
-                           const uint8* const* data,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   I420Buffers* dest = (I420Buffers*)(opaque);
-  I422ToI420(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->y, dest->y_stride,
-             dest->u, dest->u_stride,
-             dest->v, dest->v_stride,
-             dest->w, rows);
+  I422ToI420(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->u, dest->u_stride, dest->v,
+             dest->v_stride, dest->w, rows);
   dest->y += rows * dest->y_stride;
   dest->u += ((rows + 1) >> 1) * dest->u_stride;
   dest->v += ((rows + 1) >> 1) * dest->v_stride;
@@ -69,35 +61,13 @@ static void JpegI422ToI420(void* opaque,
 }
 
 static void JpegI444ToI420(void* opaque,
-                           const uint8* const* data,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   I420Buffers* dest = (I420Buffers*)(opaque);
-  I444ToI420(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->y, dest->y_stride,
-             dest->u, dest->u_stride,
-             dest->v, dest->v_stride,
-             dest->w, rows);
-  dest->y += rows * dest->y_stride;
-  dest->u += ((rows + 1) >> 1) * dest->u_stride;
-  dest->v += ((rows + 1) >> 1) * dest->v_stride;
-  dest->h -= rows;
-}
-
-static void JpegI411ToI420(void* opaque,
-                           const uint8* const* data,
-                           const int* strides,
-                           int rows) {
-  I420Buffers* dest = (I420Buffers*)(opaque);
-  I411ToI420(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->y, dest->y_stride,
-             dest->u, dest->u_stride,
-             dest->v, dest->v_stride,
-             dest->w, rows);
+  I444ToI420(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->u, dest->u_stride, dest->v,
+             dest->v_stride, dest->w, rows);
   dest->y += rows * dest->y_stride;
   dest->u += ((rows + 1) >> 1) * dest->u_stride;
   dest->v += ((rows + 1) >> 1) * dest->v_stride;
@@ -105,15 +75,12 @@ static void JpegI411ToI420(void* opaque,
 }
 
 static void JpegI400ToI420(void* opaque,
-                           const uint8* const* data,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   I420Buffers* dest = (I420Buffers*)(opaque);
-  I400ToI420(data[0], strides[0],
-             dest->y, dest->y_stride,
-             dest->u, dest->u_stride,
-             dest->v, dest->v_stride,
-             dest->w, rows);
+  I400ToI420(data[0], strides[0], dest->y, dest->y_stride, dest->u,
+             dest->u_stride, dest->v, dest->v_stride, dest->w, rows);
   dest->y += rows * dest->y_stride;
   dest->u += ((rows + 1) >> 1) * dest->u_stride;
   dest->v += ((rows + 1) >> 1) * dest->v_stride;
@@ -122,10 +89,12 @@ static void JpegI400ToI420(void* opaque,
 
 // Query size of MJPG in pixels.
 LIBYUV_API
-int MJPGSize(const uint8* sample, size_t sample_size,
-             int* width, int* height) {
+int MJPGSize(const uint8_t* src_mjpg,
+             size_t src_size_mjpg,
+             int* width,
+             int* height) {
   MJpegDecoder mjpeg_decoder;
-  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(src_mjpg, src_size_mjpg);
   if (ret) {
     *width = mjpeg_decoder.GetWidth();
     *height = mjpeg_decoder.GetHeight();
@@ -135,34 +104,40 @@ int MJPGSize(const uint8* sample, size_t sample_size,
 }
 
 // MJPG (Motion JPeg) to I420
-// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+// TODO(fbarchard): review src_width and src_height requirement. dst_width and
+// dst_height may be enough.
 LIBYUV_API
-int MJPGToI420(const uint8* sample,
-               size_t sample_size,
-               uint8* y, int y_stride,
-               uint8* u, int u_stride,
-               uint8* v, int v_stride,
-               int w, int h,
-               int dw, int dh) {
-  if (sample_size == kUnknownDataSize) {
+int MJPGToI420(const uint8_t* src_mjpg,
+               size_t src_size_mjpg,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height) {
+  if (src_size_mjpg == kUnknownDataSize) {
     // ERROR: MJPEG frame size unknown
     return -1;
   }
 
   // TODO(fbarchard): Port MJpeg to C.
   MJpegDecoder mjpeg_decoder;
-  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
-  if (ret && (mjpeg_decoder.GetWidth() != w ||
-              mjpeg_decoder.GetHeight() != h)) {
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(src_mjpg, src_size_mjpg);
+  if (ret && (mjpeg_decoder.GetWidth() != src_width ||
+              mjpeg_decoder.GetHeight() != src_height)) {
     // ERROR: MJPEG frame has unexpected dimensions
     mjpeg_decoder.UnloadFrame();
     return 1;  // runtime failure
   }
   if (ret) {
-    I420Buffers bufs = { y, y_stride, u, u_stride, v, v_stride, dw, dh };
+    I420Buffers bufs = {dst_y, dst_stride_y, dst_u,     dst_stride_u,
+                        dst_v, dst_stride_v, dst_width, dst_height};
     // YUV420
-    if (mjpeg_decoder.GetColorSpace() ==
-            MJpegDecoder::kColorSpaceYCbCr &&
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
         mjpeg_decoder.GetNumComponents() == 3 &&
         mjpeg_decoder.GetVertSampFactor(0) == 2 &&
         mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
@@ -170,8 +145,9 @@ int MJPGToI420(const uint8* sample,
         mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
         mjpeg_decoder.GetVertSampFactor(2) == 1 &&
         mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh);
-    // YUV422
+      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dst_width,
+                                           dst_height);
+      // YUV422
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceYCbCr &&
                mjpeg_decoder.GetNumComponents() == 3 &&
@@ -181,8 +157,9 @@ int MJPGToI420(const uint8* sample,
                mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
                mjpeg_decoder.GetVertSampFactor(2) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh);
-    // YUV444
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dst_width,
+                                           dst_height);
+      // YUV444
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceYCbCr &&
                mjpeg_decoder.GetNumComponents() == 3 &&
@@ -192,29 +169,292 @@ int MJPGToI420(const uint8* sample,
                mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
                mjpeg_decoder.GetVertSampFactor(2) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh);
-    // YUV411
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dst_width,
+                                           dst_height);
+      // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dst_width,
+                                           dst_height);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other
+      // colorspace/subsample factors that occur in practice. ERROR: Unable to
+      // convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+struct NV21Buffers {
+  uint8_t* y;
+  int y_stride;
+  uint8_t* vu;
+  int vu_stride;
+  int w;
+  int h;
+};
+
+static void JpegI420ToNV21(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  I420ToNV21(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToNV21(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  I422ToNV21(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToNV21(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  I444ToNV21(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToNV21(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  I400ToNV21(data[0], strides[0], dest->y, dest->y_stride, dest->vu,
+             dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPeg) to NV21
+LIBYUV_API
+int MJPGToNV21(const uint8_t* src_mjpg,
+               size_t src_size_mjpg,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height) {
+  if (src_size_mjpg == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(src_mjpg, src_size_mjpg);
+  if (ret && (mjpeg_decoder.GetWidth() != src_width ||
+              mjpeg_decoder.GetHeight() != src_height)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    NV21Buffers bufs = {dst_y,         dst_stride_y, dst_vu,
+                        dst_stride_vu, dst_width,    dst_height};
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToNV21, &bufs, dst_width,
+                                           dst_height);
+      // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToNV21, &bufs, dst_width,
+                                           dst_height);
+      // YUV444
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceYCbCr &&
                mjpeg_decoder.GetNumComponents() == 3 &&
                mjpeg_decoder.GetVertSampFactor(0) == 1 &&
-               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
                mjpeg_decoder.GetVertSampFactor(1) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
                mjpeg_decoder.GetVertSampFactor(2) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToI420, &bufs, dw, dh);
-    // YUV400
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToNV21, &bufs, dst_width,
+                                           dst_height);
+      // YUV400
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceGrayscale &&
                mjpeg_decoder.GetNumComponents() == 1 &&
                mjpeg_decoder.GetVertSampFactor(0) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(0) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh);
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToNV21, &bufs, dst_width,
+                                           dst_height);
     } else {
-      // TODO(fbarchard): Implement conversion for any other colorspace/sample
-      // factors that occur in practice. 411 is supported by libjpeg
-      // ERROR: Unable to convert MJPEG frame because format is not supported
+      // Unknown colorspace.
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+static void JpegI420ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 with VU swapped.
+  I420ToNV21(data[0], strides[0], data[2], strides[2], data[1], strides[1],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 with VU swapped.
+  I422ToNV21(data[0], strides[0], data[2], strides[2], data[1], strides[1],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 with VU swapped.
+  I444ToNV21(data[0], strides[0], data[2], strides[2], data[1], strides[1],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 since there is no UV plane.
+  I400ToNV21(data[0], strides[0], dest->y, dest->y_stride, dest->vu,
+             dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPEG) to NV12.
+LIBYUV_API
+int MJPGToNV12(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != src_width ||
+              mjpeg_decoder.GetHeight() != src_height)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    // Use NV21Buffers but with UV instead of VU.
+    NV21Buffers bufs = {dst_y,         dst_stride_y, dst_uv,
+                        dst_stride_uv, dst_width,    dst_height};
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToNV12, &bufs, dst_width,
+                                           dst_height);
+      // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToNV12, &bufs, dst_width,
+                                           dst_height);
+      // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToNV12, &bufs, dst_width,
+                                           dst_height);
+      // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToNV12, &bufs, dst_width,
+                                           dst_height);
+    } else {
+      // Unknown colorspace.
       mjpeg_decoder.UnloadFrame();
       return 1;
     }
@@ -222,109 +462,86 @@ int MJPGToI420(const uint8* sample,
   return ret ? 0 : 1;
 }
 
-#ifdef HAVE_JPEG
 struct ARGBBuffers {
-  uint8* argb;
+  uint8_t* argb;
   int argb_stride;
   int w;
   int h;
 };
 
 static void JpegI420ToARGB(void* opaque,
-                         const uint8* const* data,
-                         const int* strides,
-                         int rows) {
-  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
-  I420ToARGB(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->argb, dest->argb_stride,
-             dest->w, rows);
-  dest->argb += rows * dest->argb_stride;
-  dest->h -= rows;
-}
-
-static void JpegI422ToARGB(void* opaque,
-                           const uint8* const* data,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   ARGBBuffers* dest = (ARGBBuffers*)(opaque);
-  I422ToARGB(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->argb, dest->argb_stride,
-             dest->w, rows);
+  I420ToARGB(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->argb, dest->argb_stride, dest->w, rows);
   dest->argb += rows * dest->argb_stride;
   dest->h -= rows;
 }
 
-static void JpegI444ToARGB(void* opaque,
-                           const uint8* const* data,
+static void JpegI422ToARGB(void* opaque,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   ARGBBuffers* dest = (ARGBBuffers*)(opaque);
-  I444ToARGB(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->argb, dest->argb_stride,
-             dest->w, rows);
+  I422ToARGB(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->argb, dest->argb_stride, dest->w, rows);
   dest->argb += rows * dest->argb_stride;
   dest->h -= rows;
 }
 
-static void JpegI411ToARGB(void* opaque,
-                           const uint8* const* data,
+static void JpegI444ToARGB(void* opaque,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   ARGBBuffers* dest = (ARGBBuffers*)(opaque);
-  I411ToARGB(data[0], strides[0],
-             data[1], strides[1],
-             data[2], strides[2],
-             dest->argb, dest->argb_stride,
-             dest->w, rows);
+  I444ToARGB(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->argb, dest->argb_stride, dest->w, rows);
   dest->argb += rows * dest->argb_stride;
   dest->h -= rows;
 }
 
 static void JpegI400ToARGB(void* opaque,
-                           const uint8* const* data,
+                           const uint8_t* const* data,
                            const int* strides,
                            int rows) {
   ARGBBuffers* dest = (ARGBBuffers*)(opaque);
-  I400ToARGB(data[0], strides[0],
-             dest->argb, dest->argb_stride,
-             dest->w, rows);
+  I400ToARGB(data[0], strides[0], dest->argb, dest->argb_stride, dest->w, rows);
   dest->argb += rows * dest->argb_stride;
   dest->h -= rows;
 }
 
 // MJPG (Motion JPeg) to ARGB
-// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+// TODO(fbarchard): review src_width and src_height requirement. dst_width and
+// dst_height may be enough.
 LIBYUV_API
-int MJPGToARGB(const uint8* sample,
-               size_t sample_size,
-               uint8* argb, int argb_stride,
-               int w, int h,
-               int dw, int dh) {
-  if (sample_size == kUnknownDataSize) {
+int MJPGToARGB(const uint8_t* src_mjpg,
+               size_t src_size_mjpg,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height) {
+  if (src_size_mjpg == kUnknownDataSize) {
     // ERROR: MJPEG frame size unknown
     return -1;
   }
 
   // TODO(fbarchard): Port MJpeg to C.
   MJpegDecoder mjpeg_decoder;
-  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
-  if (ret && (mjpeg_decoder.GetWidth() != w ||
-              mjpeg_decoder.GetHeight() != h)) {
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(src_mjpg, src_size_mjpg);
+  if (ret && (mjpeg_decoder.GetWidth() != src_width ||
+              mjpeg_decoder.GetHeight() != src_height)) {
     // ERROR: MJPEG frame has unexpected dimensions
     mjpeg_decoder.UnloadFrame();
     return 1;  // runtime failure
   }
   if (ret) {
-    ARGBBuffers bufs = { argb, argb_stride, dw, dh };
+    ARGBBuffers bufs = {dst_argb, dst_stride_argb, dst_width, dst_height};
     // YUV420
-    if (mjpeg_decoder.GetColorSpace() ==
-            MJpegDecoder::kColorSpaceYCbCr &&
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
         mjpeg_decoder.GetNumComponents() == 3 &&
         mjpeg_decoder.GetVertSampFactor(0) == 2 &&
         mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
@@ -332,8 +549,9 @@ int MJPGToARGB(const uint8* sample,
         mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
         mjpeg_decoder.GetVertSampFactor(2) == 1 &&
         mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh);
-    // YUV422
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dst_width,
+                                           dst_height);
+      // YUV422
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceYCbCr &&
                mjpeg_decoder.GetNumComponents() == 3 &&
@@ -343,8 +561,9 @@ int MJPGToARGB(const uint8* sample,
                mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
                mjpeg_decoder.GetVertSampFactor(2) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh);
-    // YUV444
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dst_width,
+                                           dst_height);
+      // YUV444
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceYCbCr &&
                mjpeg_decoder.GetNumComponents() == 3 &&
@@ -354,38 +573,28 @@ int MJPGToARGB(const uint8* sample,
                mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
                mjpeg_decoder.GetVertSampFactor(2) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh);
-    // YUV411
-    } else if (mjpeg_decoder.GetColorSpace() ==
-                   MJpegDecoder::kColorSpaceYCbCr &&
-               mjpeg_decoder.GetNumComponents() == 3 &&
-               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
-               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
-               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
-               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
-               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
-               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToARGB, &bufs, dw, dh);
-    // YUV400
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dst_width,
+                                           dst_height);
+      // YUV400
     } else if (mjpeg_decoder.GetColorSpace() ==
                    MJpegDecoder::kColorSpaceGrayscale &&
                mjpeg_decoder.GetNumComponents() == 1 &&
                mjpeg_decoder.GetVertSampFactor(0) == 1 &&
                mjpeg_decoder.GetHorizSampFactor(0) == 1) {
-      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh);
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dst_width,
+                                           dst_height);
     } else {
-      // TODO(fbarchard): Implement conversion for any other colorspace/sample
-      // factors that occur in practice. 411 is supported by libjpeg
-      // ERROR: Unable to convert MJPEG frame because format is not supported
+      // TODO(fbarchard): Implement conversion for any other
+      // colorspace/subsample factors that occur in practice. ERROR: Unable to
+      // convert MJPEG frame because format is not supported
       mjpeg_decoder.UnloadFrame();
       return 1;
     }
   }
   return ret ? 0 : 1;
 }
-#endif
 
-#endif
+#endif  // HAVE_JPEG
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/media/libyuv/source/convert_to_argb.cc b/media/libyuv/source/convert_to_argb.cc
index aecdc80fde..84df16c8c2 100644
--- a/media/libyuv/source/convert_to_argb.cc
+++ b/media/libyuv/source/convert_to_argb.cc
@@ -28,36 +28,45 @@ extern "C" {
 // src_height is used to compute location of planes, and indicate inversion
 // sample_size is measured in bytes and is the size of the frame.
 //   With MJPEG it is the compressed size of the frame.
+
+// TODO(fbarchard): Add the following:
+// H010ToARGB
+// I010ToARGB
+
 LIBYUV_API
-int ConvertToARGB(const uint8* sample, size_t sample_size,
-                  uint8* crop_argb, int argb_stride,
-                  int crop_x, int crop_y,
-                  int src_width, int src_height,
-                  int crop_width, int crop_height,
+int ConvertToARGB(const uint8_t* sample,
+                  size_t sample_size,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int crop_x,
+                  int crop_y,
+                  int src_width,
+                  int src_height,
+                  int crop_width,
+                  int crop_height,
                   enum RotationMode rotation,
-                  uint32 fourcc) {
-  uint32 format = CanonicalFourCC(fourcc);
+                  uint32_t fourcc) {
+  uint32_t format = CanonicalFourCC(fourcc);
   int aligned_src_width = (src_width + 1) & ~1;
-  const uint8* src;
-  const uint8* src_uv;
+  const uint8_t* src;
+  const uint8_t* src_uv;
   int abs_src_height = (src_height < 0) ? -src_height : src_height;
   int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
   int r = 0;
 
   // One pass rotation is available for some formats. For the rest, convert
-  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
-  // and then rotate the I420 to the final destination buffer.
-  // For in-place conversion, if destination crop_argb is same as source sample,
+  // to ARGB (with optional vertical flipping) into a temporary ARGB buffer,
+  // and then rotate the ARGB to the final destination buffer.
+  // For in-place conversion, if destination dst_argb is same as source sample,
   // also enable temporary buffer.
-  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_ARGB) ||
-      crop_argb == sample;
-  uint8* dest_argb = crop_argb;
-  int dest_argb_stride = argb_stride;
-  uint8* rotate_buffer = NULL;
+  LIBYUV_BOOL need_buf =
+      (rotation && format != FOURCC_ARGB) || dst_argb == sample;
+  uint8_t* dest_argb = dst_argb;
+  int dest_dst_stride_argb = dst_stride_argb;
+  uint8_t* rotate_buffer = NULL;
   int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
 
-  if (crop_argb == NULL || sample == NULL ||
-      src_width <= 0 || crop_width <= 0 ||
+  if (dst_argb == NULL || sample == NULL || src_width <= 0 || crop_width <= 0 ||
       src_height == 0 || crop_height == 0) {
     return -1;
   }
@@ -67,187 +76,236 @@ int ConvertToARGB(const uint8* sample, size_t sample_size,
 
   if (need_buf) {
     int argb_size = crop_width * 4 * abs_crop_height;
-    rotate_buffer = (uint8*)malloc(argb_size);
+    rotate_buffer = (uint8_t*)malloc(argb_size); /* NOLINT */
     if (!rotate_buffer) {
       return 1;  // Out of memory runtime error.
     }
-    crop_argb = rotate_buffer;
-    argb_stride = crop_width * 4;
+    dst_argb = rotate_buffer;
+    dst_stride_argb = crop_width * 4;
   }
 
   switch (format) {
     // Single plane formats
     case FOURCC_YUY2:
       src = sample + (aligned_src_width * crop_y + crop_x) * 2;
-      r = YUY2ToARGB(src, aligned_src_width * 2,
-                     crop_argb, argb_stride,
+      r = YUY2ToARGB(src, aligned_src_width * 2, dst_argb, dst_stride_argb,
                      crop_width, inv_crop_height);
       break;
     case FOURCC_UYVY:
       src = sample + (aligned_src_width * crop_y + crop_x) * 2;
-      r = UYVYToARGB(src, aligned_src_width * 2,
-                     crop_argb, argb_stride,
+      r = UYVYToARGB(src, aligned_src_width * 2, dst_argb, dst_stride_argb,
                      crop_width, inv_crop_height);
       break;
     case FOURCC_24BG:
       src = sample + (src_width * crop_y + crop_x) * 3;
-      r = RGB24ToARGB(src, src_width * 3,
-                      crop_argb, argb_stride,
-                      crop_width, inv_crop_height);
+      r = RGB24ToARGB(src, src_width * 3, dst_argb, dst_stride_argb, crop_width,
+                      inv_crop_height);
       break;
     case FOURCC_RAW:
       src = sample + (src_width * crop_y + crop_x) * 3;
-      r = RAWToARGB(src, src_width * 3,
-                    crop_argb, argb_stride,
-                    crop_width, inv_crop_height);
+      r = RAWToARGB(src, src_width * 3, dst_argb, dst_stride_argb, crop_width,
+                    inv_crop_height);
       break;
     case FOURCC_ARGB:
-      src = sample + (src_width * crop_y + crop_x) * 4;
-      r = ARGBToARGB(src, src_width * 4,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      if (!need_buf && !rotation) {
+        src = sample + (src_width * crop_y + crop_x) * 4;
+        r = ARGBToARGB(src, src_width * 4, dst_argb, dst_stride_argb,
+                       crop_width, inv_crop_height);
+      }
       break;
     case FOURCC_BGRA:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = BGRAToARGB(src, src_width * 4,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = BGRAToARGB(src, src_width * 4, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
       break;
     case FOURCC_ABGR:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = ABGRToARGB(src, src_width * 4,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = ABGRToARGB(src, src_width * 4, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
       break;
     case FOURCC_RGBA:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = RGBAToARGB(src, src_width * 4,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = RGBAToARGB(src, src_width * 4, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
+      break;
+    case FOURCC_AR30:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = AR30ToARGB(src, src_width * 4, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
+      break;
+    case FOURCC_AB30:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = AB30ToARGB(src, src_width * 4, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
       break;
     case FOURCC_RGBP:
       src = sample + (src_width * crop_y + crop_x) * 2;
-      r = RGB565ToARGB(src, src_width * 2,
-                       crop_argb, argb_stride,
+      r = RGB565ToARGB(src, src_width * 2, dst_argb, dst_stride_argb,
                        crop_width, inv_crop_height);
       break;
     case FOURCC_RGBO:
       src = sample + (src_width * crop_y + crop_x) * 2;
-      r = ARGB1555ToARGB(src, src_width * 2,
-                         crop_argb, argb_stride,
+      r = ARGB1555ToARGB(src, src_width * 2, dst_argb, dst_stride_argb,
                          crop_width, inv_crop_height);
       break;
     case FOURCC_R444:
       src = sample + (src_width * crop_y + crop_x) * 2;
-      r = ARGB4444ToARGB(src, src_width * 2,
-                         crop_argb, argb_stride,
+      r = ARGB4444ToARGB(src, src_width * 2, dst_argb, dst_stride_argb,
                          crop_width, inv_crop_height);
       break;
     case FOURCC_I400:
       src = sample + src_width * crop_y + crop_x;
-      r = I400ToARGB(src, src_width,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = I400ToARGB(src, src_width, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
+      break;
+    case FOURCC_J400:
+      src = sample + src_width * crop_y + crop_x;
+      r = J400ToARGB(src, src_width, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
       break;
 
     // Biplanar formats
     case FOURCC_NV12:
       src = sample + (src_width * crop_y + crop_x);
-      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
-      r = NV12ToARGB(src, src_width,
-                     src_uv, aligned_src_width,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      src_uv =
+          sample + aligned_src_width * (abs_src_height + crop_y / 2) + crop_x;
+      r = NV12ToARGB(src, src_width, src_uv, aligned_src_width, dst_argb,
+                     dst_stride_argb, crop_width, inv_crop_height);
       break;
     case FOURCC_NV21:
       src = sample + (src_width * crop_y + crop_x);
-      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      src_uv =
+          sample + aligned_src_width * (abs_src_height + crop_y / 2) + crop_x;
       // Call NV12 but with u and v parameters swapped.
-      r = NV21ToARGB(src, src_width,
-                     src_uv, aligned_src_width,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
-      break;
-    case FOURCC_M420:
-      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
-      r = M420ToARGB(src, src_width,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = NV21ToARGB(src, src_width, src_uv, aligned_src_width, dst_argb,
+                     dst_stride_argb, crop_width, inv_crop_height);
       break;
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YV12: {
-      const uint8* src_y = sample + (src_width * crop_y + crop_x);
-      const uint8* src_u;
-      const uint8* src_v;
+      const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8_t* src_u;
+      const uint8_t* src_v;
       int halfwidth = (src_width + 1) / 2;
       int halfheight = (abs_src_height + 1) / 2;
       if (format == FOURCC_YV12) {
         src_v = sample + src_width * abs_src_height +
-            (halfwidth * crop_y + crop_x) / 2;
+                (halfwidth * crop_y + crop_x) / 2;
         src_u = sample + src_width * abs_src_height +
-            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+                halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
       } else {
         src_u = sample + src_width * abs_src_height +
-            (halfwidth * crop_y + crop_x) / 2;
+                (halfwidth * crop_y + crop_x) / 2;
         src_v = sample + src_width * abs_src_height +
-            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+                halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
       }
-      r = I420ToARGB(src_y, src_width,
-                     src_u, halfwidth,
-                     src_v, halfwidth,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = I420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
 
     case FOURCC_J420: {
-      const uint8* src_y = sample + (src_width * crop_y + crop_x);
-      const uint8* src_u;
-      const uint8* src_v;
       int halfwidth = (src_width + 1) / 2;
       int halfheight = (abs_src_height + 1) / 2;
-      src_u = sample + src_width * abs_src_height +
-          (halfwidth * crop_y + crop_x) / 2;
-      src_v = sample + src_width * abs_src_height +
-          halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
-      r = J420ToARGB(src_y, src_width,
-                     src_u, halfwidth,
-                     src_v, halfwidth,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8_t* src_u = sample + src_width * abs_src_height +
+                             (halfwidth * crop_y + crop_x) / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = J420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_H420: {
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8_t* src_u = sample + src_width * abs_src_height +
+                             (halfwidth * crop_y + crop_x) / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = H420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_U420: {
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8_t* src_u = sample + src_width * abs_src_height +
+                             (halfwidth * crop_y + crop_x) / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = U420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
 
     case FOURCC_I422:
     case FOURCC_YV16: {
-      const uint8* src_y = sample + src_width * crop_y + crop_x;
-      const uint8* src_u;
-      const uint8* src_v;
       int halfwidth = (src_width + 1) / 2;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
       if (format == FOURCC_YV16) {
-        src_v = sample + src_width * abs_src_height +
-            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height + halfwidth * crop_y +
+                crop_x / 2;
         src_u = sample + src_width * abs_src_height +
-            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+                halfwidth * (abs_src_height + crop_y) + crop_x / 2;
       } else {
-        src_u = sample + src_width * abs_src_height +
-            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height + halfwidth * crop_y +
+                crop_x / 2;
         src_v = sample + src_width * abs_src_height +
-            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+                halfwidth * (abs_src_height + crop_y) + crop_x / 2;
       }
-      r = I422ToARGB(src_y, src_width,
-                     src_u, halfwidth,
-                     src_v, halfwidth,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = I422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
+
+    case FOURCC_J422: {
+      int halfwidth = (src_width + 1) / 2;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u =
+          sample + src_width * abs_src_height + halfwidth * crop_y + crop_x / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      r = J422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_H422: {
+      int halfwidth = (src_width + 1) / 2;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u =
+          sample + src_width * abs_src_height + halfwidth * crop_y + crop_x / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      r = H422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_U422: {
+      int halfwidth = (src_width + 1) / 2;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u =
+          sample + src_width * abs_src_height + halfwidth * crop_y + crop_x / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      r = H422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
     case FOURCC_I444:
     case FOURCC_YV24: {
-      const uint8* src_y = sample + src_width * crop_y + crop_x;
-      const uint8* src_u;
-      const uint8* src_v;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
       if (format == FOURCC_YV24) {
         src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
         src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
@@ -255,32 +313,48 @@ int ConvertToARGB(const uint8* sample, size_t sample_size,
         src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
         src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
       }
-      r = I444ToARGB(src_y, src_width,
-                     src_u, src_width,
-                     src_v, src_width,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+      r = I444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
-    case FOURCC_I411: {
-      int quarterwidth = (src_width + 3) / 4;
-      const uint8* src_y = sample + src_width * crop_y + crop_x;
-      const uint8* src_u = sample + src_width * abs_src_height +
-          quarterwidth * crop_y + crop_x / 4;
-      const uint8* src_v = sample + src_width * abs_src_height +
-          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
-      r = I411ToARGB(src_y, src_width,
-                     src_u, quarterwidth,
-                     src_v, quarterwidth,
-                     crop_argb, argb_stride,
-                     crop_width, inv_crop_height);
+
+    case FOURCC_J444: {
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
+      src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+      src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      r = J444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
+
+    case FOURCC_H444: {
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
+      src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+      src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      r = H444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_U444: {
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
+      src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+      src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      r = U444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
 #ifdef HAVE_JPEG
     case FOURCC_MJPG:
-      r = MJPGToARGB(sample, sample_size,
-                     crop_argb, argb_stride,
-                     src_width, abs_src_height, crop_width, inv_crop_height);
+      r = MJPGToARGB(sample, sample_size, dst_argb, dst_stride_argb, src_width,
+                     abs_src_height, crop_width, inv_crop_height);
       break;
 #endif
     default:
@@ -289,11 +363,14 @@ int ConvertToARGB(const uint8* sample, size_t sample_size,
 
   if (need_buf) {
     if (!r) {
-      r = ARGBRotate(crop_argb, argb_stride,
-                     dest_argb, dest_argb_stride,
+      r = ARGBRotate(dst_argb, dst_stride_argb, dest_argb, dest_dst_stride_argb,
                      crop_width, abs_crop_height, rotation);
     }
     free(rotate_buffer);
+  } else if (rotation) {
+    src = sample + (src_width * crop_y + crop_x) * 4;
+    r = ARGBRotate(src, src_width * 4, dst_argb, dst_stride_argb, crop_width,
+                   inv_crop_height, rotation);
   }
 
   return r;
diff --git a/media/libyuv/source/convert_to_i420.cc b/media/libyuv/source/convert_to_i420.cc
index e5f307c446..5869ecd7b9 100644
--- a/media/libyuv/source/convert_to_i420.cc
+++ b/media/libyuv/source/convert_to_i420.cc
@@ -25,251 +25,219 @@ extern "C" {
 // sample_size is measured in bytes and is the size of the frame.
 //   With MJPEG it is the compressed size of the frame.
 LIBYUV_API
-int ConvertToI420(const uint8* sample,
+int ConvertToI420(const uint8_t* sample,
                   size_t sample_size,
-                  uint8* y, int y_stride,
-                  uint8* u, int u_stride,
-                  uint8* v, int v_stride,
-                  int crop_x, int crop_y,
-                  int src_width, int src_height,
-                  int crop_width, int crop_height,
+                  uint8_t* dst_y,
+                  int dst_stride_y,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int crop_x,
+                  int crop_y,
+                  int src_width,
+                  int src_height,
+                  int crop_width,
+                  int crop_height,
                   enum RotationMode rotation,
-                  uint32 fourcc) {
-  uint32 format = CanonicalFourCC(fourcc);
+                  uint32_t fourcc) {
+  uint32_t format = CanonicalFourCC(fourcc);
   int aligned_src_width = (src_width + 1) & ~1;
-  const uint8* src;
-  const uint8* src_uv;
+  const uint8_t* src;
+  const uint8_t* src_uv;
   const int abs_src_height = (src_height < 0) ? -src_height : src_height;
   // TODO(nisse): Why allow crop_height < 0?
   const int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
   int r = 0;
-  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_I420 &&
-      format != FOURCC_NV12 && format != FOURCC_NV21 &&
-      format != FOURCC_YV12) || y == sample;
-  uint8* tmp_y = y;
-  uint8* tmp_u = u;
-  uint8* tmp_v = v;
-  int tmp_y_stride = y_stride;
-  int tmp_u_stride = u_stride;
-  int tmp_v_stride = v_stride;
-  uint8* rotate_buffer = NULL;
+  LIBYUV_BOOL need_buf =
+      (rotation && format != FOURCC_I420 && format != FOURCC_NV12 &&
+       format != FOURCC_NV21 && format != FOURCC_YV12) ||
+      dst_y == sample;
+  uint8_t* tmp_y = dst_y;
+  uint8_t* tmp_u = dst_u;
+  uint8_t* tmp_v = dst_v;
+  int tmp_y_stride = dst_stride_y;
+  int tmp_u_stride = dst_stride_u;
+  int tmp_v_stride = dst_stride_v;
+  uint8_t* rotate_buffer = NULL;
   const int inv_crop_height =
       (src_height < 0) ? -abs_crop_height : abs_crop_height;
 
-  if (!y || !u || !v || !sample ||
-      src_width <= 0 || crop_width <= 0  ||
-      src_height == 0 || crop_height == 0) {
+  if (!dst_y || !dst_u || !dst_v || !sample || src_width <= 0 ||
+      crop_width <= 0 || src_height == 0 || crop_height == 0) {
     return -1;
   }
 
   // One pass rotation is available for some formats. For the rest, convert
   // to I420 (with optional vertical flipping) into a temporary I420 buffer,
   // and then rotate the I420 to the final destination buffer.
-  // For in-place conversion, if destination y is same as source sample,
+  // For in-place conversion, if destination dst_y is same as source sample,
   // also enable temporary buffer.
   if (need_buf) {
     int y_size = crop_width * abs_crop_height;
     int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2);
-    rotate_buffer = (uint8*)malloc(y_size + uv_size * 2);
+    rotate_buffer = (uint8_t*)malloc(y_size + uv_size * 2); /* NOLINT */
     if (!rotate_buffer) {
       return 1;  // Out of memory runtime error.
     }
-    y = rotate_buffer;
-    u = y + y_size;
-    v = u + uv_size;
-    y_stride = crop_width;
-    u_stride = v_stride = ((crop_width + 1) / 2);
+    dst_y = rotate_buffer;
+    dst_u = dst_y + y_size;
+    dst_v = dst_u + uv_size;
+    dst_stride_y = crop_width;
+    dst_stride_u = dst_stride_v = ((crop_width + 1) / 2);
   }
 
   switch (format) {
     // Single plane formats
-    case FOURCC_YUY2:
+    case FOURCC_YUY2: {  // TODO(fbarchard): Find better odd crop fix.
+      uint8_t* u = (crop_x & 1) ? dst_v : dst_u;
+      uint8_t* v = (crop_x & 1) ? dst_u : dst_v;
+      int stride_u = (crop_x & 1) ? dst_stride_v : dst_stride_u;
+      int stride_v = (crop_x & 1) ? dst_stride_u : dst_stride_v;
       src = sample + (aligned_src_width * crop_y + crop_x) * 2;
-      r = YUY2ToI420(src, aligned_src_width * 2,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = YUY2ToI420(src, aligned_src_width * 2, dst_y, dst_stride_y, u,
+                     stride_u, v, stride_v, crop_width, inv_crop_height);
       break;
-    case FOURCC_UYVY:
+    }
+    case FOURCC_UYVY: {
+      uint8_t* u = (crop_x & 1) ? dst_v : dst_u;
+      uint8_t* v = (crop_x & 1) ? dst_u : dst_v;
+      int stride_u = (crop_x & 1) ? dst_stride_v : dst_stride_u;
+      int stride_v = (crop_x & 1) ? dst_stride_u : dst_stride_v;
       src = sample + (aligned_src_width * crop_y + crop_x) * 2;
-      r = UYVYToI420(src, aligned_src_width * 2,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = UYVYToI420(src, aligned_src_width * 2, dst_y, dst_stride_y, u,
+                     stride_u, v, stride_v, crop_width, inv_crop_height);
       break;
+    }
     case FOURCC_RGBP:
       src = sample + (src_width * crop_y + crop_x) * 2;
-      r = RGB565ToI420(src, src_width * 2,
-                       y, y_stride,
-                       u, u_stride,
-                       v, v_stride,
-                       crop_width, inv_crop_height);
+      r = RGB565ToI420(src, src_width * 2, dst_y, dst_stride_y, dst_u,
+                       dst_stride_u, dst_v, dst_stride_v, crop_width,
+                       inv_crop_height);
       break;
     case FOURCC_RGBO:
       src = sample + (src_width * crop_y + crop_x) * 2;
-      r = ARGB1555ToI420(src, src_width * 2,
-                         y, y_stride,
-                         u, u_stride,
-                         v, v_stride,
-                         crop_width, inv_crop_height);
+      r = ARGB1555ToI420(src, src_width * 2, dst_y, dst_stride_y, dst_u,
+                         dst_stride_u, dst_v, dst_stride_v, crop_width,
+                         inv_crop_height);
       break;
     case FOURCC_R444:
       src = sample + (src_width * crop_y + crop_x) * 2;
-      r = ARGB4444ToI420(src, src_width * 2,
-                         y, y_stride,
-                         u, u_stride,
-                         v, v_stride,
-                         crop_width, inv_crop_height);
+      r = ARGB4444ToI420(src, src_width * 2, dst_y, dst_stride_y, dst_u,
+                         dst_stride_u, dst_v, dst_stride_v, crop_width,
+                         inv_crop_height);
       break;
     case FOURCC_24BG:
       src = sample + (src_width * crop_y + crop_x) * 3;
-      r = RGB24ToI420(src, src_width * 3,
-                      y, y_stride,
-                      u, u_stride,
-                      v, v_stride,
-                      crop_width, inv_crop_height);
+      r = RGB24ToI420(src, src_width * 3, dst_y, dst_stride_y, dst_u,
+                      dst_stride_u, dst_v, dst_stride_v, crop_width,
+                      inv_crop_height);
       break;
     case FOURCC_RAW:
       src = sample + (src_width * crop_y + crop_x) * 3;
-      r = RAWToI420(src, src_width * 3,
-                    y, y_stride,
-                    u, u_stride,
-                    v, v_stride,
-                    crop_width, inv_crop_height);
+      r = RAWToI420(src, src_width * 3, dst_y, dst_stride_y, dst_u,
+                    dst_stride_u, dst_v, dst_stride_v, crop_width,
+                    inv_crop_height);
       break;
     case FOURCC_ARGB:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = ARGBToI420(src, src_width * 4,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = ARGBToI420(src, src_width * 4, dst_y, dst_stride_y, dst_u,
+                     dst_stride_u, dst_v, dst_stride_v, crop_width,
+                     inv_crop_height);
       break;
     case FOURCC_BGRA:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = BGRAToI420(src, src_width * 4,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = BGRAToI420(src, src_width * 4, dst_y, dst_stride_y, dst_u,
+                     dst_stride_u, dst_v, dst_stride_v, crop_width,
+                     inv_crop_height);
       break;
     case FOURCC_ABGR:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = ABGRToI420(src, src_width * 4,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = ABGRToI420(src, src_width * 4, dst_y, dst_stride_y, dst_u,
+                     dst_stride_u, dst_v, dst_stride_v, crop_width,
+                     inv_crop_height);
       break;
     case FOURCC_RGBA:
       src = sample + (src_width * crop_y + crop_x) * 4;
-      r = RGBAToI420(src, src_width * 4,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = RGBAToI420(src, src_width * 4, dst_y, dst_stride_y, dst_u,
+                     dst_stride_u, dst_v, dst_stride_v, crop_width,
+                     inv_crop_height);
       break;
+    // TODO(fbarchard): Add AR30 and AB30
     case FOURCC_I400:
       src = sample + src_width * crop_y + crop_x;
-      r = I400ToI420(src, src_width,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = I400ToI420(src, src_width, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                     dst_v, dst_stride_v, crop_width, inv_crop_height);
       break;
     // Biplanar formats
     case FOURCC_NV12:
       src = sample + (src_width * crop_y + crop_x);
-      src_uv = sample + (src_width * src_height) +
-        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
-      r = NV12ToI420Rotate(src, src_width,
-                           src_uv, aligned_src_width,
-                           y, y_stride,
-                           u, u_stride,
-                           v, v_stride,
-                           crop_width, inv_crop_height, rotation);
+      src_uv = sample + (src_width * abs_src_height) +
+               ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      r = NV12ToI420Rotate(src, src_width, src_uv, aligned_src_width, dst_y,
+                           dst_stride_y, dst_u, dst_stride_u, dst_v,
+                           dst_stride_v, crop_width, inv_crop_height, rotation);
       break;
     case FOURCC_NV21:
       src = sample + (src_width * crop_y + crop_x);
-      src_uv = sample + (src_width * src_height) +
-        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
-      // Call NV12 but with u and v parameters swapped.
-      r = NV12ToI420Rotate(src, src_width,
-                           src_uv, aligned_src_width,
-                           y, y_stride,
-                           v, v_stride,
-                           u, u_stride,
-                           crop_width, inv_crop_height, rotation);
-      break;
-    case FOURCC_M420:
-      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
-      r = M420ToI420(src, src_width,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      src_uv = sample + (src_width * abs_src_height) +
+               ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      // Call NV12 but with dst_u and dst_v parameters swapped.
+      r = NV12ToI420Rotate(src, src_width, src_uv, aligned_src_width, dst_y,
+                           dst_stride_y, dst_v, dst_stride_v, dst_u,
+                           dst_stride_u, crop_width, inv_crop_height, rotation);
       break;
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YV12: {
-      const uint8* src_y = sample + (src_width * crop_y + crop_x);
-      const uint8* src_u;
-      const uint8* src_v;
+      const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8_t* src_u;
+      const uint8_t* src_v;
       int halfwidth = (src_width + 1) / 2;
       int halfheight = (abs_src_height + 1) / 2;
       if (format == FOURCC_YV12) {
-        src_v = sample + src_width * abs_src_height +
-            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height + halfwidth * (crop_y / 2) +
+                (crop_x / 2);
         src_u = sample + src_width * abs_src_height +
-            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+                halfwidth * (halfheight + (crop_y / 2)) + (crop_x / 2);
       } else {
-        src_u = sample + src_width * abs_src_height +
-            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height + halfwidth * (crop_y / 2) +
+                (crop_x / 2);
         src_v = sample + src_width * abs_src_height +
-            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+                halfwidth * (halfheight + (crop_y / 2)) + (crop_x / 2);
       }
-      r = I420Rotate(src_y, src_width,
-                     src_u, halfwidth,
-                     src_v, halfwidth,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height, rotation);
+      r = I420Rotate(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_y, dst_stride_y, dst_u, dst_stride_u, dst_v,
+                     dst_stride_v, crop_width, inv_crop_height, rotation);
       break;
     }
     case FOURCC_I422:
     case FOURCC_YV16: {
-      const uint8* src_y = sample + src_width * crop_y + crop_x;
-      const uint8* src_u;
-      const uint8* src_v;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
       int halfwidth = (src_width + 1) / 2;
       if (format == FOURCC_YV16) {
-        src_v = sample + src_width * abs_src_height +
-            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height + halfwidth * crop_y +
+                (crop_x / 2);
         src_u = sample + src_width * abs_src_height +
-            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+                halfwidth * (abs_src_height + crop_y) + (crop_x / 2);
       } else {
-        src_u = sample + src_width * abs_src_height +
-            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height + halfwidth * crop_y +
+                (crop_x / 2);
         src_v = sample + src_width * abs_src_height +
-            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+                halfwidth * (abs_src_height + crop_y) + (crop_x / 2);
       }
-      r = I422ToI420(src_y, src_width,
-                     src_u, halfwidth,
-                     src_v, halfwidth,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = I422ToI420(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_y, dst_stride_y, dst_u, dst_stride_u, dst_v,
+                     dst_stride_v, crop_width, inv_crop_height);
       break;
     }
     case FOURCC_I444:
     case FOURCC_YV24: {
-      const uint8* src_y = sample + src_width * crop_y + crop_x;
-      const uint8* src_u;
-      const uint8* src_v;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
       if (format == FOURCC_YV24) {
         src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
         src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
@@ -277,38 +245,16 @@ int ConvertToI420(const uint8* sample,
         src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
         src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
       }
-      r = I444ToI420(src_y, src_width,
-                     src_u, src_width,
-                     src_v, src_width,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
-      break;
-    }
-    case FOURCC_I411: {
-      int quarterwidth = (src_width + 3) / 4;
-      const uint8* src_y = sample + src_width * crop_y + crop_x;
-      const uint8* src_u = sample + src_width * abs_src_height +
-          quarterwidth * crop_y + crop_x / 4;
-      const uint8* src_v = sample + src_width * abs_src_height +
-          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
-      r = I411ToI420(src_y, src_width,
-                     src_u, quarterwidth,
-                     src_v, quarterwidth,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     crop_width, inv_crop_height);
+      r = I444ToI420(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_y, dst_stride_y, dst_u, dst_stride_u, dst_v,
+                     dst_stride_v, crop_width, inv_crop_height);
       break;
     }
 #ifdef HAVE_JPEG
     case FOURCC_MJPG:
-      r = MJPGToI420(sample, sample_size,
-                     y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     src_width, abs_src_height, crop_width, inv_crop_height);
+      r = MJPGToI420(sample, sample_size, dst_y, dst_stride_y, dst_u,
+                     dst_stride_u, dst_v, dst_stride_v, src_width,
+                     abs_src_height, crop_width, inv_crop_height);
       break;
 #endif
     default:
@@ -317,13 +263,10 @@ int ConvertToI420(const uint8* sample,
 
   if (need_buf) {
     if (!r) {
-      r = I420Rotate(y, y_stride,
-                     u, u_stride,
-                     v, v_stride,
-                     tmp_y, tmp_y_stride,
-                     tmp_u, tmp_u_stride,
-                     tmp_v, tmp_v_stride,
-                     crop_width, abs_crop_height, rotation);
+      r = I420Rotate(dst_y, dst_stride_y, dst_u, dst_stride_u, dst_v,
+                     dst_stride_v, tmp_y, tmp_y_stride, tmp_u, tmp_u_stride,
+                     tmp_v, tmp_v_stride, crop_width, abs_crop_height,
+                     rotation);
     }
     free(rotate_buffer);
   }
diff --git a/media/libyuv/source/cpu_id.cc b/media/libyuv/source/cpu_id.cc
index 84927ebc3e..13e3da7bb1 100644
--- a/media/libyuv/source/cpu_id.cc
+++ b/media/libyuv/source/cpu_id.cc
@@ -13,22 +13,16 @@
 #if defined(_MSC_VER)
 #include <intrin.h>  // For __cpuidex()
 #endif
-#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+#if !defined(__pnacl__) && !defined(__CLR_VER) &&                           \
     !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \
     defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
 #include <immintrin.h>  // For _xgetbv()
 #endif
 
-#if !defined(__native_client__)
-#include <stdlib.h>  // For getenv()
-#endif
-
 // For ArmCpuCaps() but unittested on all platforms
-#include <stdio.h>
+#include <stdio.h>  // For fopen()
 #include <string.h>
 
-#include "libyuv/basic_types.h"  // For CPU_X86
-
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
@@ -43,16 +37,20 @@ extern "C" {
 #define SAFEBUFFERS
 #endif
 
+// cpu_info_ variable for SIMD instruction sets detected.
+LIBYUV_API int cpu_info_ = 0;
+
+// TODO(fbarchard): Consider using int for cpuid so casting is not needed.
 // Low level cpuid for X86.
-#if (defined(_M_IX86) || defined(_M_X64) || \
-    defined(__i386__) || defined(__x86_64__)) && \
+#if (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \
+     defined(__x86_64__)) &&                                     \
     !defined(__pnacl__) && !defined(__CLR_VER)
 LIBYUV_API
-void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
+void CpuId(int info_eax, int info_ecx, int* cpu_info) {
 #if defined(_MSC_VER)
 // Visual C version uses intrinsic or inline x86 assembly.
 #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
-  __cpuidex((int*)(cpu_info), info_eax, info_ecx);
+  __cpuidex(cpu_info, info_eax, info_ecx);
 #elif defined(_M_IX86)
   __asm {
     mov        eax, info_eax
@@ -66,26 +64,26 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
   }
 #else  // Visual C but not x86
   if (info_ecx == 0) {
-    __cpuid((int*)(cpu_info), info_eax);
+    __cpuid(cpu_info, info_eax);
   } else {
-    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0u;
   }
 #endif
 // GCC version uses inline x86 assembly.
 #else  // defined(_MSC_VER)
-  uint32 info_ebx, info_edx;
-  asm volatile (
-#if defined( __i386__) && defined(__PIC__)
-    // Preserve ebx for fpic 32 bit.
-    "mov %%ebx, %%edi                          \n"
-    "cpuid                                     \n"
-    "xchg %%edi, %%ebx                         \n"
-    : "=D" (info_ebx),
+  int info_ebx, info_edx;
+  asm volatile(
+#if defined(__i386__) && defined(__PIC__)
+      // Preserve ebx for fpic 32 bit.
+      "mov         %%ebx, %%edi                  \n"
+      "cpuid                                     \n"
+      "xchg        %%edi, %%ebx                  \n"
+      : "=D"(info_ebx),
 #else
-    "cpuid                                     \n"
-    : "=b" (info_ebx),
+      "cpuid                                     \n"
+      : "=b"(info_ebx),
 #endif  //  defined( __i386__) && defined(__PIC__)
-      "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
+        "+a"(info_eax), "+c"(info_ecx), "=d"(info_edx));
   cpu_info[0] = info_eax;
   cpu_info[1] = info_ebx;
   cpu_info[2] = info_ecx;
@@ -94,7 +92,9 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
 }
 #else  // (defined(_M_IX86) || defined(_M_X64) ...
 LIBYUV_API
-void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
+void CpuId(int eax, int ecx, int* cpu_info) {
+  (void)eax;
+  (void)ecx;
   cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
 }
 #endif
@@ -108,33 +108,34 @@ void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
 //  }
 // For VS2013 and earlier 32 bit, the _xgetbv(0) optimizer produces bad code.
 // https://code.google.com/p/libyuv/issues/detail?id=529
-#if defined(_M_IX86) && (_MSC_VER < 1900)
+#if defined(_M_IX86) && defined(_MSC_VER) && (_MSC_VER < 1900)
 #pragma optimize("g", off)
 #endif
-#if (defined(_M_IX86) || defined(_M_X64) || \
-    defined(__i386__) || defined(__x86_64__)) && \
+#if (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \
+     defined(__x86_64__)) &&                                     \
     !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__)
-#define HAS_XGETBV
 // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
-int GetXCR0() {
-  uint32 xcr0 = 0u;
+static int GetXCR0() {
+  int xcr0 = 0;
 #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
-  xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+  xcr0 = (int)_xgetbv(0);  // VS2010 SP1 required.  NOLINT
 #elif defined(__i386__) || defined(__x86_64__)
-  asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a"(xcr0) : "c"(0) : "%edx");
 #endif  // defined(__i386__) || defined(__x86_64__)
   return xcr0;
 }
+#else
+// xgetbv unavailable to query for OSSave support.  Return 0.
+#define GetXCR0() 0
 #endif  // defined(_M_IX86) || defined(_M_X64) ..
 // Return optimization to previous setting.
-#if defined(_M_IX86) && (_MSC_VER < 1900)
+#if defined(_M_IX86) && defined(_MSC_VER) && (_MSC_VER < 1900)
 #pragma optimize("g", on)
 #endif
 
-// based on libvpx arm_cpudetect.c
+// Based on libvpx arm_cpudetect.c
 // For Arm, but public to allow testing on any CPU
-LIBYUV_API SAFEBUFFERS
-int ArmCpuCaps(const char* cpuinfo_name) {
+LIBYUV_API SAFEBUFFERS int ArmCpuCaps(const char* cpuinfo_name) {
   char cpuinfo_line[512];
   FILE* f = fopen(cpuinfo_name, "r");
   if (!f) {
@@ -151,7 +152,7 @@ int ArmCpuCaps(const char* cpuinfo_name) {
       }
       // aarch64 uses asimd for Neon.
       p = strstr(cpuinfo_line, " asimd");
-      if (p && (p[6] == ' ' || p[6] == '\n')) {
+      if (p) {
         fclose(f);
         return kCpuHasNEON;
       }
@@ -161,103 +162,102 @@ int ArmCpuCaps(const char* cpuinfo_name) {
   return 0;
 }
 
-// CPU detect function for SIMD instruction sets.
-LIBYUV_API
-int cpu_info_ = 0;  // cpu_info is not initialized yet.
-
-// Test environment variable for disabling CPU features. Any non-zero value
-// to disable. Zero ignored to make it easy to set the variable on/off.
-#if !defined(__native_client__) && !defined(_M_ARM)
-
-static LIBYUV_BOOL TestEnv(const char* name) {
-  const char* var = getenv(name);
-  if (var) {
-    if (var[0] != '0') {
-      return LIBYUV_TRUE;
+// TODO(fbarchard): Consider read_msa_ir().
+LIBYUV_API SAFEBUFFERS int MipsCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  int flag = 0x0;
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume nothing if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return 0;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "cpu model", 9) == 0) {
+      // Workaround early kernel without MSA in ASEs line.
+      if (strstr(cpuinfo_line, "Loongson-2K")) {
+        flag |= kCpuHasMSA;
+      }
+    }
+    if (memcmp(cpuinfo_line, "ASEs implemented", 16) == 0) {
+      if (strstr(cpuinfo_line, "msa")) {
+        flag |= kCpuHasMSA;
+      }
+      // ASEs is the last line, so we can break here.
+      break;
     }
   }
-  return LIBYUV_FALSE;
+  fclose(f);
+  return flag;
 }
-#else  // nacl does not support getenv().
-static LIBYUV_BOOL TestEnv(const char*) {
-  return LIBYUV_FALSE;
+
+// TODO(fbarchard): Consider read_loongarch_ir().
+#define LOONGARCH_CFG2 0x2
+#define LOONGARCH_CFG2_LSX (1 << 6)
+#define LOONGARCH_CFG2_LASX (1 << 7)
+
+#if defined(__loongarch__)
+LIBYUV_API SAFEBUFFERS int LoongarchCpuCaps(void) {
+  int flag = 0x0;
+  uint32_t cfg2 = 0;
+
+  __asm__ volatile("cpucfg %0, %1 \n\t" : "+&r"(cfg2) : "r"(LOONGARCH_CFG2));
+
+  if (cfg2 & LOONGARCH_CFG2_LSX)
+    flag |= kCpuHasLSX;
+
+  if (cfg2 & LOONGARCH_CFG2_LASX)
+    flag |= kCpuHasLASX;
+  return flag;
 }
 #endif
 
-LIBYUV_API SAFEBUFFERS
-int InitCpuFlags(void) {
-  // TODO(fbarchard): swap kCpuInit logic so 0 means uninitialized.
+static SAFEBUFFERS int GetCpuFlags(void) {
   int cpu_info = 0;
-#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
-  uint32 cpu_info0[4] = { 0, 0, 0, 0 };
-  uint32 cpu_info1[4] = { 0, 0, 0, 0 };
-  uint32 cpu_info7[4] = { 0, 0, 0, 0 };
+#if !defined(__pnacl__) && !defined(__CLR_VER) &&                   \
+    (defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \
+     defined(_M_IX86))
+  int cpu_info0[4] = {0, 0, 0, 0};
+  int cpu_info1[4] = {0, 0, 0, 0};
+  int cpu_info7[4] = {0, 0, 0, 0};
   CpuId(0, 0, cpu_info0);
   CpuId(1, 0, cpu_info1);
   if (cpu_info0[0] >= 7) {
     CpuId(7, 0, cpu_info7);
   }
-  cpu_info = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+  cpu_info = kCpuHasX86 | ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
-             ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
-             ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
-             kCpuHasX86;
+             ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0);
 
-#ifdef HAS_XGETBV
-  // AVX requires CPU has AVX, XSAVE and OSXSave for xgetbv
+  // AVX requires OS saves YMM registers.
   if (((cpu_info1[2] & 0x1c000000) == 0x1c000000) &&  // AVX and OSXSave
       ((GetXCR0() & 6) == 6)) {  // Test OS saves YMM registers
-    cpu_info |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | kCpuHasAVX;
+    cpu_info |= kCpuHasAVX | ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+                ((cpu_info1[2] & 0x20000000) ? kCpuHasF16C : 0);
 
     // Detect AVX512bw
     if ((GetXCR0() & 0xe0) == 0xe0) {
-      cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX3 : 0;
+      cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX512BW : 0;
+      cpu_info |= (cpu_info7[1] & 0x80000000) ? kCpuHasAVX512VL : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000002) ? kCpuHasAVX512VBMI : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000040) ? kCpuHasAVX512VBMI2 : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000800) ? kCpuHasAVX512VNNI : 0;
+      cpu_info |= (cpu_info7[2] & 0x00001000) ? kCpuHasAVX512VBITALG : 0;
+      cpu_info |= (cpu_info7[2] & 0x00004000) ? kCpuHasAVX512VPOPCNTDQ : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000100) ? kCpuHasGFNI : 0;
     }
   }
 #endif
-
-  // Environment variable overrides for testing.
-  if (TestEnv("LIBYUV_DISABLE_X86")) {
-    cpu_info &= ~kCpuHasX86;
-  }
-  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
-    cpu_info &= ~kCpuHasSSE2;
-  }
-  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
-    cpu_info &= ~kCpuHasSSSE3;
-  }
-  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
-    cpu_info &= ~kCpuHasSSE41;
-  }
-  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
-    cpu_info &= ~kCpuHasSSE42;
-  }
-  if (TestEnv("LIBYUV_DISABLE_AVX")) {
-    cpu_info &= ~kCpuHasAVX;
-  }
-  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
-    cpu_info &= ~kCpuHasAVX2;
-  }
-  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
-    cpu_info &= ~kCpuHasERMS;
-  }
-  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
-    cpu_info &= ~kCpuHasFMA3;
-  }
-  if (TestEnv("LIBYUV_DISABLE_AVX3")) {
-    cpu_info &= ~kCpuHasAVX3;
-  }
-#endif
 #if defined(__mips__) && defined(__linux__)
-#if defined(__mips_dspr2)
-  cpu_info |= kCpuHasDSPR2;
-#endif
+  cpu_info = MipsCpuCaps("/proc/cpuinfo");
   cpu_info |= kCpuHasMIPS;
-  if (getenv("LIBYUV_DISABLE_DSPR2")) {
-    cpu_info &= ~kCpuHasDSPR2;
-  }
+#endif
+#if defined(__loongarch__) && defined(__linux__)
+  cpu_info = LoongarchCpuCaps();
+  cpu_info |= kCpuHasLOONGARCH;
 #endif
 #if defined(__arm__) || defined(__aarch64__)
 // gcc -mfpu=neon defines __ARM_NEON__
@@ -276,22 +276,22 @@ int InitCpuFlags(void) {
   cpu_info = ArmCpuCaps("/proc/cpuinfo");
 #endif
   cpu_info |= kCpuHasARM;
-  if (TestEnv("LIBYUV_DISABLE_NEON")) {
-    cpu_info &= ~kCpuHasNEON;
-  }
 #endif  // __arm__
-  if (TestEnv("LIBYUV_DISABLE_ASM")) {
-    cpu_info = 0;
-  }
-  cpu_info  |= kCpuInitialized;
-  cpu_info_ = cpu_info;
+  cpu_info |= kCpuInitialized;
   return cpu_info;
 }
 
 // Note that use of this function is not thread safe.
 LIBYUV_API
-void MaskCpuFlags(int enable_flags) {
-  cpu_info_ = InitCpuFlags() & enable_flags;
+int MaskCpuFlags(int enable_flags) {
+  int cpu_info = GetCpuFlags() & enable_flags;
+  SetCpuFlags(cpu_info);
+  return cpu_info;
+}
+
+LIBYUV_API
+int InitCpuFlags(void) {
+  return MaskCpuFlags(-1);
 }
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/mjpeg_decoder.cc b/media/libyuv/source/mjpeg_decoder.cc
index 59d969c549..0141da8a1d 100644
--- a/media/libyuv/source/mjpeg_decoder.cc
+++ b/media/libyuv/source/mjpeg_decoder.cc
@@ -21,11 +21,12 @@
 
 #if defined(_MSC_VER)
 // disable warning 4324: structure was padded due to __declspec(align())
-#pragma warning(disable:4324)
+#pragma warning(disable : 4324)
 #endif
 
 #endif
-struct FILE;  // For jpeglib.h.
+
+#include <stdio.h>  // For jpeglib.h.
 
 // C++ build requires extern C for jpeg internals.
 #ifdef __cplusplus
@@ -78,10 +79,8 @@ MJpegDecoder::MJpegDecoder()
   decompress_struct_->err = jpeg_std_error(&error_mgr_->base);
   // Override standard exit()-based error handler.
   error_mgr_->base.error_exit = &ErrorHandler;
-#ifndef DEBUG_MJPEG
   error_mgr_->base.output_message = &OutputHandler;
 #endif
-#endif
   decompress_struct_->client_data = NULL;
   source_mgr_->init_source = &init_source;
   source_mgr_->fill_input_buffer = &fill_input_buffer;
@@ -104,13 +103,13 @@ MJpegDecoder::~MJpegDecoder() {
   DestroyOutputBuffers();
 }
 
-LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
+LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8_t* src, size_t src_len) {
   if (!ValidateJpeg(src, src_len)) {
     return LIBYUV_FALSE;
   }
 
   buf_.data = src;
-  buf_.len = static_cast<int>(src_len);
+  buf_.len = (int)src_len;
   buf_vec_.pos = 0;
   decompress_struct_->client_data = &buf_vec_;
 #ifdef HAVE_SETJMP
@@ -131,7 +130,7 @@ LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
       if (scanlines_[i]) {
         delete scanlines_[i];
       }
-      scanlines_[i] = new uint8* [scanlines_size];
+      scanlines_[i] = new uint8_t*[scanlines_size];
       scanlines_sizes_[i] = scanlines_size;
     }
 
@@ -147,7 +146,7 @@ LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
       if (databuf_[i]) {
         delete databuf_[i];
       }
-      databuf_[i] = new uint8[databuf_size];
+      databuf_[i] = new uint8_t[databuf_size];
       databuf_strides_[i] = databuf_stride;
     }
 
@@ -197,13 +196,11 @@ int MJpegDecoder::GetVertSampFactor(int component) {
 }
 
 int MJpegDecoder::GetHorizSubSampFactor(int component) {
-  return decompress_struct_->max_h_samp_factor /
-      GetHorizSampFactor(component);
+  return decompress_struct_->max_h_samp_factor / GetHorizSampFactor(component);
 }
 
 int MJpegDecoder::GetVertSubSampFactor(int component) {
-  return decompress_struct_->max_v_samp_factor /
-      GetVertSampFactor(component);
+  return decompress_struct_->max_v_samp_factor / GetVertSampFactor(component);
 }
 
 int MJpegDecoder::GetImageScanlinesPerImcuRow() {
@@ -247,10 +244,10 @@ LIBYUV_BOOL MJpegDecoder::UnloadFrame() {
 }
 
 // TODO(fbarchard): Allow rectangle to be specified: x, y, width, height.
-LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
-    uint8** planes, int dst_width, int dst_height) {
-  if (dst_width != GetWidth() ||
-      dst_height > GetHeight()) {
+LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(uint8_t** planes,
+                                          int dst_width,
+                                          int dst_height) {
+  if (dst_width != GetWidth() || dst_height > GetHeight()) {
     // ERROR: Bad dimensions
     return LIBYUV_FALSE;
   }
@@ -291,14 +288,13 @@ LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
       for (int i = 0; i < num_outbufs_; ++i) {
         // TODO(fbarchard): Compute skip to avoid this
         assert(skip % GetVertSubSampFactor(i) == 0);
-        int rows_to_skip =
-            DivideAndRoundDown(skip, GetVertSubSampFactor(i));
-        int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i) -
-                                rows_to_skip;
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int scanlines_to_copy =
+            GetComponentScanlinesPerImcuRow(i) - rows_to_skip;
         int data_to_skip = rows_to_skip * GetComponentStride(i);
-        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i),
-                  planes[i], GetComponentWidth(i),
-                  GetComponentWidth(i), scanlines_to_copy);
+        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i), planes[i],
+                  GetComponentWidth(i), GetComponentWidth(i),
+                  scanlines_to_copy);
         planes[i] += scanlines_to_copy * GetComponentWidth(i);
       }
       lines_left -= (GetImageScanlinesPerImcuRow() - skip);
@@ -307,16 +303,15 @@ LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
 
   // Read full MCUs but cropped horizontally
   for (; lines_left > GetImageScanlinesPerImcuRow();
-         lines_left -= GetImageScanlinesPerImcuRow()) {
+       lines_left -= GetImageScanlinesPerImcuRow()) {
     if (!DecodeImcuRow()) {
       FinishDecode();
       return LIBYUV_FALSE;
     }
     for (int i = 0; i < num_outbufs_; ++i) {
       int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i);
-      CopyPlane(databuf_[i], GetComponentStride(i),
-                planes[i], GetComponentWidth(i),
-                GetComponentWidth(i), scanlines_to_copy);
+      CopyPlane(databuf_[i], GetComponentStride(i), planes[i],
+                GetComponentWidth(i), GetComponentWidth(i), scanlines_to_copy);
       planes[i] += scanlines_to_copy * GetComponentWidth(i);
     }
   }
@@ -330,19 +325,19 @@ LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
     for (int i = 0; i < num_outbufs_; ++i) {
       int scanlines_to_copy =
           DivideAndRoundUp(lines_left, GetVertSubSampFactor(i));
-      CopyPlane(databuf_[i], GetComponentStride(i),
-                planes[i], GetComponentWidth(i),
-                GetComponentWidth(i), scanlines_to_copy);
+      CopyPlane(databuf_[i], GetComponentStride(i), planes[i],
+                GetComponentWidth(i), GetComponentWidth(i), scanlines_to_copy);
       planes[i] += scanlines_to_copy * GetComponentWidth(i);
     }
   }
   return FinishDecode();
 }
 
-LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque,
-    int dst_width, int dst_height) {
-  if (dst_width != GetWidth() ||
-      dst_height > GetHeight()) {
+LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn,
+                                           void* opaque,
+                                           int dst_width,
+                                           int dst_height) {
+  if (dst_width != GetWidth() || dst_height > GetHeight()) {
     // ERROR: Bad dimensions
     return LIBYUV_FALSE;
   }
@@ -397,7 +392,7 @@ LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque,
   }
   // Read full MCUs until we get to the crop point.
   for (; lines_left >= GetImageScanlinesPerImcuRow();
-         lines_left -= GetImageScanlinesPerImcuRow()) {
+       lines_left -= GetImageScanlinesPerImcuRow()) {
     if (!DecodeImcuRow()) {
       FinishDecode();
       return LIBYUV_FALSE;
@@ -422,7 +417,6 @@ void init_source(j_decompress_ptr cinfo) {
 boolean fill_input_buffer(j_decompress_ptr cinfo) {
   BufferVector* buf_vec = reinterpret_cast<BufferVector*>(cinfo->client_data);
   if (buf_vec->pos >= buf_vec->len) {
-    assert(0 && "No more data");
     // ERROR: No more data
     return FALSE;
   }
@@ -433,26 +427,34 @@ boolean fill_input_buffer(j_decompress_ptr cinfo) {
 }
 
 void skip_input_data(j_decompress_ptr cinfo, long num_bytes) {  // NOLINT
-  cinfo->src->next_input_byte += num_bytes;
+  jpeg_source_mgr* src = cinfo->src;
+  size_t bytes = (size_t)num_bytes;
+  if (bytes > src->bytes_in_buffer) {
+    src->next_input_byte = nullptr;
+    src->bytes_in_buffer = 0;
+  } else {
+    src->next_input_byte += bytes;
+    src->bytes_in_buffer -= bytes;
+  }
 }
 
 void term_source(j_decompress_ptr cinfo) {
-  // Nothing to do.
+  (void)cinfo;  // Nothing to do.
 }
 
 #ifdef HAVE_SETJMP
 void ErrorHandler(j_common_ptr cinfo) {
-  // This is called when a jpeglib command experiences an error. Unfortunately
-  // jpeglib's error handling model is not very flexible, because it expects the
-  // error handler to not return--i.e., it wants the program to terminate. To
-  // recover from errors we use setjmp() as shown in their example. setjmp() is
-  // C's implementation for the "call with current continuation" functionality
-  // seen in some functional programming languages.
-  // A formatted message can be output, but is unsafe for release.
+// This is called when a jpeglib command experiences an error. Unfortunately
+// jpeglib's error handling model is not very flexible, because it expects the
+// error handler to not return--i.e., it wants the program to terminate. To
+// recover from errors we use setjmp() as shown in their example. setjmp() is
+// C's implementation for the "call with current continuation" functionality
+// seen in some functional programming languages.
+// A formatted message can be output, but is unsafe for release.
 #ifdef DEBUG
   char buf[JMSG_LENGTH_MAX];
   (*cinfo->err->format_message)(cinfo, buf);
-  // ERROR: Error in jpeglib: buf
+// ERROR: Error in jpeglib: buf
 #endif
 
   SetJmpErrorMgr* mgr = reinterpret_cast<SetJmpErrorMgr*>(cinfo->err);
@@ -461,12 +463,12 @@ void ErrorHandler(j_common_ptr cinfo) {
   longjmp(mgr->setjmp_buffer, 1);
 }
 
-#ifndef DEBUG_MJPEG
+// Suppress fprintf warnings.
 void OutputHandler(j_common_ptr cinfo) {
-  // silently eat messages
+  (void)cinfo;
 }
-#endif
-#endif // HAVE_SETJMP
+
+#endif  // HAVE_SETJMP
 
 void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
   if (num_outbufs != num_outbufs_) {
@@ -475,9 +477,9 @@ void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
     // it.
     DestroyOutputBuffers();
 
-    scanlines_ = new uint8** [num_outbufs];
+    scanlines_ = new uint8_t**[num_outbufs];
     scanlines_sizes_ = new int[num_outbufs];
-    databuf_ = new uint8* [num_outbufs];
+    databuf_ = new uint8_t*[num_outbufs];
     databuf_strides_ = new int[num_outbufs];
 
     for (int i = 0; i < num_outbufs; ++i) {
@@ -493,13 +495,13 @@ void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
 
 void MJpegDecoder::DestroyOutputBuffers() {
   for (int i = 0; i < num_outbufs_; ++i) {
-    delete [] scanlines_[i];
-    delete [] databuf_[i];
+    delete[] scanlines_[i];
+    delete[] databuf_[i];
   }
-  delete [] scanlines_;
-  delete [] databuf_;
-  delete [] scanlines_sizes_;
-  delete [] databuf_strides_;
+  delete[] scanlines_;
+  delete[] databuf_;
+  delete[] scanlines_sizes_;
+  delete[] databuf_strides_;
   scanlines_ = NULL;
   databuf_ = NULL;
   scanlines_sizes_ = NULL;
@@ -533,9 +535,9 @@ LIBYUV_BOOL MJpegDecoder::FinishDecode() {
   return LIBYUV_TRUE;
 }
 
-void MJpegDecoder::SetScanlinePointers(uint8** data) {
+void MJpegDecoder::SetScanlinePointers(uint8_t** data) {
   for (int i = 0; i < num_outbufs_; ++i) {
-    uint8* data_i = data[i];
+    uint8_t* data_i = data[i];
     for (int j = 0; j < scanlines_sizes_[i]; ++j) {
       scanlines_[i][j] = data_i;
       data_i += GetComponentStride(i);
@@ -545,26 +547,26 @@ void MJpegDecoder::SetScanlinePointers(uint8** data) {
 
 inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() {
   return (unsigned int)(GetImageScanlinesPerImcuRow()) ==
-      jpeg_read_raw_data(decompress_struct_,
-                         scanlines_,
-                         GetImageScanlinesPerImcuRow());
+         jpeg_read_raw_data(decompress_struct_, scanlines_,
+                            GetImageScanlinesPerImcuRow());
 }
 
 // The helper function which recognizes the jpeg sub-sampling type.
 JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
-    int* subsample_x, int* subsample_y, int number_of_components) {
+    int* subsample_x,
+    int* subsample_y,
+    int number_of_components) {
   if (number_of_components == 3) {  // Color images.
-    if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
-        subsample_x[1] == 2 && subsample_y[1] == 2 &&
-        subsample_x[2] == 2 && subsample_y[2] == 2) {
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 && subsample_x[1] == 2 &&
+        subsample_y[1] == 2 && subsample_x[2] == 2 && subsample_y[2] == 2) {
       return kJpegYuv420;
-    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
-        subsample_x[1] == 2 && subsample_y[1] == 1 &&
-        subsample_x[2] == 2 && subsample_y[2] == 1) {
+    }
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 && subsample_x[1] == 2 &&
+        subsample_y[1] == 1 && subsample_x[2] == 2 && subsample_y[2] == 1) {
       return kJpegYuv422;
-    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
-        subsample_x[1] == 1 && subsample_y[1] == 1 &&
-        subsample_x[2] == 1 && subsample_y[2] == 1) {
+    }
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 && subsample_x[1] == 1 &&
+        subsample_y[1] == 1 && subsample_x[2] == 1 && subsample_y[2] == 1) {
       return kJpegYuv444;
     }
   } else if (number_of_components == 1) {  // Grey-scale images.
@@ -577,4 +579,3 @@ JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
 
 }  // namespace libyuv
 #endif  // HAVE_JPEG
-
diff --git a/media/libyuv/source/mjpeg_validate.cc b/media/libyuv/source/mjpeg_validate.cc
index 9c48832045..ba0a03ab9e 100644
--- a/media/libyuv/source/mjpeg_validate.cc
+++ b/media/libyuv/source/mjpeg_validate.cc
@@ -18,13 +18,13 @@ extern "C" {
 #endif
 
 // Helper function to scan for EOI marker (0xff 0xd9).
-static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) {
-  if (sample_size >= 2) {
-    const uint8* end = sample + sample_size - 1;
-    const uint8* it = sample;
+static LIBYUV_BOOL ScanEOI(const uint8_t* src_mjpg, size_t src_size_mjpg) {
+  if (src_size_mjpg >= 2) {
+    const uint8_t* end = src_mjpg + src_size_mjpg - 1;
+    const uint8_t* it = src_mjpg;
     while (it < end) {
       // TODO(fbarchard): scan for 0xd9 instead.
-      it = static_cast<const uint8 *>(memchr(it, 0xff, end - it));
+      it = (const uint8_t*)(memchr(it, 0xff, end - it));
       if (it == NULL) {
         break;
       }
@@ -34,38 +34,38 @@ static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) {
       ++it;  // Skip over current 0xff.
     }
   }
-  // ERROR: Invalid jpeg end code not found. Size sample_size
+  // ERROR: Invalid jpeg end code not found. Size src_size_mjpg
   return LIBYUV_FALSE;
 }
 
 // Helper function to validate the jpeg appears intact.
-LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
+LIBYUV_BOOL ValidateJpeg(const uint8_t* src_mjpg, size_t src_size_mjpg) {
   // Maximum size that ValidateJpeg will consider valid.
   const size_t kMaxJpegSize = 0x7fffffffull;
   const size_t kBackSearchSize = 1024;
-  if (sample_size < 64 || sample_size > kMaxJpegSize || !sample) {
-    // ERROR: Invalid jpeg size: sample_size
+  if (src_size_mjpg < 64 || src_size_mjpg > kMaxJpegSize || !src_mjpg) {
+    // ERROR: Invalid jpeg size: src_size_mjpg
     return LIBYUV_FALSE;
   }
-  if (sample[0] != 0xff || sample[1] != 0xd8) {  // SOI marker
+  // SOI marker
+  if (src_mjpg[0] != 0xff || src_mjpg[1] != 0xd8 || src_mjpg[2] != 0xff) {
     // ERROR: Invalid jpeg initial start code
     return LIBYUV_FALSE;
   }
 
   // Look for the End Of Image (EOI) marker near the end of the buffer.
-  if (sample_size > kBackSearchSize) {
-    if (ScanEOI(sample + sample_size - kBackSearchSize, kBackSearchSize)) {
+  if (src_size_mjpg > kBackSearchSize) {
+    if (ScanEOI(src_mjpg + src_size_mjpg - kBackSearchSize, kBackSearchSize)) {
       return LIBYUV_TRUE;  // Success: Valid jpeg.
     }
     // Reduce search size for forward search.
-    sample_size = sample_size - kBackSearchSize + 1;
+    src_size_mjpg = src_size_mjpg - kBackSearchSize + 1;
   }
   // Step over SOI marker and scan for EOI.
-  return ScanEOI(sample + 2, sample_size - 2);
+  return ScanEOI(src_mjpg + 2, src_size_mjpg - 2);
 }
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
-
diff --git a/media/libyuv/source/planar_functions.cc b/media/libyuv/source/planar_functions.cc
index 237ab68315..e3452f58e9 100644
--- a/media/libyuv/source/planar_functions.cc
+++ b/media/libyuv/source/planar_functions.cc
@@ -10,6 +10,7 @@
 
 #include "libyuv/planar_functions.h"
 
+#include <assert.h>
 #include <string.h>  // for memset()
 
 #include "libyuv/cpu_id.h"
@@ -26,14 +27,25 @@ extern "C" {
 
 // Copy a plane of data
 LIBYUV_API
-void CopyPlane(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height) {
+void CopyPlane(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height) {
   int y;
-  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  void (*CopyRow)(const uint8_t* src, uint8_t* dst, int width) = CopyRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      dst_stride_y == width) {
+  if (src_stride_y == width && dst_stride_y == width) {
     width *= height;
     height = 1;
     src_stride_y = dst_stride_y = 0;
@@ -42,6 +54,7 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
   if (src_y == dst_y && src_stride_y == dst_stride_y) {
     return;
   }
+
 #if defined(HAS_COPYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
@@ -62,11 +75,6 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
     CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
   }
 #endif
-#if defined(HAS_COPYROW_MIPS)
-  if (TestCpuFlag(kCpuHasMIPS)) {
-    CopyRow = CopyRow_MIPS;
-  }
-#endif
 
   // Copy plane
   for (y = 0; y < height; ++y) {
@@ -77,42 +85,125 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
 }
 
 LIBYUV_API
-void CopyPlane_16(const uint16* src_y, int src_stride_y,
-                  uint16* dst_y, int dst_stride_y,
-                  int width, int height) {
+void CopyPlane_16(const uint16_t* src_y,
+                  int src_stride_y,
+                  uint16_t* dst_y,
+                  int dst_stride_y,
+                  int width,
+                  int height) {
+  CopyPlane((const uint8_t*)src_y, src_stride_y * 2, (uint8_t*)dst_y,
+            dst_stride_y * 2, width * 2, height);
+}
+
+// Convert a plane of 16 bit data to 8 bit
+LIBYUV_API
+void Convert16To8Plane(const uint16_t* src_y,
+                       int src_stride_y,
+                       uint8_t* dst_y,
+                       int dst_stride_y,
+                       int scale,  // 16384 for 10 bits
+                       int width,
+                       int height) {
   int y;
-  void (*CopyRow)(const uint16* src, uint16* dst, int width) = CopyRow_16_C;
+  void (*Convert16To8Row)(const uint16_t* src_y, uint8_t* dst_y, int scale,
+                          int width) = Convert16To8Row_C;
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      dst_stride_y == width) {
+  if (src_stride_y == width && dst_stride_y == width) {
     width *= height;
     height = 1;
     src_stride_y = dst_stride_y = 0;
   }
-#if defined(HAS_COPYROW_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_16_SSE2;
+#if defined(HAS_CONVERT16TO8ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Convert16To8Row = Convert16To8Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      Convert16To8Row = Convert16To8Row_NEON;
+    }
   }
 #endif
-#if defined(HAS_COPYROW_16_ERMS)
-  if (TestCpuFlag(kCpuHasERMS)) {
-    CopyRow = CopyRow_16_ERMS;
+#if defined(HAS_CONVERT16TO8ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Convert16To8Row = Convert16To8Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      Convert16To8Row = Convert16To8Row_SSSE3;
+    }
   }
 #endif
-#if defined(HAS_COPYROW_16_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_16_NEON;
+#if defined(HAS_CONVERT16TO8ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Convert16To8Row = Convert16To8Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      Convert16To8Row = Convert16To8Row_AVX2;
+    }
+  }
+#endif
+
+  // Convert plane
+  for (y = 0; y < height; ++y) {
+    Convert16To8Row(src_y, dst_y, scale, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert a plane of 8 bit data to 16 bit
+LIBYUV_API
+void Convert8To16Plane(const uint8_t* src_y,
+                       int src_stride_y,
+                       uint16_t* dst_y,
+                       int dst_stride_y,
+                       int scale,  // 1024 for 10 bits
+                       int width,
+                       int height) {
+  int y;
+  void (*Convert8To16Row)(const uint8_t* src_y, uint16_t* dst_y, int scale,
+                          int width) = Convert8To16Row_C;
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_CONVERT8TO16ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Convert8To16Row = Convert8To16Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      Convert8To16Row = Convert8To16Row_SSE2;
+    }
   }
 #endif
-#if defined(HAS_COPYROW_16_MIPS)
-  if (TestCpuFlag(kCpuHasMIPS)) {
-    CopyRow = CopyRow_16_MIPS;
+#if defined(HAS_CONVERT8TO16ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Convert8To16Row = Convert8To16Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      Convert8To16Row = Convert8To16Row_AVX2;
+    }
   }
 #endif
 
-  // Copy plane
+  // Convert plane
   for (y = 0; y < height; ++y) {
-    CopyRow(src_y, dst_y, width);
+    Convert8To16Row(src_y, dst_y, scale, width);
     src_y += src_stride_y;
     dst_y += dst_stride_y;
   }
@@ -120,19 +211,27 @@ void CopyPlane_16(const uint16* src_y, int src_stride_y,
 
 // Copy I422.
 LIBYUV_API
-int I422Copy(const uint8* src_y, int src_stride_y,
-             const uint8* src_u, int src_stride_u,
-             const uint8* src_v, int src_stride_v,
-             uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int width, int height) {
+int I422Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_u,
+             int src_stride_u,
+             const uint8_t* src_v,
+             int src_stride_v,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
   int halfwidth = (width + 1) >> 1;
-  if (!src_y || !src_u || !src_v ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -143,7 +242,10 @@ int I422Copy(const uint8* src_y, int src_stride_y,
     src_stride_u = -src_stride_u;
     src_stride_v = -src_stride_v;
   }
-  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
   CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
   CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
   return 0;
@@ -151,16 +253,22 @@ int I422Copy(const uint8* src_y, int src_stride_y,
 
 // Copy I444.
 LIBYUV_API
-int I444Copy(const uint8* src_y, int src_stride_y,
-             const uint8* src_u, int src_stride_u,
-             const uint8* src_v, int src_stride_v,
-             uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int width, int height) {
-  if (!src_y || !src_u || !src_v ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+int I444Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_u,
+             int src_stride_u,
+             const uint8_t* src_v,
+             int src_stride_v,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -174,17 +282,104 @@ int I444Copy(const uint8* src_y, int src_stride_y,
     src_stride_v = -src_stride_v;
   }
 
-  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
   CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
   CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
   return 0;
 }
 
+// Copy I210.
+LIBYUV_API
+int I210Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
+  int halfwidth = (width + 1) >> 1;
+
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
+// Copy I410.
+LIBYUV_API
+int I410Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+  CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+  return 0;
+}
+
 // Copy I400.
 LIBYUV_API
-int I400ToI400(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height) {
+int I400ToI400(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height) {
   if (!src_y || !dst_y || width <= 0 || height == 0) {
     return -1;
   }
@@ -200,11 +395,20 @@ int I400ToI400(const uint8* src_y, int src_stride_y,
 
 // Convert I420 to I400.
 LIBYUV_API
-int I420ToI400(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height) {
+int I420ToI400(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height) {
+  (void)src_u;
+  (void)src_stride_u;
+  (void)src_v;
+  (void)src_stride_v;
   if (!src_y || !dst_y || width <= 0 || height == 0) {
     return -1;
   }
@@ -214,76 +418,1605 @@ int I420ToI400(const uint8* src_y, int src_stride_y,
     src_y = src_y + (height - 1) * src_stride_y;
     src_stride_y = -src_stride_y;
   }
+
   CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
   return 0;
 }
 
-// Mirror a plane of data.
-void MirrorPlane(const uint8* src_y, int src_stride_y,
-                 uint8* dst_y, int dst_stride_y,
-                 int width, int height) {
-  int y;
-  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+// Copy NV12. Supports inverting.
+LIBYUV_API
+int NV12Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_uv,
+             int src_stride_uv,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_uv,
+             int dst_stride_uv,
+             int width,
+             int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+
+  if (!src_y || !dst_y || !src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
+    halfheight = (height + 1) >> 1;
     src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
     src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
   }
-#if defined(HAS_MIRRORROW_NEON)
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, halfwidth * 2,
+            halfheight);
+  return 0;
+}
+
+// Copy NV21. Supports inverting.
+LIBYUV_API
+int NV21Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_vu,
+             int src_stride_vu,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_vu,
+             int dst_stride_vu,
+             int width,
+             int height) {
+  return NV12Copy(src_y, src_stride_y, src_vu, src_stride_vu, dst_y,
+                  dst_stride_y, dst_vu, dst_stride_vu, width, height);
+}
+
+// Support function for NV12 etc UV channels.
+// Width and height are plane sizes (typically half pixel width).
+LIBYUV_API
+void SplitUVPlane(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int width,
+                  int height) {
+  int y;
+  void (*SplitUVRow)(const uint8_t* src_uv, uint8_t* dst_u, uint8_t* dst_v,
+                     int width) = SplitUVRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_u = dst_u + (height - 1) * dst_stride_u;
+    dst_v = dst_v + (height - 1) * dst_stride_v;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == width * 2 && dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    MirrorRow = MirrorRow_Any_NEON;
+    SplitUVRow = SplitUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
-      MirrorRow = MirrorRow_NEON;
+      SplitUVRow = SplitUVRow_NEON;
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    MirrorRow = MirrorRow_Any_SSSE3;
+#if defined(HAS_SPLITUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SplitUVRow = SplitUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SplitUVRow = SplitUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_LSX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Copy a row of UV.
+    SplitUVRow(src_uv, dst_u, dst_v, width);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+}
+
+LIBYUV_API
+void MergeUVPlane(const uint8_t* src_u,
+                  int src_stride_u,
+                  const uint8_t* src_v,
+                  int src_stride_v,
+                  uint8_t* dst_uv,
+                  int dst_stride_uv,
+                  int width,
+                  int height) {
+  int y;
+  void (*MergeUVRow)(const uint8_t* src_u, const uint8_t* src_v,
+                     uint8_t* dst_uv, int width) = MergeUVRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uv = dst_uv + (height - 1) * dst_stride_uv;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  // Coalesce rows.
+  if (src_stride_u == width && src_stride_v == width &&
+      dst_stride_uv == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow = MergeUVRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      MirrorRow = MirrorRow_SSSE3;
+      MergeUVRow = MergeUVRow_SSE2;
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_AVX2)
+#if defined(HAS_MERGEUVROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    MirrorRow = MirrorRow_Any_AVX2;
+    MergeUVRow = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    MergeUVRow = MergeUVRow_Any_AVX512BW;
     if (IS_ALIGNED(width, 32)) {
-      MirrorRow = MirrorRow_AVX2;
+      MergeUVRow = MergeUVRow_AVX512BW;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow = MergeUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow = MergeUVRow_MSA;
     }
   }
 #endif
-// TODO(fbarchard): Mirror on mips handle unaligned memory.
-#if defined(HAS_MIRRORROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(dst_y, 4) && IS_ALIGNED(dst_stride_y, 4)) {
-    MirrorRow = MirrorRow_DSPR2;
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow = MergeUVRow_LSX;
+    }
   }
 #endif
 
-  // Mirror plane
   for (y = 0; y < height; ++y) {
-    MirrorRow(src_y, dst_y, width);
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow(src_u, src_v, dst_uv, width);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+}
+
+// Support function for P010 etc UV channels.
+// Width and height are plane sizes (typically half pixel width).
+LIBYUV_API
+void SplitUVPlane_16(const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint16_t* dst_u,
+                     int dst_stride_u,
+                     uint16_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
+                     int depth) {
+  int y;
+  void (*SplitUVRow_16)(const uint16_t* src_uv, uint16_t* dst_u,
+                        uint16_t* dst_v, int depth, int width) =
+      SplitUVRow_16_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_u = dst_u + (height - 1) * dst_stride_u;
+    dst_v = dst_v + (height - 1) * dst_stride_v;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == width * 2 && dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_SPLITUVROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow_16 = SplitUVRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow_16 = SplitUVRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow_16 = SplitUVRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      SplitUVRow_16 = SplitUVRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Copy a row of UV.
+    SplitUVRow_16(src_uv, dst_u, dst_v, depth, width);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+}
+
+LIBYUV_API
+void MergeUVPlane_16(const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint16_t* dst_uv,
+                     int dst_stride_uv,
+                     int width,
+                     int height,
+                     int depth) {
+  int y;
+  void (*MergeUVRow_16)(const uint16_t* src_u, const uint16_t* src_v,
+                        uint16_t* dst_uv, int depth, int width) =
+      MergeUVRow_16_C;
+  assert(depth >= 8);
+  assert(depth <= 16);
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uv = dst_uv + (height - 1) * dst_stride_uv;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  // Coalesce rows.
+  if (src_stride_u == width && src_stride_v == width &&
+      dst_stride_uv == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+#if defined(HAS_MERGEUVROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_16 = MergeUVRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      MergeUVRow_16 = MergeUVRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_16 = MergeUVRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeUVRow_16 = MergeUVRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow_16(src_u, src_v, dst_uv, depth, width);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+}
+
+// Convert plane from lsb to msb
+LIBYUV_API
+void ConvertToMSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth) {
+  int y;
+  int scale = 1 << (16 - depth);
+  void (*MultiplyRow_16)(const uint16_t* src_y, uint16_t* dst_y, int scale,
+                         int width) = MultiplyRow_16_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+
+#if defined(HAS_MULTIPLYROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MultiplyRow_16 = MultiplyRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MultiplyRow_16 = MultiplyRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MULTIPLYROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MultiplyRow_16 = MultiplyRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MultiplyRow_16 = MultiplyRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MultiplyRow_16(src_y, dst_y, scale, width);
     src_y += src_stride_y;
     dst_y += dst_stride_y;
   }
 }
 
+// Convert plane from msb to lsb
+LIBYUV_API
+void ConvertToLSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth) {
+  int y;
+  int scale = 1 << depth;
+  void (*DivideRow)(const uint16_t* src_y, uint16_t* dst_y, int scale,
+                    int width) = DivideRow_16_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+
+#if defined(HAS_DIVIDEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    DivideRow = DivideRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      DivideRow = DivideRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_DIVIDEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DivideRow = DivideRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DivideRow = DivideRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    DivideRow(src_y, dst_y, scale, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Swap U and V channels in interleaved UV plane.
+LIBYUV_API
+void SwapUVPlane(const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_vu,
+                 int dst_stride_vu,
+                 int width,
+                 int height) {
+  int y;
+  void (*SwapUVRow)(const uint8_t* src_uv, uint8_t* dst_vu, int width) =
+      SwapUVRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == width * 2 && dst_stride_vu == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_uv = dst_stride_vu = 0;
+  }
+
+#if defined(HAS_SWAPUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SwapUVRow = SwapUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      SwapUVRow = SwapUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SWAPUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SwapUVRow = SwapUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SwapUVRow = SwapUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SWAPUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SwapUVRow = SwapUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SwapUVRow = SwapUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    SwapUVRow(src_uv, dst_vu, width);
+    src_uv += src_stride_uv;
+    dst_vu += dst_stride_vu;
+  }
+}
+
+// Convert NV21 to NV12.
+LIBYUV_API
+int NV21ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_vu,
+               int src_stride_vu,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+
+  if (!src_vu || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_vu = src_vu + (halfheight - 1) * src_stride_vu;
+    src_stride_vu = -src_stride_vu;
+  }
+
+  SwapUVPlane(src_vu, src_stride_vu, dst_uv, dst_stride_uv, halfwidth,
+              halfheight);
+  return 0;
+}
+
+// Test if tile_height is a power of 2 (16 or 32)
+#define IS_POWEROFTWO(x) (!((x) & ((x)-1)))
+
+// Detile a plane of data
+// tile width is 16 and assumed.
+// tile_height is 16 or 32 for MM21.
+// src_stride_y is bytes per row of source ignoring tiling. e.g. 640
+// TODO: More detile row functions.
+LIBYUV_API
+int DetilePlane(const uint8_t* src_y,
+                int src_stride_y,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                int width,
+                int height,
+                int tile_height) {
+  const ptrdiff_t src_tile_stride = 16 * tile_height;
+  int y;
+  void (*DetileRow)(const uint8_t* src, ptrdiff_t src_tile_stride, uint8_t* dst,
+                    int width) = DetileRow_C;
+  if (!src_y || !dst_y || width <= 0 || height == 0 ||
+      !IS_POWEROFTWO(tile_height)) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+
+#if defined(HAS_DETILEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    DetileRow = DetileRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow = DetileRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_DETILEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DetileRow = DetileRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow = DetileRow_NEON;
+    }
+  }
+#endif
+
+  // Detile plane
+  for (y = 0; y < height; ++y) {
+    DetileRow(src_y, src_tile_stride, dst_y, width);
+    dst_y += dst_stride_y;
+    src_y += 16;
+    // Advance to next row of tiles.
+    if ((y & (tile_height - 1)) == (tile_height - 1)) {
+      src_y = src_y - src_tile_stride + src_stride_y * tile_height;
+    }
+  }
+  return 0;
+}
+
+// Convert a plane of 16 bit tiles of 16 x H to linear.
+// tile width is 16 and assumed.
+// tile_height is 16 or 32 for MT2T.
+LIBYUV_API
+int DetilePlane_16(const uint16_t* src_y,
+                   int src_stride_y,
+                   uint16_t* dst_y,
+                   int dst_stride_y,
+                   int width,
+                   int height,
+                   int tile_height) {
+  const ptrdiff_t src_tile_stride = 16 * tile_height;
+  int y;
+  void (*DetileRow_16)(const uint16_t* src, ptrdiff_t src_tile_stride,
+                       uint16_t* dst, int width) = DetileRow_16_C;
+  if (!src_y || !dst_y || width <= 0 || height == 0 ||
+      !IS_POWEROFTWO(tile_height)) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+
+#if defined(HAS_DETILEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    DetileRow_16 = DetileRow_16_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow_16 = DetileRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_DETILEROW_16_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    DetileRow_16 = DetileRow_16_Any_AVX;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow_16 = DetileRow_16_AVX;
+    }
+  }
+#endif
+#if defined(HAS_DETILEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DetileRow_16 = DetileRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow_16 = DetileRow_16_NEON;
+    }
+  }
+#endif
+
+  // Detile plane
+  for (y = 0; y < height; ++y) {
+    DetileRow_16(src_y, src_tile_stride, dst_y, width);
+    dst_y += dst_stride_y;
+    src_y += 16;
+    // Advance to next row of tiles.
+    if ((y & (tile_height - 1)) == (tile_height - 1)) {
+      src_y = src_y - src_tile_stride + src_stride_y * tile_height;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+void DetileSplitUVPlane(const uint8_t* src_uv,
+                        int src_stride_uv,
+                        uint8_t* dst_u,
+                        int dst_stride_u,
+                        uint8_t* dst_v,
+                        int dst_stride_v,
+                        int width,
+                        int height,
+                        int tile_height) {
+  const ptrdiff_t src_tile_stride = 16 * tile_height;
+  int y;
+  void (*DetileSplitUVRow)(const uint8_t* src, ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u, uint8_t* dst_v, int width) =
+      DetileSplitUVRow_C;
+  assert(src_stride_uv >= 0);
+  assert(tile_height > 0);
+  assert(src_stride_uv > 0);
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_u = dst_u + (height - 1) * dst_stride_u;
+    dst_stride_u = -dst_stride_u;
+    dst_v = dst_v + (height - 1) * dst_stride_v;
+    dst_stride_v = -dst_stride_v;
+  }
+
+#if defined(HAS_DETILESPLITUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    DetileSplitUVRow = DetileSplitUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      DetileSplitUVRow = DetileSplitUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_DETILESPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DetileSplitUVRow = DetileSplitUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DetileSplitUVRow = DetileSplitUVRow_NEON;
+    }
+  }
+#endif
+
+  // Detile plane
+  for (y = 0; y < height; ++y) {
+    DetileSplitUVRow(src_uv, src_tile_stride, dst_u, dst_v, width);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += 16;
+    // Advance to next row of tiles.
+    if ((y & (tile_height - 1)) == (tile_height - 1)) {
+      src_uv = src_uv - src_tile_stride + src_stride_uv * tile_height;
+    }
+  }
+}
+
+LIBYUV_API
+void DetileToYUY2(const uint8_t* src_y,
+                  int src_stride_y,
+                  const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_yuy2,
+                  int dst_stride_yuy2,
+                  int width,
+                  int height,
+                  int tile_height) {
+  const ptrdiff_t src_y_tile_stride = 16 * tile_height;
+  const ptrdiff_t src_uv_tile_stride = src_y_tile_stride / 2;
+  int y;
+  void (*DetileToYUY2)(const uint8_t* src_y, ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv, ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2, int width) = DetileToYUY2_C;
+  assert(src_stride_y >= 0);
+  assert(src_stride_y > 0);
+  assert(src_stride_uv >= 0);
+  assert(src_stride_uv > 0);
+  assert(tile_height > 0);
+
+  if (width <= 0 || height == 0 || tile_height <= 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+
+#if defined(HAS_DETILETOYUY2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DetileToYUY2 = DetileToYUY2_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DetileToYUY2 = DetileToYUY2_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_DETILETOYUY2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    DetileToYUY2 = DetileToYUY2_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      DetileToYUY2 = DetileToYUY2_SSE2;
+    }
+  }
+#endif
+
+  // Detile plane
+  for (y = 0; y < height; ++y) {
+    DetileToYUY2(src_y, src_y_tile_stride, src_uv, src_uv_tile_stride, dst_yuy2,
+                 width);
+    dst_yuy2 += dst_stride_yuy2;
+    src_y += 16;
+
+    if (y & 0x1)
+      src_uv += 16;
+
+    // Advance to next row of tiles.
+    if ((y & (tile_height - 1)) == (tile_height - 1)) {
+      src_y = src_y - src_y_tile_stride + src_stride_y * tile_height;
+      src_uv = src_uv - src_uv_tile_stride + src_stride_uv * (tile_height / 2);
+    }
+  }
+}
+
+// Support function for NV12 etc RGB channels.
+// Width and height are plane sizes (typically half pixel width).
+LIBYUV_API
+void SplitRGBPlane(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_r,
+                   int dst_stride_r,
+                   uint8_t* dst_g,
+                   int dst_stride_g,
+                   uint8_t* dst_b,
+                   int dst_stride_b,
+                   int width,
+                   int height) {
+  int y;
+  void (*SplitRGBRow)(const uint8_t* src_rgb, uint8_t* dst_r, uint8_t* dst_g,
+                      uint8_t* dst_b, int width) = SplitRGBRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_r = dst_r + (height - 1) * dst_stride_r;
+    dst_g = dst_g + (height - 1) * dst_stride_g;
+    dst_b = dst_b + (height - 1) * dst_stride_b;
+    dst_stride_r = -dst_stride_r;
+    dst_stride_g = -dst_stride_g;
+    dst_stride_b = -dst_stride_b;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb == width * 3 && dst_stride_r == width &&
+      dst_stride_g == width && dst_stride_b == width) {
+    width *= height;
+    height = 1;
+    src_stride_rgb = dst_stride_r = dst_stride_g = dst_stride_b = 0;
+  }
+#if defined(HAS_SPLITRGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SplitRGBRow = SplitRGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      SplitRGBRow = SplitRGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SPLITRGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitRGBRow = SplitRGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitRGBRow = SplitRGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Copy a row of RGB.
+    SplitRGBRow(src_rgb, dst_r, dst_g, dst_b, width);
+    dst_r += dst_stride_r;
+    dst_g += dst_stride_g;
+    dst_b += dst_stride_b;
+    src_rgb += src_stride_rgb;
+  }
+}
+
+LIBYUV_API
+void MergeRGBPlane(const uint8_t* src_r,
+                   int src_stride_r,
+                   const uint8_t* src_g,
+                   int src_stride_g,
+                   const uint8_t* src_b,
+                   int src_stride_b,
+                   uint8_t* dst_rgb,
+                   int dst_stride_rgb,
+                   int width,
+                   int height) {
+  int y;
+  void (*MergeRGBRow)(const uint8_t* src_r, const uint8_t* src_g,
+                      const uint8_t* src_b, uint8_t* dst_rgb, int width) =
+      MergeRGBRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Coalesce rows.
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb = dst_rgb + (height - 1) * dst_stride_rgb;
+    dst_stride_rgb = -dst_stride_rgb;
+  }
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_rgb == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_rgb = 0;
+  }
+#if defined(HAS_MERGERGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MergeRGBRow = MergeRGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MergeRGBRow = MergeRGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MERGERGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeRGBRow = MergeRGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MergeRGBRow = MergeRGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Merge a row of U and V into a row of RGB.
+    MergeRGBRow(src_r, src_g, src_b, dst_rgb, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_rgb += dst_stride_rgb;
+  }
+}
+
+LIBYUV_NOINLINE
+static void SplitARGBPlaneAlpha(const uint8_t* src_argb,
+                                int src_stride_argb,
+                                uint8_t* dst_r,
+                                int dst_stride_r,
+                                uint8_t* dst_g,
+                                int dst_stride_g,
+                                uint8_t* dst_b,
+                                int dst_stride_b,
+                                uint8_t* dst_a,
+                                int dst_stride_a,
+                                int width,
+                                int height) {
+  int y;
+  void (*SplitARGBRow)(const uint8_t* src_rgb, uint8_t* dst_r, uint8_t* dst_g,
+                       uint8_t* dst_b, uint8_t* dst_a, int width) =
+      SplitARGBRow_C;
+
+  assert(height > 0);
+
+  if (src_stride_argb == width * 4 && dst_stride_r == width &&
+      dst_stride_g == width && dst_stride_b == width && dst_stride_a == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_r = dst_stride_g = dst_stride_b =
+        dst_stride_a = 0;
+  }
+
+#if defined(HAS_SPLITARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitARGBRow = SplitARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      SplitARGBRow = SplitARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SplitARGBRow = SplitARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      SplitARGBRow = SplitARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SPLITARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitARGBRow = SplitARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitARGBRow = SplitARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitARGBRow = SplitARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitARGBRow = SplitARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    SplitARGBRow(src_argb, dst_r, dst_g, dst_b, dst_a, width);
+    dst_r += dst_stride_r;
+    dst_g += dst_stride_g;
+    dst_b += dst_stride_b;
+    dst_a += dst_stride_a;
+    src_argb += src_stride_argb;
+  }
+}
+
+LIBYUV_NOINLINE
+static void SplitARGBPlaneOpaque(const uint8_t* src_argb,
+                                 int src_stride_argb,
+                                 uint8_t* dst_r,
+                                 int dst_stride_r,
+                                 uint8_t* dst_g,
+                                 int dst_stride_g,
+                                 uint8_t* dst_b,
+                                 int dst_stride_b,
+                                 int width,
+                                 int height) {
+  int y;
+  void (*SplitXRGBRow)(const uint8_t* src_rgb, uint8_t* dst_r, uint8_t* dst_g,
+                       uint8_t* dst_b, int width) = SplitXRGBRow_C;
+  assert(height > 0);
+
+  if (src_stride_argb == width * 4 && dst_stride_r == width &&
+      dst_stride_g == width && dst_stride_b == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_r = dst_stride_g = dst_stride_b = 0;
+  }
+
+#if defined(HAS_SPLITXRGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitXRGBRow = SplitXRGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      SplitXRGBRow = SplitXRGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITXRGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SplitXRGBRow = SplitXRGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      SplitXRGBRow = SplitXRGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SPLITXRGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitXRGBRow = SplitXRGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitXRGBRow = SplitXRGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITXRGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitXRGBRow = SplitXRGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitXRGBRow = SplitXRGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    SplitXRGBRow(src_argb, dst_r, dst_g, dst_b, width);
+    dst_r += dst_stride_r;
+    dst_g += dst_stride_g;
+    dst_b += dst_stride_b;
+    src_argb += src_stride_argb;
+  }
+}
+
+LIBYUV_API
+void SplitARGBPlane(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_r,
+                    int dst_stride_r,
+                    uint8_t* dst_g,
+                    int dst_stride_g,
+                    uint8_t* dst_b,
+                    int dst_stride_b,
+                    uint8_t* dst_a,
+                    int dst_stride_a,
+                    int width,
+                    int height) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_r = dst_r + (height - 1) * dst_stride_r;
+    dst_g = dst_g + (height - 1) * dst_stride_g;
+    dst_b = dst_b + (height - 1) * dst_stride_b;
+    dst_a = dst_a + (height - 1) * dst_stride_a;
+    dst_stride_r = -dst_stride_r;
+    dst_stride_g = -dst_stride_g;
+    dst_stride_b = -dst_stride_b;
+    dst_stride_a = -dst_stride_a;
+  }
+
+  if (dst_a == NULL) {
+    SplitARGBPlaneOpaque(src_argb, src_stride_argb, dst_r, dst_stride_r, dst_g,
+                         dst_stride_g, dst_b, dst_stride_b, width, height);
+  } else {
+    SplitARGBPlaneAlpha(src_argb, src_stride_argb, dst_r, dst_stride_r, dst_g,
+                        dst_stride_g, dst_b, dst_stride_b, dst_a, dst_stride_a,
+                        width, height);
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeARGBPlaneAlpha(const uint8_t* src_r,
+                                int src_stride_r,
+                                const uint8_t* src_g,
+                                int src_stride_g,
+                                const uint8_t* src_b,
+                                int src_stride_b,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                int width,
+                                int height) {
+  int y;
+  void (*MergeARGBRow)(const uint8_t* src_r, const uint8_t* src_g,
+                       const uint8_t* src_b, const uint8_t* src_a,
+                       uint8_t* dst_argb, int width) = MergeARGBRow_C;
+
+  assert(height > 0);
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      src_stride_a == width && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = src_stride_a =
+        dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeARGBRow = MergeARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      MergeARGBRow = MergeARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeARGBRow = MergeARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeARGBRow = MergeARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeARGBRow = MergeARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MergeARGBRow = MergeARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeARGBRow(src_r, src_g, src_b, src_a, dst_argb, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    src_a += src_stride_a;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeARGBPlaneOpaque(const uint8_t* src_r,
+                                 int src_stride_r,
+                                 const uint8_t* src_g,
+                                 int src_stride_g,
+                                 const uint8_t* src_b,
+                                 int src_stride_b,
+                                 uint8_t* dst_argb,
+                                 int dst_stride_argb,
+                                 int width,
+                                 int height) {
+  int y;
+  void (*MergeXRGBRow)(const uint8_t* src_r, const uint8_t* src_g,
+                       const uint8_t* src_b, uint8_t* dst_argb, int width) =
+      MergeXRGBRow_C;
+
+  assert(height > 0);
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEXRGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeXRGBRow = MergeXRGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      MergeXRGBRow = MergeXRGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXRGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXRGBRow = MergeXRGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXRGBRow = MergeXRGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXRGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeXRGBRow = MergeXRGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXRGBRow = MergeXRGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXRGBRow(src_r, src_g, src_b, dst_argb, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_API
+void MergeARGBPlane(const uint8_t* src_r,
+                    int src_stride_r,
+                    const uint8_t* src_g,
+                    int src_stride_g,
+                    const uint8_t* src_b,
+                    int src_stride_b,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+
+  if (src_a == NULL) {
+    MergeARGBPlaneOpaque(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                         src_stride_b, dst_argb, dst_stride_argb, width,
+                         height);
+  } else {
+    MergeARGBPlaneAlpha(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                        src_stride_b, src_a, src_stride_a, dst_argb,
+                        dst_stride_argb, width, height);
+  }
+}
+
+// TODO(yuan): Support 2 bit alpha channel.
+LIBYUV_API
+void MergeXR30Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    uint8_t* dst_ar30,
+                    int dst_stride_ar30,
+                    int width,
+                    int height,
+                    int depth) {
+  int y;
+  void (*MergeXR30Row)(const uint16_t* src_r, const uint16_t* src_g,
+                       const uint16_t* src_b, uint8_t* dst_ar30, int depth,
+                       int width) = MergeXR30Row_C;
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_ar30 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_ar30 = 0;
+  }
+#if defined(HAS_MERGEXR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXR30Row = MergeXR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXR30Row = MergeXR30Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXR30ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (depth == 10) {
+      MergeXR30Row = MergeXR30Row_10_Any_NEON;
+      if (IS_ALIGNED(width, 8)) {
+        MergeXR30Row = MergeXR30Row_10_NEON;
+      }
+    } else {
+      MergeXR30Row = MergeXR30Row_Any_NEON;
+      if (IS_ALIGNED(width, 8)) {
+        MergeXR30Row = MergeXR30Row_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXR30Row(src_r, src_g, src_b, dst_ar30, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_ar30 += dst_stride_ar30;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeAR64PlaneAlpha(const uint16_t* src_r,
+                                int src_stride_r,
+                                const uint16_t* src_g,
+                                int src_stride_g,
+                                const uint16_t* src_b,
+                                int src_stride_b,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint16_t* dst_ar64,
+                                int dst_stride_ar64,
+                                int width,
+                                int height,
+                                int depth) {
+  int y;
+  void (*MergeAR64Row)(const uint16_t* src_r, const uint16_t* src_g,
+                       const uint16_t* src_b, const uint16_t* src_a,
+                       uint16_t* dst_argb, int depth, int width) =
+      MergeAR64Row_C;
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      src_stride_a == width && dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = src_stride_a =
+        dst_stride_ar64 = 0;
+  }
+#if defined(HAS_MERGEAR64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeAR64Row = MergeAR64Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeAR64Row = MergeAR64Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEAR64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeAR64Row = MergeAR64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeAR64Row = MergeAR64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeAR64Row(src_r, src_g, src_b, src_a, dst_ar64, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    src_a += src_stride_a;
+    dst_ar64 += dst_stride_ar64;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeAR64PlaneOpaque(const uint16_t* src_r,
+                                 int src_stride_r,
+                                 const uint16_t* src_g,
+                                 int src_stride_g,
+                                 const uint16_t* src_b,
+                                 int src_stride_b,
+                                 uint16_t* dst_ar64,
+                                 int dst_stride_ar64,
+                                 int width,
+                                 int height,
+                                 int depth) {
+  int y;
+  void (*MergeXR64Row)(const uint16_t* src_r, const uint16_t* src_g,
+                       const uint16_t* src_b, uint16_t* dst_argb, int depth,
+                       int width) = MergeXR64Row_C;
+
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_ar64 = 0;
+  }
+#if defined(HAS_MERGEXR64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXR64Row = MergeXR64Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXR64Row = MergeXR64Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXR64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeXR64Row = MergeXR64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeXR64Row = MergeXR64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXR64Row(src_r, src_g, src_b, dst_ar64, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_ar64 += dst_stride_ar64;
+  }
+}
+
+LIBYUV_API
+void MergeAR64Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    const uint16_t* src_a,
+                    int src_stride_a,
+                    uint16_t* dst_ar64,
+                    int dst_stride_ar64,
+                    int width,
+                    int height,
+                    int depth) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar64 = dst_ar64 + (height - 1) * dst_stride_ar64;
+    dst_stride_ar64 = -dst_stride_ar64;
+  }
+
+  if (src_a == NULL) {
+    MergeAR64PlaneOpaque(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                         src_stride_b, dst_ar64, dst_stride_ar64, width, height,
+                         depth);
+  } else {
+    MergeAR64PlaneAlpha(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                        src_stride_b, src_a, src_stride_a, dst_ar64,
+                        dst_stride_ar64, width, height, depth);
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeARGB16To8PlaneAlpha(const uint16_t* src_r,
+                                     int src_stride_r,
+                                     const uint16_t* src_g,
+                                     int src_stride_g,
+                                     const uint16_t* src_b,
+                                     int src_stride_b,
+                                     const uint16_t* src_a,
+                                     int src_stride_a,
+                                     uint8_t* dst_argb,
+                                     int dst_stride_argb,
+                                     int width,
+                                     int height,
+                                     int depth) {
+  int y;
+  void (*MergeARGB16To8Row)(const uint16_t* src_r, const uint16_t* src_g,
+                            const uint16_t* src_b, const uint16_t* src_a,
+                            uint8_t* dst_argb, int depth, int width) =
+      MergeARGB16To8Row_C;
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      src_stride_a == width && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = src_stride_a =
+        dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEARGB16TO8ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeARGB16To8Row = MergeARGB16To8Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeARGB16To8Row = MergeARGB16To8Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEARGB16TO8ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeARGB16To8Row = MergeARGB16To8Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeARGB16To8Row = MergeARGB16To8Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeARGB16To8Row(src_r, src_g, src_b, src_a, dst_argb, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    src_a += src_stride_a;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeARGB16To8PlaneOpaque(const uint16_t* src_r,
+                                      int src_stride_r,
+                                      const uint16_t* src_g,
+                                      int src_stride_g,
+                                      const uint16_t* src_b,
+                                      int src_stride_b,
+                                      uint8_t* dst_argb,
+                                      int dst_stride_argb,
+                                      int width,
+                                      int height,
+                                      int depth) {
+  int y;
+  void (*MergeXRGB16To8Row)(const uint16_t* src_r, const uint16_t* src_g,
+                            const uint16_t* src_b, uint8_t* dst_argb, int depth,
+                            int width) = MergeXRGB16To8Row_C;
+
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEXRGB16TO8ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXRGB16To8Row = MergeXRGB16To8Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXRGB16To8Row = MergeXRGB16To8Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXRGB16TO8ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeXRGB16To8Row = MergeXRGB16To8Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeXRGB16To8Row = MergeXRGB16To8Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXRGB16To8Row(src_r, src_g, src_b, dst_argb, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_API
+void MergeARGB16To8Plane(const uint16_t* src_r,
+                         int src_stride_r,
+                         const uint16_t* src_g,
+                         int src_stride_g,
+                         const uint16_t* src_b,
+                         int src_stride_b,
+                         const uint16_t* src_a,
+                         int src_stride_a,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height,
+                         int depth) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+
+  if (src_a == NULL) {
+    MergeARGB16To8PlaneOpaque(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                              src_stride_b, dst_argb, dst_stride_argb, width,
+                              height, depth);
+  } else {
+    MergeARGB16To8PlaneAlpha(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                             src_stride_b, src_a, src_stride_a, dst_argb,
+                             dst_stride_argb, width, height, depth);
+  }
+}
+
 // Convert YUY2 to I422.
 LIBYUV_API
-int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int YUY2ToI422(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int y;
-  void (*YUY2ToUV422Row)(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width) =
-      YUY2ToUV422Row_C;
-  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int width) =
+  void (*YUY2ToUV422Row)(const uint8_t* src_yuy2, uint8_t* dst_u,
+                         uint8_t* dst_v, int width) = YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8_t* src_yuy2, uint8_t* dst_y, int width) =
       YUY2ToYRow_C;
+  if (!src_yuy2 || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -291,10 +2024,9 @@ int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
     src_stride_yuy2 = -src_stride_yuy2;
   }
   // Coalesce rows.
-  if (src_stride_yuy2 == width * 2 &&
-      dst_stride_y == width &&
-      dst_stride_u * 2 == width &&
-      dst_stride_v * 2 == width) {
+  if (src_stride_yuy2 == width * 2 && dst_stride_y == width &&
+      dst_stride_u * 2 == width && dst_stride_v * 2 == width &&
+      width * height <= 32768) {
     width *= height;
     height = 1;
     src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
@@ -322,15 +2054,33 @@ int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
 #if defined(HAS_YUY2TOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     YUY2ToYRow = YUY2ToYRow_Any_NEON;
-    if (width >= 16) {
-      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
-    }
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       YUY2ToYRow = YUY2ToYRow_NEON;
       YUY2ToUV422Row = YUY2ToUV422Row_NEON;
     }
   }
 #endif
+#if defined(HAS_YUY2TOYROW_MSA) && defined(HAS_YUY2TOUV422ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    YUY2ToYRow = YUY2ToYRow_Any_MSA;
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_MSA;
+      YUY2ToUV422Row = YUY2ToUV422Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_LASX) && defined(HAS_YUY2TOUV422ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    YUY2ToYRow = YUY2ToYRow_Any_LASX;
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_LASX;
+      YUY2ToUV422Row = YUY2ToUV422Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
@@ -345,17 +2095,24 @@ int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
 
 // Convert UYVY to I422.
 LIBYUV_API
-int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
-  int y;
-  void (*UYVYToUV422Row)(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width) =
-      UYVYToUV422Row_C;
-  void (*UYVYToYRow)(const uint8* src_uyvy,
-                     uint8* dst_y, int width) = UYVYToYRow_C;
+int UYVYToI422(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int y;
+  void (*UYVYToUV422Row)(const uint8_t* src_uyvy, uint8_t* dst_u,
+                         uint8_t* dst_v, int width) = UYVYToUV422Row_C;
+  void (*UYVYToYRow)(const uint8_t* src_uyvy, uint8_t* dst_y, int width) =
+      UYVYToYRow_C;
+  if (!src_uyvy || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -363,10 +2120,9 @@ int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
     src_stride_uyvy = -src_stride_uyvy;
   }
   // Coalesce rows.
-  if (src_stride_uyvy == width * 2 &&
-      dst_stride_y == width &&
-      dst_stride_u * 2 == width &&
-      dst_stride_v * 2 == width) {
+  if (src_stride_uyvy == width * 2 && dst_stride_y == width &&
+      dst_stride_u * 2 == width && dst_stride_v * 2 == width &&
+      width * height <= 32768) {
     width *= height;
     height = 1;
     src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
@@ -394,15 +2150,33 @@ int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
 #if defined(HAS_UYVYTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     UYVYToYRow = UYVYToYRow_Any_NEON;
-    if (width >= 16) {
-      UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
-    }
+    UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       UYVYToYRow = UYVYToYRow_NEON;
       UYVYToUV422Row = UYVYToUV422Row_NEON;
     }
   }
 #endif
+#if defined(HAS_UYVYTOYROW_MSA) && defined(HAS_UYVYTOUV422ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    UYVYToYRow = UYVYToYRow_Any_MSA;
+    UYVYToUV422Row = UYVYToUV422Row_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_MSA;
+      UYVYToUV422Row = UYVYToUV422Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_LASX) && defined(HAS_UYVYTOUV422ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    UYVYToYRow = UYVYToYRow_Any_LASX;
+    UYVYToUV422Row = UYVYToUV422Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_LASX;
+      UYVYToUV422Row = UYVYToUV422Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     UYVYToUV422Row(src_uyvy, dst_u, dst_v, width);
@@ -415,13 +2189,281 @@ int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
   return 0;
 }
 
+// Convert YUY2 to Y.
+LIBYUV_API
+int YUY2ToY(const uint8_t* src_yuy2,
+            int src_stride_yuy2,
+            uint8_t* dst_y,
+            int dst_stride_y,
+            int width,
+            int height) {
+  int y;
+  void (*YUY2ToYRow)(const uint8_t* src_yuy2, uint8_t* dst_y, int width) =
+      YUY2ToYRow_C;
+  if (!src_yuy2 || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_y = 0;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    YUY2ToYRow = YUY2ToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Convert UYVY to Y.
+LIBYUV_API
+int UYVYToY(const uint8_t* src_uyvy,
+            int src_stride_uyvy,
+            uint8_t* dst_y,
+            int dst_stride_y,
+            int width,
+            int height) {
+  int y;
+  void (*UYVYToYRow)(const uint8_t* src_uyvy, uint8_t* dst_y, int width) =
+      UYVYToYRow_C;
+  if (!src_uyvy || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    UYVYToYRow = UYVYToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Mirror a plane of data.
+// See Also I400Mirror
+LIBYUV_API
+void MirrorPlane(const uint8_t* src_y,
+                 int src_stride_y,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 int width,
+                 int height) {
+  int y;
+  void (*MirrorRow)(const uint8_t* src, uint8_t* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MirrorRow = MirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    MirrorRow = MirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_LASX;
+    }
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Mirror a plane of UV data.
+LIBYUV_API
+void MirrorUVPlane(const uint8_t* src_uv,
+                   int src_stride_uv,
+                   uint8_t* dst_uv,
+                   int dst_stride_uv,
+                   int width,
+                   int height) {
+  int y;
+  void (*MirrorUVRow)(const uint8_t* src, uint8_t* dst, int width) =
+      MirrorUVRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorUVRow = MirrorUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorUVRow = MirrorUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorUVRow = MirrorUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      MirrorUVRow = MirrorUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorUVRow = MirrorUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorUVRow = MirrorUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MirrorUVRow = MirrorUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      MirrorUVRow = MirrorUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    MirrorUVRow = MirrorUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorUVRow = MirrorUVRow_LASX;
+    }
+  }
+#endif
+
+  // MirrorUV plane
+  for (y = 0; y < height; ++y) {
+    MirrorUVRow(src_uv, dst_uv, width);
+    src_uv += src_stride_uv;
+    dst_uv += dst_stride_uv;
+  }
+}
+
 // Mirror I400 with optional flipping
 LIBYUV_API
-int I400Mirror(const uint8* src_y, int src_stride_y,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height) {
-  if (!src_y || !dst_y ||
-      width <= 0 || height == 0) {
+int I400Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -437,19 +2479,28 @@ int I400Mirror(const uint8* src_y, int src_stride_y,
 
 // Mirror I420 with optional flipping
 LIBYUV_API
-int I420Mirror(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height) {
+int I420Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+
+  if (!src_y || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -470,13 +2521,53 @@ int I420Mirror(const uint8* src_y, int src_stride_y,
   return 0;
 }
 
+// NV12 mirror.
+LIBYUV_API
+int NV12Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+
+  if (!src_y || !src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorUVPlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, halfwidth,
+                halfheight);
+  return 0;
+}
+
 // ARGB mirror.
 LIBYUV_API
-int ARGBMirror(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int ARGBMirror(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
   int y;
-  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+  void (*ARGBMirrorRow)(const uint8_t* src, uint8_t* dst, int width) =
       ARGBMirrorRow_C;
   if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
@@ -490,7 +2581,7 @@ int ARGBMirror(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_ARGBMIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
-    if (IS_ALIGNED(width, 4)) {
+    if (IS_ALIGNED(width, 8)) {
       ARGBMirrorRow = ARGBMirrorRow_NEON;
     }
   }
@@ -511,6 +2602,22 @@ int ARGBMirror(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBMIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_LASX;
+    }
+  }
+#endif
 
   // Mirror plane
   for (y = 0; y < height; ++y) {
@@ -521,13 +2628,59 @@ int ARGBMirror(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// RGB24 mirror.
+LIBYUV_API
+int RGB24Mirror(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  int y;
+  void (*RGB24MirrorRow)(const uint8_t* src, uint8_t* dst, int width) =
+      RGB24MirrorRow_C;
+  if (!src_rgb24 || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+#if defined(HAS_RGB24MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24MirrorRow = RGB24MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24MirrorRow = RGB24MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGB24MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24MirrorRow = RGB24MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24MirrorRow = RGB24MirrorRow_SSSE3;
+    }
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    RGB24MirrorRow(src_rgb24, dst_rgb24, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_rgb24 += dst_stride_rgb24;
+  }
+  return 0;
+}
+
 // Get a blender that optimized for the CPU and pixel count.
 // As there are 6 blenders to choose from, the caller should try to use
 // the same blend function for all pixels if possible.
 LIBYUV_API
 ARGBBlendRow GetARGBBlend() {
-  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
-                       uint8* dst_argb, int width) = ARGBBlendRow_C;
+  void (*ARGBBlendRow)(const uint8_t* src_argb, const uint8_t* src_argb1,
+                       uint8_t* dst_argb, int width) = ARGBBlendRow_C;
 #if defined(HAS_ARGBBLENDROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBBlendRow = ARGBBlendRow_SSSE3;
@@ -539,18 +2692,32 @@ ARGBBlendRow GetARGBBlend() {
     ARGBBlendRow = ARGBBlendRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBBLENDROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBBlendRow = ARGBBlendRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBBlendRow = ARGBBlendRow_LSX;
+  }
+#endif
   return ARGBBlendRow;
 }
 
 // Alpha Blend 2 ARGB images and store to destination.
 LIBYUV_API
-int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
-              const uint8* src_argb1, int src_stride_argb1,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height) {
+int ARGBBlend(const uint8_t* src_argb0,
+              int src_stride_argb0,
+              const uint8_t* src_argb1,
+              int src_stride_argb1,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height) {
   int y;
-  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
-                       uint8* dst_argb, int width) = GetARGBBlend();
+  void (*ARGBBlendRow)(const uint8_t* src_argb, const uint8_t* src_argb1,
+                       uint8_t* dst_argb, int width) = GetARGBBlend();
   if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -561,8 +2728,7 @@ int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb0 == width * 4 &&
-      src_stride_argb1 == width * 4 &&
+  if (src_stride_argb0 == width * 4 && src_stride_argb1 == width * 4 &&
       dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
@@ -580,14 +2746,20 @@ int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
 
 // Alpha Blend plane and store to destination.
 LIBYUV_API
-int BlendPlane(const uint8* src_y0, int src_stride_y0,
-               const uint8* src_y1, int src_stride_y1,
-               const uint8* alpha, int alpha_stride,
-               uint8* dst_y, int dst_stride_y,
-               int width, int height) {
+int BlendPlane(const uint8_t* src_y0,
+               int src_stride_y0,
+               const uint8_t* src_y1,
+               int src_stride_y1,
+               const uint8_t* alpha,
+               int alpha_stride,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               int width,
+               int height) {
   int y;
-  void (*BlendPlaneRow)(const uint8* src0, const uint8* src1,
-      const uint8* alpha, uint8* dst, int width) = BlendPlaneRow_C;
+  void (*BlendPlaneRow)(const uint8_t* src0, const uint8_t* src1,
+                        const uint8_t* alpha, uint8_t* dst, int width) =
+      BlendPlaneRow_C;
   if (!src_y0 || !src_y1 || !alpha || !dst_y || width <= 0 || height == 0) {
     return -1;
   }
@@ -599,10 +2771,8 @@ int BlendPlane(const uint8* src_y0, int src_stride_y0,
   }
 
   // Coalesce rows for Y plane.
-  if (src_stride_y0 == width &&
-      src_stride_y1 == width &&
-      alpha_stride == width &&
-      dst_stride_y == width) {
+  if (src_stride_y0 == width && src_stride_y1 == width &&
+      alpha_stride == width && dst_stride_y == width) {
     width *= height;
     height = 1;
     src_stride_y0 = src_stride_y1 = alpha_stride = dst_stride_y = 0;
@@ -610,7 +2780,7 @@ int BlendPlane(const uint8* src_y0, int src_stride_y0,
 
 #if defined(HAS_BLENDPLANEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-  BlendPlaneRow = BlendPlaneRow_Any_SSSE3;
+    BlendPlaneRow = BlendPlaneRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       BlendPlaneRow = BlendPlaneRow_SSSE3;
     }
@@ -618,7 +2788,7 @@ int BlendPlane(const uint8* src_y0, int src_stride_y0,
 #endif
 #if defined(HAS_BLENDPLANEROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-  BlendPlaneRow = BlendPlaneRow_Any_AVX2;
+    BlendPlaneRow = BlendPlaneRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
       BlendPlaneRow = BlendPlaneRow_AVX2;
     }
@@ -638,24 +2808,37 @@ int BlendPlane(const uint8* src_y0, int src_stride_y0,
 #define MAXTWIDTH 2048
 // Alpha Blend YUV images and store to destination.
 LIBYUV_API
-int I420Blend(const uint8* src_y0, int src_stride_y0,
-              const uint8* src_u0, int src_stride_u0,
-              const uint8* src_v0, int src_stride_v0,
-              const uint8* src_y1, int src_stride_y1,
-              const uint8* src_u1, int src_stride_u1,
-              const uint8* src_v1, int src_stride_v1,
-              const uint8* alpha, int alpha_stride,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int width, int height) {
+int I420Blend(const uint8_t* src_y0,
+              int src_stride_y0,
+              const uint8_t* src_u0,
+              int src_stride_u0,
+              const uint8_t* src_v0,
+              int src_stride_v0,
+              const uint8_t* src_y1,
+              int src_stride_y1,
+              const uint8_t* src_u1,
+              int src_stride_u1,
+              const uint8_t* src_v1,
+              int src_stride_v1,
+              const uint8_t* alpha,
+              int alpha_stride,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height) {
   int y;
   // Half width/height for UV.
   int halfwidth = (width + 1) >> 1;
-  void (*BlendPlaneRow)(const uint8* src0, const uint8* src1,
-      const uint8* alpha, uint8* dst, int width) = BlendPlaneRow_C;
-  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) = ScaleRowDown2Box_C;
+  void (*BlendPlaneRow)(const uint8_t* src0, const uint8_t* src1,
+                        const uint8_t* alpha, uint8_t* dst, int width) =
+      BlendPlaneRow_C;
+  void (*ScaleRowDown2)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                        uint8_t* dst_ptr, int dst_width) = ScaleRowDown2Box_C;
+
   if (!src_y0 || !src_u0 || !src_v0 || !src_y1 || !src_u1 || !src_v1 ||
       !alpha || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
@@ -669,11 +2852,8 @@ int I420Blend(const uint8* src_y0, int src_stride_y0,
   }
 
   // Blend Y plane.
-  BlendPlane(src_y0, src_stride_y0,
-             src_y1, src_stride_y1,
-             alpha, alpha_stride,
-             dst_y, dst_stride_y,
-             width, height);
+  BlendPlane(src_y0, src_stride_y0, src_y1, src_stride_y1, alpha, alpha_stride,
+             dst_y, dst_stride_y, width, height);
 
 #if defined(HAS_BLENDPLANEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
@@ -753,13 +2933,17 @@ int I420Blend(const uint8* src_y0, int src_stride_y0,
 
 // Multiply 2 ARGB images and store to destination.
 LIBYUV_API
-int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
-                 const uint8* src_argb1, int src_stride_argb1,
-                 uint8* dst_argb, int dst_stride_argb,
-                 int width, int height) {
+int ARGBMultiply(const uint8_t* src_argb0,
+                 int src_stride_argb0,
+                 const uint8_t* src_argb1,
+                 int src_stride_argb1,
+                 uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int width,
+                 int height) {
   int y;
-  void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst,
-                          int width) = ARGBMultiplyRow_C;
+  void (*ARGBMultiplyRow)(const uint8_t* src0, const uint8_t* src1,
+                          uint8_t* dst, int width) = ARGBMultiplyRow_C;
   if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -770,8 +2954,7 @@ int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb0 == width * 4 &&
-      src_stride_argb1 == width * 4 &&
+  if (src_stride_argb0 == width * 4 && src_stride_argb1 == width * 4 &&
       dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
@@ -801,6 +2984,22 @@ int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
     }
   }
 #endif
+#if defined(HAS_ARGBMULTIPLYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_MSA;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_LASX;
+    }
+  }
+#endif
 
   // Multiply plane
   for (y = 0; y < height; ++y) {
@@ -814,12 +3013,16 @@ int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
 
 // Add 2 ARGB images and store to destination.
 LIBYUV_API
-int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
-            const uint8* src_argb1, int src_stride_argb1,
-            uint8* dst_argb, int dst_stride_argb,
-            int width, int height) {
+int ARGBAdd(const uint8_t* src_argb0,
+            int src_stride_argb0,
+            const uint8_t* src_argb1,
+            int src_stride_argb1,
+            uint8_t* dst_argb,
+            int dst_stride_argb,
+            int width,
+            int height) {
   int y;
-  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+  void (*ARGBAddRow)(const uint8_t* src0, const uint8_t* src1, uint8_t* dst,
                      int width) = ARGBAddRow_C;
   if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
     return -1;
@@ -831,19 +3034,18 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb0 == width * 4 &&
-      src_stride_argb1 == width * 4 &&
+  if (src_stride_argb0 == width * 4 && src_stride_argb1 == width * 4 &&
       dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
   }
-#if defined(HAS_ARGBADDROW_SSE2) && (defined(_MSC_VER) && !defined(__clang__))
+#if defined(HAS_ARGBADDROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBAddRow = ARGBAddRow_SSE2;
   }
 #endif
-#if defined(HAS_ARGBADDROW_SSE2) && !(defined(_MSC_VER) && !defined(__clang__))
+#if defined(HAS_ARGBADDROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBAddRow = ARGBAddRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
@@ -867,6 +3069,22 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
     }
   }
 #endif
+#if defined(HAS_ARGBADDROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAddRow = ARGBAddRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBAddRow = ARGBAddRow_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_LASX;
+    }
+  }
+#endif
 
   // Add plane
   for (y = 0; y < height; ++y) {
@@ -880,13 +3098,17 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
 
 // Subtract 2 ARGB images and store to destination.
 LIBYUV_API
-int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
-                 const uint8* src_argb1, int src_stride_argb1,
-                 uint8* dst_argb, int dst_stride_argb,
-                 int width, int height) {
+int ARGBSubtract(const uint8_t* src_argb0,
+                 int src_stride_argb0,
+                 const uint8_t* src_argb1,
+                 int src_stride_argb1,
+                 uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int width,
+                 int height) {
   int y;
-  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
-                          int width) = ARGBSubtractRow_C;
+  void (*ARGBSubtractRow)(const uint8_t* src0, const uint8_t* src1,
+                          uint8_t* dst, int width) = ARGBSubtractRow_C;
   if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -897,8 +3119,7 @@ int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
     dst_stride_argb = -dst_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb0 == width * 4 &&
-      src_stride_argb1 == width * 4 &&
+  if (src_stride_argb0 == width * 4 && src_stride_argb1 == width * 4 &&
       dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
@@ -928,182 +3149,45 @@ int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
     }
   }
 #endif
-
-  // Subtract plane
-  for (y = 0; y < height; ++y) {
-    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
-    src_argb0 += src_stride_argb0;
-    src_argb1 += src_stride_argb1;
-    dst_argb += dst_stride_argb;
-  }
-  return 0;
-}
-// Convert I422 to RGBA with matrix
-static int I422ToRGBAMatrix(const uint8* src_y, int src_stride_y,
-                            const uint8* src_u, int src_stride_u,
-                            const uint8* src_v, int src_stride_v,
-                            uint8* dst_rgba, int dst_stride_rgba,
-                            const struct YuvConstants* yuvconstants,
-                            int width, int height) {
-  int y;
-  void (*I422ToRGBARow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        const struct YuvConstants* yuvconstants,
-                        int width) = I422ToRGBARow_C;
-  if (!src_y || !src_u || !src_v || !dst_rgba ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
-    dst_stride_rgba = -dst_stride_rgba;
-  }
-#if defined(HAS_I422TORGBAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGBARow = I422ToRGBARow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGBARow = I422ToRGBARow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
-      IS_ALIGNED(dst_rgba, 4) && IS_ALIGNED(dst_stride_rgba, 4)) {
-    I422ToRGBARow = I422ToRGBARow_DSPR2;
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
-    dst_rgba += dst_stride_rgba;
-    src_y += src_stride_y;
-    src_u += src_stride_u;
-    src_v += src_stride_v;
-  }
-  return 0;
-}
-
-// Convert I422 to RGBA.
-LIBYUV_API
-int I422ToRGBA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_rgba, int dst_stride_rgba,
-               int width, int height) {
-  return I422ToRGBAMatrix(src_y, src_stride_y,
-                          src_u, src_stride_u,
-                          src_v, src_stride_v,
-                          dst_rgba, dst_stride_rgba,
-                          &kYuvI601Constants,
-                          width, height);
-}
-
-// Convert I422 to BGRA.
-LIBYUV_API
-int I422ToBGRA(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_bgra, int dst_stride_bgra,
-               int width, int height) {
-  return I422ToRGBAMatrix(src_y, src_stride_y,
-                          src_v, src_stride_v,  // Swap U and V
-                          src_u, src_stride_u,
-                          dst_bgra, dst_stride_bgra,
-                          &kYvuI601Constants,  // Use Yvu matrix
-                          width, height);
-}
-
-// Convert NV12 to RGB565.
-LIBYUV_API
-int NV12ToRGB565(const uint8* src_y, int src_stride_y,
-                 const uint8* src_uv, int src_stride_uv,
-                 uint8* dst_rgb565, int dst_stride_rgb565,
-                 int width, int height) {
-  int y;
-  void (*NV12ToRGB565Row)(const uint8* y_buf,
-                          const uint8* uv_buf,
-                          uint8* rgb_buf,
-                          const struct YuvConstants* yuvconstants,
-                          int width) = NV12ToRGB565Row_C;
-  if (!src_y || !src_uv || !dst_rgb565 ||
-      width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-#if defined(HAS_NV12TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+#if defined(HAS_ARGBSUBTRACTROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_MSA;
     if (IS_ALIGNED(width, 8)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_NV12TORGB565ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
+      ARGBSubtractRow = ARGBSubtractRow_MSA;
     }
   }
 #endif
-#if defined(HAS_NV12TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+#if defined(HAS_ARGBSUBTRACTROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_LASX;
     if (IS_ALIGNED(width, 8)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+      ARGBSubtractRow = ARGBSubtractRow_LASX;
     }
   }
 #endif
 
+  // Subtract plane
   for (y = 0; y < height; ++y) {
-    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, &kYuvI601Constants, width);
-    dst_rgb565 += dst_stride_rgb565;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_uv += src_stride_uv;
-    }
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
   }
   return 0;
 }
 
 // Convert RAW to RGB24.
 LIBYUV_API
-int RAWToRGB24(const uint8* src_raw, int src_stride_raw,
-               uint8* dst_rgb24, int dst_stride_rgb24,
-               int width, int height) {
+int RAWToRGB24(const uint8_t* src_raw,
+               int src_stride_raw,
+               uint8_t* dst_rgb24,
+               int dst_stride_rgb24,
+               int width,
+               int height) {
   int y;
-  void (*RAWToRGB24Row)(const uint8* src_rgb, uint8* dst_rgb24, int width) =
+  void (*RAWToRGB24Row)(const uint8_t* src_rgb, uint8_t* dst_rgb24, int width) =
       RAWToRGB24Row_C;
-  if (!src_raw || !dst_rgb24 ||
-      width <= 0 || height == 0) {
+  if (!src_raw || !dst_rgb24 || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1113,8 +3197,7 @@ int RAWToRGB24(const uint8* src_raw, int src_stride_raw,
     src_stride_raw = -src_stride_raw;
   }
   // Coalesce rows.
-  if (src_stride_raw == width * 3 &&
-      dst_stride_rgb24 == width * 3) {
+  if (src_stride_raw == width * 3 && dst_stride_rgb24 == width * 3) {
     width *= height;
     height = 1;
     src_stride_raw = dst_stride_rgb24 = 0;
@@ -1135,6 +3218,22 @@ int RAWToRGB24(const uint8* src_raw, int src_stride_raw,
     }
   }
 #endif
+#if defined(HAS_RAWTORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToRGB24Row = RAWToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToRGB24Row = RAWToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_RAWTORGB24ROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToRGB24Row = RAWToRGB24Row_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToRGB24Row = RAWToRGB24Row_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RAWToRGB24Row(src_raw, dst_rgb24, width);
@@ -1144,12 +3243,19 @@ int RAWToRGB24(const uint8* src_raw, int src_stride_raw,
   return 0;
 }
 
+// TODO(fbarchard): Consider uint8_t value
 LIBYUV_API
-void SetPlane(uint8* dst_y, int dst_stride_y,
-              int width, int height,
-              uint32 value) {
+void SetPlane(uint8_t* dst_y,
+              int dst_stride_y,
+              int width,
+              int height,
+              uint32_t value) {
   int y;
-  void (*SetRow)(uint8* dst, uint8 value, int width) = SetRow_C;
+  void (*SetRow)(uint8_t* dst, uint8_t value, int width) = SetRow_C;
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
   if (height < 0) {
     height = -height;
     dst_y = dst_y + (height - 1) * dst_stride_y;
@@ -1182,32 +3288,50 @@ void SetPlane(uint8* dst_y, int dst_stride_y,
     SetRow = SetRow_ERMS;
   }
 #endif
+#if defined(HAS_SETROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 16)) {
+    SetRow = SetRow_MSA;
+  }
+#endif
+#if defined(HAS_SETROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SetRow = SetRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      SetRow = SetRow_LSX;
+    }
+  }
+#endif
 
   // Set plane
   for (y = 0; y < height; ++y) {
-    SetRow(dst_y, value, width);
+    SetRow(dst_y, (uint8_t)value, width);
     dst_y += dst_stride_y;
   }
 }
 
 // Draw a rectangle into I420
 LIBYUV_API
-int I420Rect(uint8* dst_y, int dst_stride_y,
-             uint8* dst_u, int dst_stride_u,
-             uint8* dst_v, int dst_stride_v,
-             int x, int y,
-             int width, int height,
-             int value_y, int value_u, int value_v) {
+int I420Rect(uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_u,
+             int dst_stride_u,
+             uint8_t* dst_v,
+             int dst_stride_v,
+             int x,
+             int y,
+             int width,
+             int height,
+             int value_y,
+             int value_u,
+             int value_v) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  uint8* start_y = dst_y + y * dst_stride_y + x;
-  uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
-  uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
-  if (!dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0 ||
-      x < 0 || y < 0 ||
-      value_y < 0 || value_y > 255 ||
-      value_u < 0 || value_u > 255 ||
+  uint8_t* start_y = dst_y + y * dst_stride_y + x;
+  uint8_t* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
+  uint8_t* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+
+  if (!dst_y || !dst_u || !dst_v || width <= 0 || height == 0 || x < 0 ||
+      y < 0 || value_y < 0 || value_y > 255 || value_u < 0 || value_u > 255 ||
       value_v < 0 || value_v > 255) {
     return -1;
   }
@@ -1220,15 +3344,17 @@ int I420Rect(uint8* dst_y, int dst_stride_y,
 
 // Draw a rectangle into ARGB
 LIBYUV_API
-int ARGBRect(uint8* dst_argb, int dst_stride_argb,
-             int dst_x, int dst_y,
-             int width, int height,
-             uint32 value) {
+int ARGBRect(uint8_t* dst_argb,
+             int dst_stride_argb,
+             int dst_x,
+             int dst_y,
+             int width,
+             int height,
+             uint32_t value) {
   int y;
-  void (*ARGBSetRow)(uint8* dst_argb, uint32 value, int width) = ARGBSetRow_C;
-  if (!dst_argb ||
-      width <= 0 || height == 0 ||
-      dst_x < 0 || dst_y < 0) {
+  void (*ARGBSetRow)(uint8_t* dst_argb, uint32_t value, int width) =
+      ARGBSetRow_C;
+  if (!dst_argb || width <= 0 || height == 0 || dst_x < 0 || dst_y < 0) {
     return -1;
   }
   if (height < 0) {
@@ -1257,6 +3383,22 @@ int ARGBRect(uint8* dst_argb, int dst_stride_argb,
     ARGBSetRow = ARGBSetRow_X86;
   }
 #endif
+#if defined(HAS_ARGBSETROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBSetRow = ARGBSetRow_Any_MSA;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSETROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBSetRow = ARGBSetRow_Any_LSX;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_LSX;
+    }
+  }
+#endif
 
   // Set plane
   for (y = 0; y < height; ++y) {
@@ -1280,11 +3422,14 @@ int ARGBRect(uint8* dst_argb, int dst_stride_argb,
 //   f is foreground pixel premultiplied by alpha
 
 LIBYUV_API
-int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int width, int height) {
+int ARGBAttenuate(const uint8_t* src_argb,
+                  int src_stride_argb,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int width,
+                  int height) {
   int y;
-  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
                            int width) = ARGBAttenuateRow_C;
   if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
@@ -1295,8 +3440,7 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -1325,6 +3469,22 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBAttenuateRow(src_argb, dst_argb, width);
@@ -1336,11 +3496,14 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
 
 // Convert preattentuated ARGB to unattenuated ARGB.
 LIBYUV_API
-int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
-                    uint8* dst_argb, int dst_stride_argb,
-                    int width, int height) {
+int ARGBUnattenuate(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height) {
   int y;
-  void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
+  void (*ARGBUnattenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
                              int width) = ARGBUnattenuateRow_C;
   if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
@@ -1351,8 +3514,7 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -1373,7 +3535,7 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
-// TODO(fbarchard): Neon version.
+  // TODO(fbarchard): Neon version.
 
   for (y = 0; y < height; ++y) {
     ARGBUnattenuateRow(src_argb, dst_argb, width);
@@ -1385,12 +3547,15 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
 
 // Convert ARGB to Grayed ARGB.
 LIBYUV_API
-int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_argb, int dst_stride_argb,
-               int width, int height) {
+int ARGBGrayTo(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
   int y;
-  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
-                      int width) = ARGBGrayRow_C;
+  void (*ARGBGrayRow)(const uint8_t* src_argb, uint8_t* dst_argb, int width) =
+      ARGBGrayRow_C;
   if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -1400,8 +3565,7 @@ int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -1416,6 +3580,16 @@ int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
     ARGBGrayRow = ARGBGrayRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBGRAYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBGRAYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 16)) {
+    ARGBGrayRow = ARGBGrayRow_LASX;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBGrayRow(src_argb, dst_argb, width);
@@ -1427,13 +3601,16 @@ int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
 
 // Make a rectangle of ARGB gray scale.
 LIBYUV_API
-int ARGBGray(uint8* dst_argb, int dst_stride_argb,
-             int dst_x, int dst_y,
-             int width, int height) {
+int ARGBGray(uint8_t* dst_argb,
+             int dst_stride_argb,
+             int dst_x,
+             int dst_y,
+             int width,
+             int height) {
   int y;
-  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
-                      int width) = ARGBGrayRow_C;
-  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  void (*ARGBGrayRow)(const uint8_t* src_argb, uint8_t* dst_argb, int width) =
+      ARGBGrayRow_C;
+  uint8_t* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
   if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
     return -1;
   }
@@ -1453,6 +3630,17 @@ int ARGBGray(uint8* dst_argb, int dst_stride_argb,
     ARGBGrayRow = ARGBGrayRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBGRAYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBGRAYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 16)) {
+    ARGBGrayRow = ARGBGrayRow_LASX;
+  }
+#endif
+
   for (y = 0; y < height; ++y) {
     ARGBGrayRow(dst, dst, width);
     dst += dst_stride_argb;
@@ -1462,11 +3650,15 @@ int ARGBGray(uint8* dst_argb, int dst_stride_argb,
 
 // Make a rectangle of ARGB Sepia tone.
 LIBYUV_API
-int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
-              int dst_x, int dst_y, int width, int height) {
+int ARGBSepia(uint8_t* dst_argb,
+              int dst_stride_argb,
+              int dst_x,
+              int dst_y,
+              int width,
+              int height) {
   int y;
-  void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C;
-  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  void (*ARGBSepiaRow)(uint8_t* dst_argb, int width) = ARGBSepiaRow_C;
+  uint8_t* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
   if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
     return -1;
   }
@@ -1486,6 +3678,17 @@ int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
     ARGBSepiaRow = ARGBSepiaRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBSEPIAROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBSEPIAROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 16)) {
+    ARGBSepiaRow = ARGBSepiaRow_LASX;
+  }
+#endif
+
   for (y = 0; y < height; ++y) {
     ARGBSepiaRow(dst, width);
     dst += dst_stride_argb;
@@ -1496,13 +3699,17 @@ int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
 // Apply a 4x4 matrix to each ARGB pixel.
 // Note: Normally for shading, but can be used to swizzle or invert.
 LIBYUV_API
-int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
-                    uint8* dst_argb, int dst_stride_argb,
-                    const int8* matrix_argb,
-                    int width, int height) {
+int ARGBColorMatrix(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    const int8_t* matrix_argb,
+                    int width,
+                    int height) {
   int y;
-  void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb,
-      const int8* matrix_argb, int width) = ARGBColorMatrixRow_C;
+  void (*ARGBColorMatrixRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                             const int8_t* matrix_argb, int width) =
+      ARGBColorMatrixRow_C;
   if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -1512,8 +3719,7 @@ int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -1528,6 +3734,16 @@ int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
     ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBCOLORMATRIXROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBCOLORMATRIXROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_LSX;
+  }
+#endif
   for (y = 0; y < height; ++y) {
     ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width);
     src_argb += src_stride_argb;
@@ -1539,13 +3755,17 @@ int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
 // Apply a 4x3 matrix to each ARGB pixel.
 // Deprecated.
 LIBYUV_API
-int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
-                   const int8* matrix_rgb,
-                   int dst_x, int dst_y, int width, int height) {
-  SIMD_ALIGNED(int8 matrix_argb[16]);
-  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
-  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 ||
-      dst_x < 0 || dst_y < 0) {
+int RGBColorMatrix(uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   const int8_t* matrix_rgb,
+                   int dst_x,
+                   int dst_y,
+                   int width,
+                   int height) {
+  SIMD_ALIGNED(int8_t matrix_argb[16]);
+  uint8_t* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 || dst_x < 0 ||
+      dst_y < 0) {
     return -1;
   }
 
@@ -1565,23 +3785,26 @@ int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
   matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0;
   matrix_argb[15] = 64;  // 1.0
 
-  return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb,
-                         dst, dst_stride_argb,
-                         &matrix_argb[0], width, height);
+  return ARGBColorMatrix((const uint8_t*)(dst), dst_stride_argb, dst,
+                         dst_stride_argb, &matrix_argb[0], width, height);
 }
 
 // Apply a color table each ARGB pixel.
 // Table contains 256 ARGB values.
 LIBYUV_API
-int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
-                   const uint8* table_argb,
-                   int dst_x, int dst_y, int width, int height) {
+int ARGBColorTable(uint8_t* dst_argb,
+                   int dst_stride_argb,
+                   const uint8_t* table_argb,
+                   int dst_x,
+                   int dst_y,
+                   int width,
+                   int height) {
   int y;
-  void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+  void (*ARGBColorTableRow)(uint8_t* dst_argb, const uint8_t* table_argb,
                             int width) = ARGBColorTableRow_C;
-  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
-  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
-      dst_x < 0 || dst_y < 0) {
+  uint8_t* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 || dst_x < 0 ||
+      dst_y < 0) {
     return -1;
   }
   // Coalesce rows.
@@ -1605,15 +3828,19 @@ int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
 // Apply a color table each ARGB pixel but preserve destination alpha.
 // Table contains 256 ARGB values.
 LIBYUV_API
-int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
-                  const uint8* table_argb,
-                  int dst_x, int dst_y, int width, int height) {
+int RGBColorTable(uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  const uint8_t* table_argb,
+                  int dst_x,
+                  int dst_y,
+                  int width,
+                  int height) {
   int y;
-  void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+  void (*RGBColorTableRow)(uint8_t* dst_argb, const uint8_t* table_argb,
                            int width) = RGBColorTableRow_C;
-  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
-  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
-      dst_x < 0 || dst_y < 0) {
+  uint8_t* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 || dst_x < 0 ||
+      dst_y < 0) {
     return -1;
   }
   // Coalesce rows.
@@ -1644,13 +3871,19 @@ int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
 // Caveat - although SSE2 saturates, the C function does not and should be used
 // with care if doing anything but quantization.
 LIBYUV_API
-int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
-                 int scale, int interval_size, int interval_offset,
-                 int dst_x, int dst_y, int width, int height) {
+int ARGBQuantize(uint8_t* dst_argb,
+                 int dst_stride_argb,
+                 int scale,
+                 int interval_size,
+                 int interval_offset,
+                 int dst_x,
+                 int dst_y,
+                 int width,
+                 int height) {
   int y;
-  void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size,
+  void (*ARGBQuantizeRow)(uint8_t* dst_argb, int scale, int interval_size,
                           int interval_offset, int width) = ARGBQuantizeRow_C;
-  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  uint8_t* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
   if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 ||
       interval_size < 1 || interval_size > 255) {
     return -1;
@@ -1671,6 +3904,16 @@ int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
     ARGBQuantizeRow = ARGBQuantizeRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBQUANTIZEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBQUANTIZEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_LSX;
+  }
+#endif
   for (y = 0; y < height; ++y) {
     ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
     dst += dst_stride_argb;
@@ -1681,13 +3924,17 @@ int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
 // Computes table of cumulative sum for image where the value is the sum
 // of all values above and to the left of the entry. Used by ARGBBlur.
 LIBYUV_API
-int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
-                             int32* dst_cumsum, int dst_stride32_cumsum,
-                             int width, int height) {
+int ARGBComputeCumulativeSum(const uint8_t* src_argb,
+                             int src_stride_argb,
+                             int32_t* dst_cumsum,
+                             int dst_stride32_cumsum,
+                             int width,
+                             int height) {
   int y;
-  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
-      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
-  int32* previous_cumsum = dst_cumsum;
+  void (*ComputeCumulativeSumRow)(const uint8_t* row, int32_t* cumsum,
+                                  const int32_t* previous_cumsum, int width) =
+      ComputeCumulativeSumRow_C;
+  int32_t* previous_cumsum = dst_cumsum;
   if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) {
     return -1;
   }
@@ -1696,6 +3943,7 @@ int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
     ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
   }
 #endif
+
   memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
   for (y = 0; y < height; ++y) {
     ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width);
@@ -1711,18 +3959,25 @@ int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
 // aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory
 // as the buffer is treated as circular.
 LIBYUV_API
-int ARGBBlur(const uint8* src_argb, int src_stride_argb,
-             uint8* dst_argb, int dst_stride_argb,
-             int32* dst_cumsum, int dst_stride32_cumsum,
-             int width, int height, int radius) {
+int ARGBBlur(const uint8_t* src_argb,
+             int src_stride_argb,
+             uint8_t* dst_argb,
+             int dst_stride_argb,
+             int32_t* dst_cumsum,
+             int dst_stride32_cumsum,
+             int width,
+             int height,
+             int radius) {
   int y;
-  void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum,
-      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
-  void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft,
-      int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C;
-  int32* cumsum_bot_row;
-  int32* max_cumsum_bot_row;
-  int32* cumsum_top_row;
+  void (*ComputeCumulativeSumRow)(const uint8_t* row, int32_t* cumsum,
+                                  const int32_t* previous_cumsum, int width) =
+      ComputeCumulativeSumRow_C;
+  void (*CumulativeSumToAverageRow)(
+      const int32_t* topleft, const int32_t* botleft, int width, int area,
+      uint8_t* dst, int count) = CumulativeSumToAverageRow_C;
+  int32_t* cumsum_bot_row;
+  int32_t* max_cumsum_bot_row;
+  int32_t* cumsum_top_row;
 
   if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
@@ -1738,7 +3993,7 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb,
   if (radius > (width / 2 - 1)) {
     radius = width / 2 - 1;
   }
-  if (radius <= 0) {
+  if (radius <= 0 || height <= 1) {
     return -1;
   }
 #if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
@@ -1749,9 +4004,8 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb,
 #endif
   // Compute enough CumulativeSum for first row to be blurred. After this
   // one row of CumulativeSum is updated at a time.
-  ARGBComputeCumulativeSum(src_argb, src_stride_argb,
-                           dst_cumsum, dst_stride32_cumsum,
-                           width, radius);
+  ARGBComputeCumulativeSum(src_argb, src_stride_argb, dst_cumsum,
+                           dst_stride32_cumsum, width, radius);
 
   src_argb = src_argb + radius * src_stride_argb;
   cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
@@ -1777,7 +4031,7 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb,
     // Increment cumsum_bot_row pointer with circular buffer wrap around and
     // then fill in a row of CumulativeSum.
     if ((y + radius) < height) {
-      const int32* prev_cumsum_bot_row = cumsum_bot_row;
+      const int32_t* prev_cumsum_bot_row = cumsum_bot_row;
       cumsum_bot_row += dst_stride32_cumsum;
       if (cumsum_bot_row >= max_cumsum_bot_row) {
         cumsum_bot_row = dst_cumsum;
@@ -1789,24 +4043,24 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb,
 
     // Left clipped.
     for (x = 0; x < radius + 1; ++x) {
-      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
-                                boxwidth, area, &dst_argb[x * 4], 1);
+      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row, boxwidth, area,
+                                &dst_argb[x * 4], 1);
       area += (bot_y - top_y);
       boxwidth += 4;
     }
 
     // Middle unclipped.
     n = (width - 1) - radius - x + 1;
-    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
-                              boxwidth, area, &dst_argb[x * 4], n);
+    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row, boxwidth, area,
+                              &dst_argb[x * 4], n);
 
     // Right clipped.
     for (x += n; x <= width - 1; ++x) {
       area -= (bot_y - top_y);
       boxwidth -= 4;
       CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4,
-                                cumsum_bot_row + (x - radius - 1) * 4,
-                                boxwidth, area, &dst_argb[x * 4], 1);
+                                cumsum_bot_row + (x - radius - 1) * 4, boxwidth,
+                                area, &dst_argb[x * 4], 1);
     }
     dst_argb += dst_stride_argb;
   }
@@ -1815,12 +4069,16 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb,
 
 // Multiply ARGB image by a specified ARGB value.
 LIBYUV_API
-int ARGBShade(const uint8* src_argb, int src_stride_argb,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height, uint32 value) {
+int ARGBShade(const uint8_t* src_argb,
+              int src_stride_argb,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height,
+              uint32_t value) {
   int y;
-  void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb,
-                       int width, uint32 value) = ARGBShadeRow_C;
+  void (*ARGBShadeRow)(const uint8_t* src_argb, uint8_t* dst_argb, int width,
+                       uint32_t value) = ARGBShadeRow_C;
   if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) {
     return -1;
   }
@@ -1830,8 +4088,7 @@ int ARGBShade(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -1846,6 +4103,16 @@ int ARGBShade(const uint8* src_argb, int src_stride_argb,
     ARGBShadeRow = ARGBShadeRow_NEON;
   }
 #endif
+#if defined(HAS_ARGBSHADEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 4)) {
+    ARGBShadeRow = ARGBShadeRow_MSA;
+  }
+#endif
+#if defined(HAS_ARGBSHADEROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_LASX;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBShadeRow(src_argb, dst_argb, width, value);
@@ -1857,12 +4124,17 @@ int ARGBShade(const uint8* src_argb, int src_stride_argb,
 
 // Interpolate 2 planes by specified amount (0 to 255).
 LIBYUV_API
-int InterpolatePlane(const uint8* src0, int src_stride0,
-                     const uint8* src1, int src_stride1,
-                     uint8* dst, int dst_stride,
-                     int width, int height, int interpolation) {
+int InterpolatePlane(const uint8_t* src0,
+                     int src_stride0,
+                     const uint8_t* src1,
+                     int src_stride1,
+                     uint8_t* dst,
+                     int dst_stride,
+                     int width,
+                     int height,
+                     int interpolation) {
   int y;
-  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+  void (*InterpolateRow)(uint8_t* dst_ptr, const uint8_t* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) = InterpolateRow_C;
   if (!src0 || !src1 || !dst || width <= 0 || height == 0) {
@@ -1875,9 +4147,7 @@ int InterpolatePlane(const uint8* src0, int src_stride0,
     dst_stride = -dst_stride;
   }
   // Coalesce rows.
-  if (src_stride0 == width &&
-      src_stride1 == width &&
-      dst_stride == width) {
+  if (src_stride0 == width && src_stride1 == width && dst_stride == width) {
     width *= height;
     height = 1;
     src_stride0 = src_stride1 = dst_stride = 0;
@@ -1906,13 +4176,20 @@ int InterpolatePlane(const uint8* src0, int src_stride0,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src0, 4) && IS_ALIGNED(src_stride0, 4) &&
-      IS_ALIGNED(src1, 4) && IS_ALIGNED(src_stride1, 4) &&
-      IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4) &&
-      IS_ALIGNED(width, 4)) {
-    InterpolateRow = InterpolateRow_DSPR2;
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
   }
 #endif
 
@@ -1925,63 +4202,155 @@ int InterpolatePlane(const uint8* src0, int src_stride0,
   return 0;
 }
 
+// Interpolate 2 planes by specified amount (0 to 255).
+LIBYUV_API
+int InterpolatePlane_16(const uint16_t* src0,
+                        int src_stride0,
+                        const uint16_t* src1,
+                        int src_stride1,
+                        uint16_t* dst,
+                        int dst_stride,
+                        int width,
+                        int height,
+                        int interpolation) {
+  int y;
+  void (*InterpolateRow_16)(uint16_t* dst_ptr, const uint16_t* src_ptr,
+                            ptrdiff_t src_stride, int dst_width,
+                            int source_y_fraction) = InterpolateRow_16_C;
+  if (!src0 || !src1 || !dst || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst = dst + (height - 1) * dst_stride;
+    dst_stride = -dst_stride;
+  }
+  // Coalesce rows.
+  if (src_stride0 == width && src_stride1 == width && dst_stride == width) {
+    width *= height;
+    height = 1;
+    src_stride0 = src_stride1 = dst_stride = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow_16 = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow_16 = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow_16 = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow_16 = InterpolateRow_16_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow_16 = InterpolateRow_16_LSX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow_16(dst, src0, src1 - src0, width, interpolation);
+    src0 += src_stride0;
+    src1 += src_stride1;
+    dst += dst_stride;
+  }
+  return 0;
+}
+
 // Interpolate 2 ARGB images by specified amount (0 to 255).
 LIBYUV_API
-int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
-                    const uint8* src_argb1, int src_stride_argb1,
-                    uint8* dst_argb, int dst_stride_argb,
-                    int width, int height, int interpolation) {
-  return InterpolatePlane(src_argb0, src_stride_argb0,
-                          src_argb1, src_stride_argb1,
-                          dst_argb, dst_stride_argb,
+int ARGBInterpolate(const uint8_t* src_argb0,
+                    int src_stride_argb0,
+                    const uint8_t* src_argb1,
+                    int src_stride_argb1,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int interpolation) {
+  return InterpolatePlane(src_argb0, src_stride_argb0, src_argb1,
+                          src_stride_argb1, dst_argb, dst_stride_argb,
                           width * 4, height, interpolation);
 }
 
 // Interpolate 2 YUV images by specified amount (0 to 255).
 LIBYUV_API
-int I420Interpolate(const uint8* src0_y, int src0_stride_y,
-                    const uint8* src0_u, int src0_stride_u,
-                    const uint8* src0_v, int src0_stride_v,
-                    const uint8* src1_y, int src1_stride_y,
-                    const uint8* src1_u, int src1_stride_u,
-                    const uint8* src1_v, int src1_stride_v,
-                    uint8* dst_y, int dst_stride_y,
-                    uint8* dst_u, int dst_stride_u,
-                    uint8* dst_v, int dst_stride_v,
-                    int width, int height, int interpolation) {
+int I420Interpolate(const uint8_t* src0_y,
+                    int src0_stride_y,
+                    const uint8_t* src0_u,
+                    int src0_stride_u,
+                    const uint8_t* src0_v,
+                    int src0_stride_v,
+                    const uint8_t* src1_y,
+                    int src1_stride_y,
+                    const uint8_t* src1_u,
+                    int src1_stride_u,
+                    const uint8_t* src1_v,
+                    int src1_stride_v,
+                    uint8_t* dst_y,
+                    int dst_stride_y,
+                    uint8_t* dst_u,
+                    int dst_stride_u,
+                    uint8_t* dst_v,
+                    int dst_stride_v,
+                    int width,
+                    int height,
+                    int interpolation) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src0_y || !src0_u || !src0_v ||
-      !src1_y || !src1_u || !src1_v ||
-      !dst_y || !dst_u || !dst_v ||
-      width <= 0 || height == 0) {
+
+  if (!src0_y || !src0_u || !src0_v || !src1_y || !src1_u || !src1_v ||
+      !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
-  InterpolatePlane(src0_y, src0_stride_y,
-                   src1_y, src1_stride_y,
-                   dst_y, dst_stride_y,
-                   width, height, interpolation);
-  InterpolatePlane(src0_u, src0_stride_u,
-                   src1_u, src1_stride_u,
-                   dst_u, dst_stride_u,
-                   halfwidth, halfheight, interpolation);
-  InterpolatePlane(src0_v, src0_stride_v,
-                   src1_v, src1_stride_v,
-                   dst_v, dst_stride_v,
-                   halfwidth, halfheight, interpolation);
+
+  InterpolatePlane(src0_y, src0_stride_y, src1_y, src1_stride_y, dst_y,
+                   dst_stride_y, width, height, interpolation);
+  InterpolatePlane(src0_u, src0_stride_u, src1_u, src1_stride_u, dst_u,
+                   dst_stride_u, halfwidth, halfheight, interpolation);
+  InterpolatePlane(src0_v, src0_stride_v, src1_v, src1_stride_v, dst_v,
+                   dst_stride_v, halfwidth, halfheight, interpolation);
   return 0;
 }
 
 // Shuffle ARGB channel order.  e.g. BGRA to ARGB.
 LIBYUV_API
-int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
-                uint8* dst_argb, int dst_stride_argb,
-                const uint8* shuffler, int width, int height) {
+int ARGBShuffle(const uint8_t* src_bgra,
+                int src_stride_bgra,
+                uint8_t* dst_argb,
+                int dst_stride_argb,
+                const uint8_t* shuffler,
+                int width,
+                int height) {
   int y;
-  void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb,
-                         const uint8* shuffler, int width) = ARGBShuffleRow_C;
-  if (!src_bgra || !dst_argb ||
-      width <= 0 || height == 0) {
+  void (*ARGBShuffleRow)(const uint8_t* src_bgra, uint8_t* dst_argb,
+                         const uint8_t* shuffler, int width) = ARGBShuffleRow_C;
+  if (!src_bgra || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -1991,20 +4360,11 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
     src_stride_bgra = -src_stride_bgra;
   }
   // Coalesce rows.
-  if (src_stride_bgra == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_bgra == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_bgra = dst_stride_argb = 0;
   }
-#if defined(HAS_ARGBSHUFFLEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBShuffleRow = ARGBShuffleRow_SSE2;
-    }
-  }
-#endif
 #if defined(HAS_ARGBSHUFFLEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
@@ -2029,6 +4389,22 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
     }
   }
 #endif
+#if defined(HAS_ARGBSHUFFLEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBShuffleRow = ARGBShuffleRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBShuffleRow(src_bgra, dst_argb, shuffler, width);
@@ -2038,29 +4414,169 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
   return 0;
 }
 
+// Shuffle AR64 channel order.  e.g. AR64 to AB64.
+LIBYUV_API
+int AR64Shuffle(const uint16_t* src_ar64,
+                int src_stride_ar64,
+                uint16_t* dst_ar64,
+                int dst_stride_ar64,
+                const uint8_t* shuffler,
+                int width,
+                int height) {
+  int y;
+  void (*AR64ShuffleRow)(const uint8_t* src_ar64, uint8_t* dst_ar64,
+                         const uint8_t* shuffler, int width) = AR64ShuffleRow_C;
+  if (!src_ar64 || !dst_ar64 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar64 = src_ar64 + (height - 1) * src_stride_ar64;
+    src_stride_ar64 = -src_stride_ar64;
+  }
+  // Coalesce rows.
+  if (src_stride_ar64 == width * 4 && dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar64 = dst_stride_ar64 = 0;
+  }
+  // Assembly versions can be reused if it's implemented with shuffle.
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    AR64ShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      AR64ShuffleRow = ARGBShuffleRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AR64ShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      AR64ShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AR64ShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      AR64ShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    AR64ShuffleRow((uint8_t*)(src_ar64), (uint8_t*)(dst_ar64), shuffler,
+                   width * 2);
+    src_ar64 += src_stride_ar64;
+    dst_ar64 += dst_stride_ar64;
+  }
+  return 0;
+}
+
+// Gauss blur a float plane using Gaussian 5x5 filter with
+// coefficients of 1, 4, 6, 4, 1.
+// Each destination pixel is a blur of the 5x5
+// pixels from the source.
+// Source edges are clamped.
+// Edge is 2 pixels on each side, and interior is multiple of 4.
+LIBYUV_API
+int GaussPlane_F32(const float* src,
+                   int src_stride,
+                   float* dst,
+                   int dst_stride,
+                   int width,
+                   int height) {
+  int y;
+  void (*GaussCol_F32)(const float* src0, const float* src1, const float* src2,
+                       const float* src3, const float* src4, float* dst,
+                       int width) = GaussCol_F32_C;
+  void (*GaussRow_F32)(const float* src, float* dst, int width) =
+      GaussRow_F32_C;
+  if (!src || !dst || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+#if defined(HAS_GAUSSCOL_F32_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    GaussCol_F32 = GaussCol_F32_NEON;
+  }
+#endif
+#if defined(HAS_GAUSSROW_F32_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    GaussRow_F32 = GaussRow_F32_NEON;
+  }
+#endif
+  {
+    // 2 pixels on each side, but aligned out to 16 bytes.
+    align_buffer_64(rowbuf, (4 + width + 4) * 4);
+    memset(rowbuf, 0, 16);
+    memset(rowbuf + (4 + width) * 4, 0, 16);
+    float* row = (float*)(rowbuf + 16);
+    const float* src0 = src;
+    const float* src1 = src;
+    const float* src2 = src;
+    const float* src3 = src2 + ((height > 1) ? src_stride : 0);
+    const float* src4 = src3 + ((height > 2) ? src_stride : 0);
+
+    for (y = 0; y < height; ++y) {
+      GaussCol_F32(src0, src1, src2, src3, src4, row, width);
+
+      // Extrude edge by 2 floats
+      row[-2] = row[-1] = row[0];
+      row[width + 1] = row[width] = row[width - 1];
+
+      GaussRow_F32(row - 2, dst, width);
+
+      src0 = src1;
+      src1 = src2;
+      src2 = src3;
+      src3 = src4;
+      if ((y + 2) < (height - 1)) {
+        src4 += src_stride;
+      }
+      dst += dst_stride;
+    }
+    free_aligned_buffer_64(rowbuf);
+  }
+  return 0;
+}
+
 // Sobel ARGB effect.
-static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
-                        uint8* dst_argb, int dst_stride_argb,
-                        int width, int height,
-                        void (*SobelRow)(const uint8* src_sobelx,
-                                         const uint8* src_sobely,
-                                         uint8* dst, int width)) {
-  int y;
-  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_g, int width) =
+static int ARGBSobelize(const uint8_t* src_argb,
+                        int src_stride_argb,
+                        uint8_t* dst_argb,
+                        int dst_stride_argb,
+                        int width,
+                        int height,
+                        void (*SobelRow)(const uint8_t* src_sobelx,
+                                         const uint8_t* src_sobely,
+                                         uint8_t* dst,
+                                         int width)) {
+  int y;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_g, int width) =
       ARGBToYJRow_C;
-  void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width) = SobelYRow_C;
-  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobely, int width) =
+  void (*SobelYRow)(const uint8_t* src_y0, const uint8_t* src_y1,
+                    uint8_t* dst_sobely, int width) = SobelYRow_C;
+  void (*SobelXRow)(const uint8_t* src_y0, const uint8_t* src_y1,
+                    const uint8_t* src_y2, uint8_t* dst_sobely, int width) =
       SobelXRow_C;
   const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
-  if (!src_argb  || !dst_argb || width <= 0 || height == 0) {
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
 
@@ -2083,11 +4599,35 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
 #endif
+#if defined(HAS_ARGBTOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYJRow = ARGBToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_LASX;
+    }
+  }
+#endif
 
 #if defined(HAS_SOBELYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
@@ -2099,6 +4639,11 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
     SobelYRow = SobelYRow_NEON;
   }
 #endif
+#if defined(HAS_SOBELYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SobelYRow = SobelYRow_MSA;
+  }
+#endif
 #if defined(HAS_SOBELXROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     SobelXRow = SobelXRow_SSE2;
@@ -2109,18 +4654,23 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
     SobelXRow = SobelXRow_NEON;
   }
 #endif
+#if defined(HAS_SOBELXROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SobelXRow = SobelXRow_MSA;
+  }
+#endif
   {
     // 3 rows with edges before/after.
-    const int kRowSize = (width + kEdge + 31) & ~31;
-    align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge));
-    uint8* row_sobelx = rows;
-    uint8* row_sobely = rows + kRowSize;
-    uint8* row_y = rows + kRowSize * 2;
+    const int row_size = (width + kEdge + 31) & ~31;
+    align_buffer_64(rows, row_size * 2 + (kEdge + row_size * 3 + kEdge));
+    uint8_t* row_sobelx = rows;
+    uint8_t* row_sobely = rows + row_size;
+    uint8_t* row_y = rows + row_size * 2;
 
     // Convert first row.
-    uint8* row_y0 = row_y + kEdge;
-    uint8* row_y1 = row_y0 + kRowSize;
-    uint8* row_y2 = row_y1 + kRowSize;
+    uint8_t* row_y0 = row_y + kEdge;
+    uint8_t* row_y1 = row_y0 + row_size;
+    uint8_t* row_y2 = row_y1 + row_size;
     ARGBToYJRow(src_argb, row_y0, width);
     row_y0[-1] = row_y0[0];
     memset(row_y0 + width, row_y0[width - 1], 16);  // Extrude 16 for valgrind.
@@ -2144,7 +4694,7 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
 
       // Cycle thru circular queue of 3 row_y buffers.
       {
-        uint8* row_yt = row_y0;
+        uint8_t* row_yt = row_y0;
         row_y0 = row_y1;
         row_y1 = row_y2;
         row_y2 = row_yt;
@@ -2159,11 +4709,14 @@ static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
 
 // Sobel ARGB effect.
 LIBYUV_API
-int ARGBSobel(const uint8* src_argb, int src_stride_argb,
-              uint8* dst_argb, int dst_stride_argb,
-              int width, int height) {
-  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
-                   uint8* dst_argb, int width) = SobelRow_C;
+int ARGBSobel(const uint8_t* src_argb,
+              int src_stride_argb,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int width,
+              int height) {
+  void (*SobelRow)(const uint8_t* src_sobelx, const uint8_t* src_sobely,
+                   uint8_t* dst_argb, int width) = SobelRow_C;
 #if defined(HAS_SOBELROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     SobelRow = SobelRow_Any_SSE2;
@@ -2180,17 +4733,36 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_SOBELROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SobelRow = SobelRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      SobelRow = SobelRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SOBELROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SobelRow = SobelRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      SobelRow = SobelRow_LSX;
+    }
+  }
+#endif
   return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
                       width, height, SobelRow);
 }
 
 // Sobel ARGB effect with planar output.
 LIBYUV_API
-int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
-                     uint8* dst_y, int dst_stride_y,
-                     int width, int height) {
-  void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_, int width) = SobelToPlaneRow_C;
+int ARGBSobelToPlane(const uint8_t* src_argb,
+                     int src_stride_argb,
+                     uint8_t* dst_y,
+                     int dst_stride_y,
+                     int width,
+                     int height) {
+  void (*SobelToPlaneRow)(const uint8_t* src_sobelx, const uint8_t* src_sobely,
+                          uint8_t* dst_, int width) = SobelToPlaneRow_C;
 #if defined(HAS_SOBELTOPLANEROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     SobelToPlaneRow = SobelToPlaneRow_Any_SSE2;
@@ -2207,18 +4779,37 @@ int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
-  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
-                      width, height, SobelToPlaneRow);
+#if defined(HAS_SOBELTOPLANEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      SobelToPlaneRow = SobelToPlaneRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SOBELTOPLANEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SobelToPlaneRow = SobelToPlaneRow_LSX;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y, width,
+                      height, SobelToPlaneRow);
 }
 
 // SobelXY ARGB effect.
 // Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
 LIBYUV_API
-int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
-                uint8* dst_argb, int dst_stride_argb,
-                int width, int height) {
-  void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) = SobelXYRow_C;
+int ARGBSobelXY(const uint8_t* src_argb,
+                int src_stride_argb,
+                uint8_t* dst_argb,
+                int dst_stride_argb,
+                int width,
+                int height) {
+  void (*SobelXYRow)(const uint8_t* src_sobelx, const uint8_t* src_sobely,
+                     uint8_t* dst_argb, int width) = SobelXYRow_C;
 #if defined(HAS_SOBELXYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     SobelXYRow = SobelXYRow_Any_SSE2;
@@ -2235,32 +4826,49 @@ int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_SOBELXYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SobelXYRow = SobelXYRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      SobelXYRow = SobelXYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SOBELXYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SobelXYRow = SobelXYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      SobelXYRow = SobelXYRow_LSX;
+    }
+  }
+#endif
   return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
                       width, height, SobelXYRow);
 }
 
 // Apply a 4x4 polynomial to each ARGB pixel.
 LIBYUV_API
-int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
-                   uint8* dst_argb, int dst_stride_argb,
+int ARGBPolynomial(const uint8_t* src_argb,
+                   int src_stride_argb,
+                   uint8_t* dst_argb,
+                   int dst_stride_argb,
                    const float* poly,
-                   int width, int height) {
+                   int width,
+                   int height) {
   int y;
-  void (*ARGBPolynomialRow)(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
-                            int width) = ARGBPolynomialRow_C;
+  void (*ARGBPolynomialRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                            const float* poly, int width) = ARGBPolynomialRow_C;
   if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -2285,28 +4893,140 @@ int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// Convert plane of 16 bit shorts to half floats.
+// Source values are multiplied by scale before storing as half float.
+LIBYUV_API
+int HalfFloatPlane(const uint16_t* src_y,
+                   int src_stride_y,
+                   uint16_t* dst_y,
+                   int dst_stride_y,
+                   float scale,
+                   int width,
+                   int height) {
+  int y;
+  void (*HalfFloatRow)(const uint16_t* src, uint16_t* dst, float scale,
+                       int width) = HalfFloatRow_C;
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  src_stride_y >>= 1;
+  dst_stride_y >>= 1;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_HALFFLOATROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    HalfFloatRow = HalfFloatRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      HalfFloatRow = HalfFloatRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_HALFFLOATROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HalfFloatRow = HalfFloatRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      HalfFloatRow = HalfFloatRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_HALFFLOATROW_F16C)
+  if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasF16C)) {
+    HalfFloatRow =
+        (scale == 1.0f) ? HalfFloat1Row_Any_F16C : HalfFloatRow_Any_F16C;
+    if (IS_ALIGNED(width, 16)) {
+      HalfFloatRow = (scale == 1.0f) ? HalfFloat1Row_F16C : HalfFloatRow_F16C;
+    }
+  }
+#endif
+#if defined(HAS_HALFFLOATROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    HalfFloatRow =
+        (scale == 1.0f) ? HalfFloat1Row_Any_NEON : HalfFloatRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      HalfFloatRow = (scale == 1.0f) ? HalfFloat1Row_NEON : HalfFloatRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_HALFFLOATROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    HalfFloatRow = HalfFloatRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      HalfFloatRow = HalfFloatRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_HALFFLOATROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    HalfFloatRow = HalfFloatRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      HalfFloatRow = HalfFloatRow_LSX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    HalfFloatRow(src_y, dst_y, scale, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Convert a buffer of bytes to floats, scale the values and store as floats.
+LIBYUV_API
+int ByteToFloat(const uint8_t* src_y, float* dst_y, float scale, int width) {
+  void (*ByteToFloatRow)(const uint8_t* src, float* dst, float scale,
+                         int width) = ByteToFloatRow_C;
+  if (!src_y || !dst_y || width <= 0) {
+    return -1;
+  }
+#if defined(HAS_BYTETOFLOATROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ByteToFloatRow = ByteToFloatRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ByteToFloatRow = ByteToFloatRow_NEON;
+    }
+  }
+#endif
+
+  ByteToFloatRow(src_y, dst_y, scale, width);
+  return 0;
+}
+
 // Apply a lumacolortable to each ARGB pixel.
 LIBYUV_API
-int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_argb, int dst_stride_argb,
-                       const uint8* luma,
-                       int width, int height) {
+int ARGBLumaColorTable(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_argb,
+                       int dst_stride_argb,
+                       const uint8_t* luma,
+                       int width,
+                       int height) {
   int y;
-  void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb,
-      int width, const uint8* luma, const uint32 lumacoeff) =
-      ARGBLumaColorTableRow_C;
+  void (*ARGBLumaColorTableRow)(
+      const uint8_t* src_argb, uint8_t* dst_argb, int width,
+      const uint8_t* luma, const uint32_t lumacoeff) = ARGBLumaColorTableRow_C;
   if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -2327,12 +5047,15 @@ int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
 
 // Copy Alpha from one ARGB image to another.
 LIBYUV_API
-int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int width, int height) {
+int ARGBCopyAlpha(const uint8_t* src_argb,
+                  int src_stride_argb,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int width,
+                  int height) {
   int y;
-  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
-      ARGBCopyAlphaRow_C;
+  void (*ARGBCopyAlphaRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBCopyAlphaRow_C;
   if (!src_argb || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -2343,8 +5066,7 @@ int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride_argb == width * 4 &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_argb = dst_stride_argb = 0;
@@ -2376,55 +5098,79 @@ int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
 
 // Extract just the alpha channel from ARGB.
 LIBYUV_API
-int ARGBExtractAlpha(const uint8* src_argb, int src_stride,
-                     uint8* dst_a, int dst_stride,
-                     int width, int height) {
+int ARGBExtractAlpha(const uint8_t* src_argb,
+                     int src_stride_argb,
+                     uint8_t* dst_a,
+                     int dst_stride_a,
+                     int width,
+                     int height) {
   if (!src_argb || !dst_a || width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    src_argb += (height - 1) * src_stride;
-    src_stride = -src_stride;
+    src_argb += (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
   }
   // Coalesce rows.
-  if (src_stride == width * 4 && dst_stride == width) {
+  if (src_stride_argb == width * 4 && dst_stride_a == width) {
     width *= height;
     height = 1;
-    src_stride = dst_stride = 0;
+    src_stride_argb = dst_stride_a = 0;
   }
-  void (*ARGBExtractAlphaRow)(const uint8 *src_argb, uint8 *dst_a, int width) =
-      ARGBExtractAlphaRow_C;
+  void (*ARGBExtractAlphaRow)(const uint8_t* src_argb, uint8_t* dst_a,
+                              int width) = ARGBExtractAlphaRow_C;
 #if defined(HAS_ARGBEXTRACTALPHAROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBExtractAlphaRow = IS_ALIGNED(width, 8) ? ARGBExtractAlphaRow_SSE2
                                                : ARGBExtractAlphaRow_Any_SSE2;
   }
 #endif
+#if defined(HAS_ARGBEXTRACTALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 32) ? ARGBExtractAlphaRow_AVX2
+                                                : ARGBExtractAlphaRow_Any_AVX2;
+  }
+#endif
 #if defined(HAS_ARGBEXTRACTALPHAROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_NEON
                                                 : ARGBExtractAlphaRow_Any_NEON;
   }
 #endif
+#if defined(HAS_ARGBEXTRACTALPHAROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_MSA
+                                                : ARGBExtractAlphaRow_Any_MSA;
+  }
+#endif
+#if defined(HAS_ARGBEXTRACTALPHAROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_LSX
+                                                : ARGBExtractAlphaRow_Any_LSX;
+  }
+#endif
 
   for (int y = 0; y < height; ++y) {
     ARGBExtractAlphaRow(src_argb, dst_a, width);
-    src_argb += src_stride;
-    dst_a += dst_stride;
+    src_argb += src_stride_argb;
+    dst_a += dst_stride_a;
   }
   return 0;
 }
 
 // Copy a planar Y channel to the alpha channel of a destination ARGB image.
 LIBYUV_API
-int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
-                     uint8* dst_argb, int dst_stride_argb,
-                     int width, int height) {
+int ARGBCopyYToAlpha(const uint8_t* src_y,
+                     int src_stride_y,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     int width,
+                     int height) {
   int y;
-  void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) =
-      ARGBCopyYToAlphaRow_C;
+  void (*ARGBCopyYToAlphaRow)(const uint8_t* src_y, uint8_t* dst_argb,
+                              int width) = ARGBCopyYToAlphaRow_C;
   if (!src_y || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -2435,8 +5181,7 @@ int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
     src_stride_y = -src_stride_y;
   }
   // Coalesce rows.
-  if (src_stride_y == width &&
-      dst_stride_argb == width * 4) {
+  if (src_stride_y == width && dst_stride_argb == width * 4) {
     width *= height;
     height = 1;
     src_stride_y = dst_stride_argb = 0;
@@ -2466,124 +5211,132 @@ int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
   return 0;
 }
 
-// TODO(fbarchard): Consider if width is even Y channel can be split
-// directly. A SplitUVRow_Odd function could copy the remaining chroma.
-
 LIBYUV_API
-int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height) {
+int YUY2ToNV12(const uint8_t* src_yuy2,
+               int src_stride_yuy2,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
   int y;
-  int halfwidth = (width + 1) >> 1;
-  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) = SplitUVRow_C;
-  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
-                         ptrdiff_t src_stride, int dst_width,
-                         int source_y_fraction) = InterpolateRow_C;
-  if (!src_yuy2 ||
-      !dst_y || !dst_uv ||
-      width <= 0 || height == 0) {
+  void (*YUY2ToYRow)(const uint8_t* src_yuy2, uint8_t* dst_y, int width) =
+      YUY2ToYRow_C;
+  void (*YUY2ToNVUVRow)(const uint8_t* src_yuy2, int stride_yuy2,
+                        uint8_t* dst_uv, int width) = YUY2ToNVUVRow_C;
+  if (!src_yuy2 || !dst_y || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
     src_stride_yuy2 = -src_stride_yuy2;
   }
-#if defined(HAS_SPLITUVROW_SSE2)
+#if defined(HAS_YUY2TOYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    SplitUVRow = SplitUVRow_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      SplitUVRow = SplitUVRow_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_AVX2)
+#if defined(HAS_YUY2TOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    SplitUVRow = SplitUVRow_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      SplitUVRow = SplitUVRow_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_NEON)
+#if defined(HAS_YUY2TOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    SplitUVRow = SplitUVRow_Any_NEON;
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
-      SplitUVRow = SplitUVRow_NEON;
+      YUY2ToYRow = YUY2ToYRow_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_SSSE3;
+#if defined(HAS_YUY2TOYROW_MSA) && defined(HAS_YUY2TOUV422ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    YUY2ToYRow = YUY2ToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_LASX) && defined(HAS_YUY2TOUV422ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    YUY2ToYRow = YUY2ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_LASX;
+    }
+  }
+#endif
+
+#if defined(HAS_YUY2TONVUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    YUY2ToNVUVRow = YUY2ToNVUVRow_Any_SSE2;
     if (IS_ALIGNED(width, 16)) {
-      InterpolateRow = InterpolateRow_SSSE3;
+      YUY2ToNVUVRow = YUY2ToNVUVRow_SSE2;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_AVX2)
+#if defined(HAS_YUY2TONVUVROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_AVX2;
+    YUY2ToNVUVRow = YUY2ToNVUVRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      InterpolateRow = InterpolateRow_AVX2;
+      YUY2ToNVUVRow = YUY2ToNVUVRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_NEON)
+#if defined(HAS_YUY2TONVUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_NEON;
+    YUY2ToNVUVRow = YUY2ToNVUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
-      InterpolateRow = InterpolateRow_NEON;
+      YUY2ToNVUVRow = YUY2ToNVUVRow_NEON;
     }
   }
 #endif
 
-  {
-    int awidth = halfwidth * 2;
-    // row of y and 2 rows of uv
-    align_buffer_64(rows, awidth * 3);
-
-    for (y = 0; y < height - 1; y += 2) {
-      // Split Y from UV.
-      SplitUVRow(src_yuy2, rows, rows + awidth, awidth);
-      memcpy(dst_y, rows, width);
-      SplitUVRow(src_yuy2 + src_stride_yuy2, rows, rows + awidth * 2, awidth);
-      memcpy(dst_y + dst_stride_y, rows, width);
-      InterpolateRow(dst_uv, rows + awidth, awidth, awidth, 128);
-      src_yuy2 += src_stride_yuy2 * 2;
-      dst_y += dst_stride_y * 2;
-      dst_uv += dst_stride_uv;
-    }
-    if (height & 1) {
-      // Split Y from UV.
-      SplitUVRow(src_yuy2, rows, dst_uv, awidth);
-      memcpy(dst_y, rows, width);
-    }
-    free_aligned_buffer_64(rows);
+  for (y = 0; y < height - 1; y += 2) {
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    YUY2ToYRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, width);
+    YUY2ToNVUVRow(src_yuy2, src_stride_yuy2, dst_uv, width);
+    src_yuy2 += src_stride_yuy2 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    YUY2ToNVUVRow(src_yuy2, 0, dst_uv, width);
   }
   return 0;
 }
 
 LIBYUV_API
-int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_uv, int dst_stride_uv,
-               int width, int height) {
+int UYVYToNV12(const uint8_t* src_uyvy,
+               int src_stride_uyvy,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
   int y;
   int halfwidth = (width + 1) >> 1;
-  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+  void (*SplitUVRow)(const uint8_t* src_uv, uint8_t* dst_u, uint8_t* dst_v,
                      int width) = SplitUVRow_C;
-  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+  void (*InterpolateRow)(uint8_t* dst_ptr, const uint8_t* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) = InterpolateRow_C;
-  if (!src_uyvy ||
-      !dst_y || !dst_uv ||
-      width <= 0 || height == 0) {
+
+  if (!src_uyvy || !dst_y || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -2614,6 +5367,22 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
     }
   }
 #endif
+#if defined(HAS_SPLITUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SplitUVRow = SplitUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SplitUVRow = SplitUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
@@ -2638,6 +5407,22 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
 
   {
     int awidth = halfwidth * 2;
@@ -2665,6 +5450,56 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
   return 0;
 }
 
+// width and height are src size allowing odd size handling.
+LIBYUV_API
+void HalfMergeUVPlane(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int dst_stride_uv,
+                      int width,
+                      int height) {
+  int y;
+  void (*HalfMergeUVRow)(const uint8_t* src_u, int src_stride_u,
+                         const uint8_t* src_v, int src_stride_v,
+                         uint8_t* dst_uv, int width) = HalfMergeUVRow_C;
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+#if defined(HAS_HALFMERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    HalfMergeUVRow = HalfMergeUVRow_NEON;
+  }
+#endif
+#if defined(HAS_HALFMERGEUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
+    HalfMergeUVRow = HalfMergeUVRow_SSSE3;
+  }
+#endif
+#if defined(HAS_HALFMERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    HalfMergeUVRow = HalfMergeUVRow_AVX2;
+  }
+#endif
+  for (y = 0; y < height - 1; y += 2) {
+    // Merge a row of U and V into a row of UV.
+    HalfMergeUVRow(src_u, src_stride_u, src_v, src_stride_v, dst_uv, width);
+    src_u += src_stride_u * 2;
+    src_v += src_stride_v * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    HalfMergeUVRow(src_u, 0, src_v, 0, dst_uv, width);
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/media/libyuv/source/rotate.cc b/media/libyuv/source/rotate.cc
index 01ea5c4074..b1b4458e66 100644
--- a/media/libyuv/source/rotate.cc
+++ b/media/libyuv/source/rotate.cc
@@ -10,8 +10,8 @@
 
 #include "libyuv/rotate.h"
 
-#include "libyuv/cpu_id.h"
 #include "libyuv/convert.h"
+#include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate_row.h"
 #include "libyuv/row.h"
@@ -22,15 +22,27 @@ extern "C" {
 #endif
 
 LIBYUV_API
-void TransposePlane(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height) {
+void TransposePlane(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
   int i = height;
-  void (*TransposeWx8)(const uint8* src, int src_stride,
-                       uint8* dst, int dst_stride, int width) = TransposeWx8_C;
+#if defined(HAS_TRANSPOSEWX16_MSA) || defined(HAS_TRANSPOSEWX16_LSX)
+  void (*TransposeWx16)(const uint8_t* src, int src_stride, uint8_t* dst,
+                        int dst_stride, int width) = TransposeWx16_C;
+#else
+  void (*TransposeWx8)(const uint8_t* src, int src_stride, uint8_t* dst,
+                       int dst_stride, int width) = TransposeWx8_C;
+#endif
+
 #if defined(HAS_TRANSPOSEWX8_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    TransposeWx8 = TransposeWx8_NEON;
+    TransposeWx8 = TransposeWx8_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeWx8 = TransposeWx8_NEON;
+    }
   }
 #endif
 #if defined(HAS_TRANSPOSEWX8_SSSE3)
@@ -49,24 +61,40 @@ void TransposePlane(const uint8* src, int src_stride,
     }
   }
 #endif
-#if defined(HAS_TRANSPOSEWX8_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2)) {
-    if (IS_ALIGNED(width, 4) &&
-        IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
-      TransposeWx8 = TransposeWx8_Fast_DSPR2;
-    } else {
-      TransposeWx8 = TransposeWx8_DSPR2;
+#if defined(HAS_TRANSPOSEWX16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    TransposeWx16 = TransposeWx16_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx16 = TransposeWx16_MSA;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX16_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    TransposeWx16 = TransposeWx16_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx16 = TransposeWx16_LSX;
     }
   }
 #endif
 
+#if defined(HAS_TRANSPOSEWX16_MSA) || defined(HAS_TRANSPOSEWX16_LSX)
+  // Work across the source in 16x16 tiles
+  while (i >= 16) {
+    TransposeWx16(src, src_stride, dst, dst_stride, width);
+    src += 16 * src_stride;  // Go down 16 rows.
+    dst += 16;               // Move over 16 columns.
+    i -= 16;
+  }
+#else
   // Work across the source in 8x8 tiles
   while (i >= 8) {
     TransposeWx8(src, src_stride, dst, dst_stride, width);
-    src += 8 * src_stride;    // Go down 8 rows.
-    dst += 8;                 // Move over 8 columns.
+    src += 8 * src_stride;  // Go down 8 rows.
+    dst += 8;               // Move over 8 columns.
     i -= 8;
   }
+#endif
 
   if (i > 0) {
     TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
@@ -74,9 +102,12 @@ void TransposePlane(const uint8* src, int src_stride,
 }
 
 LIBYUV_API
-void RotatePlane90(const uint8* src, int src_stride,
-                   uint8* dst, int dst_stride,
-                   int width, int height) {
+void RotatePlane90(const uint8_t* src,
+                   int src_stride,
+                   uint8_t* dst,
+                   int dst_stride,
+                   int width,
+                   int height) {
   // Rotate by 90 is a transpose with the source read
   // from bottom to top. So set the source pointer to the end
   // of the buffer and flip the sign of the source stride.
@@ -86,9 +117,12 @@ void RotatePlane90(const uint8* src, int src_stride,
 }
 
 LIBYUV_API
-void RotatePlane270(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height) {
+void RotatePlane270(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
   // Rotate by 270 is a transpose with the destination written
   // from bottom to top. So set the destination pointer to the end
   // of the buffer and flip the sign of the destination stride.
@@ -98,21 +132,24 @@ void RotatePlane270(const uint8* src, int src_stride,
 }
 
 LIBYUV_API
-void RotatePlane180(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height) {
-  // Swap first and last row and mirror the content. Uses a temporary row.
+void RotatePlane180(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
+  // Swap top and bottom row and mirror the content. Uses a temporary row.
   align_buffer_64(row, width);
-  const uint8* src_bot = src + src_stride * (height - 1);
-  uint8* dst_bot = dst + dst_stride * (height - 1);
+  const uint8_t* src_bot = src + src_stride * (height - 1);
+  uint8_t* dst_bot = dst + dst_stride * (height - 1);
   int half_height = (height + 1) >> 1;
   int y;
-  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
-  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  void (*MirrorRow)(const uint8_t* src, uint8_t* dst, int width) = MirrorRow_C;
+  void (*CopyRow)(const uint8_t* src, uint8_t* dst, int width) = CopyRow_C;
 #if defined(HAS_MIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     MirrorRow = MirrorRow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
+    if (IS_ALIGNED(width, 32)) {
       MirrorRow = MirrorRow_NEON;
     }
   }
@@ -133,12 +170,20 @@ void RotatePlane180(const uint8* src, int src_stride,
     }
   }
 #endif
-// TODO(fbarchard): Mirror on mips handle unaligned memory.
-#if defined(HAS_MIRRORROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4)) {
-    MirrorRow = MirrorRow_DSPR2;
+#if defined(HAS_MIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MirrorRow = MirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    MirrorRow = MirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_LASX;
+    }
   }
 #endif
 #if defined(HAS_COPYROW_SSE2)
@@ -161,19 +206,14 @@ void RotatePlane180(const uint8* src, int src_stride,
     CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
   }
 #endif
-#if defined(HAS_COPYROW_MIPS)
-  if (TestCpuFlag(kCpuHasMIPS)) {
-    CopyRow = CopyRow_MIPS;
-  }
-#endif
 
   // Odd height will harmlessly mirror the middle row twice.
   for (y = 0; y < half_height; ++y) {
-    MirrorRow(src, row, width);  // Mirror first row into a buffer
+    CopyRow(src, row, width);        // Copy top row into buffer
+    MirrorRow(src_bot, dst, width);  // Mirror bottom row into top row
+    MirrorRow(row, dst_bot, width);  // Mirror buffer into bottom row
     src += src_stride;
-    MirrorRow(src_bot, dst, width);  // Mirror last row into first row
     dst += dst_stride;
-    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
     src_bot -= src_stride;
     dst_bot -= dst_stride;
   }
@@ -181,15 +221,44 @@ void RotatePlane180(const uint8* src, int src_stride,
 }
 
 LIBYUV_API
-void TransposeUV(const uint8* src, int src_stride,
-                 uint8* dst_a, int dst_stride_a,
-                 uint8* dst_b, int dst_stride_b,
-                 int width, int height) {
+void SplitTransposeUV(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   int i = height;
-  void (*TransposeUVWx8)(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b,
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  void (*TransposeUVWx16)(const uint8_t* src, int src_stride, uint8_t* dst_a,
+                          int dst_stride_a, uint8_t* dst_b, int dst_stride_b,
+                          int width) = TransposeUVWx16_C;
+#elif defined(HAS_TRANSPOSEUVWX16_LSX)
+  void (*TransposeUVWx16)(const uint8_t* src, int src_stride, uint8_t* dst_a,
+                          int dst_stride_a, uint8_t* dst_b, int dst_stride_b,
+                          int width) = TransposeUVWx16_C;
+#else
+  void (*TransposeUVWx8)(const uint8_t* src, int src_stride, uint8_t* dst_a,
+                         int dst_stride_a, uint8_t* dst_b, int dst_stride_b,
                          int width) = TransposeUVWx8_C;
+#endif
+
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    TransposeUVWx16 = TransposeUVWx16_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx16 = TransposeUVWx16_MSA;
+    }
+  }
+#elif defined(HAS_TRANSPOSEUVWX16_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    TransposeUVWx16 = TransposeUVWx16_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx16 = TransposeUVWx16_LSX;
+    }
+  }
+#else
 #if defined(HAS_TRANSPOSEUVWX8_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     TransposeUVWx8 = TransposeUVWx8_NEON;
@@ -203,86 +272,111 @@ void TransposeUV(const uint8* src, int src_stride,
     }
   }
 #endif
-#if defined(HAS_TRANSPOSEUVWX8_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 2) &&
-      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
-    TransposeUVWx8 = TransposeUVWx8_DSPR2;
-  }
-#endif
+#endif /* defined(HAS_TRANSPOSEUVWX16_MSA) */
 
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  // Work through the source in 8x8 tiles.
+  while (i >= 16) {
+    TransposeUVWx16(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                    width);
+    src += 16 * src_stride;  // Go down 16 rows.
+    dst_a += 16;             // Move over 8 columns.
+    dst_b += 16;             // Move over 8 columns.
+    i -= 16;
+  }
+#elif defined(HAS_TRANSPOSEUVWX16_LSX)
+  // Work through the source in 8x8 tiles.
+  while (i >= 16) {
+    TransposeUVWx16(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                    width);
+    src += 16 * src_stride;  // Go down 16 rows.
+    dst_a += 16;             // Move over 8 columns.
+    dst_b += 16;             // Move over 8 columns.
+    i -= 16;
+  }
+#else
   // Work through the source in 8x8 tiles.
   while (i >= 8) {
-    TransposeUVWx8(src, src_stride,
-                   dst_a, dst_stride_a,
-                   dst_b, dst_stride_b,
+    TransposeUVWx8(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
                    width);
-    src += 8 * src_stride;    // Go down 8 rows.
-    dst_a += 8;               // Move over 8 columns.
-    dst_b += 8;               // Move over 8 columns.
+    src += 8 * src_stride;  // Go down 8 rows.
+    dst_a += 8;             // Move over 8 columns.
+    dst_b += 8;             // Move over 8 columns.
     i -= 8;
   }
+#endif
 
   if (i > 0) {
-    TransposeUVWxH_C(src, src_stride,
-                     dst_a, dst_stride_a,
-                     dst_b, dst_stride_b,
+    TransposeUVWxH_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
                      width, i);
   }
 }
 
 LIBYUV_API
-void RotateUV90(const uint8* src, int src_stride,
-                uint8* dst_a, int dst_stride_a,
-                uint8* dst_b, int dst_stride_b,
-                int width, int height) {
+void SplitRotateUV90(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst_a,
+                     int dst_stride_a,
+                     uint8_t* dst_b,
+                     int dst_stride_b,
+                     int width,
+                     int height) {
   src += src_stride * (height - 1);
   src_stride = -src_stride;
 
-  TransposeUV(src, src_stride,
-              dst_a, dst_stride_a,
-              dst_b, dst_stride_b,
-              width, height);
+  SplitTransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width, height);
 }
 
 LIBYUV_API
-void RotateUV270(const uint8* src, int src_stride,
-                 uint8* dst_a, int dst_stride_a,
-                 uint8* dst_b, int dst_stride_b,
-                 int width, int height) {
+void SplitRotateUV270(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   dst_a += dst_stride_a * (width - 1);
   dst_b += dst_stride_b * (width - 1);
   dst_stride_a = -dst_stride_a;
   dst_stride_b = -dst_stride_b;
 
-  TransposeUV(src, src_stride,
-              dst_a, dst_stride_a,
-              dst_b, dst_stride_b,
-              width, height);
+  SplitTransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width, height);
 }
 
 // Rotate 180 is a horizontal and vertical flip.
 LIBYUV_API
-void RotateUV180(const uint8* src, int src_stride,
-                 uint8* dst_a, int dst_stride_a,
-                 uint8* dst_b, int dst_stride_b,
-                 int width, int height) {
+void SplitRotateUV180(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   int i;
-  void (*MirrorUVRow)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) =
-      MirrorUVRow_C;
-#if defined(HAS_MIRRORUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
-    MirrorUVRow = MirrorUVRow_NEON;
+  void (*MirrorSplitUVRow)(const uint8_t* src, uint8_t* dst_u, uint8_t* dst_v,
+                           int width) = MirrorSplitUVRow_C;
+#if defined(HAS_MIRRORSPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorSplitUVRow = MirrorSplitUVRow_NEON;
   }
 #endif
-#if defined(HAS_MIRRORUVROW_SSSE3)
+#if defined(HAS_MIRRORSPLITUVROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
-    MirrorUVRow = MirrorUVRow_SSSE3;
+    MirrorSplitUVRow = MirrorSplitUVRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORSPLITUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 32)) {
+    MirrorSplitUVRow = MirrorSplitUVRow_MSA;
   }
 #endif
-#if defined(HAS_MIRRORUVROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
-    MirrorUVRow = MirrorUVRow_DSPR2;
+#if defined(HAS_MIRRORSPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && IS_ALIGNED(width, 32)) {
+    MirrorSplitUVRow = MirrorSplitUVRow_LSX;
   }
 #endif
 
@@ -290,17 +384,66 @@ void RotateUV180(const uint8* src, int src_stride,
   dst_b += dst_stride_b * (height - 1);
 
   for (i = 0; i < height; ++i) {
-    MirrorUVRow(src, dst_a, dst_b, width);
+    MirrorSplitUVRow(src, dst_a, dst_b, width);
     src += src_stride;
     dst_a -= dst_stride_a;
     dst_b -= dst_stride_b;
   }
 }
 
+// Rotate UV and split into planar.
+// width and height expected to be half size for NV12
 LIBYUV_API
-int RotatePlane(const uint8* src, int src_stride,
-                uint8* dst, int dst_stride,
-                int width, int height,
+int SplitRotateUV(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int width,
+                  int height,
+                  enum RotationMode mode) {
+  if (!src_uv || width <= 0 || height == 0 || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      SplitUVPlane(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                   dst_stride_v, width, height);
+      return 0;
+    case kRotate90:
+      SplitRotateUV90(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                      dst_stride_v, width, height);
+      return 0;
+    case kRotate270:
+      SplitRotateUV270(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, width, height);
+      return 0;
+    case kRotate180:
+      SplitRotateUV180(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int RotatePlane(const uint8_t* src,
+                int src_stride,
+                uint8_t* dst,
+                int dst_stride,
+                int width,
+                int height,
                 enum RotationMode mode) {
   if (!src || width <= 0 || height == 0 || !dst) {
     return -1;
@@ -316,24 +459,130 @@ int RotatePlane(const uint8* src, int src_stride,
   switch (mode) {
     case kRotate0:
       // copy frame
-      CopyPlane(src, src_stride,
-                dst, dst_stride,
-                width, height);
+      CopyPlane(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+void TransposePlane_16(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width,
+                       int height) {
+  int i = height;
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8_16_C(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;  // Go down 8 rows.
+    dst += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+
+  if (i > 0) {
+    TransposeWxH_16_C(src, src_stride, dst, dst_stride, width, i);
+  }
+}
+
+static void RotatePlane90_16(const uint16_t* src,
+                             int src_stride,
+                             uint16_t* dst,
+                             int dst_stride,
+                             int width,
+                             int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane_16(src, src_stride, dst, dst_stride, width, height);
+}
+
+static void RotatePlane270_16(const uint16_t* src,
+                              int src_stride,
+                              uint16_t* dst,
+                              int dst_stride,
+                              int width,
+                              int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane_16(src, src_stride, dst, dst_stride, width, height);
+}
+
+static void RotatePlane180_16(const uint16_t* src,
+                              int src_stride,
+                              uint16_t* dst,
+                              int dst_stride,
+                              int width,
+                              int height) {
+  // Swap top and bottom row and mirror the content. Uses a temporary row.
+  align_buffer_64_16(row, width);
+  const uint16_t* src_bot = src + src_stride * (height - 1);
+  uint16_t* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    CopyRow_16_C(src, row, width);        // Copy top row into buffer
+    MirrorRow_16_C(src_bot, dst, width);  // Mirror bottom row into top row
+    MirrorRow_16_C(row, dst_bot, width);  // Mirror buffer into bottom row
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64_16(row);
+}
+
+LIBYUV_API
+int RotatePlane_16(const uint16_t* src,
+                   int src_stride,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int width,
+                   int height,
+                   enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane_16(src, src_stride, dst, dst_stride, width, height);
       return 0;
     case kRotate90:
-      RotatePlane90(src, src_stride,
-                    dst, dst_stride,
-                    width, height);
+      RotatePlane90_16(src, src_stride, dst, dst_stride, width, height);
       return 0;
     case kRotate270:
-      RotatePlane270(src, src_stride,
-                     dst, dst_stride,
-                     width, height);
+      RotatePlane270_16(src, src_stride, dst, dst_stride, width, height);
       return 0;
     case kRotate180:
-      RotatePlane180(src, src_stride,
-                     dst, dst_stride,
-                     width, height);
+      RotatePlane180_16(src, src_stride, dst, dst_stride, width, height);
       return 0;
     default:
       break;
@@ -342,17 +591,24 @@ int RotatePlane(const uint8* src, int src_stride,
 }
 
 LIBYUV_API
-int I420Rotate(const uint8* src_y, int src_stride_y,
-               const uint8* src_u, int src_stride_u,
-               const uint8* src_v, int src_stride_v,
-               uint8* dst_y, int dst_stride_y,
-               uint8* dst_u, int dst_stride_u,
-               uint8* dst_v, int dst_stride_v,
-               int width, int height,
+int I420Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
                enum RotationMode mode) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 ||
+  if ((!src_y && dst_y) || !src_u || !src_v || width <= 0 || height == 0 ||
       !dst_y || !dst_u || !dst_v) {
     return -1;
   }
@@ -372,45 +628,29 @@ int I420Rotate(const uint8* src_y, int src_stride_y,
   switch (mode) {
     case kRotate0:
       // copy frame
-      return I420Copy(src_y, src_stride_y,
-                      src_u, src_stride_u,
-                      src_v, src_stride_v,
-                      dst_y, dst_stride_y,
-                      dst_u, dst_stride_u,
-                      dst_v, dst_stride_v,
-                      width, height);
+      return I420Copy(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                      src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                      dst_v, dst_stride_v, width, height);
     case kRotate90:
-      RotatePlane90(src_y, src_stride_y,
-                    dst_y, dst_stride_y,
-                    width, height);
-      RotatePlane90(src_u, src_stride_u,
-                    dst_u, dst_stride_u,
-                    halfwidth, halfheight);
-      RotatePlane90(src_v, src_stride_v,
-                    dst_v, dst_stride_v,
-                    halfwidth, halfheight);
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                    halfheight);
+      RotatePlane90(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                    halfheight);
       return 0;
     case kRotate270:
-      RotatePlane270(src_y, src_stride_y,
-                     dst_y, dst_stride_y,
-                     width, height);
-      RotatePlane270(src_u, src_stride_u,
-                     dst_u, dst_stride_u,
-                     halfwidth, halfheight);
-      RotatePlane270(src_v, src_stride_v,
-                     dst_v, dst_stride_v,
-                     halfwidth, halfheight);
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane270(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                     halfheight);
+      RotatePlane270(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                     halfheight);
       return 0;
     case kRotate180:
-      RotatePlane180(src_y, src_stride_y,
-                     dst_y, dst_stride_y,
-                     width, height);
-      RotatePlane180(src_u, src_stride_u,
-                     dst_u, dst_stride_u,
-                     halfwidth, halfheight);
-      RotatePlane180(src_v, src_stride_v,
-                     dst_v, dst_stride_v,
-                     halfwidth, halfheight);
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane180(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                     halfheight);
+      RotatePlane180(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                     halfheight);
       return 0;
     default:
       break;
@@ -418,18 +658,171 @@ int I420Rotate(const uint8* src_y, int src_stride_y,
   return -1;
 }
 
+// I422 has half width x full height UV planes, so rotate by 90 and 270
+// require scaling to maintain 422 subsampling.
 LIBYUV_API
-int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
-                     const uint8* src_uv, int src_stride_uv,
-                     uint8* dst_y, int dst_stride_y,
-                     uint8* dst_u, int dst_stride_u,
-                     uint8* dst_v, int dst_stride_v,
-                     int width, int height,
+int I422Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // Copy frame
+      CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+      CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+      return 0;
+
+      // Note on temporary Y plane for UV.
+      // Rotation of UV first fits within the Y destination plane rows.
+      // Y plane is width x height
+      // Y plane rotated is height x width
+      // UV plane is (width / 2) x height
+      // UV plane rotated is height x (width / 2)
+      // UV plane rotated+scaled is (height / 2) x width.
+      // UV plane rotated is a temporary that fits within the Y plane rotated.
+
+    case kRotate90:
+      RotatePlane90(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                    height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
+                 halfheight, width, kFilterBilinear);
+      RotatePlane90(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                    height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
+                 halfheight, width, kFilterLinear);
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                     height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
+                 halfheight, width, kFilterBilinear);
+      RotatePlane270(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                     height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
+                 halfheight, width, kFilterLinear);
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane180(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                     height);
+      RotatePlane180(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                     height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I444Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      RotatePlane90(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane270(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      RotatePlane270(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane180(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      RotatePlane180(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int NV12ToI420Rotate(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_y,
+                     int dst_stride_y,
+                     uint8_t* dst_u,
+                     int dst_stride_u,
+                     uint8_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
                      enum RotationMode mode) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_uv || width <= 0 || height == 0 ||
-      !dst_y || !dst_u || !dst_v) {
+  if (!src_y || !src_uv || width <= 0 || height == 0 || !dst_y || !dst_u ||
+      !dst_v) {
     return -1;
   }
 
@@ -446,38 +839,337 @@ int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
   switch (mode) {
     case kRotate0:
       // copy frame
-      return NV12ToI420(src_y, src_stride_y,
-                        src_uv, src_stride_uv,
-                        dst_y, dst_stride_y,
-                        dst_u, dst_stride_u,
-                        dst_v, dst_stride_v,
+      return NV12ToI420(src_y, src_stride_y, src_uv, src_stride_uv, dst_y,
+                        dst_stride_y, dst_u, dst_stride_u, dst_v, dst_stride_v,
                         width, height);
     case kRotate90:
-      RotatePlane90(src_y, src_stride_y,
-                    dst_y, dst_stride_y,
-                    width, height);
-      RotateUV90(src_uv, src_stride_uv,
-                 dst_u, dst_stride_u,
-                 dst_v, dst_stride_v,
-                 halfwidth, halfheight);
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      SplitRotateUV90(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                      dst_stride_v, halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      SplitRotateUV270(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      SplitRotateUV180(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+static void SplitPixels(const uint8_t* src_u,
+                        int src_pixel_stride_uv,
+                        uint8_t* dst_u,
+                        int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst_u = *src_u;
+    ++dst_u;
+    src_u += src_pixel_stride_uv;
+  }
+}
+
+// Convert Android420 to I420 with Rotate
+LIBYUV_API
+int Android420ToI420Rotate(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_y,
+                           int dst_stride_y,
+                           uint8_t* dst_u,
+                           int dst_stride_u,
+                           uint8_t* dst_v,
+                           int dst_stride_v,
+                           int width,
+                           int height,
+                           enum RotationMode rotation) {
+  int y;
+  const ptrdiff_t vu_off = src_v - src_u;
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    RotatePlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height,
+                rotation);
+  }
+
+  // Copy UV planes - I420
+  if (src_pixel_stride_uv == 1) {
+    RotatePlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight,
+                rotation);
+    RotatePlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight,
+                rotation);
+    return 0;
+  }
+  // Split UV planes - NV21
+  if (src_pixel_stride_uv == 2 && vu_off == -1 &&
+      src_stride_u == src_stride_v) {
+    SplitRotateUV(src_v, src_stride_v, dst_v, dst_stride_v, dst_u, dst_stride_u,
+                  halfwidth, halfheight, rotation);
+    return 0;
+  }
+  // Split UV planes - NV12
+  if (src_pixel_stride_uv == 2 && vu_off == 1 && src_stride_u == src_stride_v) {
+    SplitRotateUV(src_u, src_stride_u, dst_u, dst_stride_u, dst_v, dst_stride_v,
+                  halfwidth, halfheight, rotation);
+    return 0;
+  }
+
+  if (rotation == 0) {
+    for (y = 0; y < halfheight; ++y) {
+      SplitPixels(src_u, src_pixel_stride_uv, dst_u, halfwidth);
+      SplitPixels(src_v, src_pixel_stride_uv, dst_v, halfwidth);
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    return 0;
+  }
+  // unsupported type and/or rotation.
+  return -1;
+}
+
+LIBYUV_API
+int I010Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v || dst_stride_y < 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I010Copy(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                      src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                      dst_v, dst_stride_v, width, height);
+    case kRotate90:
+      RotatePlane90_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                       halfheight);
+      RotatePlane90_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                       halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane270_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                        halfheight);
+      RotatePlane270_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                        halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane180_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                        halfheight);
+      RotatePlane180_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                        halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+// I210 has half width x full height UV planes, so rotate by 90 and 270
+// require scaling to maintain 422 subsampling.
+LIBYUV_API
+int I210Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // Copy frame
+      CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+      CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+      return 0;
+
+      // Note on temporary Y plane for UV.
+      // Rotation of UV first fits within the Y destination plane rows.
+      // Y plane is width x height
+      // Y plane rotated is height x width
+      // UV plane is (width / 2) x height
+      // UV plane rotated is height x (width / 2)
+      // UV plane rotated+scaled is (height / 2) x width.
+      // UV plane rotated is a temporary that fits within the Y plane rotated.
+
+    case kRotate90:
+      RotatePlane90_16(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                       height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
+                    halfheight, width, kFilterBilinear);
+      RotatePlane90_16(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                       height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
+                    halfheight, width, kFilterLinear);
+      RotatePlane90_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270_16(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                        height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
+                    halfheight, width, kFilterBilinear);
+      RotatePlane270_16(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                        height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
+                    halfheight, width, kFilterLinear);
+      RotatePlane270_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      return 0;
+    case kRotate180:
+      RotatePlane180_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane180_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                        height);
+      RotatePlane180_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                        height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I410Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v || dst_stride_y < 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90_16(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      RotatePlane90_16(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
       return 0;
     case kRotate270:
-      RotatePlane270(src_y, src_stride_y,
-                     dst_y, dst_stride_y,
-                     width, height);
-      RotateUV270(src_uv, src_stride_uv,
-                  dst_u, dst_stride_u,
-                  dst_v, dst_stride_v,
-                  halfwidth, halfheight);
+      RotatePlane270_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane270_16(src_u, src_stride_u, dst_u, dst_stride_u, width,
+                        height);
+      RotatePlane270_16(src_v, src_stride_v, dst_v, dst_stride_v, width,
+                        height);
       return 0;
     case kRotate180:
-      RotatePlane180(src_y, src_stride_y,
-                     dst_y, dst_stride_y,
-                     width, height);
-      RotateUV180(src_uv, src_stride_uv,
-                  dst_u, dst_stride_u,
-                  dst_v, dst_stride_v,
-                  halfwidth, halfheight);
+      RotatePlane180_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane180_16(src_u, src_stride_u, dst_u, dst_stride_u, width,
+                        height);
+      RotatePlane180_16(src_v, src_stride_v, dst_v, dst_stride_v, width,
+                        height);
       return 0;
     default:
       break;
diff --git a/media/libyuv/source/rotate_any.cc b/media/libyuv/source/rotate_any.cc
index 31a74c3155..88ca78765a 100644
--- a/media/libyuv/source/rotate_any.cc
+++ b/media/libyuv/source/rotate_any.cc
@@ -18,16 +18,16 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#define TANY(NAMEANY, TPOS_SIMD, MASK)                                         \
-    void NAMEANY(const uint8* src, int src_stride,                             \
-                 uint8* dst, int dst_stride, int width) {                      \
-      int r = width & MASK;                                                    \
-      int n = width - r;                                                       \
-      if (n > 0) {                                                             \
-        TPOS_SIMD(src, src_stride, dst, dst_stride, n);                        \
-      }                                                                        \
-      TransposeWx8_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r);\
-    }
+#define TANY(NAMEANY, TPOS_SIMD, MASK)                                        \
+  void NAMEANY(const uint8_t* src, int src_stride, uint8_t* dst,              \
+               int dst_stride, int width) {                                   \
+    int r = width & MASK;                                                     \
+    int n = width - r;                                                        \
+    if (n > 0) {                                                              \
+      TPOS_SIMD(src, src_stride, dst, dst_stride, n);                         \
+    }                                                                         \
+    TransposeWx8_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r); \
+  }
 
 #ifdef HAS_TRANSPOSEWX8_NEON
 TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, 7)
@@ -38,25 +38,26 @@ TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, 7)
 #ifdef HAS_TRANSPOSEWX8_FAST_SSSE3
 TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15)
 #endif
-#ifdef HAS_TRANSPOSEWX8_DSPR2
-TANY(TransposeWx8_Any_DSPR2, TransposeWx8_DSPR2, 7)
+#ifdef HAS_TRANSPOSEWX16_MSA
+TANY(TransposeWx16_Any_MSA, TransposeWx16_MSA, 15)
+#endif
+#ifdef HAS_TRANSPOSEWX16_LSX
+TANY(TransposeWx16_Any_LSX, TransposeWx16_LSX, 15)
 #endif
 #undef TANY
 
 #define TUVANY(NAMEANY, TPOS_SIMD, MASK)                                       \
-    void NAMEANY(const uint8* src, int src_stride,                             \
-                uint8* dst_a, int dst_stride_a,                                \
-                uint8* dst_b, int dst_stride_b, int width) {                   \
-      int r = width & MASK;                                                    \
-      int n = width - r;                                                       \
-      if (n > 0) {                                                             \
-        TPOS_SIMD(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,   \
-                  n);                                                          \
-      }                                                                        \
-      TransposeUVWx8_C(src + n * 2, src_stride,                                \
-                       dst_a + n * dst_stride_a, dst_stride_a,                 \
-                       dst_b + n * dst_stride_b, dst_stride_b, r);             \
-    }
+  void NAMEANY(const uint8_t* src, int src_stride, uint8_t* dst_a,             \
+               int dst_stride_a, uint8_t* dst_b, int dst_stride_b,             \
+               int width) {                                                    \
+    int r = width & MASK;                                                      \
+    int n = width - r;                                                         \
+    if (n > 0) {                                                               \
+      TPOS_SIMD(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, n); \
+    }                                                                          \
+    TransposeUVWx8_C(src + n * 2, src_stride, dst_a + n * dst_stride_a,        \
+                     dst_stride_a, dst_b + n * dst_stride_b, dst_stride_b, r); \
+  }
 
 #ifdef HAS_TRANSPOSEUVWX8_NEON
 TUVANY(TransposeUVWx8_Any_NEON, TransposeUVWx8_NEON, 7)
@@ -64,8 +65,11 @@ TUVANY(TransposeUVWx8_Any_NEON, TransposeUVWx8_NEON, 7)
 #ifdef HAS_TRANSPOSEUVWX8_SSE2
 TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7)
 #endif
-#ifdef HAS_TRANSPOSEUVWX8_DSPR2
-TUVANY(TransposeUVWx8_Any_DSPR2, TransposeUVWx8_DSPR2, 7)
+#ifdef HAS_TRANSPOSEUVWX16_MSA
+TUVANY(TransposeUVWx16_Any_MSA, TransposeUVWx16_MSA, 7)
+#endif
+#ifdef HAS_TRANSPOSEUVWX16_LSX
+TUVANY(TransposeUVWx16_Any_LSX, TransposeUVWx16_LSX, 7)
 #endif
 #undef TUVANY
 
@@ -73,8 +77,3 @@ TUVANY(TransposeUVWx8_Any_DSPR2, TransposeUVWx8_DSPR2, 7)
 }  // extern "C"
 }  // namespace libyuv
 #endif
-
-
-
-
-
diff --git a/media/libyuv/source/rotate_argb.cc b/media/libyuv/source/rotate_argb.cc
index 787c0ad1be..28226210e1 100644
--- a/media/libyuv/source/rotate_argb.cc
+++ b/media/libyuv/source/rotate_argb.cc
@@ -8,96 +8,126 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/rotate.h"
+#include "libyuv/rotate_argb.h"
 
-#include "libyuv/cpu_id.h"
 #include "libyuv/convert.h"
+#include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
 #include "libyuv/row.h"
+#include "libyuv/scale_row.h" /* for ScaleARGBRowDownEven_ */
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// ARGBScale has a function to copy pixels to a row, striding each source
-// pixel by a constant.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(_M_IX86) || \
-    (defined(__x86_64__) && !defined(__native_client__)) || defined(__i386__))
-#define HAS_SCALEARGBROWDOWNEVEN_SSE2
-void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride,
-                               int src_stepx, uint8* dst_ptr, int dst_width);
-#endif
-#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
-    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
-#define HAS_SCALEARGBROWDOWNEVEN_NEON
-void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride,
-                               int src_stepx, uint8* dst_ptr, int dst_width);
-#endif
-
-void ScaleARGBRowDownEven_C(const uint8* src_ptr, int,
-                            int src_stepx, uint8* dst_ptr, int dst_width);
-
-static void ARGBTranspose(const uint8* src, int src_stride,
-                          uint8* dst, int dst_stride, int width, int height) {
+static int ARGBTranspose(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
   int i;
-  int src_pixel_step = src_stride >> 2;
-  void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
-      int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C;
+  int src_pixel_step = src_stride_argb >> 2;
+  void (*ScaleARGBRowDownEven)(
+      const uint8_t* src_argb, ptrdiff_t src_stride_argb, int src_step,
+      uint8_t* dst_argb, int dst_width) = ScaleARGBRowDownEven_C;
+  // Check stride is a multiple of 4.
+  if (src_stride_argb & 3) {
+    return -1;
+  }
 #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4)) {  // Width of dest.
-    ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_SSE2;
+    if (IS_ALIGNED(height, 4)) {  // Width of dest.
+      ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+    }
   }
 #endif
 #if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4)) {  // Width of dest.
-    ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_NEON;
+    if (IS_ALIGNED(height, 4)) {  // Width of dest.
+      ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_MSA;
+    if (IS_ALIGNED(height, 4)) {  // Width of dest.
+      ScaleARGBRowDownEven = ScaleARGBRowDownEven_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_LSX;
+    if (IS_ALIGNED(height, 4)) {  // Width of dest.
+      ScaleARGBRowDownEven = ScaleARGBRowDownEven_LSX;
+    }
   }
 #endif
 
   for (i = 0; i < width; ++i) {  // column of source to row of dest.
-    ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height);
-    dst += dst_stride;
-    src += 4;
+    ScaleARGBRowDownEven(src_argb, 0, src_pixel_step, dst_argb, height);
+    dst_argb += dst_stride_argb;
+    src_argb += 4;
   }
+  return 0;
 }
 
-void ARGBRotate90(const uint8* src, int src_stride,
-                  uint8* dst, int dst_stride, int width, int height) {
+static int ARGBRotate90(const uint8_t* src_argb,
+                        int src_stride_argb,
+                        uint8_t* dst_argb,
+                        int dst_stride_argb,
+                        int width,
+                        int height) {
   // Rotate by 90 is a ARGBTranspose with the source read
   // from bottom to top. So set the source pointer to the end
   // of the buffer and flip the sign of the source stride.
-  src += src_stride * (height - 1);
-  src_stride = -src_stride;
-  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+  src_argb += src_stride_argb * (height - 1);
+  src_stride_argb = -src_stride_argb;
+  return ARGBTranspose(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                       width, height);
 }
 
-void ARGBRotate270(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride, int width, int height) {
+static int ARGBRotate270(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
   // Rotate by 270 is a ARGBTranspose with the destination written
   // from bottom to top. So set the destination pointer to the end
   // of the buffer and flip the sign of the destination stride.
-  dst += dst_stride * (width - 1);
-  dst_stride = -dst_stride;
-  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+  dst_argb += dst_stride_argb * (width - 1);
+  dst_stride_argb = -dst_stride_argb;
+  return ARGBTranspose(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                       width, height);
 }
 
-void ARGBRotate180(const uint8* src, int src_stride,
-                   uint8* dst, int dst_stride, int width, int height) {
+static int ARGBRotate180(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
   // Swap first and last row and mirror the content. Uses a temporary row.
   align_buffer_64(row, width * 4);
-  const uint8* src_bot = src + src_stride * (height - 1);
-  uint8* dst_bot = dst + dst_stride * (height - 1);
+  const uint8_t* src_bot = src_argb + src_stride_argb * (height - 1);
+  uint8_t* dst_bot = dst_argb + dst_stride_argb * (height - 1);
   int half_height = (height + 1) >> 1;
   int y;
-  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+  void (*ARGBMirrorRow)(const uint8_t* src_argb, uint8_t* dst_argb, int width) =
       ARGBMirrorRow_C;
-  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  void (*CopyRow)(const uint8_t* src_argb, uint8_t* dst_argb, int width) =
+      CopyRow_C;
 #if defined(HAS_ARGBMIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
-    if (IS_ALIGNED(width, 4)) {
+    if (IS_ALIGNED(width, 8)) {
       ARGBMirrorRow = ARGBMirrorRow_NEON;
     }
   }
@@ -118,6 +148,22 @@ void ARGBRotate180(const uint8* src, int src_stride,
     }
   }
 #endif
+#if defined(HAS_ARGBMIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_COPYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
@@ -138,28 +184,28 @@ void ARGBRotate180(const uint8* src, int src_stride,
     CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
   }
 #endif
-#if defined(HAS_COPYROW_MIPS)
-  if (TestCpuFlag(kCpuHasMIPS)) {
-    CopyRow = CopyRow_MIPS;
-  }
-#endif
 
   // Odd height will harmlessly mirror the middle row twice.
   for (y = 0; y < half_height; ++y) {
-    ARGBMirrorRow(src, row, width);  // Mirror first row into a buffer
-    ARGBMirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    ARGBMirrorRow(src_argb, row, width);      // Mirror first row into a buffer
+    ARGBMirrorRow(src_bot, dst_argb, width);  // Mirror last row into first row
     CopyRow(row, dst_bot, width * 4);  // Copy first mirrored row into last
-    src += src_stride;
-    dst += dst_stride;
-    src_bot -= src_stride;
-    dst_bot -= dst_stride;
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+    src_bot -= src_stride_argb;
+    dst_bot -= dst_stride_argb;
   }
   free_aligned_buffer_64(row);
+  return 0;
 }
 
 LIBYUV_API
-int ARGBRotate(const uint8* src_argb, int src_stride_argb,
-               uint8* dst_argb, int dst_stride_argb, int width, int height,
+int ARGBRotate(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height,
                enum RotationMode mode) {
   if (!src_argb || width <= 0 || height == 0 || !dst_argb) {
     return -1;
@@ -175,24 +221,17 @@ int ARGBRotate(const uint8* src_argb, int src_stride_argb,
   switch (mode) {
     case kRotate0:
       // copy frame
-      return ARGBCopy(src_argb, src_stride_argb,
-                      dst_argb, dst_stride_argb,
+      return ARGBCopy(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
                       width, height);
     case kRotate90:
-      ARGBRotate90(src_argb, src_stride_argb,
-                   dst_argb, dst_stride_argb,
-                   width, height);
-      return 0;
+      return ARGBRotate90(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                          width, height);
     case kRotate270:
-      ARGBRotate270(src_argb, src_stride_argb,
-                    dst_argb, dst_stride_argb,
-                    width, height);
-      return 0;
+      return ARGBRotate270(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                           width, height);
     case kRotate180:
-      ARGBRotate180(src_argb, src_stride_argb,
-                    dst_argb, dst_stride_argb,
-                    width, height);
-      return 0;
+      return ARGBRotate180(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                           width, height);
     default:
       break;
   }
diff --git a/media/libyuv/source/rotate_common.cc b/media/libyuv/source/rotate_common.cc
index b33a9a0c6e..4b496d1b35 100644
--- a/media/libyuv/source/rotate_common.cc
+++ b/media/libyuv/source/rotate_common.cc
@@ -8,16 +8,19 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/row.h"
 #include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-void TransposeWx8_C(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride, int width) {
+void TransposeWx8_C(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width) {
   int i;
   for (i = 0; i < width; ++i) {
     dst[0] = src[0 * src_stride];
@@ -33,9 +36,13 @@ void TransposeWx8_C(const uint8* src, int src_stride,
   }
 }
 
-void TransposeUVWx8_C(const uint8* src, int src_stride,
-                      uint8* dst_a, int dst_stride_a,
-                      uint8* dst_b, int dst_stride_b, int width) {
+void TransposeUVWx8_C(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width) {
   int i;
   for (i = 0; i < width; ++i) {
     dst_a[0] = src[0 * src_stride + 0];
@@ -60,9 +67,12 @@ void TransposeUVWx8_C(const uint8* src, int src_stride,
   }
 }
 
-void TransposeWxH_C(const uint8* src, int src_stride,
-                    uint8* dst, int dst_stride,
-                    int width, int height) {
+void TransposeWxH_C(const uint8_t* src,
+                    int src_stride,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
   int i;
   for (i = 0; i < width; ++i) {
     int j;
@@ -72,20 +82,147 @@ void TransposeWxH_C(const uint8* src, int src_stride,
   }
 }
 
-void TransposeUVWxH_C(const uint8* src, int src_stride,
-                      uint8* dst_a, int dst_stride_a,
-                      uint8* dst_b, int dst_stride_b,
-                      int width, int height) {
+void TransposeUVWxH_C(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   int i;
   for (i = 0; i < width * 2; i += 2) {
     int j;
     for (j = 0; j < height; ++j) {
-      dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)];
-      dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1];
+      dst_a[((i >> 1) * dst_stride_a) + j] = src[i + (j * src_stride)];
+      dst_b[((i >> 1) * dst_stride_b) + j] = src[i + (j * src_stride) + 1];
+    }
+  }
+}
+
+void TransposeWx8_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+void TransposeUVWx8_16_C(const uint16_t* src,
+                         int src_stride,
+                         uint16_t* dst_a,
+                         int dst_stride_a,
+                         uint16_t* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+void TransposeWxH_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width,
+                       int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
     }
   }
 }
 
+// Transpose 32 bit values (ARGB)
+void Transpose4x4_32_C(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width) {
+  const uint8_t* src1 = src + src_stride;
+  const uint8_t* src2 = src1 + src_stride;
+  const uint8_t* src3 = src2 + src_stride;
+  uint8_t* dst1 = dst + dst_stride;
+  uint8_t* dst2 = dst1 + dst_stride;
+  uint8_t* dst3 = dst2 + dst_stride;
+  int i;
+  for (i = 0; i < width; i += 4) {
+    uint32_t p00 = ((uint32_t*)(src))[0];
+    uint32_t p10 = ((uint32_t*)(src))[1];
+    uint32_t p20 = ((uint32_t*)(src))[2];
+    uint32_t p30 = ((uint32_t*)(src))[3];
+    uint32_t p01 = ((uint32_t*)(src1))[0];
+    uint32_t p11 = ((uint32_t*)(src1))[1];
+    uint32_t p21 = ((uint32_t*)(src1))[2];
+    uint32_t p31 = ((uint32_t*)(src1))[3];
+    uint32_t p02 = ((uint32_t*)(src2))[0];
+    uint32_t p12 = ((uint32_t*)(src2))[1];
+    uint32_t p22 = ((uint32_t*)(src2))[2];
+    uint32_t p32 = ((uint32_t*)(src2))[3];
+    uint32_t p03 = ((uint32_t*)(src3))[0];
+    uint32_t p13 = ((uint32_t*)(src3))[1];
+    uint32_t p23 = ((uint32_t*)(src3))[2];
+    uint32_t p33 = ((uint32_t*)(src3))[3];
+    ((uint32_t*)(dst))[0] = p00;
+    ((uint32_t*)(dst))[1] = p01;
+    ((uint32_t*)(dst))[2] = p02;
+    ((uint32_t*)(dst))[3] = p03;
+    ((uint32_t*)(dst1))[0] = p10;
+    ((uint32_t*)(dst1))[1] = p11;
+    ((uint32_t*)(dst1))[2] = p12;
+    ((uint32_t*)(dst1))[3] = p13;
+    ((uint32_t*)(dst2))[0] = p20;
+    ((uint32_t*)(dst2))[1] = p21;
+    ((uint32_t*)(dst2))[2] = p22;
+    ((uint32_t*)(dst2))[3] = p23;
+    ((uint32_t*)(dst3))[0] = p30;
+    ((uint32_t*)(dst3))[1] = p31;
+    ((uint32_t*)(dst3))[2] = p32;
+    ((uint32_t*)(dst3))[3] = p33;
+    src += src_stride * 4;  // advance 4 rows
+    src1 += src_stride * 4;
+    src2 += src_stride * 4;
+    src3 += src_stride * 4;
+    dst += 4 * 4;  // advance 4 columns
+    dst1 += 4 * 4;
+    dst2 += 4 * 4;
+    dst3 += 4 * 4;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/media/libyuv/source/rotate_gcc.cc b/media/libyuv/source/rotate_gcc.cc
index cbe870caa7..1a3f8cbbda 100644
--- a/media/libyuv/source/rotate_gcc.cc
+++ b/media/libyuv/source/rotate_gcc.cc
@@ -8,8 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/row.h"
 #include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -17,347 +17,352 @@ extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 // Transpose 8x8. 32 or 64 bit, but not NaCL for 64 bit.
 #if defined(HAS_TRANSPOSEWX8_SSSE3)
-void TransposeWx8_SSSE3(const uint8* src, int src_stride,
-                        uint8* dst, int dst_stride, int width) {
-  asm volatile (
-    // Read in the data from the source pointer.
-    // First round of bit swap.
-    LABELALIGN
-  "1:                                            \n"
-    "movq       (%0),%%xmm0                      \n"
-    "movq       (%0,%3),%%xmm1                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "punpcklbw  %%xmm1,%%xmm0                    \n"
-    "movq       (%0),%%xmm2                      \n"
-    "movdqa     %%xmm0,%%xmm1                    \n"
-    "palignr    $0x8,%%xmm1,%%xmm1               \n"
-    "movq       (%0,%3),%%xmm3                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "punpcklbw  %%xmm3,%%xmm2                    \n"
-    "movdqa     %%xmm2,%%xmm3                    \n"
-    "movq       (%0),%%xmm4                      \n"
-    "palignr    $0x8,%%xmm3,%%xmm3               \n"
-    "movq       (%0,%3),%%xmm5                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "punpcklbw  %%xmm5,%%xmm4                    \n"
-    "movdqa     %%xmm4,%%xmm5                    \n"
-    "movq       (%0),%%xmm6                      \n"
-    "palignr    $0x8,%%xmm5,%%xmm5               \n"
-    "movq       (%0,%3),%%xmm7                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "punpcklbw  %%xmm7,%%xmm6                    \n"
-    "neg        %3                               \n"
-    "movdqa     %%xmm6,%%xmm7                    \n"
-    "lea        0x8(%0,%3,8),%0                  \n"
-    "palignr    $0x8,%%xmm7,%%xmm7               \n"
-    "neg        %3                               \n"
-     // Second round of bit swap.
-    "punpcklwd  %%xmm2,%%xmm0                    \n"
-    "punpcklwd  %%xmm3,%%xmm1                    \n"
-    "movdqa     %%xmm0,%%xmm2                    \n"
-    "movdqa     %%xmm1,%%xmm3                    \n"
-    "palignr    $0x8,%%xmm2,%%xmm2               \n"
-    "palignr    $0x8,%%xmm3,%%xmm3               \n"
-    "punpcklwd  %%xmm6,%%xmm4                    \n"
-    "punpcklwd  %%xmm7,%%xmm5                    \n"
-    "movdqa     %%xmm4,%%xmm6                    \n"
-    "movdqa     %%xmm5,%%xmm7                    \n"
-    "palignr    $0x8,%%xmm6,%%xmm6               \n"
-    "palignr    $0x8,%%xmm7,%%xmm7               \n"
-    // Third round of bit swap.
-    // Write to the destination pointer.
-    "punpckldq  %%xmm4,%%xmm0                    \n"
-    "movq       %%xmm0,(%1)                      \n"
-    "movdqa     %%xmm0,%%xmm4                    \n"
-    "palignr    $0x8,%%xmm4,%%xmm4               \n"
-    "movq       %%xmm4,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm6,%%xmm2                    \n"
-    "movdqa     %%xmm2,%%xmm6                    \n"
-    "movq       %%xmm2,(%1)                      \n"
-    "palignr    $0x8,%%xmm6,%%xmm6               \n"
-    "punpckldq  %%xmm5,%%xmm1                    \n"
-    "movq       %%xmm6,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "movdqa     %%xmm1,%%xmm5                    \n"
-    "movq       %%xmm1,(%1)                      \n"
-    "palignr    $0x8,%%xmm5,%%xmm5               \n"
-    "movq       %%xmm5,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm7,%%xmm3                    \n"
-    "movq       %%xmm3,(%1)                      \n"
-    "movdqa     %%xmm3,%%xmm7                    \n"
-    "palignr    $0x8,%%xmm7,%%xmm7               \n"
-    "sub        $0x8,%2                          \n"
-    "movq       %%xmm7,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "jg         1b                               \n"
-    : "+r"(src),    // %0
-      "+r"(dst),    // %1
-      "+r"(width)   // %2
-    : "r"((intptr_t)(src_stride)),  // %3
-      "r"((intptr_t)(dst_stride))   // %4
-    : "memory", "cc",
-      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void TransposeWx8_SSSE3(const uint8_t* src,
+                        int src_stride,
+                        uint8_t* dst,
+                        int dst_stride,
+                        int width) {
+  asm volatile(
+      // Read in the data from the source pointer.
+      // First round of bit swap.
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movq        (%0,%3),%%xmm1                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "movq        (%0),%%xmm2                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm1            \n"
+      "movq        (%0,%3),%%xmm3                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movq        (%0),%%xmm4                   \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "movq        (%0,%3),%%xmm5                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm5,%%xmm4                 \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "movq        (%0),%%xmm6                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "movq        (%0,%3),%%xmm7                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm7,%%xmm6                 \n"
+      "neg         %3                            \n"
+      "movdqa      %%xmm6,%%xmm7                 \n"
+      "lea         0x8(%0,%3,8),%0               \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "neg         %3                            \n"
+      // Second round of bit swap.
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpcklwd   %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "palignr     $0x8,%%xmm2,%%xmm2            \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "punpcklwd   %%xmm6,%%xmm4                 \n"
+      "punpcklwd   %%xmm7,%%xmm5                 \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "movdqa      %%xmm5,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      // Third round of bit swap.
+      // Write to the destination pointer.
+      "punpckldq   %%xmm4,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "palignr     $0x8,%%xmm4,%%xmm4            \n"
+      "movq        %%xmm4,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm6,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "movq        %%xmm2,(%1)                   \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "punpckldq   %%xmm5,%%xmm1                 \n"
+      "movq        %%xmm6,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "movq        %%xmm1,(%1)                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "movq        %%xmm5,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm7,%%xmm3                 \n"
+      "movq        %%xmm3,(%1)                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "sub         $0x8,%2                       \n"
+      "movq        %%xmm7,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "jg          1b                            \n"
+      : "+r"(src),                    // %0
+        "+r"(dst),                    // %1
+        "+r"(width)                   // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // defined(HAS_TRANSPOSEWX8_SSSE3)
 
 // Transpose 16x8. 64 bit
 #if defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
-void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride,
-                             uint8* dst, int dst_stride, int width) {
-  asm volatile (
-    // Read in the data from the source pointer.
-    // First round of bit swap.
-    LABELALIGN
-  "1:                                            \n"
-    "movdqu     (%0),%%xmm0                      \n"
-    "movdqu     (%0,%3),%%xmm1                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "movdqa     %%xmm0,%%xmm8                    \n"
-    "punpcklbw  %%xmm1,%%xmm0                    \n"
-    "punpckhbw  %%xmm1,%%xmm8                    \n"
-    "movdqu     (%0),%%xmm2                      \n"
-    "movdqa     %%xmm0,%%xmm1                    \n"
-    "movdqa     %%xmm8,%%xmm9                    \n"
-    "palignr    $0x8,%%xmm1,%%xmm1               \n"
-    "palignr    $0x8,%%xmm9,%%xmm9               \n"
-    "movdqu     (%0,%3),%%xmm3                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "movdqa     %%xmm2,%%xmm10                   \n"
-    "punpcklbw  %%xmm3,%%xmm2                    \n"
-    "punpckhbw  %%xmm3,%%xmm10                   \n"
-    "movdqa     %%xmm2,%%xmm3                    \n"
-    "movdqa     %%xmm10,%%xmm11                  \n"
-    "movdqu     (%0),%%xmm4                      \n"
-    "palignr    $0x8,%%xmm3,%%xmm3               \n"
-    "palignr    $0x8,%%xmm11,%%xmm11             \n"
-    "movdqu     (%0,%3),%%xmm5                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "movdqa     %%xmm4,%%xmm12                   \n"
-    "punpcklbw  %%xmm5,%%xmm4                    \n"
-    "punpckhbw  %%xmm5,%%xmm12                   \n"
-    "movdqa     %%xmm4,%%xmm5                    \n"
-    "movdqa     %%xmm12,%%xmm13                  \n"
-    "movdqu     (%0),%%xmm6                      \n"
-    "palignr    $0x8,%%xmm5,%%xmm5               \n"
-    "palignr    $0x8,%%xmm13,%%xmm13             \n"
-    "movdqu     (%0,%3),%%xmm7                   \n"
-    "lea        (%0,%3,2),%0                     \n"
-    "movdqa     %%xmm6,%%xmm14                   \n"
-    "punpcklbw  %%xmm7,%%xmm6                    \n"
-    "punpckhbw  %%xmm7,%%xmm14                   \n"
-    "neg        %3                               \n"
-    "movdqa     %%xmm6,%%xmm7                    \n"
-    "movdqa     %%xmm14,%%xmm15                  \n"
-    "lea        0x10(%0,%3,8),%0                 \n"
-    "palignr    $0x8,%%xmm7,%%xmm7               \n"
-    "palignr    $0x8,%%xmm15,%%xmm15             \n"
-    "neg        %3                               \n"
-     // Second round of bit swap.
-    "punpcklwd  %%xmm2,%%xmm0                    \n"
-    "punpcklwd  %%xmm3,%%xmm1                    \n"
-    "movdqa     %%xmm0,%%xmm2                    \n"
-    "movdqa     %%xmm1,%%xmm3                    \n"
-    "palignr    $0x8,%%xmm2,%%xmm2               \n"
-    "palignr    $0x8,%%xmm3,%%xmm3               \n"
-    "punpcklwd  %%xmm6,%%xmm4                    \n"
-    "punpcklwd  %%xmm7,%%xmm5                    \n"
-    "movdqa     %%xmm4,%%xmm6                    \n"
-    "movdqa     %%xmm5,%%xmm7                    \n"
-    "palignr    $0x8,%%xmm6,%%xmm6               \n"
-    "palignr    $0x8,%%xmm7,%%xmm7               \n"
-    "punpcklwd  %%xmm10,%%xmm8                   \n"
-    "punpcklwd  %%xmm11,%%xmm9                   \n"
-    "movdqa     %%xmm8,%%xmm10                   \n"
-    "movdqa     %%xmm9,%%xmm11                   \n"
-    "palignr    $0x8,%%xmm10,%%xmm10             \n"
-    "palignr    $0x8,%%xmm11,%%xmm11             \n"
-    "punpcklwd  %%xmm14,%%xmm12                  \n"
-    "punpcklwd  %%xmm15,%%xmm13                  \n"
-    "movdqa     %%xmm12,%%xmm14                  \n"
-    "movdqa     %%xmm13,%%xmm15                  \n"
-    "palignr    $0x8,%%xmm14,%%xmm14             \n"
-    "palignr    $0x8,%%xmm15,%%xmm15             \n"
-    // Third round of bit swap.
-    // Write to the destination pointer.
-    "punpckldq  %%xmm4,%%xmm0                    \n"
-    "movq       %%xmm0,(%1)                      \n"
-    "movdqa     %%xmm0,%%xmm4                    \n"
-    "palignr    $0x8,%%xmm4,%%xmm4               \n"
-    "movq       %%xmm4,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm6,%%xmm2                    \n"
-    "movdqa     %%xmm2,%%xmm6                    \n"
-    "movq       %%xmm2,(%1)                      \n"
-    "palignr    $0x8,%%xmm6,%%xmm6               \n"
-    "punpckldq  %%xmm5,%%xmm1                    \n"
-    "movq       %%xmm6,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "movdqa     %%xmm1,%%xmm5                    \n"
-    "movq       %%xmm1,(%1)                      \n"
-    "palignr    $0x8,%%xmm5,%%xmm5               \n"
-    "movq       %%xmm5,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm7,%%xmm3                    \n"
-    "movq       %%xmm3,(%1)                      \n"
-    "movdqa     %%xmm3,%%xmm7                    \n"
-    "palignr    $0x8,%%xmm7,%%xmm7               \n"
-    "movq       %%xmm7,(%1,%4)                   \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm12,%%xmm8                   \n"
-    "movq       %%xmm8,(%1)                      \n"
-    "movdqa     %%xmm8,%%xmm12                   \n"
-    "palignr    $0x8,%%xmm12,%%xmm12             \n"
-    "movq       %%xmm12,(%1,%4)                  \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm14,%%xmm10                  \n"
-    "movdqa     %%xmm10,%%xmm14                  \n"
-    "movq       %%xmm10,(%1)                     \n"
-    "palignr    $0x8,%%xmm14,%%xmm14             \n"
-    "punpckldq  %%xmm13,%%xmm9                   \n"
-    "movq       %%xmm14,(%1,%4)                  \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "movdqa     %%xmm9,%%xmm13                   \n"
-    "movq       %%xmm9,(%1)                      \n"
-    "palignr    $0x8,%%xmm13,%%xmm13             \n"
-    "movq       %%xmm13,(%1,%4)                  \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "punpckldq  %%xmm15,%%xmm11                  \n"
-    "movq       %%xmm11,(%1)                     \n"
-    "movdqa     %%xmm11,%%xmm15                  \n"
-    "palignr    $0x8,%%xmm15,%%xmm15             \n"
-    "sub        $0x10,%2                         \n"
-    "movq       %%xmm15,(%1,%4)                  \n"
-    "lea        (%1,%4,2),%1                     \n"
-    "jg         1b                               \n"
-    : "+r"(src),    // %0
-      "+r"(dst),    // %1
-      "+r"(width)   // %2
-    : "r"((intptr_t)(src_stride)),  // %3
-      "r"((intptr_t)(dst_stride))   // %4
-    : "memory", "cc",
-      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
-      "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",  "xmm14",  "xmm15"
-  );
+void TransposeWx8_Fast_SSSE3(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst,
+                             int dst_stride,
+                             int width) {
+  asm volatile(
+      // Read in the data from the source pointer.
+      // First round of bit swap.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%0,%3),%%xmm1                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm8                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm8,%%xmm9                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm1            \n"
+      "palignr     $0x8,%%xmm9,%%xmm9            \n"
+      "movdqu      (%0,%3),%%xmm3                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm2,%%xmm10                \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"
+      "punpckhbw   %%xmm3,%%xmm10                \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm10,%%xmm11               \n"
+      "movdqu      (%0),%%xmm4                   \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "palignr     $0x8,%%xmm11,%%xmm11          \n"
+      "movdqu      (%0,%3),%%xmm5                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm4,%%xmm12                \n"
+      "punpcklbw   %%xmm5,%%xmm4                 \n"
+      "punpckhbw   %%xmm5,%%xmm12                \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "movdqa      %%xmm12,%%xmm13               \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "palignr     $0x8,%%xmm13,%%xmm13          \n"
+      "movdqu      (%0,%3),%%xmm7                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm6,%%xmm14                \n"
+      "punpcklbw   %%xmm7,%%xmm6                 \n"
+      "punpckhbw   %%xmm7,%%xmm14                \n"
+      "neg         %3                            \n"
+      "movdqa      %%xmm6,%%xmm7                 \n"
+      "movdqa      %%xmm14,%%xmm15               \n"
+      "lea         0x10(%0,%3,8),%0              \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "palignr     $0x8,%%xmm15,%%xmm15          \n"
+      "neg         %3                            \n"
+      // Second round of bit swap.
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpcklwd   %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "palignr     $0x8,%%xmm2,%%xmm2            \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "punpcklwd   %%xmm6,%%xmm4                 \n"
+      "punpcklwd   %%xmm7,%%xmm5                 \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "movdqa      %%xmm5,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "punpcklwd   %%xmm10,%%xmm8                \n"
+      "punpcklwd   %%xmm11,%%xmm9                \n"
+      "movdqa      %%xmm8,%%xmm10                \n"
+      "movdqa      %%xmm9,%%xmm11                \n"
+      "palignr     $0x8,%%xmm10,%%xmm10          \n"
+      "palignr     $0x8,%%xmm11,%%xmm11          \n"
+      "punpcklwd   %%xmm14,%%xmm12               \n"
+      "punpcklwd   %%xmm15,%%xmm13               \n"
+      "movdqa      %%xmm12,%%xmm14               \n"
+      "movdqa      %%xmm13,%%xmm15               \n"
+      "palignr     $0x8,%%xmm14,%%xmm14          \n"
+      "palignr     $0x8,%%xmm15,%%xmm15          \n"
+      // Third round of bit swap.
+      // Write to the destination pointer.
+      "punpckldq   %%xmm4,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "palignr     $0x8,%%xmm4,%%xmm4            \n"
+      "movq        %%xmm4,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm6,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "movq        %%xmm2,(%1)                   \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "punpckldq   %%xmm5,%%xmm1                 \n"
+      "movq        %%xmm6,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "movq        %%xmm1,(%1)                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "movq        %%xmm5,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm7,%%xmm3                 \n"
+      "movq        %%xmm3,(%1)                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "movq        %%xmm7,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm12,%%xmm8                \n"
+      "movq        %%xmm8,(%1)                   \n"
+      "movdqa      %%xmm8,%%xmm12                \n"
+      "palignr     $0x8,%%xmm12,%%xmm12          \n"
+      "movq        %%xmm12,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm14,%%xmm10               \n"
+      "movdqa      %%xmm10,%%xmm14               \n"
+      "movq        %%xmm10,(%1)                  \n"
+      "palignr     $0x8,%%xmm14,%%xmm14          \n"
+      "punpckldq   %%xmm13,%%xmm9                \n"
+      "movq        %%xmm14,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "movdqa      %%xmm9,%%xmm13                \n"
+      "movq        %%xmm9,(%1)                   \n"
+      "palignr     $0x8,%%xmm13,%%xmm13          \n"
+      "movq        %%xmm13,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm15,%%xmm11               \n"
+      "movq        %%xmm11,(%1)                  \n"
+      "movdqa      %%xmm11,%%xmm15               \n"
+      "palignr     $0x8,%%xmm15,%%xmm15          \n"
+      "sub         $0x10,%2                      \n"
+      "movq        %%xmm15,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "jg          1b                            \n"
+      : "+r"(src),                    // %0
+        "+r"(dst),                    // %1
+        "+r"(width)                   // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14",
+        "xmm15");
 }
 #endif  // defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
 
 // Transpose UV 8x8.  64 bit.
 #if defined(HAS_TRANSPOSEUVWX8_SSE2)
-void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b, int width) {
-  asm volatile (
-    // Read in the data from the source pointer.
-    // First round of bit swap.
-    LABELALIGN
-  "1:                                            \n"
-    "movdqu     (%0),%%xmm0                      \n"
-    "movdqu     (%0,%4),%%xmm1                   \n"
-    "lea        (%0,%4,2),%0                     \n"
-    "movdqa     %%xmm0,%%xmm8                    \n"
-    "punpcklbw  %%xmm1,%%xmm0                    \n"
-    "punpckhbw  %%xmm1,%%xmm8                    \n"
-    "movdqa     %%xmm8,%%xmm1                    \n"
-    "movdqu     (%0),%%xmm2                      \n"
-    "movdqu     (%0,%4),%%xmm3                   \n"
-    "lea        (%0,%4,2),%0                     \n"
-    "movdqa     %%xmm2,%%xmm8                    \n"
-    "punpcklbw  %%xmm3,%%xmm2                    \n"
-    "punpckhbw  %%xmm3,%%xmm8                    \n"
-    "movdqa     %%xmm8,%%xmm3                    \n"
-    "movdqu     (%0),%%xmm4                      \n"
-    "movdqu     (%0,%4),%%xmm5                   \n"
-    "lea        (%0,%4,2),%0                     \n"
-    "movdqa     %%xmm4,%%xmm8                    \n"
-    "punpcklbw  %%xmm5,%%xmm4                    \n"
-    "punpckhbw  %%xmm5,%%xmm8                    \n"
-    "movdqa     %%xmm8,%%xmm5                    \n"
-    "movdqu     (%0),%%xmm6                      \n"
-    "movdqu     (%0,%4),%%xmm7                   \n"
-    "lea        (%0,%4,2),%0                     \n"
-    "movdqa     %%xmm6,%%xmm8                    \n"
-    "punpcklbw  %%xmm7,%%xmm6                    \n"
-    "neg        %4                               \n"
-    "lea        0x10(%0,%4,8),%0                 \n"
-    "punpckhbw  %%xmm7,%%xmm8                    \n"
-    "movdqa     %%xmm8,%%xmm7                    \n"
-    "neg        %4                               \n"
-     // Second round of bit swap.
-    "movdqa     %%xmm0,%%xmm8                    \n"
-    "movdqa     %%xmm1,%%xmm9                    \n"
-    "punpckhwd  %%xmm2,%%xmm8                    \n"
-    "punpckhwd  %%xmm3,%%xmm9                    \n"
-    "punpcklwd  %%xmm2,%%xmm0                    \n"
-    "punpcklwd  %%xmm3,%%xmm1                    \n"
-    "movdqa     %%xmm8,%%xmm2                    \n"
-    "movdqa     %%xmm9,%%xmm3                    \n"
-    "movdqa     %%xmm4,%%xmm8                    \n"
-    "movdqa     %%xmm5,%%xmm9                    \n"
-    "punpckhwd  %%xmm6,%%xmm8                    \n"
-    "punpckhwd  %%xmm7,%%xmm9                    \n"
-    "punpcklwd  %%xmm6,%%xmm4                    \n"
-    "punpcklwd  %%xmm7,%%xmm5                    \n"
-    "movdqa     %%xmm8,%%xmm6                    \n"
-    "movdqa     %%xmm9,%%xmm7                    \n"
-    // Third round of bit swap.
-    // Write to the destination pointer.
-    "movdqa     %%xmm0,%%xmm8                    \n"
-    "punpckldq  %%xmm4,%%xmm0                    \n"
-    "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
-    "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
-    "punpckhdq  %%xmm4,%%xmm8                    \n"
-    "movlpd     %%xmm8,(%1,%5)                   \n"
-    "lea        (%1,%5,2),%1                     \n"
-    "movhpd     %%xmm8,(%2,%6)                   \n"
-    "lea        (%2,%6,2),%2                     \n"
-    "movdqa     %%xmm2,%%xmm8                    \n"
-    "punpckldq  %%xmm6,%%xmm2                    \n"
-    "movlpd     %%xmm2,(%1)                      \n"
-    "movhpd     %%xmm2,(%2)                      \n"
-    "punpckhdq  %%xmm6,%%xmm8                    \n"
-    "movlpd     %%xmm8,(%1,%5)                   \n"
-    "lea        (%1,%5,2),%1                     \n"
-    "movhpd     %%xmm8,(%2,%6)                   \n"
-    "lea        (%2,%6,2),%2                     \n"
-    "movdqa     %%xmm1,%%xmm8                    \n"
-    "punpckldq  %%xmm5,%%xmm1                    \n"
-    "movlpd     %%xmm1,(%1)                      \n"
-    "movhpd     %%xmm1,(%2)                      \n"
-    "punpckhdq  %%xmm5,%%xmm8                    \n"
-    "movlpd     %%xmm8,(%1,%5)                   \n"
-    "lea        (%1,%5,2),%1                     \n"
-    "movhpd     %%xmm8,(%2,%6)                   \n"
-    "lea        (%2,%6,2),%2                     \n"
-    "movdqa     %%xmm3,%%xmm8                    \n"
-    "punpckldq  %%xmm7,%%xmm3                    \n"
-    "movlpd     %%xmm3,(%1)                      \n"
-    "movhpd     %%xmm3,(%2)                      \n"
-    "punpckhdq  %%xmm7,%%xmm8                    \n"
-    "sub        $0x8,%3                          \n"
-    "movlpd     %%xmm8,(%1,%5)                   \n"
-    "lea        (%1,%5,2),%1                     \n"
-    "movhpd     %%xmm8,(%2,%6)                   \n"
-    "lea        (%2,%6,2),%2                     \n"
-    "jg         1b                               \n"
-    : "+r"(src),    // %0
-      "+r"(dst_a),  // %1
-      "+r"(dst_b),  // %2
-      "+r"(width)   // %3
-    : "r"((intptr_t)(src_stride)),    // %4
-      "r"((intptr_t)(dst_stride_a)),  // %5
-      "r"((intptr_t)(dst_stride_b))   // %6
-    : "memory", "cc",
-      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
-      "xmm8", "xmm9"
-  );
+void TransposeUVWx8_SSE2(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  asm volatile(
+      // Read in the data from the source pointer.
+      // First round of bit swap.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%0,%4),%%xmm1                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqu      (%0,%4),%%xmm3                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm2,%%xmm8                 \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"
+      "punpckhbw   %%xmm3,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm3                 \n"
+      "movdqu      (%0),%%xmm4                   \n"
+      "movdqu      (%0,%4),%%xmm5                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm4,%%xmm8                 \n"
+      "punpcklbw   %%xmm5,%%xmm4                 \n"
+      "punpckhbw   %%xmm5,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm5                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      (%0,%4),%%xmm7                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm6,%%xmm8                 \n"
+      "punpcklbw   %%xmm7,%%xmm6                 \n"
+      "neg         %4                            \n"
+      "lea         0x10(%0,%4,8),%0              \n"
+      "punpckhbw   %%xmm7,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm7                 \n"
+      "neg         %4                            \n"
+      // Second round of bit swap.
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "movdqa      %%xmm1,%%xmm9                 \n"
+      "punpckhwd   %%xmm2,%%xmm8                 \n"
+      "punpckhwd   %%xmm3,%%xmm9                 \n"
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpcklwd   %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm8,%%xmm2                 \n"
+      "movdqa      %%xmm9,%%xmm3                 \n"
+      "movdqa      %%xmm4,%%xmm8                 \n"
+      "movdqa      %%xmm5,%%xmm9                 \n"
+      "punpckhwd   %%xmm6,%%xmm8                 \n"
+      "punpckhwd   %%xmm7,%%xmm9                 \n"
+      "punpcklwd   %%xmm6,%%xmm4                 \n"
+      "punpcklwd   %%xmm7,%%xmm5                 \n"
+      "movdqa      %%xmm8,%%xmm6                 \n"
+      "movdqa      %%xmm9,%%xmm7                 \n"
+      // Third round of bit swap.
+      // Write to the destination pointer.
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "punpckldq   %%xmm4,%%xmm0                 \n"
+      "movlpd      %%xmm0,(%1)                   \n"  // Write back U channel
+      "movhpd      %%xmm0,(%2)                   \n"  // Write back V channel
+      "punpckhdq   %%xmm4,%%xmm8                 \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "movdqa      %%xmm2,%%xmm8                 \n"
+      "punpckldq   %%xmm6,%%xmm2                 \n"
+      "movlpd      %%xmm2,(%1)                   \n"
+      "movhpd      %%xmm2,(%2)                   \n"
+      "punpckhdq   %%xmm6,%%xmm8                 \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "movdqa      %%xmm1,%%xmm8                 \n"
+      "punpckldq   %%xmm5,%%xmm1                 \n"
+      "movlpd      %%xmm1,(%1)                   \n"
+      "movhpd      %%xmm1,(%2)                   \n"
+      "punpckhdq   %%xmm5,%%xmm8                 \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "movdqa      %%xmm3,%%xmm8                 \n"
+      "punpckldq   %%xmm7,%%xmm3                 \n"
+      "movlpd      %%xmm3,(%1)                   \n"
+      "movhpd      %%xmm3,(%2)                   \n"
+      "punpckhdq   %%xmm7,%%xmm8                 \n"
+      "sub         $0x8,%3                       \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "jg          1b                            \n"
+      : "+r"(src),                      // %0
+        "+r"(dst_a),                    // %1
+        "+r"(dst_b),                    // %2
+        "+r"(width)                     // %3
+      : "r"((intptr_t)(src_stride)),    // %4
+        "r"((intptr_t)(dst_stride_a)),  // %5
+        "r"((intptr_t)(dst_stride_b))   // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7", "xmm8", "xmm9");
 }
 #endif  // defined(HAS_TRANSPOSEUVWX8_SSE2)
 #endif  // defined(__x86_64__) || defined(__i386__)
diff --git a/media/libyuv/source/rotate_lsx.cc b/media/libyuv/source/rotate_lsx.cc
new file mode 100644
index 0000000000..94a2b91cd8
--- /dev/null
+++ b/media/libyuv/source/rotate_lsx.cc
@@ -0,0 +1,243 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ILVLH_B(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_b, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_b, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define ILVLH_H(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_h, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_h, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define ILVLH_W(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_w, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_w, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define ILVLH_D(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_d, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_d, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define LSX_ST_4(_dst0, _dst1, _dst2, _dst3, _dst, _stride, _stride2, \
+                 _stride3, _stride4)                                  \
+  {                                                                   \
+    __lsx_vst(_dst0, _dst, 0);                                        \
+    __lsx_vstx(_dst1, _dst, _stride);                                 \
+    __lsx_vstx(_dst2, _dst, _stride2);                                \
+    __lsx_vstx(_dst3, _dst, _stride3);                                \
+    _dst += _stride4;                                                 \
+  }
+
+#define LSX_ST_2(_dst0, _dst1, _dst, _stride, _stride2) \
+  {                                                     \
+    __lsx_vst(_dst0, _dst, 0);                          \
+    __lsx_vstx(_dst1, _dst, _stride);                   \
+    _dst += _stride2;                                   \
+  }
+
+void TransposeWx16_C(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst,
+                     int dst_stride,
+                     int width) {
+  TransposeWx8_C(src, src_stride, dst, dst_stride, width);
+  TransposeWx8_C((src + 8 * src_stride), src_stride, (dst + 8), dst_stride,
+                 width);
+}
+
+void TransposeUVWx16_C(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst_a,
+                       int dst_stride_a,
+                       uint8_t* dst_b,
+                       int dst_stride_b,
+                       int width) {
+  TransposeUVWx8_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width);
+  TransposeUVWx8_C((src + 8 * src_stride), src_stride, (dst_a + 8),
+                   dst_stride_a, (dst_b + 8), dst_stride_b, width);
+}
+
+void TransposeWx16_LSX(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width) {
+  int x;
+  int len = width / 16;
+  uint8_t* s;
+  int src_stride2 = src_stride << 1;
+  int src_stride3 = src_stride + src_stride2;
+  int src_stride4 = src_stride2 << 1;
+  int dst_stride2 = dst_stride << 1;
+  int dst_stride3 = dst_stride + dst_stride2;
+  int dst_stride4 = dst_stride2 << 1;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < len; x++) {
+    s = (uint8_t*)src;
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    ILVLH_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVLH_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    res8 = __lsx_vilvl_w(reg4, reg0);
+    res9 = __lsx_vilvh_w(reg4, reg0);
+    ILVLH_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    res8 = __lsx_vilvl_w(reg5, reg1);
+    res9 = __lsx_vilvh_w(reg5, reg1);
+    ILVLH_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    res8 = __lsx_vilvl_w(reg6, reg2);
+    res9 = __lsx_vilvh_w(reg6, reg2);
+    ILVLH_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    res8 = __lsx_vilvl_w(reg7, reg3);
+    res9 = __lsx_vilvh_w(reg7, reg3);
+    ILVLH_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    src += 16;
+  }
+}
+
+void TransposeUVWx16_LSX(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  int x;
+  int len = width / 8;
+  uint8_t* s;
+  int src_stride2 = src_stride << 1;
+  int src_stride3 = src_stride + src_stride2;
+  int src_stride4 = src_stride2 << 1;
+  int dst_stride_a2 = dst_stride_a << 1;
+  int dst_stride_b2 = dst_stride_b << 1;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < len; x++) {
+    s = (uint8_t*)src;
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    ILVLH_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVLH_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    res8 = __lsx_vilvl_w(reg4, reg0);
+    res9 = __lsx_vilvh_w(reg4, reg0);
+    ILVLH_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    res8 = __lsx_vilvl_w(reg5, reg1);
+    res9 = __lsx_vilvh_w(reg5, reg1);
+    ILVLH_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    res8 = __lsx_vilvl_w(reg6, reg2);
+    res9 = __lsx_vilvh_w(reg6, reg2);
+    ILVLH_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    res8 = __lsx_vilvl_w(reg7, reg3);
+    res9 = __lsx_vilvh_w(reg7, reg3);
+    ILVLH_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    src += 16;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
diff --git a/media/libyuv/source/rotate_mips.cc b/media/libyuv/source/rotate_mips.cc
deleted file mode 100644
index 1e8ce25197..0000000000
--- a/media/libyuv/source/rotate_mips.cc
+++ /dev/null
@@ -1,484 +0,0 @@
-/*
- *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/row.h"
-#include "libyuv/rotate_row.h"
-
-#include "libyuv/basic_types.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-#if !defined(LIBYUV_DISABLE_MIPS) && \
-    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32)
-
-void TransposeWx8_DSPR2(const uint8* src, int src_stride,
-                        uint8* dst, int dst_stride, int width) {
-   __asm__ __volatile__ (
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "sll              $t2, %[src_stride], 0x1          \n" // src_stride x 2
-      "sll              $t4, %[src_stride], 0x2          \n" // src_stride x 4
-      "sll              $t9, %[src_stride], 0x3          \n" // src_stride x 8
-      "addu             $t3, $t2, %[src_stride]          \n"
-      "addu             $t5, $t4, %[src_stride]          \n"
-      "addu             $t6, $t2, $t4                    \n"
-      "andi             $t0, %[dst], 0x3                 \n"
-      "andi             $t1, %[dst_stride], 0x3          \n"
-      "or               $t0, $t0, $t1                    \n"
-      "bnez             $t0, 11f                         \n"
-      " subu            $t7, $t9, %[src_stride]          \n"
-//dst + dst_stride word aligned
-    "1:                                                  \n"
-      "lbu              $t0, 0(%[src])                   \n"
-      "lbux             $t1, %[src_stride](%[src])       \n"
-      "lbux             $t8, $t2(%[src])                 \n"
-      "lbux             $t9, $t3(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s0, $t8, $t0                    \n"
-      "lbux             $t0, $t4(%[src])                 \n"
-      "lbux             $t1, $t5(%[src])                 \n"
-      "lbux             $t8, $t6(%[src])                 \n"
-      "lbux             $t9, $t7(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s1, $t8, $t0                    \n"
-      "sw               $s0, 0(%[dst])                   \n"
-      "addiu            %[width], -1                     \n"
-      "addiu            %[src], 1                        \n"
-      "sw               $s1, 4(%[dst])                   \n"
-      "bnez             %[width], 1b                     \n"
-      " addu            %[dst], %[dst], %[dst_stride]    \n"
-      "b                2f                               \n"
-//dst + dst_stride unaligned
-   "11:                                                  \n"
-      "lbu              $t0, 0(%[src])                   \n"
-      "lbux             $t1, %[src_stride](%[src])       \n"
-      "lbux             $t8, $t2(%[src])                 \n"
-      "lbux             $t9, $t3(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s0, $t8, $t0                    \n"
-      "lbux             $t0, $t4(%[src])                 \n"
-      "lbux             $t1, $t5(%[src])                 \n"
-      "lbux             $t8, $t6(%[src])                 \n"
-      "lbux             $t9, $t7(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s1, $t8, $t0                    \n"
-      "swr              $s0, 0(%[dst])                   \n"
-      "swl              $s0, 3(%[dst])                   \n"
-      "addiu            %[width], -1                     \n"
-      "addiu            %[src], 1                        \n"
-      "swr              $s1, 4(%[dst])                   \n"
-      "swl              $s1, 7(%[dst])                   \n"
-      "bnez             %[width], 11b                    \n"
-       "addu             %[dst], %[dst], %[dst_stride]   \n"
-    "2:                                                  \n"
-      ".set pop                                          \n"
-      :[src] "+r" (src),
-       [dst] "+r" (dst),
-       [width] "+r" (width)
-      :[src_stride] "r" (src_stride),
-       [dst_stride] "r" (dst_stride)
-      : "t0", "t1",  "t2", "t3", "t4", "t5",
-        "t6", "t7", "t8", "t9",
-        "s0", "s1"
-  );
-}
-
-void TransposeWx8_Fast_DSPR2(const uint8* src, int src_stride,
-                             uint8* dst, int dst_stride, int width) {
-  __asm__ __volatile__ (
-      ".set noat                                         \n"
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "beqz             %[width], 2f                     \n"
-      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
-      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
-      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
-      "addu             $t3, $t2, %[src_stride]          \n"
-      "addu             $t5, $t4, %[src_stride]          \n"
-      "addu             $t6, $t2, $t4                    \n"
-
-      "srl              $AT, %[width], 0x2               \n"
-      "andi             $t0, %[dst], 0x3                 \n"
-      "andi             $t1, %[dst_stride], 0x3          \n"
-      "or               $t0, $t0, $t1                    \n"
-      "bnez             $t0, 11f                         \n"
-      " subu            $t7, $t9, %[src_stride]          \n"
-//dst + dst_stride word aligned
-      "1:                                                \n"
-      "lw               $t0, 0(%[src])                   \n"
-      "lwx              $t1, %[src_stride](%[src])       \n"
-      "lwx              $t8, $t2(%[src])                 \n"
-      "lwx              $t9, $t3(%[src])                 \n"
-
-// t0 = | 30 | 20 | 10 | 00 |
-// t1 = | 31 | 21 | 11 | 01 |
-// t8 = | 32 | 22 | 12 | 02 |
-// t9 = | 33 | 23 | 13 | 03 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-  // s0 = | 21 | 01 | 20 | 00 |
-  // s1 = | 23 | 03 | 22 | 02 |
-  // s2 = | 31 | 11 | 30 | 10 |
-  // s3 = | 33 | 13 | 32 | 12 |
-
-      "precr.qb.ph     $s4, $s1, $s0                     \n"
-      "precrq.qb.ph    $s5, $s1, $s0                     \n"
-      "precr.qb.ph     $s6, $s3, $s2                     \n"
-      "precrq.qb.ph    $s7, $s3, $s2                     \n"
-
-  // s4 = | 03 | 02 | 01 | 00 |
-  // s5 = | 23 | 22 | 21 | 20 |
-  // s6 = | 13 | 12 | 11 | 10 |
-  // s7 = | 33 | 32 | 31 | 30 |
-
-      "lwx              $t0, $t4(%[src])                 \n"
-      "lwx              $t1, $t5(%[src])                 \n"
-      "lwx              $t8, $t6(%[src])                 \n"
-      "lwx              $t9, $t7(%[src])                 \n"
-
-// t0 = | 34 | 24 | 14 | 04 |
-// t1 = | 35 | 25 | 15 | 05 |
-// t8 = | 36 | 26 | 16 | 06 |
-// t9 = | 37 | 27 | 17 | 07 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-  // s0 = | 25 | 05 | 24 | 04 |
-  // s1 = | 27 | 07 | 26 | 06 |
-  // s2 = | 35 | 15 | 34 | 14 |
-  // s3 = | 37 | 17 | 36 | 16 |
-
-      "precr.qb.ph     $t0, $s1, $s0                     \n"
-      "precrq.qb.ph    $t1, $s1, $s0                     \n"
-      "precr.qb.ph     $t8, $s3, $s2                     \n"
-      "precrq.qb.ph    $t9, $s3, $s2                     \n"
-
-  // t0 = | 07 | 06 | 05 | 04 |
-  // t1 = | 27 | 26 | 25 | 24 |
-  // t8 = | 17 | 16 | 15 | 14 |
-  // t9 = | 37 | 36 | 35 | 34 |
-
-      "addu            $s0, %[dst], %[dst_stride]        \n"
-      "addu            $s1, $s0, %[dst_stride]           \n"
-      "addu            $s2, $s1, %[dst_stride]           \n"
-
-      "sw              $s4, 0(%[dst])                    \n"
-      "sw              $t0, 4(%[dst])                    \n"
-      "sw              $s6, 0($s0)                       \n"
-      "sw              $t8, 4($s0)                       \n"
-      "sw              $s5, 0($s1)                       \n"
-      "sw              $t1, 4($s1)                       \n"
-      "sw              $s7, 0($s2)                       \n"
-      "sw              $t9, 4($s2)                       \n"
-
-      "addiu            $AT, -1                          \n"
-      "addiu            %[src], 4                        \n"
-
-      "bnez             $AT, 1b                          \n"
-      " addu            %[dst], $s2, %[dst_stride]       \n"
-      "b                2f                               \n"
-//dst + dst_stride unaligned
-      "11:                                               \n"
-      "lw               $t0, 0(%[src])                   \n"
-      "lwx              $t1, %[src_stride](%[src])       \n"
-      "lwx              $t8, $t2(%[src])                 \n"
-      "lwx              $t9, $t3(%[src])                 \n"
-
-// t0 = | 30 | 20 | 10 | 00 |
-// t1 = | 31 | 21 | 11 | 01 |
-// t8 = | 32 | 22 | 12 | 02 |
-// t9 = | 33 | 23 | 13 | 03 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-  // s0 = | 21 | 01 | 20 | 00 |
-  // s1 = | 23 | 03 | 22 | 02 |
-  // s2 = | 31 | 11 | 30 | 10 |
-  // s3 = | 33 | 13 | 32 | 12 |
-
-      "precr.qb.ph     $s4, $s1, $s0                     \n"
-      "precrq.qb.ph    $s5, $s1, $s0                     \n"
-      "precr.qb.ph     $s6, $s3, $s2                     \n"
-      "precrq.qb.ph    $s7, $s3, $s2                     \n"
-
-  // s4 = | 03 | 02 | 01 | 00 |
-  // s5 = | 23 | 22 | 21 | 20 |
-  // s6 = | 13 | 12 | 11 | 10 |
-  // s7 = | 33 | 32 | 31 | 30 |
-
-      "lwx              $t0, $t4(%[src])                 \n"
-      "lwx              $t1, $t5(%[src])                 \n"
-      "lwx              $t8, $t6(%[src])                 \n"
-      "lwx              $t9, $t7(%[src])                 \n"
-
-// t0 = | 34 | 24 | 14 | 04 |
-// t1 = | 35 | 25 | 15 | 05 |
-// t8 = | 36 | 26 | 16 | 06 |
-// t9 = | 37 | 27 | 17 | 07 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-  // s0 = | 25 | 05 | 24 | 04 |
-  // s1 = | 27 | 07 | 26 | 06 |
-  // s2 = | 35 | 15 | 34 | 14 |
-  // s3 = | 37 | 17 | 36 | 16 |
-
-      "precr.qb.ph     $t0, $s1, $s0                     \n"
-      "precrq.qb.ph    $t1, $s1, $s0                     \n"
-      "precr.qb.ph     $t8, $s3, $s2                     \n"
-      "precrq.qb.ph    $t9, $s3, $s2                     \n"
-
-  // t0 = | 07 | 06 | 05 | 04 |
-  // t1 = | 27 | 26 | 25 | 24 |
-  // t8 = | 17 | 16 | 15 | 14 |
-  // t9 = | 37 | 36 | 35 | 34 |
-
-      "addu            $s0, %[dst], %[dst_stride]        \n"
-      "addu            $s1, $s0, %[dst_stride]           \n"
-      "addu            $s2, $s1, %[dst_stride]           \n"
-
-      "swr              $s4, 0(%[dst])                   \n"
-      "swl              $s4, 3(%[dst])                   \n"
-      "swr              $t0, 4(%[dst])                   \n"
-      "swl              $t0, 7(%[dst])                   \n"
-      "swr              $s6, 0($s0)                      \n"
-      "swl              $s6, 3($s0)                      \n"
-      "swr              $t8, 4($s0)                      \n"
-      "swl              $t8, 7($s0)                      \n"
-      "swr              $s5, 0($s1)                      \n"
-      "swl              $s5, 3($s1)                      \n"
-      "swr              $t1, 4($s1)                      \n"
-      "swl              $t1, 7($s1)                      \n"
-      "swr              $s7, 0($s2)                      \n"
-      "swl              $s7, 3($s2)                      \n"
-      "swr              $t9, 4($s2)                      \n"
-      "swl              $t9, 7($s2)                      \n"
-
-      "addiu            $AT, -1                          \n"
-      "addiu            %[src], 4                        \n"
-
-      "bnez             $AT, 11b                         \n"
-      " addu            %[dst], $s2, %[dst_stride]       \n"
-      "2:                                                \n"
-      ".set pop                                          \n"
-      ".set at                                           \n"
-      :[src] "+r" (src),
-       [dst] "+r" (dst),
-       [width] "+r" (width)
-      :[src_stride] "r" (src_stride),
-       [dst_stride] "r" (dst_stride)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9",
-        "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7"
-  );
-}
-
-void TransposeUVWx8_DSPR2(const uint8* src, int src_stride,
-                          uint8* dst_a, int dst_stride_a,
-                          uint8* dst_b, int dst_stride_b,
-                          int width) {
-  __asm__ __volatile__ (
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "beqz            %[width], 2f                      \n"
-      " sll            $t2, %[src_stride], 0x1           \n" // src_stride x 2
-      "sll             $t4, %[src_stride], 0x2           \n" // src_stride x 4
-      "sll             $t9, %[src_stride], 0x3           \n" // src_stride x 8
-      "addu            $t3, $t2, %[src_stride]           \n"
-      "addu            $t5, $t4, %[src_stride]           \n"
-      "addu            $t6, $t2, $t4                     \n"
-      "subu            $t7, $t9, %[src_stride]           \n"
-      "srl             $t1, %[width], 1                  \n"
-
-// check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
-      "andi            $t0, %[dst_a], 0x3                \n"
-      "andi            $t8, %[dst_b], 0x3                \n"
-      "or              $t0, $t0, $t8                     \n"
-      "andi            $t8, %[dst_stride_a], 0x3         \n"
-      "andi            $s5, %[dst_stride_b], 0x3         \n"
-      "or              $t8, $t8, $s5                     \n"
-      "or              $t0, $t0, $t8                     \n"
-      "bnez            $t0, 11f                          \n"
-      " nop                                              \n"
-// dst + dst_stride word aligned (both, a & b dst addresses)
-    "1:                                                  \n"
-      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
-      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
-      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
-      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
-      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
-      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
-      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
-
-      "sw              $s3, 0($s5)                       \n"
-      "sw              $s4, 0($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
-
-      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
-      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
-      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
-      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
-      "sw              $s3, 0(%[dst_a])                  \n"
-      "sw              $s4, 0(%[dst_b])                  \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
-      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
-      "sw              $s3, 4($s5)                       \n"
-      "sw              $s4, 4($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
-
-      "addiu           %[src], 4                         \n"
-      "addiu           $t1, -1                           \n"
-      "sll             $t0, %[dst_stride_a], 1           \n"
-      "sll             $t8, %[dst_stride_b], 1           \n"
-      "sw              $s3, 4(%[dst_a])                  \n"
-      "sw              $s4, 4(%[dst_b])                  \n"
-      "addu            %[dst_a], %[dst_a], $t0           \n"
-      "bnez            $t1, 1b                           \n"
-      " addu           %[dst_b], %[dst_b], $t8           \n"
-      "b               2f                                \n"
-      " nop                                              \n"
-
-// dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
-   "11:                                                  \n"
-      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
-      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
-      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
-      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
-      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
-      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
-      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
-
-      "swr             $s3, 0($s5)                       \n"
-      "swl             $s3, 3($s5)                       \n"
-      "swr             $s4, 0($s6)                       \n"
-      "swl             $s4, 3($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
-
-      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
-      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
-      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
-      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
-      "swr             $s3, 0(%[dst_a])                  \n"
-      "swl             $s3, 3(%[dst_a])                  \n"
-      "swr             $s4, 0(%[dst_b])                  \n"
-      "swl             $s4, 3(%[dst_b])                  \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
-      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
-
-      "swr             $s3, 4($s5)                       \n"
-      "swl             $s3, 7($s5)                       \n"
-      "swr             $s4, 4($s6)                       \n"
-      "swl             $s4, 7($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
-
-      "addiu           %[src], 4                         \n"
-      "addiu           $t1, -1                           \n"
-      "sll             $t0, %[dst_stride_a], 1           \n"
-      "sll             $t8, %[dst_stride_b], 1           \n"
-      "swr             $s3, 4(%[dst_a])                  \n"
-      "swl             $s3, 7(%[dst_a])                  \n"
-      "swr             $s4, 4(%[dst_b])                  \n"
-      "swl             $s4, 7(%[dst_b])                  \n"
-      "addu            %[dst_a], %[dst_a], $t0           \n"
-      "bnez            $t1, 11b                          \n"
-      " addu           %[dst_b], %[dst_b], $t8           \n"
-
-      "2:                                                \n"
-      ".set pop                                          \n"
-      : [src] "+r" (src),
-        [dst_a] "+r" (dst_a),
-        [dst_b] "+r" (dst_b),
-        [width] "+r" (width),
-        [src_stride] "+r" (src_stride)
-      : [dst_stride_a] "r" (dst_stride_a),
-        [dst_stride_b] "r" (dst_stride_b)
-      : "t0", "t1",  "t2", "t3",  "t4", "t5",
-        "t6", "t7", "t8", "t9",
-        "s0", "s1", "s2", "s3",
-        "s4", "s5", "s6"
-  );
-}
-
-#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/media/libyuv/source/rotate_mmi.cc b/media/libyuv/source/rotate_mmi.cc
new file mode 100644
index 0000000000..f8de60834d
--- /dev/null
+++ b/media/libyuv/source/rotate_mmi.cc
@@ -0,0 +1,291 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Mips MMI.
+#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+void TransposeWx8_MMI(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst,
+                      int dst_stride,
+                      int width) {
+  uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
+  uint64_t tmp7, tmp8, tmp9, tmp10, tmp11, tmp12, tmp13;
+  uint8_t* src_tmp = nullptr;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "ldc1       %[tmp12],        0x00(%[src])                     \n\t"
+      "dadd       %[src_tmp],      %[src],         %[src_stride]    \n\t"
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp0 = (00 10 01 11 02 12 03 13) */
+      "punpcklbh  %[tmp0],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp1 = (04 14 05 15 06 16 07 17) */
+      "punpckhbh  %[tmp1],         %[tmp12],       %[tmp13]         \n\t"
+
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp12],        0x00(%[src_tmp])                 \n\t"
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp2 = (20 30 21 31 22 32 23 33) */
+      "punpcklbh  %[tmp2],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp3 = (24 34 25 35 26 36 27 37) */
+      "punpckhbh  %[tmp3],         %[tmp12],       %[tmp13]         \n\t"
+
+      /* tmp4 = (00 10 20 30 01 11 21 31) */
+      "punpcklhw  %[tmp4],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp5 = (02 12 22 32 03 13 23 33) */
+      "punpckhhw  %[tmp5],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp6 = (04 14 24 34 05 15 25 35) */
+      "punpcklhw  %[tmp6],         %[tmp1],        %[tmp3]          \n\t"
+      /* tmp7 = (06 16 26 36 07 17 27 37) */
+      "punpckhhw  %[tmp7],         %[tmp1],        %[tmp3]          \n\t"
+
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp12],        0x00(%[src_tmp])                 \n\t"
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp0 = (40 50 41 51 42 52 43 53) */
+      "punpcklbh  %[tmp0],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp1 = (44 54 45 55 46 56 47 57) */
+      "punpckhbh  %[tmp1],         %[tmp12],       %[tmp13]         \n\t"
+
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp12],        0x00(%[src_tmp])                 \n\t"
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp2 = (60 70 61 71 62 72 63 73) */
+      "punpcklbh  %[tmp2],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp3 = (64 74 65 75 66 76 67 77) */
+      "punpckhbh  %[tmp3],         %[tmp12],       %[tmp13]         \n\t"
+
+      /* tmp8 = (40 50 60 70 41 51 61 71) */
+      "punpcklhw  %[tmp8],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp9 = (42 52 62 72 43 53 63 73) */
+      "punpckhhw  %[tmp9],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp10 = (44 54 64 74 45 55 65 75) */
+      "punpcklhw  %[tmp10],        %[tmp1],        %[tmp3]          \n\t"
+      /* tmp11 = (46 56 66 76 47 57 67 77) */
+      "punpckhhw  %[tmp11],        %[tmp1],        %[tmp3]          \n\t"
+
+      /* tmp0 = (00 10 20 30 40 50 60 70) */
+      "punpcklwd  %[tmp0],         %[tmp4],        %[tmp8]          \n\t"
+      /* tmp1 = (01 11 21 31 41 51 61 71) */
+      "punpckhwd  %[tmp1],         %[tmp4],        %[tmp8]          \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst])                     \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst])                     \n\t"
+
+      /* tmp0 = (02 12 22 32 42 52 62 72) */
+      "punpcklwd  %[tmp0],         %[tmp5],        %[tmp9]          \n\t"
+      /* tmp1 = (03 13 23 33 43 53 63 73) */
+      "punpckhwd  %[tmp1],         %[tmp5],        %[tmp9]          \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst])                     \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst])                     \n\t"
+
+      /* tmp0 = (04 14 24 34 44 54 64 74) */
+      "punpcklwd  %[tmp0],         %[tmp6],        %[tmp10]         \n\t"
+      /* tmp1 = (05 15 25 35 45 55 65 75) */
+      "punpckhwd  %[tmp1],         %[tmp6],        %[tmp10]         \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst])                     \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst])                     \n\t"
+
+      /* tmp0 = (06 16 26 36 46 56 66 76) */
+      "punpcklwd  %[tmp0],         %[tmp7],        %[tmp11]         \n\t"
+      /* tmp1 = (07 17 27 37 47 57 67 77) */
+      "punpckhwd  %[tmp1],         %[tmp7],        %[tmp11]         \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst])                     \n\t"
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst])                     \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst])                     \n\t"
+
+      "dadd       %[dst],          %[dst],         %[dst_stride]    \n\t"
+      "daddi      %[src],          %[src],          0x08            \n\t"
+      "daddi      %[width],        %[width],       -0x08            \n\t"
+      "bnez       %[width],        1b                               \n\t"
+
+      : [tmp0] "=&f"(tmp0), [tmp1] "=&f"(tmp1), [tmp2] "=&f"(tmp2),
+        [tmp3] "=&f"(tmp3), [tmp4] "=&f"(tmp4), [tmp5] "=&f"(tmp5),
+        [tmp6] "=&f"(tmp6), [tmp7] "=&f"(tmp7), [tmp8] "=&f"(tmp8),
+        [tmp9] "=&f"(tmp9), [tmp10] "=&f"(tmp10), [tmp11] "=&f"(tmp11),
+        [tmp12] "=&f"(tmp12), [tmp13] "=&f"(tmp13), [dst] "+&r"(dst),
+        [src_tmp] "+&r"(src_tmp)
+      : [src] "r"(src), [width] "r"(width), [src_stride] "r"(src_stride),
+        [dst_stride] "r"(dst_stride)
+      : "memory");
+}
+
+void TransposeUVWx8_MMI(const uint8_t* src,
+                        int src_stride,
+                        uint8_t* dst_a,
+                        int dst_stride_a,
+                        uint8_t* dst_b,
+                        int dst_stride_b,
+                        int width) {
+  uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
+  uint64_t tmp7, tmp8, tmp9, tmp10, tmp11, tmp12, tmp13;
+  uint8_t* src_tmp = nullptr;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      /* tmp12 = (u00 v00 u01 v01 u02 v02 u03 v03) */
+      "ldc1       %[tmp12],        0x00(%[src])                     \n\t"
+      "dadd       %[src_tmp],      %[src],         %[src_stride]    \n\t"
+      /* tmp13 = (u10 v10 u11 v11 u12 v12 u13 v13) */
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                  \n\t"
+
+      /* tmp0 = (u00 u10 v00 v10 u01 u11 v01 v11) */
+      "punpcklbh  %[tmp0],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp1 = (u02 u12 v02 v12 u03 u13 v03 v13) */
+      "punpckhbh  %[tmp1],         %[tmp12],       %[tmp13]         \n\t"
+
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      /* tmp12 = (u20 v20 u21 v21 u22 v22 u23 v23) */
+      "ldc1       %[tmp12],        0x00(%[src_tmp])                 \n\t"
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      /* tmp13 = (u30 v30 u31 v31 u32 v32 u33 v33) */
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp2 = (u20 u30 v20 v30 u21 u31 v21 v31) */
+      "punpcklbh  %[tmp2],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp3 = (u22 u32 v22 v32 u23 u33 v23 v33) */
+      "punpckhbh  %[tmp3],         %[tmp12],       %[tmp13]         \n\t"
+
+      /* tmp4 = (u00 u10 u20 u30 v00 v10 v20 v30) */
+      "punpcklhw  %[tmp4],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp5 = (u01 u11 u21 u31 v01 v11 v21 v31) */
+      "punpckhhw  %[tmp5],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp6 = (u02 u12 u22 u32 v02 v12 v22 v32) */
+      "punpcklhw  %[tmp6],         %[tmp1],        %[tmp3]          \n\t"
+      /* tmp7 = (u03 u13 u23 u33 v03 v13 v23 v33) */
+      "punpckhhw  %[tmp7],         %[tmp1],        %[tmp3]          \n\t"
+
+      "dadd       %[src_tmp],     %[src_tmp],      %[src_stride]    \n\t"
+      /* tmp12 = (u40 v40 u41 v41 u42 v42 u43 v43) */
+      "ldc1       %[tmp12],        0x00(%[src_tmp])                 \n\t"
+      /* tmp13 = (u50 v50 u51 v51 u52 v52 u53 v53) */
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp0 = (u40 u50 v40 v50 u41 u51 v41 v51) */
+      "punpcklbh  %[tmp0],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp1 = (u42 u52 v42 v52 u43 u53 v43 v53) */
+      "punpckhbh  %[tmp1],         %[tmp12],       %[tmp13]         \n\t"
+
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      /* tmp12 = (u60 v60 u61 v61 u62 v62 u63 v63) */
+      "ldc1       %[tmp12],        0x00(%[src_tmp])                 \n\t"
+      /* tmp13 = (u70 v70 u71 v71 u72 v72 u73 v73) */
+      "dadd       %[src_tmp],      %[src_tmp],     %[src_stride]    \n\t"
+      "ldc1       %[tmp13],        0x00(%[src_tmp])                 \n\t"
+
+      /* tmp2 = (u60 u70 v60 v70 u61 u71 v61 v71) */
+      "punpcklbh  %[tmp2],         %[tmp12],       %[tmp13]         \n\t"
+      /* tmp3 = (u62 u72 v62 v72 u63 u73 v63 v73) */
+      "punpckhbh  %[tmp3],         %[tmp12],       %[tmp13]         \n\t"
+
+      /* tmp8 = (u40 u50 u60 u70 v40 v50 v60 v70) */
+      "punpcklhw  %[tmp8],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp9 = (u41 u51 u61 u71 v41 v51 v61 v71) */
+      "punpckhhw  %[tmp9],         %[tmp0],        %[tmp2]          \n\t"
+      /* tmp10 = (u42 u52 u62 u72 v42 v52 v62 v72) */
+      "punpcklhw  %[tmp10],        %[tmp1],        %[tmp3]          \n\t"
+      /* tmp11 = (u43 u53 u63 u73 v43 v53 v63 v73) */
+      "punpckhhw  %[tmp11],        %[tmp1],        %[tmp3]          \n\t"
+
+      /* tmp0 = (u00 u10 u20 u30 u40 u50 u60 u70) */
+      "punpcklwd  %[tmp0],         %[tmp4],        %[tmp8]          \n\t"
+      /* tmp1 = (v00 v10 v20 v30 v40 v50 v60 v70) */
+      "punpckhwd  %[tmp1],         %[tmp4],        %[tmp8]          \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst_a])                   \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst_a])                   \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst_b])                   \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst_b])                   \n\t"
+
+      /* tmp0 = (u01 u11 u21 u31 u41 u51 u61 u71) */
+      "punpcklwd  %[tmp0],         %[tmp5],        %[tmp9]          \n\t"
+      /* tmp1 = (v01 v11 v21 v31 v41 v51 v61 v71) */
+      "punpckhwd  %[tmp1],         %[tmp5],        %[tmp9]          \n\t"
+      "dadd       %[dst_a],        %[dst_a],       %[dst_stride_a]  \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst_a])                   \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst_a])                   \n\t"
+      "dadd       %[dst_b],        %[dst_b],       %[dst_stride_b]  \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst_b])                   \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst_b])                   \n\t"
+
+      /* tmp0 = (u02 u12 u22 u32 u42 u52 u62 u72) */
+      "punpcklwd  %[tmp0],         %[tmp6],        %[tmp10]         \n\t"
+      /* tmp1 = (v02 v12 v22 v32 v42 v52 v62 v72) */
+      "punpckhwd  %[tmp1],         %[tmp6],        %[tmp10]         \n\t"
+      "dadd       %[dst_a],        %[dst_a],       %[dst_stride_a]  \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst_a])                   \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst_a])                   \n\t"
+      "dadd       %[dst_b],        %[dst_b],       %[dst_stride_b]  \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst_b])                   \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst_b])                   \n\t"
+
+      /* tmp0 = (u03 u13 u23 u33 u43 u53 u63 u73) */
+      "punpcklwd  %[tmp0],         %[tmp7],        %[tmp11]         \n\t"
+      /* tmp1 = (v03 v13 v23 v33 v43 v53 v63 v73) */
+      "punpckhwd  %[tmp1],         %[tmp7],        %[tmp11]         \n\t"
+      "dadd       %[dst_a],        %[dst_a],       %[dst_stride_a]  \n\t"
+      "gssdlc1    %[tmp0],         0x07(%[dst_a])                   \n\t"
+      "gssdrc1    %[tmp0],         0x00(%[dst_a])                   \n\t"
+      "dadd       %[dst_b],        %[dst_b],       %[dst_stride_b]  \n\t"
+      "gssdlc1    %[tmp1],         0x07(%[dst_b])                   \n\t"
+      "gssdrc1    %[tmp1],         0x00(%[dst_b])                   \n\t"
+
+      "dadd       %[dst_a],        %[dst_a],       %[dst_stride_a]  \n\t"
+      "dadd       %[dst_b],        %[dst_b],       %[dst_stride_b]  \n\t"
+      "daddiu     %[src],          %[src],          0x08            \n\t"
+      "daddi      %[width],        %[width],       -0x04            \n\t"
+      "bnez       %[width],        1b                               \n\t"
+
+      : [tmp0] "=&f"(tmp0), [tmp1] "=&f"(tmp1), [tmp2] "=&f"(tmp2),
+        [tmp3] "=&f"(tmp3), [tmp4] "=&f"(tmp4), [tmp5] "=&f"(tmp5),
+        [tmp6] "=&f"(tmp6), [tmp7] "=&f"(tmp7), [tmp8] "=&f"(tmp8),
+        [tmp9] "=&f"(tmp9), [tmp10] "=&f"(tmp10), [tmp11] "=&f"(tmp11),
+        [tmp12] "=&f"(tmp12), [tmp13] "=&f"(tmp13), [dst_a] "+&r"(dst_a),
+        [dst_b] "+&r"(dst_b), [src_tmp] "+&r"(src_tmp)
+      : [src] "r"(src), [width] "r"(width), [dst_stride_a] "r"(dst_stride_a),
+        [dst_stride_b] "r"(dst_stride_b), [src_stride] "r"(src_stride)
+      : "memory");
+}
+
+#endif  // !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/rotate_msa.cc b/media/libyuv/source/rotate_msa.cc
new file mode 100644
index 0000000000..99bdca65b3
--- /dev/null
+++ b/media/libyuv/source/rotate_msa.cc
@@ -0,0 +1,250 @@
+/*
+ *  Copyright 2016 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ILVRL_B(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_b((v16i8)in1, (v16i8)in0);     \
+    out1 = (v16u8)__msa_ilvl_b((v16i8)in1, (v16i8)in0);     \
+    out2 = (v16u8)__msa_ilvr_b((v16i8)in3, (v16i8)in2);     \
+    out3 = (v16u8)__msa_ilvl_b((v16i8)in3, (v16i8)in2);     \
+  }
+
+#define ILVRL_H(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_h((v8i16)in1, (v8i16)in0);     \
+    out1 = (v16u8)__msa_ilvl_h((v8i16)in1, (v8i16)in0);     \
+    out2 = (v16u8)__msa_ilvr_h((v8i16)in3, (v8i16)in2);     \
+    out3 = (v16u8)__msa_ilvl_h((v8i16)in3, (v8i16)in2);     \
+  }
+
+#define ILVRL_W(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_w((v4i32)in1, (v4i32)in0);     \
+    out1 = (v16u8)__msa_ilvl_w((v4i32)in1, (v4i32)in0);     \
+    out2 = (v16u8)__msa_ilvr_w((v4i32)in3, (v4i32)in2);     \
+    out3 = (v16u8)__msa_ilvl_w((v4i32)in3, (v4i32)in2);     \
+  }
+
+#define ILVRL_D(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_d((v2i64)in1, (v2i64)in0);     \
+    out1 = (v16u8)__msa_ilvl_d((v2i64)in1, (v2i64)in0);     \
+    out2 = (v16u8)__msa_ilvr_d((v2i64)in3, (v2i64)in2);     \
+    out3 = (v16u8)__msa_ilvl_d((v2i64)in3, (v2i64)in2);     \
+  }
+
+void TransposeWx16_C(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst,
+                     int dst_stride,
+                     int width) {
+  TransposeWx8_C(src, src_stride, dst, dst_stride, width);
+  TransposeWx8_C((src + 8 * src_stride), src_stride, (dst + 8), dst_stride,
+                 width);
+}
+
+void TransposeUVWx16_C(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst_a,
+                       int dst_stride_a,
+                       uint8_t* dst_b,
+                       int dst_stride_b,
+                       int width) {
+  TransposeUVWx8_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width);
+  TransposeUVWx8_C((src + 8 * src_stride), src_stride, (dst_a + 8),
+                   dst_stride_a, (dst_b + 8), dst_stride_b, width);
+}
+
+void TransposeWx16_MSA(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width) {
+  int x;
+  const uint8_t* s;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < width; x += 16) {
+    s = src;
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    ILVRL_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVRL_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);
+    ILVRL_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    dst += dst_stride * 4;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);
+    ILVRL_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    dst += dst_stride * 4;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);
+    ILVRL_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    dst += dst_stride * 4;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);
+    ILVRL_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    src += 16;
+    dst += dst_stride * 4;
+  }
+}
+
+void TransposeUVWx16_MSA(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  int x;
+  const uint8_t* s;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < width; x += 8) {
+    s = src;
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    ILVRL_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVRL_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);
+    ILVRL_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);
+    ILVRL_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);
+    ILVRL_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);
+    ILVRL_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    src += 16;
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
diff --git a/media/libyuv/source/rotate_neon.cc b/media/libyuv/source/rotate_neon.cc
index 1c22b472bc..844df2bf30 100644
--- a/media/libyuv/source/rotate_neon.cc
+++ b/media/libyuv/source/rotate_neon.cc
@@ -8,8 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/row.h"
 #include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
 
 #include "libyuv/basic_types.h"
 
@@ -21,509 +21,394 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
     !defined(__aarch64__)
 
-static uvec8 kVTbl4x4Transpose =
-  { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
+static const uvec8 kVTbl4x4Transpose = {0, 4, 8,  12, 1, 5, 9,  13,
+                                        2, 6, 10, 14, 3, 7, 11, 15};
 
-void TransposeWx8_NEON(const uint8* src, int src_stride,
-                       uint8* dst, int dst_stride,
+void TransposeWx8_NEON(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
                        int width) {
-  const uint8* src_temp;
-  asm volatile (
-    // loops are on blocks of 8. loop will stop when
-    // counter gets to or below 0. starting the counter
-    // at w-8 allow for this
-    "sub         %5, #8                        \n"
-
-    // handle 8x8 blocks. this should be the majority of the plane
-    "1:                                        \n"
-      "mov         %0, %1                      \n"
-
-      MEMACCESS(0)
-      "vld1.8      {d0}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d1}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d2}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d3}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d4}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d5}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d6}, [%0], %2              \n"
-      MEMACCESS(0)
-      "vld1.8      {d7}, [%0]                  \n"
-
-      "vtrn.8      d1, d0                      \n"
-      "vtrn.8      d3, d2                      \n"
-      "vtrn.8      d5, d4                      \n"
-      "vtrn.8      d7, d6                      \n"
-
-      "vtrn.16     d1, d3                      \n"
-      "vtrn.16     d0, d2                      \n"
-      "vtrn.16     d5, d7                      \n"
-      "vtrn.16     d4, d6                      \n"
-
-      "vtrn.32     d1, d5                      \n"
-      "vtrn.32     d0, d4                      \n"
-      "vtrn.32     d3, d7                      \n"
-      "vtrn.32     d2, d6                      \n"
-
-      "vrev16.8    q0, q0                      \n"
-      "vrev16.8    q1, q1                      \n"
-      "vrev16.8    q2, q2                      \n"
-      "vrev16.8    q3, q3                      \n"
-
-      "mov         %0, %3                      \n"
-
-    MEMACCESS(0)
-      "vst1.8      {d1}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d0}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d3}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d2}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d5}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d4}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d7}, [%0], %4              \n"
-    MEMACCESS(0)
-      "vst1.8      {d6}, [%0]                  \n"
-
-      "add         %1, #8                      \n"  // src += 8
-      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
-      "subs        %5,  #8                     \n"  // w   -= 8
-      "bge         1b                          \n"
-
-    // add 8 back to counter. if the result is 0 there are
-    // no residuals.
-    "adds        %5, #8                        \n"
-    "beq         4f                            \n"
-
-    // some residual, so between 1 and 7 lines left to transpose
-    "cmp         %5, #2                        \n"
-    "blt         3f                            \n"
-
-    "cmp         %5, #4                        \n"
-    "blt         2f                            \n"
-
-    // 4x8 block
-    "mov         %0, %1                        \n"
-    MEMACCESS(0)
-    "vld1.32     {d0[0]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d0[1]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d1[0]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d1[1]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d2[0]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d2[1]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d3[0]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.32     {d3[1]}, [%0]                 \n"
-
-    "mov         %0, %3                        \n"
-
-    MEMACCESS(6)
-    "vld1.8      {q3}, [%6]                    \n"
-
-    "vtbl.8      d4, {d0, d1}, d6              \n"
-    "vtbl.8      d5, {d0, d1}, d7              \n"
-    "vtbl.8      d0, {d2, d3}, d6              \n"
-    "vtbl.8      d1, {d2, d3}, d7              \n"
-
-    // TODO(frkoenig): Rework shuffle above to
-    // write out with 4 instead of 8 writes.
-    MEMACCESS(0)
-    "vst1.32     {d4[0]}, [%0], %4             \n"
-    MEMACCESS(0)
-    "vst1.32     {d4[1]}, [%0], %4             \n"
-    MEMACCESS(0)
-    "vst1.32     {d5[0]}, [%0], %4             \n"
-    MEMACCESS(0)
-    "vst1.32     {d5[1]}, [%0]                 \n"
-
-    "add         %0, %3, #4                    \n"
-    MEMACCESS(0)
-    "vst1.32     {d0[0]}, [%0], %4             \n"
-    MEMACCESS(0)
-    "vst1.32     {d0[1]}, [%0], %4             \n"
-    MEMACCESS(0)
-    "vst1.32     {d1[0]}, [%0], %4             \n"
-    MEMACCESS(0)
-    "vst1.32     {d1[1]}, [%0]                 \n"
-
-    "add         %1, #4                        \n"  // src += 4
-    "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
-    "subs        %5,  #4                       \n"  // w   -= 4
-    "beq         4f                            \n"
-
-    // some residual, check to see if it includes a 2x8 block,
-    // or less
-    "cmp         %5, #2                        \n"
-    "blt         3f                            \n"
-
-    // 2x8 block
-    "2:                                        \n"
-    "mov         %0, %1                        \n"
-    MEMACCESS(0)
-    "vld1.16     {d0[0]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d1[0]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d0[1]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d1[1]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d0[2]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d1[2]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d0[3]}, [%0], %2             \n"
-    MEMACCESS(0)
-    "vld1.16     {d1[3]}, [%0]                 \n"
-
-    "vtrn.8      d0, d1                        \n"
-
-    "mov         %0, %3                        \n"
-
-    MEMACCESS(0)
-    "vst1.64     {d0}, [%0], %4                \n"
-    MEMACCESS(0)
-    "vst1.64     {d1}, [%0]                    \n"
-
-    "add         %1, #2                        \n"  // src += 2
-    "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
-    "subs        %5,  #2                       \n"  // w   -= 2
-    "beq         4f                            \n"
-
-    // 1x8 block
-    "3:                                        \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[0]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[1]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[2]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[3]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[4]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[5]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[6]}, [%1], %2             \n"
-    MEMACCESS(1)
-    "vld1.8      {d0[7]}, [%1]                 \n"
-
-    MEMACCESS(3)
-    "vst1.64     {d0}, [%3]                    \n"
-
-    "4:                                        \n"
-
-    : "=&r"(src_temp),         // %0
-      "+r"(src),               // %1
-      "+r"(src_stride),        // %2
-      "+r"(dst),               // %3
-      "+r"(dst_stride),        // %4
-      "+r"(width)              // %5
-    : "r"(&kVTbl4x4Transpose)  // %6
-    : "memory", "cc", "q0", "q1", "q2", "q3"
-  );
+  const uint8_t* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %5, #8                        \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov         %0, %1                        \n"
+
+      "vld1.8      {d0}, [%0], %2                \n"
+      "vld1.8      {d1}, [%0], %2                \n"
+      "vld1.8      {d2}, [%0], %2                \n"
+      "vld1.8      {d3}, [%0], %2                \n"
+      "vld1.8      {d4}, [%0], %2                \n"
+      "vld1.8      {d5}, [%0], %2                \n"
+      "vld1.8      {d6}, [%0], %2                \n"
+      "vld1.8      {d7}, [%0]                    \n"
+
+      "vtrn.8      d1, d0                        \n"
+      "vtrn.8      d3, d2                        \n"
+      "vtrn.8      d5, d4                        \n"
+      "vtrn.8      d7, d6                        \n"
+
+      "vtrn.16     d1, d3                        \n"
+      "vtrn.16     d0, d2                        \n"
+      "vtrn.16     d5, d7                        \n"
+      "vtrn.16     d4, d6                        \n"
+
+      "vtrn.32     d1, d5                        \n"
+      "vtrn.32     d0, d4                        \n"
+      "vtrn.32     d3, d7                        \n"
+      "vtrn.32     d2, d6                        \n"
+
+      "vrev16.8    q0, q0                        \n"
+      "vrev16.8    q1, q1                        \n"
+      "vrev16.8    q2, q2                        \n"
+      "vrev16.8    q3, q3                        \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.8      {d1}, [%0], %4                \n"
+      "vst1.8      {d0}, [%0], %4                \n"
+      "vst1.8      {d3}, [%0], %4                \n"
+      "vst1.8      {d2}, [%0], %4                \n"
+      "vst1.8      {d5}, [%0], %4                \n"
+      "vst1.8      {d4}, [%0], %4                \n"
+      "vst1.8      {d7}, [%0], %4                \n"
+      "vst1.8      {d6}, [%0]                    \n"
+
+      "add         %1, #8                        \n"  // src += 8
+      "add         %3, %3, %4, lsl #3            \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                       \n"  // w   -= 8
+      "bge         1b                            \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %5, #8                        \n"
+      "beq         4f                            \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %5, #2                        \n"
+      "blt         3f                            \n"
+
+      "cmp         %5, #4                        \n"
+      "blt         2f                            \n"
+
+      // 4x8 block
+      "mov         %0, %1                        \n"
+      "vld1.32     {d0[0]}, [%0], %2             \n"
+      "vld1.32     {d0[1]}, [%0], %2             \n"
+      "vld1.32     {d1[0]}, [%0], %2             \n"
+      "vld1.32     {d1[1]}, [%0], %2             \n"
+      "vld1.32     {d2[0]}, [%0], %2             \n"
+      "vld1.32     {d2[1]}, [%0], %2             \n"
+      "vld1.32     {d3[0]}, [%0], %2             \n"
+      "vld1.32     {d3[1]}, [%0]                 \n"
+
+      "mov         %0, %3                        \n"
+
+      "vld1.8      {q3}, [%6]                    \n"
+
+      "vtbl.8      d4, {d0, d1}, d6              \n"
+      "vtbl.8      d5, {d0, d1}, d7              \n"
+      "vtbl.8      d0, {d2, d3}, d6              \n"
+      "vtbl.8      d1, {d2, d3}, d7              \n"
+
+      // TODO(frkoenig): Rework shuffle above to
+      // write out with 4 instead of 8 writes.
+      "vst1.32     {d4[0]}, [%0], %4             \n"
+      "vst1.32     {d4[1]}, [%0], %4             \n"
+      "vst1.32     {d5[0]}, [%0], %4             \n"
+      "vst1.32     {d5[1]}, [%0]                 \n"
+
+      "add         %0, %3, #4                    \n"
+      "vst1.32     {d0[0]}, [%0], %4             \n"
+      "vst1.32     {d0[1]}, [%0], %4             \n"
+      "vst1.32     {d1[0]}, [%0], %4             \n"
+      "vst1.32     {d1[1]}, [%0]                 \n"
+
+      "add         %1, #4                        \n"  // src += 4
+      "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
+      "subs        %5,  #4                       \n"  // w   -= 4
+      "beq         4f                            \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp         %5, #2                        \n"
+      "blt         3f                            \n"
+
+      // 2x8 block
+      "2:                                        \n"
+      "mov         %0, %1                        \n"
+      "vld1.16     {d0[0]}, [%0], %2             \n"
+      "vld1.16     {d1[0]}, [%0], %2             \n"
+      "vld1.16     {d0[1]}, [%0], %2             \n"
+      "vld1.16     {d1[1]}, [%0], %2             \n"
+      "vld1.16     {d0[2]}, [%0], %2             \n"
+      "vld1.16     {d1[2]}, [%0], %2             \n"
+      "vld1.16     {d0[3]}, [%0], %2             \n"
+      "vld1.16     {d1[3]}, [%0]                 \n"
+
+      "vtrn.8      d0, d1                        \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.64     {d0}, [%0], %4                \n"
+      "vst1.64     {d1}, [%0]                    \n"
+
+      "add         %1, #2                        \n"  // src += 2
+      "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
+      "subs        %5,  #2                       \n"  // w   -= 2
+      "beq         4f                            \n"
+
+      // 1x8 block
+      "3:                                        \n"
+      "vld1.8      {d0[0]}, [%1], %2             \n"
+      "vld1.8      {d0[1]}, [%1], %2             \n"
+      "vld1.8      {d0[2]}, [%1], %2             \n"
+      "vld1.8      {d0[3]}, [%1], %2             \n"
+      "vld1.8      {d0[4]}, [%1], %2             \n"
+      "vld1.8      {d0[5]}, [%1], %2             \n"
+      "vld1.8      {d0[6]}, [%1], %2             \n"
+      "vld1.8      {d0[7]}, [%1]                 \n"
+
+      "vst1.64     {d0}, [%3]                    \n"
+
+      "4:                                        \n"
+
+      : "=&r"(src_temp),         // %0
+        "+r"(src),               // %1
+        "+r"(src_stride),        // %2
+        "+r"(dst),               // %3
+        "+r"(dst_stride),        // %4
+        "+r"(width)              // %5
+      : "r"(&kVTbl4x4Transpose)  // %6
+      : "memory", "cc", "q0", "q1", "q2", "q3");
 }
 
-static uvec8 kVTbl4x4TransposeDi =
-  { 0,  8,  1,  9,  2, 10,  3, 11,  4, 12,  5, 13,  6, 14,  7, 15 };
+static const uvec8 kVTbl4x4TransposeDi = {0, 8,  1, 9,  2, 10, 3, 11,
+                                          4, 12, 5, 13, 6, 14, 7, 15};
 
-void TransposeUVWx8_NEON(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b,
+void TransposeUVWx8_NEON(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
                          int width) {
-  const uint8* src_temp;
-  asm volatile (
-    // loops are on blocks of 8. loop will stop when
-    // counter gets to or below 0. starting the counter
-    // at w-8 allow for this
-    "sub         %7, #8                        \n"
-
-    // handle 8x8 blocks. this should be the majority of the plane
-    "1:                                        \n"
-      "mov         %0, %1                      \n"
-
-      MEMACCESS(0)
-      "vld2.8      {d0,  d1},  [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d2,  d3},  [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d4,  d5},  [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d6,  d7},  [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d16, d17}, [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d18, d19}, [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d20, d21}, [%0], %2        \n"
-      MEMACCESS(0)
-      "vld2.8      {d22, d23}, [%0]            \n"
-
-      "vtrn.8      q1, q0                      \n"
-      "vtrn.8      q3, q2                      \n"
-      "vtrn.8      q9, q8                      \n"
-      "vtrn.8      q11, q10                    \n"
-
-      "vtrn.16     q1, q3                      \n"
-      "vtrn.16     q0, q2                      \n"
-      "vtrn.16     q9, q11                     \n"
-      "vtrn.16     q8, q10                     \n"
-
-      "vtrn.32     q1, q9                      \n"
-      "vtrn.32     q0, q8                      \n"
-      "vtrn.32     q3, q11                     \n"
-      "vtrn.32     q2, q10                     \n"
-
-      "vrev16.8    q0, q0                      \n"
-      "vrev16.8    q1, q1                      \n"
-      "vrev16.8    q2, q2                      \n"
-      "vrev16.8    q3, q3                      \n"
-      "vrev16.8    q8, q8                      \n"
-      "vrev16.8    q9, q9                      \n"
-      "vrev16.8    q10, q10                    \n"
-      "vrev16.8    q11, q11                    \n"
-
-      "mov         %0, %3                      \n"
-
-    MEMACCESS(0)
-      "vst1.8      {d2},  [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d0},  [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d6},  [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d4},  [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d18}, [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d16}, [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d22}, [%0], %4             \n"
-    MEMACCESS(0)
-      "vst1.8      {d20}, [%0]                 \n"
-
-      "mov         %0, %5                      \n"
-
-    MEMACCESS(0)
-      "vst1.8      {d3},  [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d1},  [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d7},  [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d5},  [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d19}, [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d17}, [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d23}, [%0], %6             \n"
-    MEMACCESS(0)
-      "vst1.8      {d21}, [%0]                 \n"
-
-      "add         %1, #8*2                    \n"  // src   += 8*2
-      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
-      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
-      "subs        %7,  #8                     \n"  // w     -= 8
-      "bge         1b                          \n"
-
-    // add 8 back to counter. if the result is 0 there are
-    // no residuals.
-    "adds        %7, #8                        \n"
-    "beq         4f                            \n"
-
-    // some residual, so between 1 and 7 lines left to transpose
-    "cmp         %7, #2                        \n"
-    "blt         3f                            \n"
-
-    "cmp         %7, #4                        \n"
-    "blt         2f                            \n"
-
-    // TODO(frkoenig): Clean this up
-    // 4x8 block
-    "mov         %0, %1                        \n"
-    MEMACCESS(0)
-    "vld1.64     {d0}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d1}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d2}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d3}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d4}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d5}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d6}, [%0], %2                \n"
-    MEMACCESS(0)
-    "vld1.64     {d7}, [%0]                    \n"
-
-    MEMACCESS(8)
-    "vld1.8      {q15}, [%8]                   \n"
-
-    "vtrn.8      q0, q1                        \n"
-    "vtrn.8      q2, q3                        \n"
-
-    "vtbl.8      d16, {d0, d1}, d30            \n"
-    "vtbl.8      d17, {d0, d1}, d31            \n"
-    "vtbl.8      d18, {d2, d3}, d30            \n"
-    "vtbl.8      d19, {d2, d3}, d31            \n"
-    "vtbl.8      d20, {d4, d5}, d30            \n"
-    "vtbl.8      d21, {d4, d5}, d31            \n"
-    "vtbl.8      d22, {d6, d7}, d30            \n"
-    "vtbl.8      d23, {d6, d7}, d31            \n"
-
-    "mov         %0, %3                        \n"
-
-    MEMACCESS(0)
-    "vst1.32     {d16[0]},  [%0], %4           \n"
-    MEMACCESS(0)
-    "vst1.32     {d16[1]},  [%0], %4           \n"
-    MEMACCESS(0)
-    "vst1.32     {d17[0]},  [%0], %4           \n"
-    MEMACCESS(0)
-    "vst1.32     {d17[1]},  [%0], %4           \n"
-
-    "add         %0, %3, #4                    \n"
-    MEMACCESS(0)
-    "vst1.32     {d20[0]}, [%0], %4            \n"
-    MEMACCESS(0)
-    "vst1.32     {d20[1]}, [%0], %4            \n"
-    MEMACCESS(0)
-    "vst1.32     {d21[0]}, [%0], %4            \n"
-    MEMACCESS(0)
-    "vst1.32     {d21[1]}, [%0]                \n"
-
-    "mov         %0, %5                        \n"
-
-    MEMACCESS(0)
-    "vst1.32     {d18[0]}, [%0], %6            \n"
-    MEMACCESS(0)
-    "vst1.32     {d18[1]}, [%0], %6            \n"
-    MEMACCESS(0)
-    "vst1.32     {d19[0]}, [%0], %6            \n"
-    MEMACCESS(0)
-    "vst1.32     {d19[1]}, [%0], %6            \n"
-
-    "add         %0, %5, #4                    \n"
-    MEMACCESS(0)
-    "vst1.32     {d22[0]},  [%0], %6           \n"
-    MEMACCESS(0)
-    "vst1.32     {d22[1]},  [%0], %6           \n"
-    MEMACCESS(0)
-    "vst1.32     {d23[0]},  [%0], %6           \n"
-    MEMACCESS(0)
-    "vst1.32     {d23[1]},  [%0]               \n"
-
-    "add         %1, #4*2                      \n"  // src   += 4 * 2
-    "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 * dst_stride_a
-    "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 * dst_stride_b
-    "subs        %7,  #4                       \n"  // w     -= 4
-    "beq         4f                            \n"
-
-    // some residual, check to see if it includes a 2x8 block,
-    // or less
-    "cmp         %7, #2                        \n"
-    "blt         3f                            \n"
-
-    // 2x8 block
-    "2:                                        \n"
-    "mov         %0, %1                        \n"
-    MEMACCESS(0)
-    "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
-    MEMACCESS(0)
-    "vld2.16     {d1[3], d3[3]}, [%0]          \n"
-
-    "vtrn.8      d0, d1                        \n"
-    "vtrn.8      d2, d3                        \n"
-
-    "mov         %0, %3                        \n"
-
-    MEMACCESS(0)
-    "vst1.64     {d0}, [%0], %4                \n"
-    MEMACCESS(0)
-    "vst1.64     {d2}, [%0]                    \n"
-
-    "mov         %0, %5                        \n"
-
-    MEMACCESS(0)
-    "vst1.64     {d1}, [%0], %6                \n"
-    MEMACCESS(0)
-    "vst1.64     {d3}, [%0]                    \n"
-
-    "add         %1, #2*2                      \n"  // src   += 2 * 2
-    "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 * dst_stride_a
-    "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 * dst_stride_b
-    "subs        %7,  #2                       \n"  // w     -= 2
-    "beq         4f                            \n"
-
-    // 1x8 block
-    "3:                                        \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
-    MEMACCESS(1)
-    "vld2.8      {d0[7], d1[7]}, [%1]          \n"
-
-    MEMACCESS(3)
-    "vst1.64     {d0}, [%3]                    \n"
-    MEMACCESS(5)
-    "vst1.64     {d1}, [%5]                    \n"
-
-    "4:                                        \n"
-
-    : "=&r"(src_temp),           // %0
-      "+r"(src),                 // %1
-      "+r"(src_stride),          // %2
-      "+r"(dst_a),               // %3
-      "+r"(dst_stride_a),        // %4
-      "+r"(dst_b),               // %5
-      "+r"(dst_stride_b),        // %6
-      "+r"(width)                // %7
-    : "r"(&kVTbl4x4TransposeDi)  // %8
-    : "memory", "cc",
-      "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
-  );
+  const uint8_t* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %7, #8                        \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov         %0, %1                        \n"
+
+      "vld2.8      {d0,  d1},  [%0], %2          \n"
+      "vld2.8      {d2,  d3},  [%0], %2          \n"
+      "vld2.8      {d4,  d5},  [%0], %2          \n"
+      "vld2.8      {d6,  d7},  [%0], %2          \n"
+      "vld2.8      {d16, d17}, [%0], %2          \n"
+      "vld2.8      {d18, d19}, [%0], %2          \n"
+      "vld2.8      {d20, d21}, [%0], %2          \n"
+      "vld2.8      {d22, d23}, [%0]              \n"
+
+      "vtrn.8      q1, q0                        \n"
+      "vtrn.8      q3, q2                        \n"
+      "vtrn.8      q9, q8                        \n"
+      "vtrn.8      q11, q10                      \n"
+
+      "vtrn.16     q1, q3                        \n"
+      "vtrn.16     q0, q2                        \n"
+      "vtrn.16     q9, q11                       \n"
+      "vtrn.16     q8, q10                       \n"
+
+      "vtrn.32     q1, q9                        \n"
+      "vtrn.32     q0, q8                        \n"
+      "vtrn.32     q3, q11                       \n"
+      "vtrn.32     q2, q10                       \n"
+
+      "vrev16.8    q0, q0                        \n"
+      "vrev16.8    q1, q1                        \n"
+      "vrev16.8    q2, q2                        \n"
+      "vrev16.8    q3, q3                        \n"
+      "vrev16.8    q8, q8                        \n"
+      "vrev16.8    q9, q9                        \n"
+      "vrev16.8    q10, q10                      \n"
+      "vrev16.8    q11, q11                      \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.8      {d2},  [%0], %4               \n"
+      "vst1.8      {d0},  [%0], %4               \n"
+      "vst1.8      {d6},  [%0], %4               \n"
+      "vst1.8      {d4},  [%0], %4               \n"
+      "vst1.8      {d18}, [%0], %4               \n"
+      "vst1.8      {d16}, [%0], %4               \n"
+      "vst1.8      {d22}, [%0], %4               \n"
+      "vst1.8      {d20}, [%0]                   \n"
+
+      "mov         %0, %5                        \n"
+
+      "vst1.8      {d3},  [%0], %6               \n"
+      "vst1.8      {d1},  [%0], %6               \n"
+      "vst1.8      {d7},  [%0], %6               \n"
+      "vst1.8      {d5},  [%0], %6               \n"
+      "vst1.8      {d19}, [%0], %6               \n"
+      "vst1.8      {d17}, [%0], %6               \n"
+      "vst1.8      {d23}, [%0], %6               \n"
+      "vst1.8      {d21}, [%0]                   \n"
+
+      "add         %1, #8*2                      \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3            \n"  // dst_a += 8 *
+                                                      // dst_stride_a
+      "add         %5, %5, %6, lsl #3            \n"  // dst_b += 8 *
+                                                      // dst_stride_b
+      "subs        %7,  #8                       \n"  // w     -= 8
+      "bge         1b                            \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %7, #8                        \n"
+      "beq         4f                            \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %7, #2                        \n"
+      "blt         3f                            \n"
+
+      "cmp         %7, #4                        \n"
+      "blt         2f                            \n"
+
+      // TODO(frkoenig): Clean this up
+      // 4x8 block
+      "mov         %0, %1                        \n"
+      "vld1.64     {d0}, [%0], %2                \n"
+      "vld1.64     {d1}, [%0], %2                \n"
+      "vld1.64     {d2}, [%0], %2                \n"
+      "vld1.64     {d3}, [%0], %2                \n"
+      "vld1.64     {d4}, [%0], %2                \n"
+      "vld1.64     {d5}, [%0], %2                \n"
+      "vld1.64     {d6}, [%0], %2                \n"
+      "vld1.64     {d7}, [%0]                    \n"
+
+      "vld1.8      {q15}, [%8]                   \n"
+
+      "vtrn.8      q0, q1                        \n"
+      "vtrn.8      q2, q3                        \n"
+
+      "vtbl.8      d16, {d0, d1}, d30            \n"
+      "vtbl.8      d17, {d0, d1}, d31            \n"
+      "vtbl.8      d18, {d2, d3}, d30            \n"
+      "vtbl.8      d19, {d2, d3}, d31            \n"
+      "vtbl.8      d20, {d4, d5}, d30            \n"
+      "vtbl.8      d21, {d4, d5}, d31            \n"
+      "vtbl.8      d22, {d6, d7}, d30            \n"
+      "vtbl.8      d23, {d6, d7}, d31            \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.32     {d16[0]},  [%0], %4           \n"
+      "vst1.32     {d16[1]},  [%0], %4           \n"
+      "vst1.32     {d17[0]},  [%0], %4           \n"
+      "vst1.32     {d17[1]},  [%0], %4           \n"
+
+      "add         %0, %3, #4                    \n"
+      "vst1.32     {d20[0]}, [%0], %4            \n"
+      "vst1.32     {d20[1]}, [%0], %4            \n"
+      "vst1.32     {d21[0]}, [%0], %4            \n"
+      "vst1.32     {d21[1]}, [%0]                \n"
+
+      "mov         %0, %5                        \n"
+
+      "vst1.32     {d18[0]}, [%0], %6            \n"
+      "vst1.32     {d18[1]}, [%0], %6            \n"
+      "vst1.32     {d19[0]}, [%0], %6            \n"
+      "vst1.32     {d19[1]}, [%0], %6            \n"
+
+      "add         %0, %5, #4                    \n"
+      "vst1.32     {d22[0]},  [%0], %6           \n"
+      "vst1.32     {d22[1]},  [%0], %6           \n"
+      "vst1.32     {d23[0]},  [%0], %6           \n"
+      "vst1.32     {d23[1]},  [%0]               \n"
+
+      "add         %1, #4*2                      \n"  // src   += 4 * 2
+      "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 *
+                                                      // dst_stride_a
+      "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 *
+                                                      // dst_stride_b
+      "subs        %7,  #4                       \n"  // w     -= 4
+      "beq         4f                            \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp         %7, #2                        \n"
+      "blt         3f                            \n"
+
+      // 2x8 block
+      "2:                                        \n"
+      "mov         %0, %1                        \n"
+      "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
+      "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
+      "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
+      "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
+      "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
+      "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
+      "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
+      "vld2.16     {d1[3], d3[3]}, [%0]          \n"
+
+      "vtrn.8      d0, d1                        \n"
+      "vtrn.8      d2, d3                        \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.64     {d0}, [%0], %4                \n"
+      "vst1.64     {d2}, [%0]                    \n"
+
+      "mov         %0, %5                        \n"
+
+      "vst1.64     {d1}, [%0], %6                \n"
+      "vst1.64     {d3}, [%0]                    \n"
+
+      "add         %1, #2*2                      \n"  // src   += 2 * 2
+      "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 *
+                                                      // dst_stride_a
+      "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 *
+                                                      // dst_stride_b
+      "subs        %7,  #2                       \n"  // w     -= 2
+      "beq         4f                            \n"
+
+      // 1x8 block
+      "3:                                        \n"
+      "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
+      "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
+      "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
+      "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
+      "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
+      "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
+      "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
+      "vld2.8      {d0[7], d1[7]}, [%1]          \n"
+
+      "vst1.64     {d0}, [%3]                    \n"
+      "vst1.64     {d1}, [%5]                    \n"
+
+      "4:                                        \n"
+
+      : "=&r"(src_temp),           // %0
+        "+r"(src),                 // %1
+        "+r"(src_stride),          // %2
+        "+r"(dst_a),               // %3
+        "+r"(dst_stride_a),        // %4
+        "+r"(dst_b),               // %5
+        "+r"(dst_stride_b),        // %6
+        "+r"(width)                // %7
+      : "r"(&kVTbl4x4TransposeDi)  // %8
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
 }
 #endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
diff --git a/media/libyuv/source/rotate_neon64.cc b/media/libyuv/source/rotate_neon64.cc
index 1ab448f3ab..95047fa7a5 100644
--- a/media/libyuv/source/rotate_neon64.cc
+++ b/media/libyuv/source/rotate_neon64.cc
@@ -8,8 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/row.h"
 #include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
 
 #include "libyuv/basic_types.h"
 
@@ -21,520 +21,459 @@ extern "C" {
 // This module is for GCC Neon armv8 64 bit.
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
-static uvec8 kVTbl4x4Transpose =
-  { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 };
-
-void TransposeWx8_NEON(const uint8* src, int src_stride,
-                       uint8* dst, int dst_stride, int width) {
-  const uint8* src_temp;
-  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
-  asm volatile (
-    // loops are on blocks of 8. loop will stop when
-    // counter gets to or below 0. starting the counter
-    // at w-8 allow for this
-    "sub         %3, %3, #8                      \n"
-
-    // handle 8x8 blocks. this should be the majority of the plane
-    "1:                                          \n"
+static const uvec8 kVTbl4x4Transpose = {0, 4, 8,  12, 1, 5, 9,  13,
+                                        2, 6, 10, 14, 3, 7, 11, 15};
+
+void TransposeWx8_NEON(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width) {
+  const uint8_t* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %w3, %w3, #8                  \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov         %0, %1                        \n"
+
+      "ld1         {v0.8b}, [%0], %5             \n"
+      "ld1         {v1.8b}, [%0], %5             \n"
+      "ld1         {v2.8b}, [%0], %5             \n"
+      "ld1         {v3.8b}, [%0], %5             \n"
+      "ld1         {v4.8b}, [%0], %5             \n"
+      "ld1         {v5.8b}, [%0], %5             \n"
+      "ld1         {v6.8b}, [%0], %5             \n"
+      "ld1         {v7.8b}, [%0]                 \n"
       "mov         %0, %1                        \n"
 
-      MEMACCESS(0)
-      "ld1        {v0.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v1.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v2.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v3.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v4.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v5.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v6.8b}, [%0], %5              \n"
-      MEMACCESS(0)
-      "ld1        {v7.8b}, [%0]                  \n"
-
-      "trn2     v16.8b, v0.8b, v1.8b             \n"
-      "trn1     v17.8b, v0.8b, v1.8b             \n"
-      "trn2     v18.8b, v2.8b, v3.8b             \n"
-      "trn1     v19.8b, v2.8b, v3.8b             \n"
-      "trn2     v20.8b, v4.8b, v5.8b             \n"
-      "trn1     v21.8b, v4.8b, v5.8b             \n"
-      "trn2     v22.8b, v6.8b, v7.8b             \n"
-      "trn1     v23.8b, v6.8b, v7.8b             \n"
-
-      "trn2     v3.4h, v17.4h, v19.4h            \n"
-      "trn1     v1.4h, v17.4h, v19.4h            \n"
-      "trn2     v2.4h, v16.4h, v18.4h            \n"
-      "trn1     v0.4h, v16.4h, v18.4h            \n"
-      "trn2     v7.4h, v21.4h, v23.4h            \n"
-      "trn1     v5.4h, v21.4h, v23.4h            \n"
-      "trn2     v6.4h, v20.4h, v22.4h            \n"
-      "trn1     v4.4h, v20.4h, v22.4h            \n"
-
-      "trn2     v21.2s, v1.2s, v5.2s             \n"
-      "trn1     v17.2s, v1.2s, v5.2s             \n"
-      "trn2     v20.2s, v0.2s, v4.2s             \n"
-      "trn1     v16.2s, v0.2s, v4.2s             \n"
-      "trn2     v23.2s, v3.2s, v7.2s             \n"
-      "trn1     v19.2s, v3.2s, v7.2s             \n"
-      "trn2     v22.2s, v2.2s, v6.2s             \n"
-      "trn1     v18.2s, v2.2s, v6.2s             \n"
+      "trn2        v16.8b, v0.8b, v1.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "trn1        v17.8b, v0.8b, v1.8b          \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v18.8b, v2.8b, v3.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 1
+      "trn1        v19.8b, v2.8b, v3.8b          \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v20.8b, v4.8b, v5.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 2
+      "trn1        v21.8b, v4.8b, v5.8b          \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v22.8b, v6.8b, v7.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 3
+      "trn1        v23.8b, v6.8b, v7.8b          \n"
+      "add         %0, %0, %5                    \n"
+
+      "trn2        v3.4h, v17.4h, v19.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 4
+      "trn1        v1.4h, v17.4h, v19.4h         \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v2.4h, v16.4h, v18.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 5
+      "trn1        v0.4h, v16.4h, v18.4h         \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v7.4h, v21.4h, v23.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 6
+      "trn1        v5.4h, v21.4h, v23.4h         \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v6.4h, v20.4h, v22.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 7
+      "trn1        v4.4h, v20.4h, v22.4h         \n"
+
+      "trn2        v21.2s, v1.2s, v5.2s          \n"
+      "trn1        v17.2s, v1.2s, v5.2s          \n"
+      "trn2        v20.2s, v0.2s, v4.2s          \n"
+      "trn1        v16.2s, v0.2s, v4.2s          \n"
+      "trn2        v23.2s, v3.2s, v7.2s          \n"
+      "trn1        v19.2s, v3.2s, v7.2s          \n"
+      "trn2        v22.2s, v2.2s, v6.2s          \n"
+      "trn1        v18.2s, v2.2s, v6.2s          \n"
 
       "mov         %0, %2                        \n"
 
-    MEMACCESS(0)
-      "st1      {v17.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v16.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v19.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v18.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v21.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v20.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v23.8b}, [%0], %6               \n"
-    MEMACCESS(0)
-      "st1      {v22.8b}, [%0]                   \n"
+      "st1         {v17.8b}, [%0], %6            \n"
+      "st1         {v16.8b}, [%0], %6            \n"
+      "st1         {v19.8b}, [%0], %6            \n"
+      "st1         {v18.8b}, [%0], %6            \n"
+      "st1         {v21.8b}, [%0], %6            \n"
+      "st1         {v20.8b}, [%0], %6            \n"
+      "st1         {v23.8b}, [%0], %6            \n"
+      "st1         {v22.8b}, [%0]                \n"
 
       "add         %1, %1, #8                    \n"  // src += 8
       "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
-      "subs        %3, %3, #8                    \n"  // w   -= 8
+      "subs        %w3, %w3, #8                  \n"  // w   -= 8
       "b.ge        1b                            \n"
 
-    // add 8 back to counter. if the result is 0 there are
-    // no residuals.
-    "adds        %3, %3, #8                      \n"
-    "b.eq        4f                              \n"
-
-    // some residual, so between 1 and 7 lines left to transpose
-    "cmp         %3, #2                          \n"
-    "b.lt        3f                              \n"
-
-    "cmp         %3, #4                          \n"
-    "b.lt        2f                              \n"
-
-    // 4x8 block
-    "mov         %0, %1                          \n"
-    MEMACCESS(0)
-    "ld1     {v0.s}[0], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v0.s}[1], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v0.s}[2], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v0.s}[3], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.s}[0], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.s}[1], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.s}[2], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.s}[3], [%0]                     \n"
-
-    "mov         %0, %2                          \n"
-
-    MEMACCESS(4)
-    "ld1      {v2.16b}, [%4]                     \n"
-
-    "tbl      v3.16b, {v0.16b}, v2.16b           \n"
-    "tbl      v0.16b, {v1.16b}, v2.16b           \n"
-
-    // TODO(frkoenig): Rework shuffle above to
-    // write out with 4 instead of 8 writes.
-    MEMACCESS(0)
-    "st1 {v3.s}[0], [%0], %6                     \n"
-    MEMACCESS(0)
-    "st1 {v3.s}[1], [%0], %6                     \n"
-    MEMACCESS(0)
-    "st1 {v3.s}[2], [%0], %6                     \n"
-    MEMACCESS(0)
-    "st1 {v3.s}[3], [%0]                         \n"
-
-    "add         %0, %2, #4                      \n"
-    MEMACCESS(0)
-    "st1 {v0.s}[0], [%0], %6                     \n"
-    MEMACCESS(0)
-    "st1 {v0.s}[1], [%0], %6                     \n"
-    MEMACCESS(0)
-    "st1 {v0.s}[2], [%0], %6                     \n"
-    MEMACCESS(0)
-    "st1 {v0.s}[3], [%0]                         \n"
-
-    "add         %1, %1, #4                      \n"  // src += 4
-    "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
-    "subs        %3, %3, #4                      \n"  // w   -= 4
-    "b.eq        4f                              \n"
-
-    // some residual, check to see if it includes a 2x8 block,
-    // or less
-    "cmp         %3, #2                          \n"
-    "b.lt        3f                              \n"
-
-    // 2x8 block
-    "2:                                          \n"
-    "mov         %0, %1                          \n"
-    MEMACCESS(0)
-    "ld1     {v0.h}[0], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.h}[0], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v0.h}[1], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.h}[1], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v0.h}[2], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.h}[2], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v0.h}[3], [%0], %5                 \n"
-    MEMACCESS(0)
-    "ld1     {v1.h}[3], [%0]                     \n"
-
-    "trn2    v2.8b, v0.8b, v1.8b                 \n"
-    "trn1    v3.8b, v0.8b, v1.8b                 \n"
-
-    "mov         %0, %2                          \n"
-
-    MEMACCESS(0)
-    "st1     {v3.8b}, [%0], %6                   \n"
-    MEMACCESS(0)
-    "st1     {v2.8b}, [%0]                       \n"
-
-    "add         %1, %1, #2                      \n"  // src += 2
-    "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
-    "subs        %3, %3,  #2                     \n"  // w   -= 2
-    "b.eq        4f                              \n"
-
-    // 1x8 block
-    "3:                                          \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[0], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[1], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[2], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[3], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[4], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[5], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[6], [%1], %5             \n"
-    MEMACCESS(1)
-    "ld1         {v0.b}[7], [%1]                 \n"
-
-    MEMACCESS(2)
-    "st1         {v0.8b}, [%2]                   \n"
-
-    "4:                                          \n"
-
-    : "=&r"(src_temp),                            // %0
-      "+r"(src),                                  // %1
-      "+r"(dst),                                  // %2
-      "+r"(width64)                               // %3
-    : "r"(&kVTbl4x4Transpose),                    // %4
-      "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
-      "r"(static_cast<ptrdiff_t>(dst_stride))     // %6
-    : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
-      "v17", "v18", "v19", "v20", "v21", "v22", "v23"
-  );
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %w3, %w3, #8                  \n"
+      "b.eq        4f                            \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %w3, #2                       \n"
+      "b.lt        3f                            \n"
+
+      "cmp         %w3, #4                       \n"
+      "b.lt        2f                            \n"
+
+      // 4x8 block
+      "mov         %0, %1                        \n"
+      "ld1         {v0.s}[0], [%0], %5           \n"
+      "ld1         {v0.s}[1], [%0], %5           \n"
+      "ld1         {v0.s}[2], [%0], %5           \n"
+      "ld1         {v0.s}[3], [%0], %5           \n"
+      "ld1         {v1.s}[0], [%0], %5           \n"
+      "ld1         {v1.s}[1], [%0], %5           \n"
+      "ld1         {v1.s}[2], [%0], %5           \n"
+      "ld1         {v1.s}[3], [%0]               \n"
+
+      "mov         %0, %2                        \n"
+
+      "ld1         {v2.16b}, [%4]                \n"
+
+      "tbl         v3.16b, {v0.16b}, v2.16b      \n"
+      "tbl         v0.16b, {v1.16b}, v2.16b      \n"
+
+      // TODO(frkoenig): Rework shuffle above to
+      // write out with 4 instead of 8 writes.
+      "st1 {v3.s}[0], [%0], %6                     \n"
+      "st1 {v3.s}[1], [%0], %6                     \n"
+      "st1 {v3.s}[2], [%0], %6                     \n"
+      "st1 {v3.s}[3], [%0]                         \n"
+
+      "add         %0, %2, #4                      \n"
+      "st1 {v0.s}[0], [%0], %6                     \n"
+      "st1 {v0.s}[1], [%0], %6                     \n"
+      "st1 {v0.s}[2], [%0], %6                     \n"
+      "st1 {v0.s}[3], [%0]                         \n"
+
+      "add         %1, %1, #4                      \n"  // src += 4
+      "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
+      "subs        %w3, %w3, #4                    \n"  // w   -= 4
+      "b.eq        4f                              \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp         %w3, #2                         \n"
+      "b.lt        3f                              \n"
+
+      // 2x8 block
+      "2:                                          \n"
+      "mov         %0, %1                          \n"
+      "ld1     {v0.h}[0], [%0], %5                 \n"
+      "ld1     {v1.h}[0], [%0], %5                 \n"
+      "ld1     {v0.h}[1], [%0], %5                 \n"
+      "ld1     {v1.h}[1], [%0], %5                 \n"
+      "ld1     {v0.h}[2], [%0], %5                 \n"
+      "ld1     {v1.h}[2], [%0], %5                 \n"
+      "ld1     {v0.h}[3], [%0], %5                 \n"
+      "ld1     {v1.h}[3], [%0]                     \n"
+
+      "trn2    v2.8b, v0.8b, v1.8b                 \n"
+      "trn1    v3.8b, v0.8b, v1.8b                 \n"
+
+      "mov         %0, %2                          \n"
+
+      "st1     {v3.8b}, [%0], %6                   \n"
+      "st1     {v2.8b}, [%0]                       \n"
+
+      "add         %1, %1, #2                      \n"  // src += 2
+      "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
+      "subs        %w3, %w3,  #2                   \n"  // w   -= 2
+      "b.eq        4f                              \n"
+
+      // 1x8 block
+      "3:                                          \n"
+      "ld1         {v0.b}[0], [%1], %5             \n"
+      "ld1         {v0.b}[1], [%1], %5             \n"
+      "ld1         {v0.b}[2], [%1], %5             \n"
+      "ld1         {v0.b}[3], [%1], %5             \n"
+      "ld1         {v0.b}[4], [%1], %5             \n"
+      "ld1         {v0.b}[5], [%1], %5             \n"
+      "ld1         {v0.b}[6], [%1], %5             \n"
+      "ld1         {v0.b}[7], [%1]                 \n"
+
+      "st1         {v0.8b}, [%2]                   \n"
+
+      "4:                                          \n"
+
+      : "=&r"(src_temp),             // %0
+        "+r"(src),                   // %1
+        "+r"(dst),                   // %2
+        "+r"(width)                  // %3
+      : "r"(&kVTbl4x4Transpose),     // %4
+        "r"((ptrdiff_t)src_stride),  // %5
+        "r"((ptrdiff_t)dst_stride)   // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17", "v18", "v19", "v20", "v21", "v22", "v23");
 }
 
-static uint8 kVTbl4x4TransposeDi[32] =
-  { 0,  16, 32, 48,  2, 18, 34, 50,  4, 20, 36, 52,  6, 22, 38, 54,
-    1,  17, 33, 49,  3, 19, 35, 51,  5, 21, 37, 53,  7, 23, 39, 55};
+static const uint8_t kVTbl4x4TransposeDi[32] = {
+    0, 16, 32, 48, 2, 18, 34, 50, 4, 20, 36, 52, 6, 22, 38, 54,
+    1, 17, 33, 49, 3, 19, 35, 51, 5, 21, 37, 53, 7, 23, 39, 55};
 
-void TransposeUVWx8_NEON(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b,
+void TransposeUVWx8_NEON(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
                          int width) {
-  const uint8* src_temp;
-  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
-  asm volatile (
-    // loops are on blocks of 8. loop will stop when
-    // counter gets to or below 0. starting the counter
-    // at w-8 allow for this
-    "sub       %4, %4, #8                      \n"
-
-    // handle 8x8 blocks. this should be the majority of the plane
-    "1:                                        \n"
-    "mov       %0, %1                          \n"
-
-    MEMACCESS(0)
-    "ld1       {v0.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v1.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v2.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v3.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v4.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v5.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v6.16b}, [%0], %5              \n"
-    MEMACCESS(0)
-    "ld1       {v7.16b}, [%0]                  \n"
-
-    "trn1      v16.16b, v0.16b, v1.16b         \n"
-    "trn2      v17.16b, v0.16b, v1.16b         \n"
-    "trn1      v18.16b, v2.16b, v3.16b         \n"
-    "trn2      v19.16b, v2.16b, v3.16b         \n"
-    "trn1      v20.16b, v4.16b, v5.16b         \n"
-    "trn2      v21.16b, v4.16b, v5.16b         \n"
-    "trn1      v22.16b, v6.16b, v7.16b         \n"
-    "trn2      v23.16b, v6.16b, v7.16b         \n"
-
-    "trn1      v0.8h, v16.8h, v18.8h           \n"
-    "trn2      v1.8h, v16.8h, v18.8h           \n"
-    "trn1      v2.8h, v20.8h, v22.8h           \n"
-    "trn2      v3.8h, v20.8h, v22.8h           \n"
-    "trn1      v4.8h, v17.8h, v19.8h           \n"
-    "trn2      v5.8h, v17.8h, v19.8h           \n"
-    "trn1      v6.8h, v21.8h, v23.8h           \n"
-    "trn2      v7.8h, v21.8h, v23.8h           \n"
-
-    "trn1      v16.4s, v0.4s, v2.4s            \n"
-    "trn2      v17.4s, v0.4s, v2.4s            \n"
-    "trn1      v18.4s, v1.4s, v3.4s            \n"
-    "trn2      v19.4s, v1.4s, v3.4s            \n"
-    "trn1      v20.4s, v4.4s, v6.4s            \n"
-    "trn2      v21.4s, v4.4s, v6.4s            \n"
-    "trn1      v22.4s, v5.4s, v7.4s            \n"
-    "trn2      v23.4s, v5.4s, v7.4s            \n"
-
-    "mov       %0, %2                          \n"
-
-    MEMACCESS(0)
-    "st1       {v16.d}[0], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v18.d}[0], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v17.d}[0], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v19.d}[0], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v16.d}[1], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v18.d}[1], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v17.d}[1], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v19.d}[1], [%0]                \n"
-
-    "mov       %0, %3                          \n"
-
-    MEMACCESS(0)
-    "st1       {v20.d}[0], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v22.d}[0], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v21.d}[0], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v23.d}[0], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v20.d}[1], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v22.d}[1], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v21.d}[1], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v23.d}[1], [%0]                \n"
-
-    "add       %1, %1, #16                     \n"  // src   += 8*2
-    "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 * dst_stride_a
-    "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 * dst_stride_b
-    "subs      %4, %4,  #8                     \n"  // w     -= 8
-    "b.ge      1b                              \n"
-
-    // add 8 back to counter. if the result is 0 there are
-    // no residuals.
-    "adds      %4, %4, #8                      \n"
-    "b.eq      4f                              \n"
-
-    // some residual, so between 1 and 7 lines left to transpose
-    "cmp       %4, #2                          \n"
-    "b.lt      3f                              \n"
-
-    "cmp       %4, #4                          \n"
-    "b.lt      2f                              \n"
-
-    // TODO(frkoenig): Clean this up
-    // 4x8 block
-    "mov       %0, %1                          \n"
-    MEMACCESS(0)
-    "ld1       {v0.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v1.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v2.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v3.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v4.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v5.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v6.8b}, [%0], %5               \n"
-    MEMACCESS(0)
-    "ld1       {v7.8b}, [%0]                   \n"
-
-    MEMACCESS(8)
-    "ld1       {v30.16b}, [%8], #16            \n"
-    "ld1       {v31.16b}, [%8]                 \n"
-
-    "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
-    "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
-    "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
-    "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"
-
-    "mov       %0, %2                          \n"
-
-    MEMACCESS(0)
-    "st1       {v16.s}[0],  [%0], %6           \n"
-    MEMACCESS(0)
-    "st1       {v16.s}[1],  [%0], %6           \n"
-    MEMACCESS(0)
-    "st1       {v16.s}[2],  [%0], %6           \n"
-    MEMACCESS(0)
-    "st1       {v16.s}[3],  [%0], %6           \n"
-
-    "add       %0, %2, #4                      \n"
-    MEMACCESS(0)
-    "st1       {v18.s}[0], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v18.s}[1], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v18.s}[2], [%0], %6            \n"
-    MEMACCESS(0)
-    "st1       {v18.s}[3], [%0]                \n"
-
-    "mov       %0, %3                          \n"
-
-    MEMACCESS(0)
-    "st1       {v17.s}[0], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v17.s}[1], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v17.s}[2], [%0], %7            \n"
-    MEMACCESS(0)
-    "st1       {v17.s}[3], [%0], %7            \n"
-
-    "add       %0, %3, #4                      \n"
-    MEMACCESS(0)
-    "st1       {v19.s}[0],  [%0], %7           \n"
-    MEMACCESS(0)
-    "st1       {v19.s}[1],  [%0], %7           \n"
-    MEMACCESS(0)
-    "st1       {v19.s}[2],  [%0], %7           \n"
-    MEMACCESS(0)
-    "st1       {v19.s}[3],  [%0]               \n"
-
-    "add       %1, %1, #8                      \n"  // src   += 4 * 2
-    "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 * dst_stride_a
-    "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 * dst_stride_b
-    "subs      %4,  %4,  #4                    \n"  // w     -= 4
-    "b.eq      4f                              \n"
-
-    // some residual, check to see if it includes a 2x8 block,
-    // or less
-    "cmp       %4, #2                          \n"
-    "b.lt      3f                              \n"
-
-    // 2x8 block
-    "2:                                        \n"
-    "mov       %0, %1                          \n"
-    MEMACCESS(0)
-    "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
-    MEMACCESS(0)
-    "ld2       {v2.h, v3.h}[3], [%0]           \n"
-
-    "trn1      v4.8b, v0.8b, v2.8b             \n"
-    "trn2      v5.8b, v0.8b, v2.8b             \n"
-    "trn1      v6.8b, v1.8b, v3.8b             \n"
-    "trn2      v7.8b, v1.8b, v3.8b             \n"
-
-    "mov       %0, %2                          \n"
-
-    MEMACCESS(0)
-    "st1       {v4.d}[0], [%0], %6             \n"
-    MEMACCESS(0)
-    "st1       {v6.d}[0], [%0]                 \n"
-
-    "mov       %0, %3                          \n"
-
-    MEMACCESS(0)
-    "st1       {v5.d}[0], [%0], %7             \n"
-    MEMACCESS(0)
-    "st1       {v7.d}[0], [%0]                 \n"
-
-    "add       %1, %1, #4                      \n"  // src   += 2 * 2
-    "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 * dst_stride_a
-    "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 * dst_stride_b
-    "subs      %4,  %4,  #2                    \n"  // w     -= 2
-    "b.eq      4f                              \n"
-
-    // 1x8 block
-    "3:                                        \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
-    MEMACCESS(1)
-    "ld2       {v0.b, v1.b}[7], [%1]           \n"
-
-    MEMACCESS(2)
-    "st1       {v0.d}[0], [%2]                 \n"
-    MEMACCESS(3)
-    "st1       {v1.d}[0], [%3]                 \n"
-
-    "4:                                        \n"
-
-    : "=&r"(src_temp),                            // %0
-      "+r"(src),                                  // %1
-      "+r"(dst_a),                                // %2
-      "+r"(dst_b),                                // %3
-      "+r"(width64)                               // %4
-    : "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
-      "r"(static_cast<ptrdiff_t>(dst_stride_a)),  // %6
-      "r"(static_cast<ptrdiff_t>(dst_stride_b)),  // %7
-      "r"(&kVTbl4x4TransposeDi)                   // %8
-    : "memory", "cc",
-      "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
-      "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
-      "v30", "v31"
-  );
+  const uint8_t* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %w4, %w4, #8                  \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov         %0, %1                        \n"
+
+      "ld1         {v0.16b}, [%0], %5            \n"
+      "ld1         {v1.16b}, [%0], %5            \n"
+      "ld1         {v2.16b}, [%0], %5            \n"
+      "ld1         {v3.16b}, [%0], %5            \n"
+      "ld1         {v4.16b}, [%0], %5            \n"
+      "ld1         {v5.16b}, [%0], %5            \n"
+      "ld1         {v6.16b}, [%0], %5            \n"
+      "ld1         {v7.16b}, [%0]                \n"
+      "mov         %0, %1                        \n"
+
+      "trn1        v16.16b, v0.16b, v1.16b       \n"
+      "trn2        v17.16b, v0.16b, v1.16b       \n"
+      "trn1        v18.16b, v2.16b, v3.16b       \n"
+      "trn2        v19.16b, v2.16b, v3.16b       \n"
+      "trn1        v20.16b, v4.16b, v5.16b       \n"
+      "trn2        v21.16b, v4.16b, v5.16b       \n"
+      "trn1        v22.16b, v6.16b, v7.16b       \n"
+      "trn2        v23.16b, v6.16b, v7.16b       \n"
+
+      "trn1        v0.8h, v16.8h, v18.8h         \n"
+      "trn2        v1.8h, v16.8h, v18.8h         \n"
+      "trn1        v2.8h, v20.8h, v22.8h         \n"
+      "trn2        v3.8h, v20.8h, v22.8h         \n"
+      "trn1        v4.8h, v17.8h, v19.8h         \n"
+      "trn2        v5.8h, v17.8h, v19.8h         \n"
+      "trn1        v6.8h, v21.8h, v23.8h         \n"
+      "trn2        v7.8h, v21.8h, v23.8h         \n"
+
+      "trn1        v16.4s, v0.4s, v2.4s          \n"
+      "trn2        v17.4s, v0.4s, v2.4s          \n"
+      "trn1        v18.4s, v1.4s, v3.4s          \n"
+      "trn2        v19.4s, v1.4s, v3.4s          \n"
+      "trn1        v20.4s, v4.4s, v6.4s          \n"
+      "trn2        v21.4s, v4.4s, v6.4s          \n"
+      "trn1        v22.4s, v5.4s, v7.4s          \n"
+      "trn2        v23.4s, v5.4s, v7.4s          \n"
+
+      "mov         %0, %2                        \n"
+
+      "st1         {v16.d}[0], [%0], %6          \n"
+      "st1         {v18.d}[0], [%0], %6          \n"
+      "st1         {v17.d}[0], [%0], %6          \n"
+      "st1         {v19.d}[0], [%0], %6          \n"
+      "st1         {v16.d}[1], [%0], %6          \n"
+      "st1         {v18.d}[1], [%0], %6          \n"
+      "st1         {v17.d}[1], [%0], %6          \n"
+      "st1         {v19.d}[1], [%0]              \n"
+
+      "mov         %0, %3                        \n"
+
+      "st1         {v20.d}[0], [%0], %7          \n"
+      "st1         {v22.d}[0], [%0], %7          \n"
+      "st1         {v21.d}[0], [%0], %7          \n"
+      "st1         {v23.d}[0], [%0], %7          \n"
+      "st1         {v20.d}[1], [%0], %7          \n"
+      "st1         {v22.d}[1], [%0], %7          \n"
+      "st1         {v21.d}[1], [%0], %7          \n"
+      "st1         {v23.d}[1], [%0]              \n"
+
+      "add         %1, %1, #16                   \n"  // src   += 8*2
+      "add         %2, %2, %6, lsl #3            \n"  // dst_a += 8 *
+                                                      // dst_stride_a
+      "add         %3, %3, %7, lsl #3            \n"  // dst_b += 8 *
+                                                      // dst_stride_b
+      "subs        %w4, %w4,  #8                 \n"  // w     -= 8
+      "b.ge        1b                            \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %w4, %w4, #8                  \n"
+      "b.eq        4f                            \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %w4, #2                       \n"
+      "b.lt        3f                            \n"
+
+      "cmp         %w4, #4                       \n"
+      "b.lt        2f                            \n"
+
+      // TODO(frkoenig): Clean this up
+      // 4x8 block
+      "mov       %0, %1                          \n"
+      "ld1       {v0.8b}, [%0], %5               \n"
+      "ld1       {v1.8b}, [%0], %5               \n"
+      "ld1       {v2.8b}, [%0], %5               \n"
+      "ld1       {v3.8b}, [%0], %5               \n"
+      "ld1       {v4.8b}, [%0], %5               \n"
+      "ld1       {v5.8b}, [%0], %5               \n"
+      "ld1       {v6.8b}, [%0], %5               \n"
+      "ld1       {v7.8b}, [%0]                   \n"
+
+      "ld1       {v30.16b}, [%8], #16            \n"
+      "ld1       {v31.16b}, [%8]                 \n"
+
+      "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
+      "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
+      "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
+      "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"
+
+      "mov       %0, %2                          \n"
+
+      "st1       {v16.s}[0],  [%0], %6           \n"
+      "st1       {v16.s}[1],  [%0], %6           \n"
+      "st1       {v16.s}[2],  [%0], %6           \n"
+      "st1       {v16.s}[3],  [%0], %6           \n"
+
+      "add       %0, %2, #4                      \n"
+      "st1       {v18.s}[0], [%0], %6            \n"
+      "st1       {v18.s}[1], [%0], %6            \n"
+      "st1       {v18.s}[2], [%0], %6            \n"
+      "st1       {v18.s}[3], [%0]                \n"
+
+      "mov       %0, %3                          \n"
+
+      "st1       {v17.s}[0], [%0], %7            \n"
+      "st1       {v17.s}[1], [%0], %7            \n"
+      "st1       {v17.s}[2], [%0], %7            \n"
+      "st1       {v17.s}[3], [%0], %7            \n"
+
+      "add       %0, %3, #4                      \n"
+      "st1       {v19.s}[0],  [%0], %7           \n"
+      "st1       {v19.s}[1],  [%0], %7           \n"
+      "st1       {v19.s}[2],  [%0], %7           \n"
+      "st1       {v19.s}[3],  [%0]               \n"
+
+      "add       %1, %1, #8                      \n"  // src   += 4 * 2
+      "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 *
+                                                      // dst_stride_a
+      "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 *
+                                                      // dst_stride_b
+      "subs      %w4,  %w4,  #4                  \n"  // w     -= 4
+      "b.eq      4f                              \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp       %w4, #2                         \n"
+      "b.lt      3f                              \n"
+
+      // 2x8 block
+      "2:                                        \n"
+      "mov       %0, %1                          \n"
+      "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
+      "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
+      "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
+      "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[3], [%0]           \n"
+
+      "trn1      v4.8b, v0.8b, v2.8b             \n"
+      "trn2      v5.8b, v0.8b, v2.8b             \n"
+      "trn1      v6.8b, v1.8b, v3.8b             \n"
+      "trn2      v7.8b, v1.8b, v3.8b             \n"
+
+      "mov       %0, %2                          \n"
+
+      "st1       {v4.d}[0], [%0], %6             \n"
+      "st1       {v6.d}[0], [%0]                 \n"
+
+      "mov       %0, %3                          \n"
+
+      "st1       {v5.d}[0], [%0], %7             \n"
+      "st1       {v7.d}[0], [%0]                 \n"
+
+      "add       %1, %1, #4                      \n"  // src   += 2 * 2
+      "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 *
+                                                      // dst_stride_a
+      "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 *
+                                                      // dst_stride_b
+      "subs      %w4,  %w4,  #2                  \n"  // w     -= 2
+      "b.eq      4f                              \n"
+
+      // 1x8 block
+      "3:                                        \n"
+      "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[7], [%1]           \n"
+
+      "st1       {v0.d}[0], [%2]                 \n"
+      "st1       {v1.d}[0], [%3]                 \n"
+
+      "4:                                        \n"
+
+      : "=&r"(src_temp),               // %0
+        "+r"(src),                     // %1
+        "+r"(dst_a),                   // %2
+        "+r"(dst_b),                   // %3
+        "+r"(width)                    // %4
+      : "r"((ptrdiff_t)src_stride),    // %5
+        "r"((ptrdiff_t)dst_stride_a),  // %6
+        "r"((ptrdiff_t)dst_stride_b),  // %7
+        "r"(&kVTbl4x4TransposeDi)      // %8
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v30", "v31");
 }
+
+// Transpose 32 bit values (ARGB)
+void Transpose4x4_32_NEON(const uint8_t* src,
+                          int src_stride,
+                          uint8_t* dst,
+                          int dst_stride,
+                          int width) {
+  const uint8_t* src1 = src + src_stride;
+  const uint8_t* src2 = src1 + src_stride;
+  const uint8_t* src3 = src2 + src_stride;
+  uint8_t* dst1 = dst + dst_stride;
+  uint8_t* dst2 = dst1 + dst_stride;
+  uint8_t* dst3 = dst2 + dst_stride;
+  asm volatile(
+      // Main loop transpose 4x4.  Read a column, write a row.
+      "1:                                        \n"
+      "ld4         {v0.s, v1.s, v2.s, v3.s}[0], [%0], %9 \n"
+      "ld4         {v0.s, v1.s, v2.s, v3.s}[1], [%1], %9 \n"
+      "ld4         {v0.s, v1.s, v2.s, v3.s}[2], [%2], %9 \n"
+      "ld4         {v0.s, v1.s, v2.s, v3.s}[3], [%3], %9 \n"
+      "subs        %w8, %w8, #4                  \n"  // w -= 4
+      "st1         {v0.4s}, [%4], 16             \n"
+      "st1         {v1.4s}, [%5], 16             \n"
+      "st1         {v2.4s}, [%6], 16             \n"
+      "st1         {v3.4s}, [%7], 16             \n"
+      "b.gt        1b                            \n"
+      : "+r"(src),                        // %0
+        "+r"(src1),                       // %1
+        "+r"(src2),                       // %2
+        "+r"(src3),                       // %3
+        "+r"(dst),                        // %4
+        "+r"(dst1),                       // %5
+        "+r"(dst2),                       // %6
+        "+r"(dst3),                       // %7
+        "+r"(width)                       // %8
+      : "r"((ptrdiff_t)(src_stride * 4))  // %9
+      : "memory", "cc", "v0", "v1", "v2", "v3");
+}
+
 #endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/rotate_win.cc b/media/libyuv/source/rotate_win.cc
index 1300fc0feb..a78873f843 100644
--- a/media/libyuv/source/rotate_win.cc
+++ b/media/libyuv/source/rotate_win.cc
@@ -8,26 +8,29 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/row.h"
 #include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for 32 bit Visual C x86 and clangcl
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+// This module is for 32 bit Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(_M_IX86)
 
-__declspec(naked)
-void TransposeWx8_SSSE3(const uint8* src, int src_stride,
-                        uint8* dst, int dst_stride, int width) {
+__declspec(naked) void TransposeWx8_SSSE3(const uint8_t* src,
+                                          int src_stride,
+                                          uint8_t* dst,
+                                          int dst_stride,
+                                          int width) {
   __asm {
     push      edi
     push      esi
     push      ebp
-    mov       eax, [esp + 12 + 4]   // src
-    mov       edi, [esp + 12 + 8]   // src_stride
+    mov       eax, [esp + 12 + 4]  // src
+    mov       edi, [esp + 12 + 8]  // src_stride
     mov       edx, [esp + 12 + 12]  // dst
     mov       esi, [esp + 12 + 16]  // dst_stride
     mov       ecx, [esp + 12 + 20]  // width
@@ -110,18 +113,20 @@ void TransposeWx8_SSSE3(const uint8* src, int src_stride,
   }
 }
 
-__declspec(naked)
-void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
-                         uint8* dst_a, int dst_stride_a,
-                         uint8* dst_b, int dst_stride_b,
-                         int w) {
+__declspec(naked) void TransposeUVWx8_SSE2(const uint8_t* src,
+                                           int src_stride,
+                                           uint8_t* dst_a,
+                                           int dst_stride_a,
+                                           uint8_t* dst_b,
+                                           int dst_stride_b,
+                                           int w) {
   __asm {
     push      ebx
     push      esi
     push      edi
     push      ebp
-    mov       eax, [esp + 16 + 4]   // src
-    mov       edi, [esp + 16 + 8]   // src_stride
+    mov       eax, [esp + 16 + 4]  // src
+    mov       edi, [esp + 16 + 8]  // src_stride
     mov       edx, [esp + 16 + 12]  // dst_a
     mov       esi, [esp + 16 + 16]  // dst_stride_a
     mov       ebx, [esp + 16 + 20]  // dst_b
@@ -133,9 +138,9 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     mov       ecx, [ecx + 16 + 28]  // w
 
     align      4
- convertloop:
     // Read in the data from the source pointer.
     // First round of bit swap.
+  convertloop:
     movdqu    xmm0, [eax]
     movdqu    xmm1, [eax + edi]
     lea       eax, [eax + 2 * edi]
@@ -162,13 +167,13 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     lea       eax, [eax + 2 * edi]
     movdqu    [esp], xmm5  // backup xmm5
     neg       edi
-    movdqa    xmm5, xmm6   // use xmm5 as temp register.
+    movdqa    xmm5, xmm6  // use xmm5 as temp register.
     punpcklbw xmm6, xmm7
     punpckhbw xmm5, xmm7
     movdqa    xmm7, xmm5
     lea       eax, [eax + 8 * edi + 16]
     neg       edi
-    // Second round of bit swap.
+        // Second round of bit swap.
     movdqa    xmm5, xmm0
     punpcklwd xmm0, xmm2
     punpckhwd xmm5, xmm2
@@ -183,12 +188,13 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     movdqa    xmm6, xmm5
     movdqu    xmm5, [esp]  // restore xmm5
     movdqu    [esp], xmm6  // backup xmm6
-    movdqa    xmm6, xmm5    // use xmm6 as temp register.
+    movdqa    xmm6, xmm5  // use xmm6 as temp register.
     punpcklwd xmm5, xmm7
     punpckhwd xmm6, xmm7
     movdqa    xmm7, xmm6
-    // Third round of bit swap.
-    // Write to the destination pointer.
+
+        // Third round of bit swap.
+        // Write to the destination pointer.
     movdqa    xmm6, xmm0
     punpckldq xmm0, xmm4
     punpckhdq xmm6, xmm4
@@ -200,7 +206,7 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     lea       edx, [edx + 2 * esi]
     movhpd    qword ptr [ebx + ebp], xmm4
     lea       ebx, [ebx + 2 * ebp]
-    movdqa    xmm0, xmm2   // use xmm0 as the temp register.
+    movdqa    xmm0, xmm2  // use xmm0 as the temp register.
     punpckldq xmm2, xmm6
     movlpd    qword ptr [edx], xmm2
     movhpd    qword ptr [ebx], xmm2
@@ -209,7 +215,7 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     lea       edx, [edx + 2 * esi]
     movhpd    qword ptr [ebx + ebp], xmm0
     lea       ebx, [ebx + 2 * ebp]
-    movdqa    xmm0, xmm1   // use xmm0 as the temp register.
+    movdqa    xmm0, xmm1  // use xmm0 as the temp register.
     punpckldq xmm1, xmm5
     movlpd    qword ptr [edx], xmm1
     movhpd    qword ptr [ebx], xmm1
@@ -218,7 +224,7 @@ void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
     lea       edx, [edx + 2 * esi]
     movhpd    qword ptr [ebx + ebp], xmm0
     lea       ebx, [ebx + 2 * ebp]
-    movdqa    xmm0, xmm3   // use xmm0 as the temp register.
+    movdqa    xmm0, xmm3  // use xmm0 as the temp register.
     punpckldq xmm3, xmm7
     movlpd    qword ptr [edx], xmm3
     movhpd    qword ptr [ebx], xmm3
diff --git a/media/libyuv/source/row_any.cc b/media/libyuv/source/row_any.cc
index 494164fd02..0168061fff 100644
--- a/media/libyuv/source/row_any.cc
+++ b/media/libyuv/source/row_any.cc
@@ -19,70 +19,284 @@ namespace libyuv {
 extern "C" {
 #endif
 
+// memset for temp is meant to clear the source buffer (not dest) so that
+// SIMD that reads full multiple of 16 bytes will not trigger msan errors.
+// memset is not needed for production, as the garbage values are processed but
+// not used, although there may be edge cases for subsampling.
+// The size of the buffer is based on the largest read, which can be inferred
+// by the source type (e.g. ARGB) and the mask (last parameter), or by examining
+// the source code for how much the source pointers are advanced.
+
 // Subsampled source needs to be increase by 1 of not even.
 #define SS(width, shift) (((width) + (1 << (shift)) - 1) >> (shift))
 
+// Any 4 planes to 1
+#define ANY41(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)               \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,                   \
+               const uint8_t* v_buf, const uint8_t* a_buf, uint8_t* dst_ptr, \
+               int width) {                                                  \
+    SIMD_ALIGNED(uint8_t temp[64 * 5]);                                      \
+    memset(temp, 0, 64 * 4); /* for msan */                                  \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(y_buf, u_buf, v_buf, a_buf, dst_ptr, n);                      \
+    }                                                                        \
+    memcpy(temp, y_buf + n, r);                                              \
+    memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
+    memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
+    memcpy(temp + 192, a_buf + n, r);                                        \
+    ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, temp + 256, MASK + 1); \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 256,                      \
+           SS(r, DUVSHIFT) * BPP);                                           \
+  }
+
+#ifdef HAS_MERGEARGBROW_SSE2
+ANY41(MergeARGBRow_Any_SSE2, MergeARGBRow_SSE2, 0, 0, 4, 7)
+#endif
+#ifdef HAS_MERGEARGBROW_AVX2
+ANY41(MergeARGBRow_Any_AVX2, MergeARGBRow_AVX2, 0, 0, 4, 15)
+#endif
+#ifdef HAS_MERGEARGBROW_NEON
+ANY41(MergeARGBRow_Any_NEON, MergeARGBRow_NEON, 0, 0, 4, 15)
+#endif
+
+// Note that odd width replication includes 444 due to implementation
+// on arm that subsamples 444 to 422 internally.
 // Any 4 planes to 1 with yuvconstants
-#define ANY41C(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)                \
-    void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf,   \
-                 const uint8* a_buf, uint8* dst_ptr,                           \
-                 const struct YuvConstants* yuvconstants,  int width) {        \
-      SIMD_ALIGNED(uint8 temp[64 * 5]);                                        \
-      memset(temp, 0, 64 * 4);  /* for msan */                                 \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(y_buf, u_buf, v_buf, a_buf, dst_ptr, yuvconstants, n);        \
-      }                                                                        \
-      memcpy(temp, y_buf + n, r);                                              \
-      memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
-      memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
-      memcpy(temp + 192, a_buf + n, r);                                        \
-      ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, temp + 256,            \
-               yuvconstants, MASK + 1);                                        \
-      memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 256,                      \
-             SS(r, DUVSHIFT) * BPP);                                           \
-    }
+#define ANY41C(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)              \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,                   \
+               const uint8_t* v_buf, const uint8_t* a_buf, uint8_t* dst_ptr, \
+               const struct YuvConstants* yuvconstants, int width) {         \
+    SIMD_ALIGNED(uint8_t temp[64 * 5]);                                      \
+    memset(temp, 0, 64 * 4); /* for msan */                                  \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(y_buf, u_buf, v_buf, a_buf, dst_ptr, yuvconstants, n);        \
+    }                                                                        \
+    memcpy(temp, y_buf + n, r);                                              \
+    memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
+    memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
+    memcpy(temp + 192, a_buf + n, r);                                        \
+    if (width & 1) {                                                         \
+      temp[64 + SS(r, UVSHIFT)] = temp[64 + SS(r, UVSHIFT) - 1];             \
+      temp[128 + SS(r, UVSHIFT)] = temp[128 + SS(r, UVSHIFT) - 1];           \
+    }                                                                        \
+    ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, temp + 256,            \
+             yuvconstants, MASK + 1);                                        \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 256,                      \
+           SS(r, DUVSHIFT) * BPP);                                           \
+  }
 
+#ifdef HAS_I444ALPHATOARGBROW_SSSE3
+ANY41C(I444AlphaToARGBRow_Any_SSSE3, I444AlphaToARGBRow_SSSE3, 0, 0, 4, 7)
+#endif
+#ifdef HAS_I444ALPHATOARGBROW_AVX2
+ANY41C(I444AlphaToARGBRow_Any_AVX2, I444AlphaToARGBRow_AVX2, 0, 0, 4, 15)
+#endif
 #ifdef HAS_I422ALPHATOARGBROW_SSSE3
 ANY41C(I422AlphaToARGBRow_Any_SSSE3, I422AlphaToARGBRow_SSSE3, 1, 0, 4, 7)
 #endif
 #ifdef HAS_I422ALPHATOARGBROW_AVX2
 ANY41C(I422AlphaToARGBRow_Any_AVX2, I422AlphaToARGBRow_AVX2, 1, 0, 4, 15)
 #endif
+#ifdef HAS_I444ALPHATOARGBROW_NEON
+ANY41C(I444AlphaToARGBRow_Any_NEON, I444AlphaToARGBRow_NEON, 0, 0, 4, 7)
+#endif
 #ifdef HAS_I422ALPHATOARGBROW_NEON
 ANY41C(I422AlphaToARGBRow_Any_NEON, I422AlphaToARGBRow_NEON, 1, 0, 4, 7)
 #endif
+#ifdef HAS_I444ALPHATOARGBROW_MSA
+ANY41C(I444AlphaToARGBRow_Any_MSA, I444AlphaToARGBRow_MSA, 0, 0, 4, 7)
+#endif
+#ifdef HAS_I422ALPHATOARGBROW_MSA
+ANY41C(I422AlphaToARGBRow_Any_MSA, I422AlphaToARGBRow_MSA, 1, 0, 4, 7)
+#endif
+#ifdef HAS_I422ALPHATOARGBROW_LASX
+ANY41C(I422AlphaToARGBRow_Any_LASX, I422AlphaToARGBRow_LASX, 1, 0, 4, 15)
+#endif
 #undef ANY41C
 
+// Any 4 planes to 1 plane of 8 bit with yuvconstants
+#define ANY41CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK)      \
+  void NAMEANY(const T* y_buf, const T* u_buf, const T* v_buf, const T* a_buf, \
+               uint8_t* dst_ptr, const struct YuvConstants* yuvconstants,      \
+               int width) {                                                    \
+    SIMD_ALIGNED(T temp[16 * 4]);                                              \
+    SIMD_ALIGNED(uint8_t out[64]);                                             \
+    memset(temp, 0, 16 * 4 * SBPP); /* for YUY2 and msan */                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(y_buf, u_buf, v_buf, a_buf, dst_ptr, yuvconstants, n);          \
+    }                                                                          \
+    memcpy(temp, y_buf + n, r * SBPP);                                         \
+    memcpy(temp + 16, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP);          \
+    memcpy(temp + 32, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP);          \
+    memcpy(temp + 48, a_buf + n, r * SBPP);                                    \
+    ANY_SIMD(temp, temp + 16, temp + 32, temp + 48, out, yuvconstants,         \
+             MASK + 1);                                                        \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, out, SS(r, DUVSHIFT) * BPP);       \
+  }
+
+#ifdef HAS_I210ALPHATOARGBROW_SSSE3
+ANY41CT(I210AlphaToARGBRow_Any_SSSE3,
+        I210AlphaToARGBRow_SSSE3,
+        1,
+        0,
+        uint16_t,
+        2,
+        4,
+        7)
+#endif
+
+#ifdef HAS_I210ALPHATOARGBROW_AVX2
+ANY41CT(I210AlphaToARGBRow_Any_AVX2,
+        I210AlphaToARGBRow_AVX2,
+        1,
+        0,
+        uint16_t,
+        2,
+        4,
+        15)
+#endif
+
+#ifdef HAS_I410ALPHATOARGBROW_SSSE3
+ANY41CT(I410AlphaToARGBRow_Any_SSSE3,
+        I410AlphaToARGBRow_SSSE3,
+        0,
+        0,
+        uint16_t,
+        2,
+        4,
+        7)
+#endif
+
+#ifdef HAS_I410ALPHATOARGBROW_AVX2
+ANY41CT(I410AlphaToARGBRow_Any_AVX2,
+        I410AlphaToARGBRow_AVX2,
+        0,
+        0,
+        uint16_t,
+        2,
+        4,
+        15)
+#endif
+
+#undef ANY41CT
+
+// Any 4 planes to 1 plane with parameter
+#define ANY41PT(NAMEANY, ANY_SIMD, STYPE, SBPP, DTYPE, BPP, MASK)          \
+  void NAMEANY(const STYPE* r_buf, const STYPE* g_buf, const STYPE* b_buf, \
+               const STYPE* a_buf, DTYPE* dst_ptr, int depth, int width) { \
+    SIMD_ALIGNED(STYPE temp[16 * 4]);                                      \
+    SIMD_ALIGNED(DTYPE out[64]);                                           \
+    memset(temp, 0, 16 * 4 * SBPP); /* for YUY2 and msan */                \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(r_buf, g_buf, b_buf, a_buf, dst_ptr, depth, n);             \
+    }                                                                      \
+    memcpy(temp, r_buf + n, r * SBPP);                                     \
+    memcpy(temp + 16, g_buf + n, r * SBPP);                                \
+    memcpy(temp + 32, b_buf + n, r * SBPP);                                \
+    memcpy(temp + 48, a_buf + n, r * SBPP);                                \
+    ANY_SIMD(temp, temp + 16, temp + 32, temp + 48, out, depth, MASK + 1); \
+    memcpy((uint8_t*)dst_ptr + n * BPP, out, r * BPP);                     \
+  }
+
+#ifdef HAS_MERGEAR64ROW_AVX2
+ANY41PT(MergeAR64Row_Any_AVX2, MergeAR64Row_AVX2, uint16_t, 2, uint16_t, 8, 15)
+#endif
+
+#ifdef HAS_MERGEAR64ROW_NEON
+ANY41PT(MergeAR64Row_Any_NEON, MergeAR64Row_NEON, uint16_t, 2, uint16_t, 8, 7)
+#endif
+
+#ifdef HAS_MERGEARGB16TO8ROW_AVX2
+ANY41PT(MergeARGB16To8Row_Any_AVX2,
+        MergeARGB16To8Row_AVX2,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        15)
+#endif
+
+#ifdef HAS_MERGEARGB16TO8ROW_NEON
+ANY41PT(MergeARGB16To8Row_Any_NEON,
+        MergeARGB16To8Row_NEON,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        7)
+#endif
+
+#undef ANY41PT
+
 // Any 3 planes to 1.
-#define ANY31(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)                 \
-    void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf,   \
-                 uint8* dst_ptr, int width) {                                  \
-      SIMD_ALIGNED(uint8 temp[64 * 4]);                                        \
-      memset(temp, 0, 64 * 3);  /* for YUY2 and msan */                        \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, n);                             \
-      }                                                                        \
-      memcpy(temp, y_buf + n, r);                                              \
-      memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
-      memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
-      ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, MASK + 1);             \
-      memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 192,                      \
-             SS(r, DUVSHIFT) * BPP);                                           \
-    }
+#define ANY31(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)      \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,          \
+               const uint8_t* v_buf, uint8_t* dst_ptr, int width) { \
+    SIMD_ALIGNED(uint8_t temp[64 * 4]);                             \
+    memset(temp, 0, 64 * 3); /* for YUY2 and msan */                \
+    int r = width & MASK;                                           \
+    int n = width & ~MASK;                                          \
+    if (n > 0) {                                                    \
+      ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, n);                    \
+    }                                                               \
+    memcpy(temp, y_buf + n, r);                                     \
+    memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));      \
+    memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));     \
+    ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, MASK + 1);    \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 192,             \
+           SS(r, DUVSHIFT) * BPP);                                  \
+  }
+
+// Merge functions.
+#ifdef HAS_MERGERGBROW_SSSE3
+ANY31(MergeRGBRow_Any_SSSE3, MergeRGBRow_SSSE3, 0, 0, 3, 15)
+#endif
+#ifdef HAS_MERGERGBROW_NEON
+ANY31(MergeRGBRow_Any_NEON, MergeRGBRow_NEON, 0, 0, 3, 15)
+#endif
+#ifdef HAS_MERGEXRGBROW_SSE2
+ANY31(MergeXRGBRow_Any_SSE2, MergeXRGBRow_SSE2, 0, 0, 4, 7)
+#endif
+#ifdef HAS_MERGEXRGBROW_AVX2
+ANY31(MergeXRGBRow_Any_AVX2, MergeXRGBRow_AVX2, 0, 0, 4, 15)
+#endif
+#ifdef HAS_MERGEXRGBROW_NEON
+ANY31(MergeXRGBRow_Any_NEON, MergeXRGBRow_NEON, 0, 0, 4, 15)
+#endif
 #ifdef HAS_I422TOYUY2ROW_SSE2
 ANY31(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, 1, 1, 4, 15)
 ANY31(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, 1, 1, 4, 15)
 #endif
+#ifdef HAS_I422TOYUY2ROW_AVX2
+ANY31(I422ToYUY2Row_Any_AVX2, I422ToYUY2Row_AVX2, 1, 1, 4, 31)
+ANY31(I422ToUYVYRow_Any_AVX2, I422ToUYVYRow_AVX2, 1, 1, 4, 31)
+#endif
 #ifdef HAS_I422TOYUY2ROW_NEON
 ANY31(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, 1, 1, 4, 15)
 #endif
+#ifdef HAS_I422TOYUY2ROW_MSA
+ANY31(I422ToYUY2Row_Any_MSA, I422ToYUY2Row_MSA, 1, 1, 4, 31)
+#endif
+#ifdef HAS_I422TOYUY2ROW_LASX
+ANY31(I422ToYUY2Row_Any_LASX, I422ToYUY2Row_LASX, 1, 1, 4, 31)
+#endif
 #ifdef HAS_I422TOUYVYROW_NEON
 ANY31(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, 1, 1, 4, 15)
 #endif
+#ifdef HAS_I422TOUYVYROW_MSA
+ANY31(I422ToUYVYRow_Any_MSA, I422ToUYVYRow_MSA, 1, 1, 4, 31)
+#endif
+#ifdef HAS_I422TOUYVYROW_LASX
+ANY31(I422ToUYVYRow_Any_LASX, I422ToUYVYRow_LASX, 1, 1, 4, 31)
+#endif
 #ifdef HAS_BLENDPLANEROW_AVX2
 ANY31(BlendPlaneRow_Any_AVX2, BlendPlaneRow_AVX2, 0, 0, 1, 31)
 #endif
@@ -94,109 +308,290 @@ ANY31(BlendPlaneRow_Any_SSSE3, BlendPlaneRow_SSSE3, 0, 0, 1, 7)
 // Note that odd width replication includes 444 due to implementation
 // on arm that subsamples 444 to 422 internally.
 // Any 3 planes to 1 with yuvconstants
-#define ANY31C(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)                \
-    void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf,   \
-                 uint8* dst_ptr, const struct YuvConstants* yuvconstants,      \
-                 int width) {                                                  \
-      SIMD_ALIGNED(uint8 temp[64 * 4]);                                        \
-      memset(temp, 0, 64 * 3);  /* for YUY2 and msan */                        \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, yuvconstants, n);               \
-      }                                                                        \
-      memcpy(temp, y_buf + n, r);                                              \
-      memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
-      memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
-      if (width & 1) {                                                         \
-        temp[64 + SS(r, UVSHIFT)] = temp[64 + SS(r, UVSHIFT) - 1];             \
-        temp[128 + SS(r, UVSHIFT)] = temp[128 + SS(r, UVSHIFT) - 1];           \
-      }                                                                        \
-      ANY_SIMD(temp, temp + 64, temp + 128, temp + 192,                        \
-               yuvconstants, MASK + 1);                                        \
-      memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 192,                      \
-             SS(r, DUVSHIFT) * BPP);                                           \
-    }
+#define ANY31C(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)      \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,           \
+               const uint8_t* v_buf, uint8_t* dst_ptr,               \
+               const struct YuvConstants* yuvconstants, int width) { \
+    SIMD_ALIGNED(uint8_t temp[128 * 4]);                             \
+    memset(temp, 0, 128 * 3); /* for YUY2 and msan */                \
+    int r = width & MASK;                                            \
+    int n = width & ~MASK;                                           \
+    if (n > 0) {                                                     \
+      ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, yuvconstants, n);       \
+    }                                                                \
+    memcpy(temp, y_buf + n, r);                                      \
+    memcpy(temp + 128, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));      \
+    memcpy(temp + 256, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));      \
+    if (width & 1) {                                                 \
+      temp[128 + SS(r, UVSHIFT)] = temp[128 + SS(r, UVSHIFT) - 1];   \
+      temp[256 + SS(r, UVSHIFT)] = temp[256 + SS(r, UVSHIFT) - 1];   \
+    }                                                                \
+    ANY_SIMD(temp, temp + 128, temp + 256, temp + 384, yuvconstants, \
+             MASK + 1);                                              \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 384,              \
+           SS(r, DUVSHIFT) * BPP);                                   \
+  }
 
 #ifdef HAS_I422TOARGBROW_SSSE3
 ANY31C(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, 1, 0, 4, 7)
 #endif
-#ifdef HAS_I411TOARGBROW_SSSE3
-ANY31C(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, 2, 0, 4, 7)
-#endif
-#ifdef HAS_I444TOARGBROW_SSSE3
-ANY31C(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7)
+#ifdef HAS_I422TORGBAROW_SSSE3
 ANY31C(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_I422TOARGB4444ROW_SSSE3
 ANY31C(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGB1555ROW_SSSE3
 ANY31C(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TORGB565ROW_SSSE3
 ANY31C(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, 1, 0, 2, 7)
-ANY31C(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 7)
-#endif  // HAS_I444TOARGBROW_SSSE3
+#endif
+#ifdef HAS_I422TORGB24ROW_SSSE3
+ANY31C(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 15)
+#endif
+#ifdef HAS_I422TOAR30ROW_SSSE3
+ANY31C(I422ToAR30Row_Any_SSSE3, I422ToAR30Row_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_I422TOAR30ROW_AVX2
+ANY31C(I422ToAR30Row_Any_AVX2, I422ToAR30Row_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I444TOARGBROW_SSSE3
+ANY31C(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7)
+#endif
+#ifdef HAS_I444TORGB24ROW_SSSE3
+ANY31C(I444ToRGB24Row_Any_SSSE3, I444ToRGB24Row_SSSE3, 0, 0, 3, 15)
+#endif
 #ifdef HAS_I422TORGB24ROW_AVX2
-ANY31C(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 15)
+ANY31C(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 31)
 #endif
 #ifdef HAS_I422TOARGBROW_AVX2
 ANY31C(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, 1, 0, 4, 15)
 #endif
+#ifdef HAS_I422TOARGBROW_AVX512BW
+ANY31C(I422ToARGBRow_Any_AVX512BW, I422ToARGBRow_AVX512BW, 1, 0, 4, 31)
+#endif
 #ifdef HAS_I422TORGBAROW_AVX2
 ANY31C(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, 1, 0, 4, 15)
 #endif
 #ifdef HAS_I444TOARGBROW_AVX2
 ANY31C(I444ToARGBRow_Any_AVX2, I444ToARGBRow_AVX2, 0, 0, 4, 15)
 #endif
-#ifdef HAS_I411TOARGBROW_AVX2
-ANY31C(I411ToARGBRow_Any_AVX2, I411ToARGBRow_AVX2, 2, 0, 4, 15)
+#ifdef HAS_I444TORGB24ROW_AVX2
+ANY31C(I444ToRGB24Row_Any_AVX2, I444ToRGB24Row_AVX2, 0, 0, 3, 31)
 #endif
 #ifdef HAS_I422TOARGB4444ROW_AVX2
-ANY31C(I422ToARGB4444Row_Any_AVX2, I422ToARGB4444Row_AVX2, 1, 0, 2, 7)
+ANY31C(I422ToARGB4444Row_Any_AVX2, I422ToARGB4444Row_AVX2, 1, 0, 2, 15)
 #endif
 #ifdef HAS_I422TOARGB1555ROW_AVX2
-ANY31C(I422ToARGB1555Row_Any_AVX2, I422ToARGB1555Row_AVX2, 1, 0, 2, 7)
+ANY31C(I422ToARGB1555Row_Any_AVX2, I422ToARGB1555Row_AVX2, 1, 0, 2, 15)
 #endif
 #ifdef HAS_I422TORGB565ROW_AVX2
-ANY31C(I422ToRGB565Row_Any_AVX2, I422ToRGB565Row_AVX2, 1, 0, 2, 7)
+ANY31C(I422ToRGB565Row_Any_AVX2, I422ToRGB565Row_AVX2, 1, 0, 2, 15)
+#endif
+#ifdef HAS_I444TORGB24ROW_NEON
+ANY31C(I444ToRGB24Row_Any_NEON, I444ToRGB24Row_NEON, 0, 0, 3, 7)
 #endif
 #ifdef HAS_I422TOARGBROW_NEON
 ANY31C(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, 0, 0, 4, 7)
 ANY31C(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, 1, 0, 4, 7)
-ANY31C(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, 2, 0, 4, 7)
 ANY31C(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, 1, 0, 4, 7)
 ANY31C(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, 1, 0, 3, 7)
 ANY31C(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, 1, 0, 2, 7)
 ANY31C(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, 1, 0, 2, 7)
 ANY31C(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, 1, 0, 2, 7)
 #endif
+#ifdef HAS_I422TOARGBROW_MSA
+ANY31C(I444ToARGBRow_Any_MSA, I444ToARGBRow_MSA, 0, 0, 4, 7)
+ANY31C(I422ToARGBRow_Any_MSA, I422ToARGBRow_MSA, 1, 0, 4, 7)
+ANY31C(I422ToRGBARow_Any_MSA, I422ToRGBARow_MSA, 1, 0, 4, 7)
+ANY31C(I422ToRGB24Row_Any_MSA, I422ToRGB24Row_MSA, 1, 0, 3, 15)
+ANY31C(I422ToARGB4444Row_Any_MSA, I422ToARGB4444Row_MSA, 1, 0, 2, 7)
+ANY31C(I422ToARGB1555Row_Any_MSA, I422ToARGB1555Row_MSA, 1, 0, 2, 7)
+ANY31C(I422ToRGB565Row_Any_MSA, I422ToRGB565Row_MSA, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGBROW_LASX
+ANY31C(I422ToARGBRow_Any_LASX, I422ToARGBRow_LASX, 1, 0, 4, 31)
+ANY31C(I422ToRGBARow_Any_LASX, I422ToRGBARow_LASX, 1, 0, 4, 31)
+ANY31C(I422ToRGB24Row_Any_LASX, I422ToRGB24Row_LASX, 1, 0, 3, 31)
+ANY31C(I422ToRGB565Row_Any_LASX, I422ToRGB565Row_LASX, 1, 0, 2, 31)
+ANY31C(I422ToARGB4444Row_Any_LASX, I422ToARGB4444Row_LASX, 1, 0, 2, 31)
+ANY31C(I422ToARGB1555Row_Any_LASX, I422ToARGB1555Row_LASX, 1, 0, 2, 31)
+#endif
+#ifdef HAS_I444TOARGBROW_LSX
+ANY31C(I444ToARGBRow_Any_LSX, I444ToARGBRow_LSX, 0, 0, 4, 15)
+#endif
 #undef ANY31C
 
+// Any 3 planes of 16 bit to 1 with yuvconstants
+// TODO(fbarchard): consider sharing this code with ANY31C
+#define ANY31CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK) \
+  void NAMEANY(const T* y_buf, const T* u_buf, const T* v_buf,            \
+               uint8_t* dst_ptr, const struct YuvConstants* yuvconstants, \
+               int width) {                                               \
+    SIMD_ALIGNED(T temp[16 * 3]);                                         \
+    SIMD_ALIGNED(uint8_t out[64]);                                        \
+    memset(temp, 0, 16 * 3 * SBPP); /* for YUY2 and msan */               \
+    int r = width & MASK;                                                 \
+    int n = width & ~MASK;                                                \
+    if (n > 0) {                                                          \
+      ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, yuvconstants, n);            \
+    }                                                                     \
+    memcpy(temp, y_buf + n, r * SBPP);                                    \
+    memcpy(temp + 16, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP);     \
+    memcpy(temp + 32, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP);     \
+    ANY_SIMD(temp, temp + 16, temp + 32, out, yuvconstants, MASK + 1);    \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, out, SS(r, DUVSHIFT) * BPP);  \
+  }
+
+#ifdef HAS_I210TOAR30ROW_SSSE3
+ANY31CT(I210ToAR30Row_Any_SSSE3, I210ToAR30Row_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I210TOARGBROW_SSSE3
+ANY31CT(I210ToARGBRow_Any_SSSE3, I210ToARGBRow_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I210TOARGBROW_AVX2
+ANY31CT(I210ToARGBRow_Any_AVX2, I210ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I210TOAR30ROW_AVX2
+ANY31CT(I210ToAR30Row_Any_AVX2, I210ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I410TOAR30ROW_SSSE3
+ANY31CT(I410ToAR30Row_Any_SSSE3, I410ToAR30Row_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I410TOARGBROW_SSSE3
+ANY31CT(I410ToARGBRow_Any_SSSE3, I410ToARGBRow_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I410TOARGBROW_AVX2
+ANY31CT(I410ToARGBRow_Any_AVX2, I410ToARGBRow_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I410TOAR30ROW_AVX2
+ANY31CT(I410ToAR30Row_Any_AVX2, I410ToAR30Row_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I212TOAR30ROW_SSSE3
+ANY31CT(I212ToAR30Row_Any_SSSE3, I212ToAR30Row_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I212TOARGBROW_SSSE3
+ANY31CT(I212ToARGBRow_Any_SSSE3, I212ToARGBRow_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I212TOARGBROW_AVX2
+ANY31CT(I212ToARGBRow_Any_AVX2, I212ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I212TOAR30ROW_AVX2
+ANY31CT(I212ToAR30Row_Any_AVX2, I212ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#undef ANY31CT
+
+// Any 3 planes to 1 plane with parameter
+#define ANY31PT(NAMEANY, ANY_SIMD, STYPE, SBPP, DTYPE, BPP, MASK)          \
+  void NAMEANY(const STYPE* r_buf, const STYPE* g_buf, const STYPE* b_buf, \
+               DTYPE* dst_ptr, int depth, int width) {                     \
+    SIMD_ALIGNED(STYPE temp[16 * 3]);                                      \
+    SIMD_ALIGNED(DTYPE out[64]);                                           \
+    memset(temp, 0, 16 * 3 * SBPP); /* for YUY2 and msan */                \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(r_buf, g_buf, b_buf, dst_ptr, depth, n);                    \
+    }                                                                      \
+    memcpy(temp, r_buf + n, r * SBPP);                                     \
+    memcpy(temp + 16, g_buf + n, r * SBPP);                                \
+    memcpy(temp + 32, b_buf + n, r * SBPP);                                \
+    ANY_SIMD(temp, temp + 16, temp + 32, out, depth, MASK + 1);            \
+    memcpy((uint8_t*)dst_ptr + n * BPP, out, r * BPP);                     \
+  }
+
+#ifdef HAS_MERGEXR30ROW_AVX2
+ANY31PT(MergeXR30Row_Any_AVX2, MergeXR30Row_AVX2, uint16_t, 2, uint8_t, 4, 15)
+#endif
+
+#ifdef HAS_MERGEXR30ROW_NEON
+ANY31PT(MergeXR30Row_Any_NEON, MergeXR30Row_NEON, uint16_t, 2, uint8_t, 4, 3)
+ANY31PT(MergeXR30Row_10_Any_NEON,
+        MergeXR30Row_10_NEON,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        3)
+#endif
+
+#ifdef HAS_MERGEXR64ROW_AVX2
+ANY31PT(MergeXR64Row_Any_AVX2, MergeXR64Row_AVX2, uint16_t, 2, uint16_t, 8, 15)
+#endif
+
+#ifdef HAS_MERGEXR64ROW_NEON
+ANY31PT(MergeXR64Row_Any_NEON, MergeXR64Row_NEON, uint16_t, 2, uint16_t, 8, 7)
+#endif
+
+#ifdef HAS_MERGEXRGB16TO8ROW_AVX2
+ANY31PT(MergeXRGB16To8Row_Any_AVX2,
+        MergeXRGB16To8Row_AVX2,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        15)
+#endif
+
+#ifdef HAS_MERGEXRGB16TO8ROW_NEON
+ANY31PT(MergeXRGB16To8Row_Any_NEON,
+        MergeXRGB16To8Row_NEON,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        7)
+#endif
+
+#undef ANY31PT
+
 // Any 2 planes to 1.
-#define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)              \
-    void NAMEANY(const uint8* y_buf, const uint8* uv_buf,                      \
-                 uint8* dst_ptr, int width) {                                  \
-      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
-      memset(temp, 0, 64 * 2);  /* for msan */                                 \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(y_buf, uv_buf, dst_ptr, n);                                   \
-      }                                                                        \
-      memcpy(temp, y_buf + n * SBPP, r * SBPP);                                \
-      memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
-             SS(r, UVSHIFT) * SBPP2);                                          \
-      ANY_SIMD(temp, temp + 64, temp + 128, MASK + 1);                         \
-      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
-    }
+#define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)             \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* dst_ptr, \
+               int width) {                                                   \
+    SIMD_ALIGNED(uint8_t temp[128 * 3]);                                      \
+    memset(temp, 0, 128 * 2); /* for msan */                                  \
+    int r = width & MASK;                                                     \
+    int n = width & ~MASK;                                                    \
+    if (n > 0) {                                                              \
+      ANY_SIMD(y_buf, uv_buf, dst_ptr, n);                                    \
+    }                                                                         \
+    memcpy(temp, y_buf + n * SBPP, r * SBPP);                                 \
+    memcpy(temp + 128, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
+           SS(r, UVSHIFT) * SBPP2);                                           \
+    ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1);                         \
+    memcpy(dst_ptr + n * BPP, temp + 256, r * BPP);                           \
+  }
 
 // Merge functions.
 #ifdef HAS_MERGEUVROW_SSE2
 ANY21(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, 0, 1, 1, 2, 15)
 #endif
 #ifdef HAS_MERGEUVROW_AVX2
-ANY21(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, 0, 1, 1, 2, 31)
+ANY21(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, 0, 1, 1, 2, 15)
+#endif
+#ifdef HAS_MERGEUVROW_AVX512BW
+ANY21(MergeUVRow_Any_AVX512BW, MergeUVRow_AVX512BW, 0, 1, 1, 2, 31)
 #endif
 #ifdef HAS_MERGEUVROW_NEON
 ANY21(MergeUVRow_Any_NEON, MergeUVRow_NEON, 0, 1, 1, 2, 15)
 #endif
-
+#ifdef HAS_MERGEUVROW_MSA
+ANY21(MergeUVRow_Any_MSA, MergeUVRow_MSA, 0, 1, 1, 2, 15)
+#endif
+#ifdef HAS_MERGEUVROW_LSX
+ANY21(MergeUVRow_Any_LSX, MergeUVRow_LSX, 0, 1, 1, 2, 15)
+#endif
+#ifdef HAS_NV21TOYUV24ROW_NEON
+ANY21(NV21ToYUV24Row_Any_NEON, NV21ToYUV24Row_NEON, 1, 1, 2, 3, 15)
+#endif
+#ifdef HAS_NV21TOYUV24ROW_SSSE3
+ANY21(NV21ToYUV24Row_Any_SSSE3, NV21ToYUV24Row_SSSE3, 1, 1, 2, 3, 15)
+#endif
+#ifdef HAS_NV21TOYUV24ROW_AVX2
+ANY21(NV21ToYUV24Row_Any_AVX2, NV21ToYUV24Row_AVX2, 1, 1, 2, 3, 31)
+#endif
 // Math functions.
 #ifdef HAS_ARGBMULTIPLYROW_SSE2
 ANY21(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, 0, 4, 4, 4, 3)
@@ -225,44 +620,108 @@ ANY21(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, 0, 4, 4, 4, 7)
 #ifdef HAS_ARGBSUBTRACTROW_NEON
 ANY21(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, 0, 4, 4, 4, 7)
 #endif
+#ifdef HAS_ARGBMULTIPLYROW_MSA
+ANY21(ARGBMultiplyRow_Any_MSA, ARGBMultiplyRow_MSA, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_LASX
+ANY21(ARGBMultiplyRow_Any_LASX, ARGBMultiplyRow_LASX, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBADDROW_MSA
+ANY21(ARGBAddRow_Any_MSA, ARGBAddRow_MSA, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBADDROW_LASX
+ANY21(ARGBAddRow_Any_LASX, ARGBAddRow_LASX, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_MSA
+ANY21(ARGBSubtractRow_Any_MSA, ARGBSubtractRow_MSA, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_LASX
+ANY21(ARGBSubtractRow_Any_LASX, ARGBSubtractRow_LASX, 0, 4, 4, 4, 7)
+#endif
 #ifdef HAS_SOBELROW_SSE2
 ANY21(SobelRow_Any_SSE2, SobelRow_SSE2, 0, 1, 1, 4, 15)
 #endif
 #ifdef HAS_SOBELROW_NEON
 ANY21(SobelRow_Any_NEON, SobelRow_NEON, 0, 1, 1, 4, 7)
 #endif
+#ifdef HAS_SOBELROW_MSA
+ANY21(SobelRow_Any_MSA, SobelRow_MSA, 0, 1, 1, 4, 15)
+#endif
+#ifdef HAS_SOBELROW_LSX
+ANY21(SobelRow_Any_LSX, SobelRow_LSX, 0, 1, 1, 4, 15)
+#endif
 #ifdef HAS_SOBELTOPLANEROW_SSE2
 ANY21(SobelToPlaneRow_Any_SSE2, SobelToPlaneRow_SSE2, 0, 1, 1, 1, 15)
 #endif
 #ifdef HAS_SOBELTOPLANEROW_NEON
 ANY21(SobelToPlaneRow_Any_NEON, SobelToPlaneRow_NEON, 0, 1, 1, 1, 15)
 #endif
+#ifdef HAS_SOBELTOPLANEROW_MSA
+ANY21(SobelToPlaneRow_Any_MSA, SobelToPlaneRow_MSA, 0, 1, 1, 1, 31)
+#endif
+#ifdef HAS_SOBELTOPLANEROW_LSX
+ANY21(SobelToPlaneRow_Any_LSX, SobelToPlaneRow_LSX, 0, 1, 1, 1, 31)
+#endif
 #ifdef HAS_SOBELXYROW_SSE2
 ANY21(SobelXYRow_Any_SSE2, SobelXYRow_SSE2, 0, 1, 1, 4, 15)
 #endif
 #ifdef HAS_SOBELXYROW_NEON
 ANY21(SobelXYRow_Any_NEON, SobelXYRow_NEON, 0, 1, 1, 4, 7)
 #endif
+#ifdef HAS_SOBELXYROW_MSA
+ANY21(SobelXYRow_Any_MSA, SobelXYRow_MSA, 0, 1, 1, 4, 15)
+#endif
+#ifdef HAS_SOBELXYROW_LSX
+ANY21(SobelXYRow_Any_LSX, SobelXYRow_LSX, 0, 1, 1, 4, 15)
+#endif
 #undef ANY21
 
+// Any 2 planes to 1 with stride
+// width is measured in source pixels. 4 bytes contains 2 pixels
+#define ANY21S(NAMEANY, ANY_SIMD, SBPP, BPP, MASK)                        \
+  void NAMEANY(const uint8_t* src_yuy2, int stride_yuy2, uint8_t* dst_uv, \
+               int width) {                                               \
+    SIMD_ALIGNED(uint8_t temp[32 * 3]);                                   \
+    memset(temp, 0, 32 * 2); /* for msan */                               \
+    int awidth = (width + 1) / 2;                                         \
+    int r = awidth & MASK;                                                \
+    int n = awidth & ~MASK;                                               \
+    if (n > 0) {                                                          \
+      ANY_SIMD(src_yuy2, stride_yuy2, dst_uv, n * 2);                     \
+    }                                                                     \
+    memcpy(temp, src_yuy2 + n * SBPP, r * SBPP);                          \
+    memcpy(temp + 32, src_yuy2 + stride_yuy2 + n * SBPP, r * SBPP);       \
+    ANY_SIMD(temp, 32, temp + 64, MASK + 1);                              \
+    memcpy(dst_uv + n * BPP, temp + 64, r * BPP);                         \
+  }
+
+#ifdef HAS_YUY2TONVUVROW_NEON
+ANY21S(YUY2ToNVUVRow_Any_NEON, YUY2ToNVUVRow_NEON, 4, 2, 7)
+#endif
+#ifdef HAS_YUY2TONVUVROW_SSE2
+ANY21S(YUY2ToNVUVRow_Any_SSE2, YUY2ToNVUVRow_SSE2, 4, 2, 7)
+#endif
+#ifdef HAS_YUY2TONVUVROW_AVX2
+ANY21S(YUY2ToNVUVRow_Any_AVX2, YUY2ToNVUVRow_AVX2, 4, 2, 15)
+#endif
+
 // Any 2 planes to 1 with yuvconstants
-#define ANY21C(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)             \
-    void NAMEANY(const uint8* y_buf, const uint8* uv_buf,                      \
-                 uint8* dst_ptr, const struct YuvConstants* yuvconstants,      \
-                 int width) {                                                  \
-      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
-      memset(temp, 0, 64 * 2);  /* for msan */                                 \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(y_buf, uv_buf, dst_ptr, yuvconstants, n);                     \
-      }                                                                        \
-      memcpy(temp, y_buf + n * SBPP, r * SBPP);                                \
-      memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
-             SS(r, UVSHIFT) * SBPP2);                                          \
-      ANY_SIMD(temp, temp + 64, temp + 128, yuvconstants, MASK + 1);           \
-      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
-    }
+#define ANY21C(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)            \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* dst_ptr, \
+               const struct YuvConstants* yuvconstants, int width) {          \
+    SIMD_ALIGNED(uint8_t temp[128 * 3]);                                      \
+    memset(temp, 0, 128 * 2); /* for msan */                                  \
+    int r = width & MASK;                                                     \
+    int n = width & ~MASK;                                                    \
+    if (n > 0) {                                                              \
+      ANY_SIMD(y_buf, uv_buf, dst_ptr, yuvconstants, n);                      \
+    }                                                                         \
+    memcpy(temp, y_buf + n * SBPP, r * SBPP);                                 \
+    memcpy(temp + 128, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
+           SS(r, UVSHIFT) * SBPP2);                                           \
+    ANY_SIMD(temp, temp + 128, temp + 256, yuvconstants, MASK + 1);           \
+    memcpy(dst_ptr + n * BPP, temp + 256, r * BPP);                           \
+  }
 
 // Biplanar to RGB.
 #ifdef HAS_NV12TOARGBROW_SSSE3
@@ -274,6 +733,15 @@ ANY21C(NV12ToARGBRow_Any_AVX2, NV12ToARGBRow_AVX2, 1, 1, 2, 4, 15)
 #ifdef HAS_NV12TOARGBROW_NEON
 ANY21C(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, 1, 1, 2, 4, 7)
 #endif
+#ifdef HAS_NV12TOARGBROW_MSA
+ANY21C(NV12ToARGBRow_Any_MSA, NV12ToARGBRow_MSA, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TOARGBROW_LSX
+ANY21C(NV12ToARGBRow_Any_LSX, NV12ToARGBRow_LSX, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TOARGBROW_LASX
+ANY21C(NV12ToARGBRow_Any_LASX, NV12ToARGBRow_LASX, 1, 1, 2, 4, 15)
+#endif
 #ifdef HAS_NV21TOARGBROW_SSSE3
 ANY21C(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, 1, 1, 2, 4, 7)
 #endif
@@ -283,6 +751,33 @@ ANY21C(NV21ToARGBRow_Any_AVX2, NV21ToARGBRow_AVX2, 1, 1, 2, 4, 15)
 #ifdef HAS_NV21TOARGBROW_NEON
 ANY21C(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, 1, 1, 2, 4, 7)
 #endif
+#ifdef HAS_NV21TOARGBROW_MSA
+ANY21C(NV21ToARGBRow_Any_MSA, NV21ToARGBRow_MSA, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV21TOARGBROW_LSX
+ANY21C(NV21ToARGBRow_Any_LSX, NV21ToARGBRow_LSX, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV21TOARGBROW_LASX
+ANY21C(NV21ToARGBRow_Any_LASX, NV21ToARGBRow_LASX, 1, 1, 2, 4, 15)
+#endif
+#ifdef HAS_NV12TORGB24ROW_NEON
+ANY21C(NV12ToRGB24Row_Any_NEON, NV12ToRGB24Row_NEON, 1, 1, 2, 3, 7)
+#endif
+#ifdef HAS_NV21TORGB24ROW_NEON
+ANY21C(NV21ToRGB24Row_Any_NEON, NV21ToRGB24Row_NEON, 1, 1, 2, 3, 7)
+#endif
+#ifdef HAS_NV12TORGB24ROW_SSSE3
+ANY21C(NV12ToRGB24Row_Any_SSSE3, NV12ToRGB24Row_SSSE3, 1, 1, 2, 3, 15)
+#endif
+#ifdef HAS_NV21TORGB24ROW_SSSE3
+ANY21C(NV21ToRGB24Row_Any_SSSE3, NV21ToRGB24Row_SSSE3, 1, 1, 2, 3, 15)
+#endif
+#ifdef HAS_NV12TORGB24ROW_AVX2
+ANY21C(NV12ToRGB24Row_Any_AVX2, NV12ToRGB24Row_AVX2, 1, 1, 2, 3, 31)
+#endif
+#ifdef HAS_NV21TORGB24ROW_AVX2
+ANY21C(NV21ToRGB24Row_Any_AVX2, NV21ToRGB24Row_AVX2, 1, 1, 2, 3, 31)
+#endif
 #ifdef HAS_NV12TORGB565ROW_SSSE3
 ANY21C(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, 1, 1, 2, 2, 7)
 #endif
@@ -292,22 +787,102 @@ ANY21C(NV12ToRGB565Row_Any_AVX2, NV12ToRGB565Row_AVX2, 1, 1, 2, 2, 15)
 #ifdef HAS_NV12TORGB565ROW_NEON
 ANY21C(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, 1, 1, 2, 2, 7)
 #endif
+#ifdef HAS_NV12TORGB565ROW_MSA
+ANY21C(NV12ToRGB565Row_Any_MSA, NV12ToRGB565Row_MSA, 1, 1, 2, 2, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_LSX
+ANY21C(NV12ToRGB565Row_Any_LSX, NV12ToRGB565Row_LSX, 1, 1, 2, 2, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_LASX
+ANY21C(NV12ToRGB565Row_Any_LASX, NV12ToRGB565Row_LASX, 1, 1, 2, 2, 15)
+#endif
 #undef ANY21C
 
+// Any 2 planes of 16 bit to 1 with yuvconstants
+#define ANY21CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK)      \
+  void NAMEANY(const T* y_buf, const T* uv_buf, uint8_t* dst_ptr,              \
+               const struct YuvConstants* yuvconstants, int width) {           \
+    SIMD_ALIGNED(T temp[16 * 3]);                                              \
+    SIMD_ALIGNED(uint8_t out[64]);                                             \
+    memset(temp, 0, 16 * 3 * SBPP); /* for YUY2 and msan */                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(y_buf, uv_buf, dst_ptr, yuvconstants, n);                       \
+    }                                                                          \
+    memcpy(temp, y_buf + n, r * SBPP);                                         \
+    memcpy(temp + 16, uv_buf + 2 * (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP * 2); \
+    ANY_SIMD(temp, temp + 16, out, yuvconstants, MASK + 1);                    \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, out, SS(r, DUVSHIFT) * BPP);       \
+  }
+
+#ifdef HAS_P210TOAR30ROW_SSSE3
+ANY21CT(P210ToAR30Row_Any_SSSE3, P210ToAR30Row_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P210TOARGBROW_SSSE3
+ANY21CT(P210ToARGBRow_Any_SSSE3, P210ToARGBRow_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P210TOARGBROW_AVX2
+ANY21CT(P210ToARGBRow_Any_AVX2, P210ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_P210TOAR30ROW_AVX2
+ANY21CT(P210ToAR30Row_Any_AVX2, P210ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_P410TOAR30ROW_SSSE3
+ANY21CT(P410ToAR30Row_Any_SSSE3, P410ToAR30Row_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P410TOARGBROW_SSSE3
+ANY21CT(P410ToARGBRow_Any_SSSE3, P410ToARGBRow_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P410TOARGBROW_AVX2
+ANY21CT(P410ToARGBRow_Any_AVX2, P410ToARGBRow_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_P410TOAR30ROW_AVX2
+ANY21CT(P410ToAR30Row_Any_AVX2, P410ToAR30Row_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+
+#undef ANY21CT
+
+// Any 2 16 bit planes with parameter to 1
+#define ANY21PT(NAMEANY, ANY_SIMD, T, BPP, MASK)                     \
+  void NAMEANY(const T* src_u, const T* src_v, T* dst_uv, int depth, \
+               int width) {                                          \
+    SIMD_ALIGNED(T temp[16 * 4]);                                    \
+    memset(temp, 0, 16 * 4 * BPP); /* for msan */                    \
+    int r = width & MASK;                                            \
+    int n = width & ~MASK;                                           \
+    if (n > 0) {                                                     \
+      ANY_SIMD(src_u, src_v, dst_uv, depth, n);                      \
+    }                                                                \
+    memcpy(temp, src_u + n, r * BPP);                                \
+    memcpy(temp + 16, src_v + n, r * BPP);                           \
+    ANY_SIMD(temp, temp + 16, temp + 32, depth, MASK + 1);           \
+    memcpy(dst_uv + n * 2, temp + 32, r * BPP * 2);                  \
+  }
+
+#ifdef HAS_MERGEUVROW_16_AVX2
+ANY21PT(MergeUVRow_16_Any_AVX2, MergeUVRow_16_AVX2, uint16_t, 2, 7)
+#endif
+#ifdef HAS_MERGEUVROW_16_NEON
+ANY21PT(MergeUVRow_16_Any_NEON, MergeUVRow_16_NEON, uint16_t, 2, 7)
+#endif
+
+#undef ANY21CT
+
 // Any 1 to 1.
-#define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                     \
-    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
-      SIMD_ALIGNED(uint8 temp[128 * 2]);                                       \
-      memset(temp, 0, 128);  /* for YUY2 and msan */                           \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr, dst_ptr, n);                                         \
-      }                                                                        \
-      memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP);    \
-      ANY_SIMD(temp, temp + 128, MASK + 1);                                    \
-      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
-    }
+#define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_ptr, int width) {     \
+    SIMD_ALIGNED(uint8_t temp[128 * 2]);                                  \
+    memset(temp, 0, 128); /* for YUY2 and msan */                         \
+    int r = width & MASK;                                                 \
+    int n = width & ~MASK;                                                \
+    if (n > 0) {                                                          \
+      ANY_SIMD(src_ptr, dst_ptr, n);                                      \
+    }                                                                     \
+    memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP); \
+    ANY_SIMD(temp, temp + 128, MASK + 1);                                 \
+    memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                       \
+  }
 
 #ifdef HAS_COPYROW_AVX
 ANY11(CopyRow_Any_AVX, CopyRow_AVX, 0, 1, 1, 63)
@@ -325,6 +900,15 @@ ANY11(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, 0, 4, 2, 3)
 ANY11(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, 0, 4, 2, 3)
 ANY11(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, 0, 4, 2, 3)
 #endif
+#if defined(HAS_ARGBTORGB24ROW_AVX2)
+ANY11(ARGBToRGB24Row_Any_AVX2, ARGBToRGB24Row_AVX2, 0, 4, 3, 31)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_AVX512VBMI)
+ANY11(ARGBToRGB24Row_Any_AVX512VBMI, ARGBToRGB24Row_AVX512VBMI, 0, 4, 3, 31)
+#endif
+#if defined(HAS_ARGBTORAWROW_AVX2)
+ANY11(ARGBToRAWRow_Any_AVX2, ARGBToRAWRow_AVX2, 0, 4, 3, 31)
+#endif
 #if defined(HAS_ARGBTORGB565ROW_AVX2)
 ANY11(ARGBToRGB565Row_Any_AVX2, ARGBToRGB565Row_AVX2, 0, 4, 2, 7)
 #endif
@@ -332,18 +916,24 @@ ANY11(ARGBToRGB565Row_Any_AVX2, ARGBToRGB565Row_AVX2, 0, 4, 2, 7)
 ANY11(ARGBToARGB1555Row_Any_AVX2, ARGBToARGB1555Row_AVX2, 0, 4, 2, 7)
 ANY11(ARGBToARGB4444Row_Any_AVX2, ARGBToARGB4444Row_AVX2, 0, 4, 2, 7)
 #endif
+#if defined(HAS_ABGRTOAR30ROW_SSSE3)
+ANY11(ABGRToAR30Row_Any_SSSE3, ABGRToAR30Row_SSSE3, 0, 4, 4, 3)
+#endif
+#if defined(HAS_ARGBTOAR30ROW_SSSE3)
+ANY11(ARGBToAR30Row_Any_SSSE3, ARGBToAR30Row_SSSE3, 0, 4, 4, 3)
+#endif
+#if defined(HAS_ABGRTOAR30ROW_AVX2)
+ANY11(ABGRToAR30Row_Any_AVX2, ABGRToAR30Row_AVX2, 0, 4, 4, 7)
+#endif
+#if defined(HAS_ARGBTOAR30ROW_AVX2)
+ANY11(ARGBToAR30Row_Any_AVX2, ARGBToAR30Row_AVX2, 0, 4, 4, 7)
+#endif
 #if defined(HAS_J400TOARGBROW_SSE2)
 ANY11(J400ToARGBRow_Any_SSE2, J400ToARGBRow_SSE2, 0, 1, 4, 7)
 #endif
 #if defined(HAS_J400TOARGBROW_AVX2)
 ANY11(J400ToARGBRow_Any_AVX2, J400ToARGBRow_AVX2, 0, 1, 4, 15)
 #endif
-#if defined(HAS_I400TOARGBROW_SSE2)
-ANY11(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, 0, 1, 4, 7)
-#endif
-#if defined(HAS_I400TOARGBROW_AVX2)
-ANY11(I400ToARGBRow_Any_AVX2, I400ToARGBRow_AVX2, 0, 1, 4, 15)
-#endif
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
 ANY11(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, 0, 3, 4, 15)
 ANY11(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, 0, 3, 4, 15)
@@ -351,6 +941,9 @@ ANY11(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, 0, 2, 4, 7)
 ANY11(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, 0, 2, 4, 7)
 ANY11(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, 0, 2, 4, 7)
 #endif
+#if defined(HAS_RAWTORGBAROW_SSSE3)
+ANY11(RAWToRGBARow_Any_SSSE3, RAWToRGBARow_SSSE3, 0, 3, 4, 15)
+#endif
 #if defined(HAS_RAWTORGB24ROW_SSSE3)
 ANY11(RAWToRGB24Row_Any_SSSE3, RAWToRGB24Row_SSSE3, 0, 3, 3, 7)
 #endif
@@ -364,23 +957,55 @@ ANY11(ARGB1555ToARGBRow_Any_AVX2, ARGB1555ToARGBRow_AVX2, 0, 2, 4, 15)
 ANY11(ARGB4444ToARGBRow_Any_AVX2, ARGB4444ToARGBRow_AVX2, 0, 2, 4, 15)
 #endif
 #if defined(HAS_ARGBTORGB24ROW_NEON)
-ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 7)
+ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 15)
 ANY11(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, 0, 4, 3, 7)
 ANY11(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, 0, 4, 2, 7)
 ANY11(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, 0, 4, 2, 7)
 ANY11(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, 0, 4, 2, 7)
 ANY11(J400ToARGBRow_Any_NEON, J400ToARGBRow_NEON, 0, 1, 4, 7)
-ANY11(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, 0, 1, 4, 7)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_MSA)
+ANY11(ARGBToRGB24Row_Any_MSA, ARGBToRGB24Row_MSA, 0, 4, 3, 15)
+ANY11(ARGBToRAWRow_Any_MSA, ARGBToRAWRow_MSA, 0, 4, 3, 15)
+ANY11(ARGBToRGB565Row_Any_MSA, ARGBToRGB565Row_MSA, 0, 4, 2, 7)
+ANY11(ARGBToARGB1555Row_Any_MSA, ARGBToARGB1555Row_MSA, 0, 4, 2, 7)
+ANY11(ARGBToARGB4444Row_Any_MSA, ARGBToARGB4444Row_MSA, 0, 4, 2, 7)
+ANY11(J400ToARGBRow_Any_MSA, J400ToARGBRow_MSA, 0, 1, 4, 15)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_LASX)
+ANY11(ARGBToRGB24Row_Any_LASX, ARGBToRGB24Row_LASX, 0, 4, 3, 31)
+ANY11(ARGBToRAWRow_Any_LASX, ARGBToRAWRow_LASX, 0, 4, 3, 31)
+ANY11(ARGBToRGB565Row_Any_LASX, ARGBToRGB565Row_LASX, 0, 4, 2, 15)
+ANY11(ARGBToARGB1555Row_Any_LASX, ARGBToARGB1555Row_LASX, 0, 4, 2, 15)
+ANY11(ARGBToARGB4444Row_Any_LASX, ARGBToARGB4444Row_LASX, 0, 4, 2, 15)
+#endif
+#if defined(HAS_J400TOARGBROW_LSX)
+ANY11(J400ToARGBRow_Any_LSX, J400ToARGBRow_LSX, 0, 1, 4, 15)
 #endif
 #if defined(HAS_RAWTORGB24ROW_NEON)
 ANY11(RAWToRGB24Row_Any_NEON, RAWToRGB24Row_NEON, 0, 3, 3, 7)
 #endif
+#if defined(HAS_RAWTORGB24ROW_MSA)
+ANY11(RAWToRGB24Row_Any_MSA, RAWToRGB24Row_MSA, 0, 3, 3, 15)
+#endif
+#if defined(HAS_RAWTORGB24ROW_LSX)
+ANY11(RAWToRGB24Row_Any_LSX, RAWToRGB24Row_LSX, 0, 3, 3, 15)
+#endif
 #ifdef HAS_ARGBTOYROW_AVX2
 ANY11(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 0, 4, 1, 31)
 #endif
+#ifdef HAS_ABGRTOYROW_AVX2
+ANY11(ABGRToYRow_Any_AVX2, ABGRToYRow_AVX2, 0, 4, 1, 31)
+#endif
 #ifdef HAS_ARGBTOYJROW_AVX2
 ANY11(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 0, 4, 1, 31)
 #endif
+#ifdef HAS_ABGRTOYJROW_AVX2
+ANY11(ABGRToYJRow_Any_AVX2, ABGRToYJRow_AVX2, 0, 4, 1, 31)
+#endif
+#ifdef HAS_RGBATOYJROW_AVX2
+ANY11(RGBAToYJRow_Any_AVX2, RGBAToYJRow_AVX2, 0, 4, 1, 31)
+#endif
 #ifdef HAS_UYVYTOYROW_AVX2
 ANY11(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 0, 2, 1, 31)
 #endif
@@ -394,39 +1019,176 @@ ANY11(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 0, 4, 1, 15)
 ANY11(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 0, 4, 1, 15)
 ANY11(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 0, 4, 1, 15)
 ANY11(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOYROW_SSE2
 ANY11(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 1, 4, 1, 15)
 ANY11(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 1, 4, 1, 15)
 #endif
 #ifdef HAS_ARGBTOYJROW_SSSE3
 ANY11(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 0, 4, 1, 15)
 #endif
+#ifdef HAS_ABGRTOYJROW_SSSE3
+ANY11(ABGRToYJRow_Any_SSSE3, ABGRToYJRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYJROW_SSSE3
+ANY11(RGBAToYJRow_Any_SSSE3, RGBAToYJRow_SSSE3, 0, 4, 1, 15)
+#endif
 #ifdef HAS_ARGBTOYROW_NEON
-ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYROW_MSA
+ANY11(ARGBToYRow_Any_MSA, ARGBToYRow_MSA, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYROW_LSX
+ANY11(ARGBToYRow_Any_LSX, ARGBToYRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYROW_LASX
+ANY11(ARGBToYRow_Any_LASX, ARGBToYRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_ARGBTOYJROW_NEON
-ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ABGRTOYJROW_NEON
+ANY11(ABGRToYJRow_Any_NEON, ABGRToYJRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYJROW_NEON
+ANY11(RGBAToYJRow_Any_NEON, RGBAToYJRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYJROW_MSA
+ANY11(ARGBToYJRow_Any_MSA, ARGBToYJRow_MSA, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYJROW_LSX
+ANY11(ARGBToYJRow_Any_LSX, ARGBToYJRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYJROW_LSX
+ANY11(RGBAToYJRow_Any_LSX, RGBAToYJRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ABGRTOYJROW_LSX
+ANY11(ABGRToYJRow_Any_LSX, ABGRToYJRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYJROW_LASX
+ANY11(RGBAToYJRow_Any_LASX, RGBAToYJRow_LASX, 0, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOYJROW_LASX
+ANY11(ARGBToYJRow_Any_LASX, ARGBToYJRow_LASX, 0, 4, 1, 31)
+#endif
+#ifdef HAS_ABGRTOYJROW_LASX
+ANY11(ABGRToYJRow_Any_LASX, ABGRToYJRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_BGRATOYROW_NEON
-ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 7)
+ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_BGRATOYROW_MSA
+ANY11(BGRAToYRow_Any_MSA, BGRAToYRow_MSA, 0, 4, 1, 15)
+#endif
+#ifdef HAS_BGRATOYROW_LSX
+ANY11(BGRAToYRow_Any_LSX, BGRAToYRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_BGRATOYROW_LASX
+ANY11(BGRAToYRow_Any_LASX, BGRAToYRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_ABGRTOYROW_NEON
-ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 7)
+ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ABGRTOYROW_MSA
+ANY11(ABGRToYRow_Any_MSA, ABGRToYRow_MSA, 0, 4, 1, 7)
+#endif
+#ifdef HAS_ABGRTOYROW_LSX
+ANY11(ABGRToYRow_Any_LSX, ABGRToYRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ABGRTOYROW_LASX
+ANY11(ABGRToYRow_Any_LASX, ABGRToYRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_RGBATOYROW_NEON
-ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 7)
+ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYROW_MSA
+ANY11(RGBAToYRow_Any_MSA, RGBAToYRow_MSA, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYROW_LSX
+ANY11(RGBAToYRow_Any_LSX, RGBAToYRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYROW_LASX
+ANY11(RGBAToYRow_Any_LASX, RGBAToYRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_RGB24TOYROW_NEON
-ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7)
+ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYJROW_AVX2
+ANY11(RGB24ToYJRow_Any_AVX2, RGB24ToYJRow_AVX2, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RGB24TOYJROW_SSSE3
+ANY11(RGB24ToYJRow_Any_SSSE3, RGB24ToYJRow_SSSE3, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYJROW_NEON
+ANY11(RGB24ToYJRow_Any_NEON, RGB24ToYJRow_NEON, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYROW_MSA
+ANY11(RGB24ToYRow_Any_MSA, RGB24ToYRow_MSA, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYROW_LSX
+ANY11(RGB24ToYRow_Any_LSX, RGB24ToYRow_LSX, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYJROW_LSX
+ANY11(RGB24ToYJRow_Any_LSX, RGB24ToYJRow_LSX, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYJROW_LASX
+ANY11(RGB24ToYJRow_Any_LASX, RGB24ToYJRow_LASX, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RGB24TOYROW_LASX
+ANY11(RGB24ToYRow_Any_LASX, RGB24ToYRow_LASX, 0, 3, 1, 31)
 #endif
 #ifdef HAS_RAWTOYROW_NEON
-ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7)
+ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYJROW_AVX2
+ANY11(RAWToYJRow_Any_AVX2, RAWToYJRow_AVX2, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RAWTOYJROW_SSSE3
+ANY11(RAWToYJRow_Any_SSSE3, RAWToYJRow_SSSE3, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYJROW_NEON
+ANY11(RAWToYJRow_Any_NEON, RAWToYJRow_NEON, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYROW_MSA
+ANY11(RAWToYRow_Any_MSA, RAWToYRow_MSA, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYROW_LSX
+ANY11(RAWToYRow_Any_LSX, RAWToYRow_LSX, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYROW_LASX
+ANY11(RAWToYRow_Any_LASX, RAWToYRow_LASX, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RAWTOYJROW_LSX
+ANY11(RAWToYJRow_Any_LSX, RAWToYJRow_LSX, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYJROW_LASX
+ANY11(RAWToYJRow_Any_LASX, RAWToYJRow_LASX, 0, 3, 1, 31)
 #endif
 #ifdef HAS_RGB565TOYROW_NEON
 ANY11(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 0, 2, 1, 7)
 #endif
+#ifdef HAS_RGB565TOYROW_MSA
+ANY11(RGB565ToYRow_Any_MSA, RGB565ToYRow_MSA, 0, 2, 1, 15)
+#endif
+#ifdef HAS_RGB565TOYROW_LSX
+ANY11(RGB565ToYRow_Any_LSX, RGB565ToYRow_LSX, 0, 2, 1, 15)
+#endif
+#ifdef HAS_RGB565TOYROW_LASX
+ANY11(RGB565ToYRow_Any_LASX, RGB565ToYRow_LASX, 0, 2, 1, 31)
+#endif
 #ifdef HAS_ARGB1555TOYROW_NEON
 ANY11(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 0, 2, 1, 7)
 #endif
+#ifdef HAS_ARGB1555TOYROW_MSA
+ANY11(ARGB1555ToYRow_Any_MSA, ARGB1555ToYRow_MSA, 0, 2, 1, 15)
+#endif
+#ifdef HAS_ARGB1555TOYROW_LSX
+ANY11(ARGB1555ToYRow_Any_LSX, ARGB1555ToYRow_LSX, 0, 2, 1, 15)
+#endif
+#ifdef HAS_ARGB1555TOYROW_LASX
+ANY11(ARGB1555ToYRow_Any_LASX, ARGB1555ToYRow_LASX, 0, 2, 1, 31)
+#endif
 #ifdef HAS_ARGB4444TOYROW_NEON
 ANY11(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 0, 2, 1, 7)
 #endif
@@ -434,23 +1196,95 @@ ANY11(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 0, 2, 1, 7)
 ANY11(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 1, 4, 1, 15)
 #endif
 #ifdef HAS_UYVYTOYROW_NEON
-ANY11(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 0, 2, 1, 15)
+ANY11(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 1, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOYROW_MSA
+ANY11(YUY2ToYRow_Any_MSA, YUY2ToYRow_MSA, 1, 4, 1, 31)
+#endif
+#ifdef HAS_YUY2TOYROW_LASX
+ANY11(YUY2ToYRow_Any_LASX, YUY2ToYRow_LASX, 1, 4, 1, 31)
+#endif
+#ifdef HAS_UYVYTOYROW_MSA
+ANY11(UYVYToYRow_Any_MSA, UYVYToYRow_MSA, 1, 4, 1, 31)
+#endif
+#ifdef HAS_UYVYTOYROW_LASX
+ANY11(UYVYToYRow_Any_LASX, UYVYToYRow_LASX, 1, 4, 1, 31)
+#endif
+#ifdef HAS_AYUVTOYROW_NEON
+ANY11(AYUVToYRow_Any_NEON, AYUVToYRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_SWAPUVROW_SSSE3
+ANY11(SwapUVRow_Any_SSSE3, SwapUVRow_SSSE3, 0, 2, 2, 15)
+#endif
+#ifdef HAS_SWAPUVROW_AVX2
+ANY11(SwapUVRow_Any_AVX2, SwapUVRow_AVX2, 0, 2, 2, 31)
+#endif
+#ifdef HAS_SWAPUVROW_NEON
+ANY11(SwapUVRow_Any_NEON, SwapUVRow_NEON, 0, 2, 2, 15)
 #endif
 #ifdef HAS_RGB24TOARGBROW_NEON
 ANY11(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 0, 3, 4, 7)
 #endif
+#ifdef HAS_RGB24TOARGBROW_MSA
+ANY11(RGB24ToARGBRow_Any_MSA, RGB24ToARGBRow_MSA, 0, 3, 4, 15)
+#endif
+#ifdef HAS_RGB24TOARGBROW_LSX
+ANY11(RGB24ToARGBRow_Any_LSX, RGB24ToARGBRow_LSX, 0, 3, 4, 15)
+#endif
+#ifdef HAS_RGB24TOARGBROW_LASX
+ANY11(RGB24ToARGBRow_Any_LASX, RGB24ToARGBRow_LASX, 0, 3, 4, 31)
+#endif
 #ifdef HAS_RAWTOARGBROW_NEON
 ANY11(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 0, 3, 4, 7)
 #endif
+#ifdef HAS_RAWTORGBAROW_NEON
+ANY11(RAWToRGBARow_Any_NEON, RAWToRGBARow_NEON, 0, 3, 4, 7)
+#endif
+#ifdef HAS_RAWTOARGBROW_MSA
+ANY11(RAWToARGBRow_Any_MSA, RAWToARGBRow_MSA, 0, 3, 4, 15)
+#endif
+#ifdef HAS_RAWTOARGBROW_LSX
+ANY11(RAWToARGBRow_Any_LSX, RAWToARGBRow_LSX, 0, 3, 4, 15)
+#endif
+#ifdef HAS_RAWTOARGBROW_LASX
+ANY11(RAWToARGBRow_Any_LASX, RAWToARGBRow_LASX, 0, 3, 4, 31)
+#endif
 #ifdef HAS_RGB565TOARGBROW_NEON
 ANY11(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 0, 2, 4, 7)
 #endif
+#ifdef HAS_RGB565TOARGBROW_MSA
+ANY11(RGB565ToARGBRow_Any_MSA, RGB565ToARGBRow_MSA, 0, 2, 4, 15)
+#endif
+#ifdef HAS_RGB565TOARGBROW_LSX
+ANY11(RGB565ToARGBRow_Any_LSX, RGB565ToARGBRow_LSX, 0, 2, 4, 15)
+#endif
+#ifdef HAS_RGB565TOARGBROW_LASX
+ANY11(RGB565ToARGBRow_Any_LASX, RGB565ToARGBRow_LASX, 0, 2, 4, 31)
+#endif
 #ifdef HAS_ARGB1555TOARGBROW_NEON
 ANY11(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 0, 2, 4, 7)
 #endif
+#ifdef HAS_ARGB1555TOARGBROW_MSA
+ANY11(ARGB1555ToARGBRow_Any_MSA, ARGB1555ToARGBRow_MSA, 0, 2, 4, 15)
+#endif
+#ifdef HAS_ARGB1555TOARGBROW_LSX
+ANY11(ARGB1555ToARGBRow_Any_LSX, ARGB1555ToARGBRow_LSX, 0, 2, 4, 15)
+#endif
+#ifdef HAS_ARGB1555TOARGBROW_LASX
+ANY11(ARGB1555ToARGBRow_Any_LASX, ARGB1555ToARGBRow_LASX, 0, 2, 4, 31)
+#endif
 #ifdef HAS_ARGB4444TOARGBROW_NEON
 ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7)
 #endif
+#ifdef HAS_ARGB4444TOARGBROW_MSA
+ANY11(ARGB4444ToARGBRow_Any_MSA, ARGB4444ToARGBRow_MSA, 0, 2, 4, 15)
+#endif
+#ifdef HAS_ARGB4444TOARGBROW_LSX
+ANY11(ARGB4444ToARGBRow_Any_LSX, ARGB4444ToARGBRow_LSX, 0, 2, 4, 15)
+#endif
+#ifdef HAS_ARGB4444TOARGBROW_LASX
+ANY11(ARGB4444ToARGBRow_Any_LASX, ARGB4444ToARGBRow_LASX, 0, 2, 4, 31)
+#endif
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
 ANY11(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, 0, 4, 4, 3)
 #endif
@@ -466,29 +1300,44 @@ ANY11(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, 0, 4, 4, 7)
 #ifdef HAS_ARGBATTENUATEROW_NEON
 ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
 #endif
+#ifdef HAS_ARGBATTENUATEROW_MSA
+ANY11(ARGBAttenuateRow_Any_MSA, ARGBAttenuateRow_MSA, 0, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_LASX
+ANY11(ARGBAttenuateRow_Any_LASX, ARGBAttenuateRow_LASX, 0, 4, 4, 15)
+#endif
 #ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
 ANY11(ARGBExtractAlphaRow_Any_SSE2, ARGBExtractAlphaRow_SSE2, 0, 4, 1, 7)
 #endif
+#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
+ANY11(ARGBExtractAlphaRow_Any_AVX2, ARGBExtractAlphaRow_AVX2, 0, 4, 1, 31)
+#endif
 #ifdef HAS_ARGBEXTRACTALPHAROW_NEON
 ANY11(ARGBExtractAlphaRow_Any_NEON, ARGBExtractAlphaRow_NEON, 0, 4, 1, 15)
 #endif
+#ifdef HAS_ARGBEXTRACTALPHAROW_MSA
+ANY11(ARGBExtractAlphaRow_Any_MSA, ARGBExtractAlphaRow_MSA, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBEXTRACTALPHAROW_LSX
+ANY11(ARGBExtractAlphaRow_Any_LSX, ARGBExtractAlphaRow_LSX, 0, 4, 1, 15)
+#endif
 #undef ANY11
 
 // Any 1 to 1 blended.  Destination is read, modify, write.
-#define ANY11B(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                    \
-    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
-      SIMD_ALIGNED(uint8 temp[128 * 2]);                                       \
-      memset(temp, 0, 128 * 2);  /* for YUY2 and msan */                       \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr, dst_ptr, n);                                         \
-      }                                                                        \
-      memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP);    \
-      memcpy(temp + 128, dst_ptr + n * BPP, r * BPP);                          \
-      ANY_SIMD(temp, temp + 128, MASK + 1);                                    \
-      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
-    }
+#define ANY11B(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)               \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_ptr, int width) {     \
+    SIMD_ALIGNED(uint8_t temp[64 * 2]);                                   \
+    memset(temp, 0, 64 * 2); /* for msan */                               \
+    int r = width & MASK;                                                 \
+    int n = width & ~MASK;                                                \
+    if (n > 0) {                                                          \
+      ANY_SIMD(src_ptr, dst_ptr, n);                                      \
+    }                                                                     \
+    memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP); \
+    memcpy(temp + 64, dst_ptr + n * BPP, r * BPP);                        \
+    ANY_SIMD(temp, temp + 64, MASK + 1);                                  \
+    memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                        \
+  }
 
 #ifdef HAS_ARGBCOPYALPHAROW_AVX2
 ANY11B(ARGBCopyAlphaRow_Any_AVX2, ARGBCopyAlphaRow_AVX2, 0, 4, 4, 15)
@@ -506,61 +1355,339 @@ ANY11B(ARGBCopyYToAlphaRow_Any_SSE2, ARGBCopyYToAlphaRow_SSE2, 0, 1, 4, 7)
 
 // Any 1 to 1 with parameter.
 #define ANY11P(NAMEANY, ANY_SIMD, T, SBPP, BPP, MASK)                          \
-    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr,                         \
-                 T shuffler, int width) {                                      \
-      SIMD_ALIGNED(uint8 temp[64 * 2]);                                        \
-      memset(temp, 0, 64);  /* for msan */                                     \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr, dst_ptr, shuffler, n);                               \
-      }                                                                        \
-      memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
-      ANY_SIMD(temp, temp + 64, shuffler, MASK + 1);                           \
-      memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                           \
-    }
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_ptr, T param, int width) { \
+    SIMD_ALIGNED(uint8_t temp[64 * 2]);                                        \
+    memset(temp, 0, 64); /* for msan */                                        \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(src_ptr, dst_ptr, param, n);                                    \
+    }                                                                          \
+    memcpy(temp, src_ptr + n * SBPP, r * SBPP);                                \
+    ANY_SIMD(temp, temp + 64, param, MASK + 1);                                \
+    memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                             \
+  }
+
+#if defined(HAS_I400TOARGBROW_SSE2)
+ANY11P(I400ToARGBRow_Any_SSE2,
+       I400ToARGBRow_SSE2,
+       const struct YuvConstants*,
+       1,
+       4,
+       7)
+#endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+ANY11P(I400ToARGBRow_Any_AVX2,
+       I400ToARGBRow_AVX2,
+       const struct YuvConstants*,
+       1,
+       4,
+       15)
+#endif
+#if defined(HAS_I400TOARGBROW_NEON)
+ANY11P(I400ToARGBRow_Any_NEON,
+       I400ToARGBRow_NEON,
+       const struct YuvConstants*,
+       1,
+       4,
+       7)
+#endif
+#if defined(HAS_I400TOARGBROW_MSA)
+ANY11P(I400ToARGBRow_Any_MSA,
+       I400ToARGBRow_MSA,
+       const struct YuvConstants*,
+       1,
+       4,
+       15)
+#endif
+#if defined(HAS_I400TOARGBROW_LSX)
+ANY11P(I400ToARGBRow_Any_LSX,
+       I400ToARGBRow_LSX,
+       const struct YuvConstants*,
+       1,
+       4,
+       15)
+#endif
 
 #if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
-ANY11P(ARGBToRGB565DitherRow_Any_SSE2, ARGBToRGB565DitherRow_SSE2,
-       const uint32, 4, 2, 3)
+ANY11P(ARGBToRGB565DitherRow_Any_SSE2,
+       ARGBToRGB565DitherRow_SSE2,
+       const uint32_t,
+       4,
+       2,
+       3)
 #endif
 #if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
-ANY11P(ARGBToRGB565DitherRow_Any_AVX2, ARGBToRGB565DitherRow_AVX2,
-       const uint32, 4, 2, 7)
+ANY11P(ARGBToRGB565DitherRow_Any_AVX2,
+       ARGBToRGB565DitherRow_AVX2,
+       const uint32_t,
+       4,
+       2,
+       7)
 #endif
 #if defined(HAS_ARGBTORGB565DITHERROW_NEON)
-ANY11P(ARGBToRGB565DitherRow_Any_NEON, ARGBToRGB565DitherRow_NEON,
-       const uint32, 4, 2, 7)
-#endif
-#ifdef HAS_ARGBSHUFFLEROW_SSE2
-ANY11P(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2, const uint8*, 4, 4, 3)
+ANY11P(ARGBToRGB565DitherRow_Any_NEON,
+       ARGBToRGB565DitherRow_NEON,
+       const uint32_t,
+       4,
+       2,
+       7)
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_MSA)
+ANY11P(ARGBToRGB565DitherRow_Any_MSA,
+       ARGBToRGB565DitherRow_MSA,
+       const uint32_t,
+       4,
+       2,
+       7)
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_LASX)
+ANY11P(ARGBToRGB565DitherRow_Any_LASX,
+       ARGBToRGB565DitherRow_LASX,
+       const uint32_t,
+       4,
+       2,
+       15)
 #endif
 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
-ANY11P(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, const uint8*, 4, 4, 7)
+ANY11P(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, const uint8_t*, 4, 4, 7)
 #endif
 #ifdef HAS_ARGBSHUFFLEROW_AVX2
-ANY11P(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2, const uint8*, 4, 4, 15)
+ANY11P(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2, const uint8_t*, 4, 4, 15)
 #endif
 #ifdef HAS_ARGBSHUFFLEROW_NEON
-ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8*, 4, 4, 3)
+ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8_t*, 4, 4, 3)
 #endif
+#ifdef HAS_ARGBSHUFFLEROW_MSA
+ANY11P(ARGBShuffleRow_Any_MSA, ARGBShuffleRow_MSA, const uint8_t*, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_LASX
+ANY11P(ARGBShuffleRow_Any_LASX, ARGBShuffleRow_LASX, const uint8_t*, 4, 4, 15)
+#endif
+#undef ANY11P
 #undef ANY11P
 
+// Any 1 to 1 with type
+#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, STYPE, DTYPE, MASK)  \
+  void NAMEANY(const STYPE* src_ptr, DTYPE* dst_ptr, int width) { \
+    SIMD_ALIGNED(uint8_t temp[(MASK + 1) * SBPP]);                \
+    SIMD_ALIGNED(uint8_t out[(MASK + 1) * BPP]);                  \
+    memset(temp, 0, (MASK + 1) * SBPP); /* for msan */            \
+    int r = width & MASK;                                         \
+    int n = width & ~MASK;                                        \
+    if (n > 0) {                                                  \
+      ANY_SIMD(src_ptr, dst_ptr, n);                              \
+    }                                                             \
+    memcpy(temp, (uint8_t*)(src_ptr) + n * SBPP, r * SBPP);       \
+    ANY_SIMD((STYPE*)temp, (DTYPE*)out, MASK + 1);                \
+    memcpy((uint8_t*)(dst_ptr) + n * BPP, out, r * BPP);          \
+  }
+
+#ifdef HAS_ARGBTOAR64ROW_SSSE3
+ANY11T(ARGBToAR64Row_Any_SSSE3, ARGBToAR64Row_SSSE3, 4, 8, uint8_t, uint16_t, 3)
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_SSSE3
+ANY11T(ARGBToAB64Row_Any_SSSE3, ARGBToAB64Row_SSSE3, 4, 8, uint8_t, uint16_t, 3)
+#endif
+
+#ifdef HAS_AR64TOARGBROW_SSSE3
+ANY11T(AR64ToARGBRow_Any_SSSE3, AR64ToARGBRow_SSSE3, 8, 4, uint16_t, uint8_t, 3)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_SSSE3
+ANY11T(AB64ToARGBRow_Any_SSSE3, AB64ToARGBRow_SSSE3, 8, 4, uint16_t, uint8_t, 3)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_AVX2
+ANY11T(ARGBToAR64Row_Any_AVX2, ARGBToAR64Row_AVX2, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_AVX2
+ANY11T(ARGBToAB64Row_Any_AVX2, ARGBToAB64Row_AVX2, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_AR64TOARGBROW_AVX2
+ANY11T(AR64ToARGBRow_Any_AVX2, AR64ToARGBRow_AVX2, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_AVX2
+ANY11T(AB64ToARGBRow_Any_AVX2, AB64ToARGBRow_AVX2, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_NEON
+ANY11T(ARGBToAR64Row_Any_NEON, ARGBToAR64Row_NEON, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_NEON
+ANY11T(ARGBToAB64Row_Any_NEON, ARGBToAB64Row_NEON, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_AR64TOARGBROW_NEON
+ANY11T(AR64ToARGBRow_Any_NEON, AR64ToARGBRow_NEON, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_NEON
+ANY11T(AB64ToARGBRow_Any_NEON, AB64ToARGBRow_NEON, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#undef ANY11T
+
+// Any 1 to 1 with parameter and shorts.  BPP measures in shorts.
+#define ANY11C(NAMEANY, ANY_SIMD, SBPP, BPP, STYPE, DTYPE, MASK)             \
+  void NAMEANY(const STYPE* src_ptr, DTYPE* dst_ptr, int scale, int width) { \
+    SIMD_ALIGNED(STYPE temp[32]);                                            \
+    SIMD_ALIGNED(DTYPE out[32]);                                             \
+    memset(temp, 0, 32 * SBPP); /* for msan */                               \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(src_ptr, dst_ptr, scale, n);                                  \
+    }                                                                        \
+    memcpy(temp, src_ptr + n, r * SBPP);                                     \
+    ANY_SIMD(temp, out, scale, MASK + 1);                                    \
+    memcpy(dst_ptr + n, out, r * BPP);                                       \
+  }
+
+#ifdef HAS_CONVERT16TO8ROW_SSSE3
+ANY11C(Convert16To8Row_Any_SSSE3,
+       Convert16To8Row_SSSE3,
+       2,
+       1,
+       uint16_t,
+       uint8_t,
+       15)
+#endif
+#ifdef HAS_CONVERT16TO8ROW_AVX2
+ANY11C(Convert16To8Row_Any_AVX2,
+       Convert16To8Row_AVX2,
+       2,
+       1,
+       uint16_t,
+       uint8_t,
+       31)
+#endif
+#ifdef HAS_CONVERT16TO8ROW_NEON
+ANY11C(Convert16To8Row_Any_NEON,
+       Convert16To8Row_NEON,
+       2,
+       1,
+       uint16_t,
+       uint8_t,
+       15)
+#endif
+#ifdef HAS_CONVERT8TO16ROW_SSE2
+ANY11C(Convert8To16Row_Any_SSE2,
+       Convert8To16Row_SSE2,
+       1,
+       2,
+       uint8_t,
+       uint16_t,
+       15)
+#endif
+#ifdef HAS_CONVERT8TO16ROW_AVX2
+ANY11C(Convert8To16Row_Any_AVX2,
+       Convert8To16Row_AVX2,
+       1,
+       2,
+       uint8_t,
+       uint16_t,
+       31)
+#endif
+#ifdef HAS_MULTIPLYROW_16_AVX2
+ANY11C(MultiplyRow_16_Any_AVX2,
+       MultiplyRow_16_AVX2,
+       2,
+       2,
+       uint16_t,
+       uint16_t,
+       31)
+#endif
+#ifdef HAS_MULTIPLYROW_16_NEON
+ANY11C(MultiplyRow_16_Any_NEON,
+       MultiplyRow_16_NEON,
+       2,
+       2,
+       uint16_t,
+       uint16_t,
+       15)
+#endif
+#ifdef HAS_DIVIDEROW_16_AVX2
+ANY11C(DivideRow_16_Any_AVX2, DivideRow_16_AVX2, 2, 2, uint16_t, uint16_t, 31)
+#endif
+#ifdef HAS_DIVIDEROW_16_NEON
+ANY11C(DivideRow_16_Any_NEON, DivideRow_16_NEON, 2, 2, uint16_t, uint16_t, 15)
+#endif
+#undef ANY11C
+
+// Any 1 to 1 with parameter and shorts to byte.  BPP measures in shorts.
+#define ANY11P16(NAMEANY, ANY_SIMD, ST, T, SBPP, BPP, MASK)             \
+  void NAMEANY(const ST* src_ptr, T* dst_ptr, float param, int width) { \
+    SIMD_ALIGNED(ST temp[32]);                                          \
+    SIMD_ALIGNED(T out[32]);                                            \
+    memset(temp, 0, SBPP * 32); /* for msan */                          \
+    int r = width & MASK;                                               \
+    int n = width & ~MASK;                                              \
+    if (n > 0) {                                                        \
+      ANY_SIMD(src_ptr, dst_ptr, param, n);                             \
+    }                                                                   \
+    memcpy(temp, src_ptr + n, r * SBPP);                                \
+    ANY_SIMD(temp, out, param, MASK + 1);                               \
+    memcpy(dst_ptr + n, out, r * BPP);                                  \
+  }
+
+#ifdef HAS_HALFFLOATROW_SSE2
+ANY11P16(HalfFloatRow_Any_SSE2, HalfFloatRow_SSE2, uint16_t, uint16_t, 2, 2, 7)
+#endif
+#ifdef HAS_HALFFLOATROW_AVX2
+ANY11P16(HalfFloatRow_Any_AVX2, HalfFloatRow_AVX2, uint16_t, uint16_t, 2, 2, 15)
+#endif
+#ifdef HAS_HALFFLOATROW_F16C
+ANY11P16(HalfFloatRow_Any_F16C, HalfFloatRow_F16C, uint16_t, uint16_t, 2, 2, 15)
+ANY11P16(HalfFloat1Row_Any_F16C,
+         HalfFloat1Row_F16C,
+         uint16_t,
+         uint16_t,
+         2,
+         2,
+         15)
+#endif
+#ifdef HAS_HALFFLOATROW_NEON
+ANY11P16(HalfFloatRow_Any_NEON, HalfFloatRow_NEON, uint16_t, uint16_t, 2, 2, 7)
+ANY11P16(HalfFloat1Row_Any_NEON,
+         HalfFloat1Row_NEON,
+         uint16_t,
+         uint16_t,
+         2,
+         2,
+         7)
+#endif
+#ifdef HAS_HALFFLOATROW_MSA
+ANY11P16(HalfFloatRow_Any_MSA, HalfFloatRow_MSA, uint16_t, uint16_t, 2, 2, 31)
+#endif
+#ifdef HAS_BYTETOFLOATROW_NEON
+ANY11P16(ByteToFloatRow_Any_NEON, ByteToFloatRow_NEON, uint8_t, float, 1, 3, 7)
+#endif
+#ifdef HAS_HALFFLOATROW_LSX
+ANY11P16(HalfFloatRow_Any_LSX, HalfFloatRow_LSX, uint16_t, uint16_t, 2, 2, 31)
+#endif
+#undef ANY11P16
+
 // Any 1 to 1 with yuvconstants
-#define ANY11C(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                    \
-    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr,                         \
-                 const struct YuvConstants* yuvconstants, int width) {         \
-      SIMD_ALIGNED(uint8 temp[128 * 2]);                                       \
-      memset(temp, 0, 128);  /* for YUY2 and msan */                           \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr, dst_ptr, yuvconstants, n);                           \
-      }                                                                        \
-      memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP);    \
-      ANY_SIMD(temp, temp + 128, yuvconstants, MASK + 1);                      \
-      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
-    }
+#define ANY11C(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)               \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_ptr,                  \
+               const struct YuvConstants* yuvconstants, int width) {      \
+    SIMD_ALIGNED(uint8_t temp[128 * 2]);                                  \
+    memset(temp, 0, 128); /* for YUY2 and msan */                         \
+    int r = width & MASK;                                                 \
+    int n = width & ~MASK;                                                \
+    if (n > 0) {                                                          \
+      ANY_SIMD(src_ptr, dst_ptr, yuvconstants, n);                        \
+    }                                                                     \
+    memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP); \
+    ANY_SIMD(temp, temp + 128, yuvconstants, MASK + 1);                   \
+    memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                       \
+  }
 #if defined(HAS_YUY2TOARGBROW_SSSE3)
 ANY11C(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, 1, 4, 4, 15)
 ANY11C(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, 1, 4, 4, 15)
@@ -573,54 +1700,126 @@ ANY11C(UYVYToARGBRow_Any_AVX2, UYVYToARGBRow_AVX2, 1, 4, 4, 31)
 ANY11C(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, 1, 4, 4, 7)
 ANY11C(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, 1, 4, 4, 7)
 #endif
+#if defined(HAS_YUY2TOARGBROW_MSA)
+ANY11C(YUY2ToARGBRow_Any_MSA, YUY2ToARGBRow_MSA, 1, 4, 4, 7)
+ANY11C(UYVYToARGBRow_Any_MSA, UYVYToARGBRow_MSA, 1, 4, 4, 7)
+#endif
+#if defined(HAS_YUY2TOARGBROW_LSX)
+ANY11C(YUY2ToARGBRow_Any_LSX, YUY2ToARGBRow_LSX, 1, 4, 4, 7)
+ANY11C(UYVYToARGBRow_Any_LSX, UYVYToARGBRow_LSX, 1, 4, 4, 7)
+#endif
 #undef ANY11C
 
 // Any 1 to 1 interpolate.  Takes 2 rows of source via stride.
-#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, MASK)                             \
-    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
-                 ptrdiff_t src_stride_ptr, int width,                          \
-                 int source_y_fraction) {                                      \
-      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
-      memset(temp, 0, 64 * 2);  /* for msan */                                 \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(dst_ptr, src_ptr, src_stride_ptr, n, source_y_fraction);      \
-      }                                                                        \
-      memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
-      memcpy(temp + 64, src_ptr + src_stride_ptr + n * SBPP, r * SBPP);        \
-      ANY_SIMD(temp + 128, temp, 64, MASK + 1, source_y_fraction);             \
-      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
-    }
+#define ANY11I(NAMEANY, ANY_SIMD, TD, TS, SBPP, BPP, MASK)           \
+  void NAMEANY(TD* dst_ptr, const TS* src_ptr, ptrdiff_t src_stride, \
+               int width, int source_y_fraction) {                   \
+    SIMD_ALIGNED(TS temps[64 * 2]);                                  \
+    SIMD_ALIGNED(TD tempd[64]);                                      \
+    memset(temps, 0, sizeof(temps)); /* for msan */                  \
+    int r = width & MASK;                                            \
+    int n = width & ~MASK;                                           \
+    if (n > 0) {                                                     \
+      ANY_SIMD(dst_ptr, src_ptr, src_stride, n, source_y_fraction);  \
+    }                                                                \
+    memcpy(temps, src_ptr + n * SBPP, r * SBPP * sizeof(TS));        \
+    if (source_y_fraction) {                                         \
+      memcpy(temps + 64, src_ptr + src_stride + n * SBPP,            \
+             r * SBPP * sizeof(TS));                                 \
+    }                                                                \
+    ANY_SIMD(tempd, temps, 64, MASK + 1, source_y_fraction);         \
+    memcpy(dst_ptr + n * BPP, tempd, r * BPP * sizeof(TD));          \
+  }
 
 #ifdef HAS_INTERPOLATEROW_AVX2
-ANY11T(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, 1, 1, 31)
+ANY11I(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, uint8_t, uint8_t, 1, 1, 31)
 #endif
 #ifdef HAS_INTERPOLATEROW_SSSE3
-ANY11T(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, 1, 1, 15)
+ANY11I(InterpolateRow_Any_SSSE3,
+       InterpolateRow_SSSE3,
+       uint8_t,
+       uint8_t,
+       1,
+       1,
+       15)
 #endif
 #ifdef HAS_INTERPOLATEROW_NEON
-ANY11T(InterpolateRow_Any_NEON, InterpolateRow_NEON, 1, 1, 15)
+ANY11I(InterpolateRow_Any_NEON, InterpolateRow_NEON, uint8_t, uint8_t, 1, 1, 15)
 #endif
-#ifdef HAS_INTERPOLATEROW_DSPR2
-ANY11T(InterpolateRow_Any_DSPR2, InterpolateRow_DSPR2, 1, 1, 3)
+#ifdef HAS_INTERPOLATEROW_MSA
+ANY11I(InterpolateRow_Any_MSA, InterpolateRow_MSA, uint8_t, uint8_t, 1, 1, 31)
 #endif
-#undef ANY11T
+#ifdef HAS_INTERPOLATEROW_LSX
+ANY11I(InterpolateRow_Any_LSX, InterpolateRow_LSX, uint8_t, uint8_t, 1, 1, 31)
+#endif
+
+#ifdef HAS_INTERPOLATEROW_16_NEON
+ANY11I(InterpolateRow_16_Any_NEON,
+       InterpolateRow_16_NEON,
+       uint16_t,
+       uint16_t,
+       1,
+       1,
+       7)
+#endif
+#undef ANY11I
+
+// Any 1 to 1 interpolate with scale param
+#define ANY11IS(NAMEANY, ANY_SIMD, TD, TS, SBPP, BPP, MASK)                \
+  void NAMEANY(TD* dst_ptr, const TS* src_ptr, ptrdiff_t src_stride,       \
+               int scale, int width, int source_y_fraction) {              \
+    SIMD_ALIGNED(TS temps[64 * 2]);                                        \
+    SIMD_ALIGNED(TD tempd[64]);                                            \
+    memset(temps, 0, sizeof(temps)); /* for msan */                        \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(dst_ptr, src_ptr, src_stride, scale, n, source_y_fraction); \
+    }                                                                      \
+    memcpy(temps, src_ptr + n * SBPP, r * SBPP * sizeof(TS));              \
+    if (source_y_fraction) {                                               \
+      memcpy(temps + 64, src_ptr + src_stride + n * SBPP,                  \
+             r * SBPP * sizeof(TS));                                       \
+    }                                                                      \
+    ANY_SIMD(tempd, temps, 64, scale, MASK + 1, source_y_fraction);        \
+    memcpy(dst_ptr + n * BPP, tempd, r * BPP * sizeof(TD));                \
+  }
+
+#ifdef HAS_INTERPOLATEROW_16TO8_NEON
+ANY11IS(InterpolateRow_16To8_Any_NEON,
+        InterpolateRow_16To8_NEON,
+        uint8_t,
+        uint16_t,
+        1,
+        1,
+        7)
+#endif
+#ifdef HAS_INTERPOLATEROW_16TO8_AVX2
+ANY11IS(InterpolateRow_16To8_Any_AVX2,
+        InterpolateRow_16To8_AVX2,
+        uint8_t,
+        uint16_t,
+        1,
+        1,
+        31)
+#endif
+
+#undef ANY11IS
 
 // Any 1 to 1 mirror.
-#define ANY11M(NAMEANY, ANY_SIMD, BPP, MASK)                                   \
-    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
-      SIMD_ALIGNED(uint8 temp[64 * 2]);                                        \
-      memset(temp, 0, 64);  /* for msan */                                     \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr + r * BPP, dst_ptr, n);                               \
-      }                                                                        \
-      memcpy(temp, src_ptr, r * BPP);                                          \
-      ANY_SIMD(temp, temp + 64, MASK + 1);                                     \
-      memcpy(dst_ptr + n * BPP, temp + 64 + (MASK + 1 - r) * BPP, r * BPP);    \
-    }
+#define ANY11M(NAMEANY, ANY_SIMD, BPP, MASK)                              \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_ptr, int width) {     \
+    SIMD_ALIGNED(uint8_t temp[64 * 2]);                                   \
+    memset(temp, 0, 64); /* for msan */                                   \
+    int r = width & MASK;                                                 \
+    int n = width & ~MASK;                                                \
+    if (n > 0) {                                                          \
+      ANY_SIMD(src_ptr + r * BPP, dst_ptr, n);                            \
+    }                                                                     \
+    memcpy(temp, src_ptr, r* BPP);                                        \
+    ANY_SIMD(temp, temp + 64, MASK + 1);                                  \
+    memcpy(dst_ptr + n * BPP, temp + 64 + (MASK + 1 - r) * BPP, r * BPP); \
+  }
 
 #ifdef HAS_MIRRORROW_AVX2
 ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31)
@@ -629,7 +1828,28 @@ ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31)
 ANY11M(MirrorRow_Any_SSSE3, MirrorRow_SSSE3, 1, 15)
 #endif
 #ifdef HAS_MIRRORROW_NEON
-ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 15)
+ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 31)
+#endif
+#ifdef HAS_MIRRORROW_MSA
+ANY11M(MirrorRow_Any_MSA, MirrorRow_MSA, 1, 63)
+#endif
+#ifdef HAS_MIRRORROW_LASX
+ANY11M(MirrorRow_Any_LASX, MirrorRow_LASX, 1, 63)
+#endif
+#ifdef HAS_MIRRORUVROW_AVX2
+ANY11M(MirrorUVRow_Any_AVX2, MirrorUVRow_AVX2, 2, 15)
+#endif
+#ifdef HAS_MIRRORUVROW_SSSE3
+ANY11M(MirrorUVRow_Any_SSSE3, MirrorUVRow_SSSE3, 2, 7)
+#endif
+#ifdef HAS_MIRRORUVROW_NEON
+ANY11M(MirrorUVRow_Any_NEON, MirrorUVRow_NEON, 2, 31)
+#endif
+#ifdef HAS_MIRRORUVROW_MSA
+ANY11M(MirrorUVRow_Any_MSA, MirrorUVRow_MSA, 2, 7)
+#endif
+#ifdef HAS_MIRRORUVROW_LASX
+ANY11M(MirrorUVRow_Any_LASX, MirrorUVRow_LASX, 2, 15)
 #endif
 #ifdef HAS_ARGBMIRRORROW_AVX2
 ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7)
@@ -638,69 +1858,72 @@ ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7)
 ANY11M(ARGBMirrorRow_Any_SSE2, ARGBMirrorRow_SSE2, 4, 3)
 #endif
 #ifdef HAS_ARGBMIRRORROW_NEON
-ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 3)
+ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 7)
+#endif
+#ifdef HAS_ARGBMIRRORROW_MSA
+ANY11M(ARGBMirrorRow_Any_MSA, ARGBMirrorRow_MSA, 4, 15)
+#endif
+#ifdef HAS_ARGBMIRRORROW_LASX
+ANY11M(ARGBMirrorRow_Any_LASX, ARGBMirrorRow_LASX, 4, 15)
+#endif
+#ifdef HAS_RGB24MIRRORROW_SSSE3
+ANY11M(RGB24MirrorRow_Any_SSSE3, RGB24MirrorRow_SSSE3, 3, 15)
+#endif
+#ifdef HAS_RGB24MIRRORROW_NEON
+ANY11M(RGB24MirrorRow_Any_NEON, RGB24MirrorRow_NEON, 3, 15)
 #endif
 #undef ANY11M
 
 // Any 1 plane. (memset)
-#define ANY1(NAMEANY, ANY_SIMD, T, BPP, MASK)                                  \
-    void NAMEANY(uint8* dst_ptr, T v32, int width) {                           \
-      SIMD_ALIGNED(uint8 temp[64]);                                            \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(dst_ptr, v32, n);                                             \
-      }                                                                        \
-      ANY_SIMD(temp, v32, MASK + 1);                                           \
-      memcpy(dst_ptr + n * BPP, temp, r * BPP);                                \
-    }
+#define ANY1(NAMEANY, ANY_SIMD, T, BPP, MASK)        \
+  void NAMEANY(uint8_t* dst_ptr, T v32, int width) { \
+    SIMD_ALIGNED(uint8_t temp[64]);                  \
+    memset(temp, 0, 64); /* for msan */              \
+    int r = width & MASK;                            \
+    int n = width & ~MASK;                           \
+    if (n > 0) {                                     \
+      ANY_SIMD(dst_ptr, v32, n);                     \
+    }                                                \
+    ANY_SIMD(temp, v32, MASK + 1);                   \
+    memcpy(dst_ptr + n * BPP, temp, r * BPP);        \
+  }
 
 #ifdef HAS_SETROW_X86
-ANY1(SetRow_Any_X86, SetRow_X86, uint8, 1, 3)
+ANY1(SetRow_Any_X86, SetRow_X86, uint8_t, 1, 3)
 #endif
 #ifdef HAS_SETROW_NEON
-ANY1(SetRow_Any_NEON, SetRow_NEON, uint8, 1, 15)
+ANY1(SetRow_Any_NEON, SetRow_NEON, uint8_t, 1, 15)
+#endif
+#ifdef HAS_SETROW_LSX
+ANY1(SetRow_Any_LSX, SetRow_LSX, uint8_t, 1, 15)
 #endif
 #ifdef HAS_ARGBSETROW_NEON
-ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32, 4, 3)
+ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32_t, 4, 3)
+#endif
+#ifdef HAS_ARGBSETROW_MSA
+ANY1(ARGBSetRow_Any_MSA, ARGBSetRow_MSA, uint32_t, 4, 3)
+#endif
+#ifdef HAS_ARGBSETROW_LSX
+ANY1(ARGBSetRow_Any_LSX, ARGBSetRow_LSX, uint32_t, 4, 3)
 #endif
 #undef ANY1
 
 // Any 1 to 2.  Outputs UV planes.
-#define ANY12(NAMEANY, ANY_SIMD, UVSHIFT, BPP, DUVSHIFT, MASK)                 \
-    void NAMEANY(const uint8* src_ptr, uint8* dst_u, uint8* dst_v, int width) {\
-      SIMD_ALIGNED(uint8 temp[128 * 3]);                                       \
-      memset(temp, 0, 128);  /* for msan */                                    \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr, dst_u, dst_v, n);                                    \
-      }                                                                        \
-      memcpy(temp, src_ptr  + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);     \
-      /* repeat last 4 bytes for 422 subsampler */                             \
-      if ((width & 1) && BPP == 4 && DUVSHIFT == 1) {                          \
-        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
-               temp + SS(r, UVSHIFT) * BPP - BPP, BPP);                        \
-      }                                                                        \
-      /* repeat last 4 - 12 bytes for 411 subsampler */                        \
-      if (((width & 3) == 1) && BPP == 4 && DUVSHIFT == 2) {                   \
-        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
-               temp + SS(r, UVSHIFT) * BPP - BPP, BPP);                        \
-        memcpy(temp + SS(r, UVSHIFT) * BPP + BPP,                              \
-               temp + SS(r, UVSHIFT) * BPP - BPP, BPP * 2);                    \
-      }                                                                        \
-      if (((width & 3) == 2) && BPP == 4 && DUVSHIFT == 2) {                   \
-        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
-               temp + SS(r, UVSHIFT) * BPP - BPP * 2, BPP * 2);                \
-      }                                                                        \
-      if (((width & 3) == 3) && BPP == 4 && DUVSHIFT == 2) {                   \
-        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
-               temp + SS(r, UVSHIFT) * BPP - BPP, BPP);                        \
-      }                                                                        \
-      ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1);                        \
-      memcpy(dst_u + (n >> DUVSHIFT), temp + 128, SS(r, DUVSHIFT));            \
-      memcpy(dst_v + (n >> DUVSHIFT), temp + 256, SS(r, DUVSHIFT));            \
-    }
+#define ANY12(NAMEANY, ANY_SIMD, UVSHIFT, BPP, DUVSHIFT, MASK)          \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_u, uint8_t* dst_v,  \
+               int width) {                                             \
+    SIMD_ALIGNED(uint8_t temp[128 * 3]);                                \
+    memset(temp, 0, 128); /* for msan */                                \
+    int r = width & MASK;                                               \
+    int n = width & ~MASK;                                              \
+    if (n > 0) {                                                        \
+      ANY_SIMD(src_ptr, dst_u, dst_v, n);                               \
+    }                                                                   \
+    memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP); \
+    ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1);                   \
+    memcpy(dst_u + (n >> DUVSHIFT), temp + 128, SS(r, DUVSHIFT));       \
+    memcpy(dst_v + (n >> DUVSHIFT), temp + 256, SS(r, DUVSHIFT));       \
+  }
 
 #ifdef HAS_SPLITUVROW_SSE2
 ANY12(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, 0, 2, 0, 15)
@@ -711,8 +1934,11 @@ ANY12(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, 0, 2, 0, 31)
 #ifdef HAS_SPLITUVROW_NEON
 ANY12(SplitUVRow_Any_NEON, SplitUVRow_NEON, 0, 2, 0, 15)
 #endif
-#ifdef HAS_SPLITUVROW_DSPR2
-ANY12(SplitUVRow_Any_DSPR2, SplitUVRow_DSPR2, 0, 2, 0, 15)
+#ifdef HAS_SPLITUVROW_MSA
+ANY12(SplitUVRow_Any_MSA, SplitUVRow_MSA, 0, 2, 0, 31)
+#endif
+#ifdef HAS_SPLITUVROW_LSX
+ANY12(SplitUVRow_Any_LSX, SplitUVRow_LSX, 0, 2, 0, 31)
 #endif
 #ifdef HAS_ARGBTOUV444ROW_SSSE3
 ANY12(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3, 0, 4, 0, 15)
@@ -727,47 +1953,163 @@ ANY12(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2, 1, 4, 1, 15)
 #endif
 #ifdef HAS_YUY2TOUV422ROW_NEON
 ANY12(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON, 0, 4, 0, 7)
-ANY12(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON, 0, 4, 2, 31)
 ANY12(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON, 1, 4, 1, 15)
 ANY12(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON, 1, 4, 1, 15)
 #endif
+#ifdef HAS_YUY2TOUV422ROW_MSA
+ANY12(ARGBToUV444Row_Any_MSA, ARGBToUV444Row_MSA, 0, 4, 0, 15)
+ANY12(YUY2ToUV422Row_Any_MSA, YUY2ToUV422Row_MSA, 1, 4, 1, 31)
+ANY12(UYVYToUV422Row_Any_MSA, UYVYToUV422Row_MSA, 1, 4, 1, 31)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_LASX
+ANY12(ARGBToUV444Row_Any_LASX, ARGBToUV444Row_LASX, 0, 4, 0, 31)
+ANY12(YUY2ToUV422Row_Any_LASX, YUY2ToUV422Row_LASX, 1, 4, 1, 31)
+ANY12(UYVYToUV422Row_Any_LASX, UYVYToUV422Row_LASX, 1, 4, 1, 31)
+#endif
 #undef ANY12
 
+// Any 2 16 bit planes with parameter to 1
+#define ANY12PT(NAMEANY, ANY_SIMD, T, BPP, MASK)                            \
+  void NAMEANY(const T* src_uv, T* dst_u, T* dst_v, int depth, int width) { \
+    SIMD_ALIGNED(T temp[16 * 4]);                                           \
+    memset(temp, 0, 16 * 4 * BPP); /* for msan */                           \
+    int r = width & MASK;                                                   \
+    int n = width & ~MASK;                                                  \
+    if (n > 0) {                                                            \
+      ANY_SIMD(src_uv, dst_u, dst_v, depth, n);                             \
+    }                                                                       \
+    memcpy(temp, src_uv + n * 2, r * BPP * 2);                              \
+    ANY_SIMD(temp, temp + 32, temp + 48, depth, MASK + 1);                  \
+    memcpy(dst_u + n, temp + 32, r * BPP);                                  \
+    memcpy(dst_v + n, temp + 48, r * BPP);                                  \
+  }
+
+#ifdef HAS_SPLITUVROW_16_AVX2
+ANY12PT(SplitUVRow_16_Any_AVX2, SplitUVRow_16_AVX2, uint16_t, 2, 15)
+#endif
+
+#ifdef HAS_SPLITUVROW_16_NEON
+ANY12PT(SplitUVRow_16_Any_NEON, SplitUVRow_16_NEON, uint16_t, 2, 7)
+#endif
+
+#undef ANY21CT
+
+// Any 1 to 3.  Outputs RGB planes.
+#define ANY13(NAMEANY, ANY_SIMD, BPP, MASK)                                \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_r, uint8_t* dst_g,     \
+               uint8_t* dst_b, int width) {                                \
+    SIMD_ALIGNED(uint8_t temp[16 * 6]);                                    \
+    memset(temp, 0, 16 * 3); /* for msan */                                \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(src_ptr, dst_r, dst_g, dst_b, n);                           \
+    }                                                                      \
+    memcpy(temp, src_ptr + n * BPP, r * BPP);                              \
+    ANY_SIMD(temp, temp + 16 * 3, temp + 16 * 4, temp + 16 * 5, MASK + 1); \
+    memcpy(dst_r + n, temp + 16 * 3, r);                                   \
+    memcpy(dst_g + n, temp + 16 * 4, r);                                   \
+    memcpy(dst_b + n, temp + 16 * 5, r);                                   \
+  }
+
+#ifdef HAS_SPLITRGBROW_SSSE3
+ANY13(SplitRGBRow_Any_SSSE3, SplitRGBRow_SSSE3, 3, 15)
+#endif
+#ifdef HAS_SPLITRGBROW_NEON
+ANY13(SplitRGBRow_Any_NEON, SplitRGBRow_NEON, 3, 15)
+#endif
+#ifdef HAS_SPLITXRGBROW_SSE2
+ANY13(SplitXRGBRow_Any_SSE2, SplitXRGBRow_SSE2, 4, 7)
+#endif
+#ifdef HAS_SPLITXRGBROW_SSSE3
+ANY13(SplitXRGBRow_Any_SSSE3, SplitXRGBRow_SSSE3, 4, 7)
+#endif
+#ifdef HAS_SPLITXRGBROW_AVX2
+ANY13(SplitXRGBRow_Any_AVX2, SplitXRGBRow_AVX2, 4, 15)
+#endif
+#ifdef HAS_SPLITXRGBROW_NEON
+ANY13(SplitXRGBRow_Any_NEON, SplitXRGBRow_NEON, 4, 15)
+#endif
+
+// Any 1 to 4.  Outputs ARGB planes.
+#define ANY14(NAMEANY, ANY_SIMD, BPP, MASK)                                    \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_r, uint8_t* dst_g,         \
+               uint8_t* dst_b, uint8_t* dst_a, int width) {                    \
+    SIMD_ALIGNED(uint8_t temp[16 * 8]);                                        \
+    memset(temp, 0, 16 * 4); /* for msan */                                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(src_ptr, dst_r, dst_g, dst_b, dst_a, n);                        \
+    }                                                                          \
+    memcpy(temp, src_ptr + n * BPP, r * BPP);                                  \
+    ANY_SIMD(temp, temp + 16 * 4, temp + 16 * 5, temp + 16 * 6, temp + 16 * 7, \
+             MASK + 1);                                                        \
+    memcpy(dst_r + n, temp + 16 * 4, r);                                       \
+    memcpy(dst_g + n, temp + 16 * 5, r);                                       \
+    memcpy(dst_b + n, temp + 16 * 6, r);                                       \
+    memcpy(dst_a + n, temp + 16 * 7, r);                                       \
+  }
+
+#ifdef HAS_SPLITARGBROW_SSE2
+ANY14(SplitARGBRow_Any_SSE2, SplitARGBRow_SSE2, 4, 7)
+#endif
+#ifdef HAS_SPLITARGBROW_SSSE3
+ANY14(SplitARGBRow_Any_SSSE3, SplitARGBRow_SSSE3, 4, 7)
+#endif
+#ifdef HAS_SPLITARGBROW_AVX2
+ANY14(SplitARGBRow_Any_AVX2, SplitARGBRow_AVX2, 4, 15)
+#endif
+#ifdef HAS_SPLITARGBROW_NEON
+ANY14(SplitARGBRow_Any_NEON, SplitARGBRow_NEON, 4, 15)
+#endif
+
 // Any 1 to 2 with source stride (2 rows of source).  Outputs UV planes.
 // 128 byte row allows for 32 avx ARGB pixels.
-#define ANY12S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                          \
-    void NAMEANY(const uint8* src_ptr, int src_stride_ptr,                     \
-                 uint8* dst_u, uint8* dst_v, int width) {                      \
-      SIMD_ALIGNED(uint8 temp[128 * 4]);                                       \
-      memset(temp, 0, 128 * 2);  /* for msan */                                \
-      int r = width & MASK;                                                    \
-      int n = width & ~MASK;                                                   \
-      if (n > 0) {                                                             \
-        ANY_SIMD(src_ptr, src_stride_ptr, dst_u, dst_v, n);                    \
-      }                                                                        \
-      memcpy(temp, src_ptr  + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);     \
-      memcpy(temp + 128, src_ptr  + src_stride_ptr + (n >> UVSHIFT) * BPP,     \
-             SS(r, UVSHIFT) * BPP);                                            \
-      if ((width & 1) && UVSHIFT == 0) {  /* repeat last pixel for subsample */\
-        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
-               temp + SS(r, UVSHIFT) * BPP - BPP, BPP);                        \
-        memcpy(temp + 128 + SS(r, UVSHIFT) * BPP,                              \
-               temp + 128 + SS(r, UVSHIFT) * BPP - BPP, BPP);                  \
-      }                                                                        \
-      ANY_SIMD(temp, 128, temp + 256, temp + 384, MASK + 1);                   \
-      memcpy(dst_u + (n >> 1), temp + 256, SS(r, 1));                          \
-      memcpy(dst_v + (n >> 1), temp + 384, SS(r, 1));                          \
-    }
+#define ANY12S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                        \
+  void NAMEANY(const uint8_t* src_ptr, int src_stride, uint8_t* dst_u,       \
+               uint8_t* dst_v, int width) {                                  \
+    SIMD_ALIGNED(uint8_t temp[128 * 4]);                                     \
+    memset(temp, 0, 128 * 2); /* for msan */                                 \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(src_ptr, src_stride, dst_u, dst_v, n);                        \
+    }                                                                        \
+    memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);      \
+    memcpy(temp + 128, src_ptr + src_stride + (n >> UVSHIFT) * BPP,          \
+           SS(r, UVSHIFT) * BPP);                                            \
+    if ((width & 1) && UVSHIFT == 0) { /* repeat last pixel for subsample */ \
+      memcpy(temp + SS(r, UVSHIFT) * BPP, temp + SS(r, UVSHIFT) * BPP - BPP, \
+             BPP);                                                           \
+      memcpy(temp + 128 + SS(r, UVSHIFT) * BPP,                              \
+             temp + 128 + SS(r, UVSHIFT) * BPP - BPP, BPP);                  \
+    }                                                                        \
+    ANY_SIMD(temp, 128, temp + 256, temp + 384, MASK + 1);                   \
+    memcpy(dst_u + (n >> 1), temp + 256, SS(r, 1));                          \
+    memcpy(dst_v + (n >> 1), temp + 384, SS(r, 1));                          \
+  }
 
 #ifdef HAS_ARGBTOUVROW_AVX2
 ANY12S(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, 0, 4, 31)
 #endif
+#ifdef HAS_ABGRTOUVROW_AVX2
+ANY12S(ABGRToUVRow_Any_AVX2, ABGRToUVRow_AVX2, 0, 4, 31)
+#endif
 #ifdef HAS_ARGBTOUVJROW_AVX2
 ANY12S(ARGBToUVJRow_Any_AVX2, ARGBToUVJRow_AVX2, 0, 4, 31)
 #endif
+#ifdef HAS_ABGRTOUVJROW_AVX2
+ANY12S(ABGRToUVJRow_Any_AVX2, ABGRToUVJRow_AVX2, 0, 4, 31)
+#endif
+#ifdef HAS_ARGBTOUVJROW_SSSE3
+ANY12S(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, 0, 4, 15)
+#endif
+#ifdef HAS_ABGRTOUVJROW_SSSE3
+ANY12S(ABGRToUVJRow_Any_SSSE3, ABGRToUVJRow_SSSE3, 0, 4, 15)
+#endif
 #ifdef HAS_ARGBTOUVROW_SSSE3
 ANY12S(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, 0, 4, 15)
-ANY12S(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, 0, 4, 15)
 ANY12S(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, 0, 4, 15)
 ANY12S(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, 0, 4, 15)
 ANY12S(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, 0, 4, 15)
@@ -783,30 +2125,108 @@ ANY12S(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, 1, 4, 15)
 #ifdef HAS_ARGBTOUVROW_NEON
 ANY12S(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, 0, 4, 15)
 #endif
+#ifdef HAS_ARGBTOUVROW_MSA
+ANY12S(ARGBToUVRow_Any_MSA, ARGBToUVRow_MSA, 0, 4, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_LASX
+ANY12S(ARGBToUVRow_Any_LASX, ARGBToUVRow_LASX, 0, 4, 31)
+#endif
 #ifdef HAS_ARGBTOUVJROW_NEON
 ANY12S(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, 0, 4, 15)
 #endif
+#ifdef HAS_ABGRTOUVJROW_NEON
+ANY12S(ABGRToUVJRow_Any_NEON, ABGRToUVJRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVJROW_MSA
+ANY12S(ARGBToUVJRow_Any_MSA, ARGBToUVJRow_MSA, 0, 4, 31)
+#endif
+#ifdef HAS_ARGBTOUVJROW_LSX
+ANY12S(ARGBToUVJRow_Any_LSX, ARGBToUVJRow_LSX, 0, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVJROW_LASX
+ANY12S(ARGBToUVJRow_Any_LASX, ARGBToUVJRow_LASX, 0, 4, 31)
+#endif
 #ifdef HAS_BGRATOUVROW_NEON
 ANY12S(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, 0, 4, 15)
 #endif
+#ifdef HAS_BGRATOUVROW_MSA
+ANY12S(BGRAToUVRow_Any_MSA, BGRAToUVRow_MSA, 0, 4, 15)
+#endif
+#ifdef HAS_BGRATOUVROW_LSX
+ANY12S(BGRAToUVRow_Any_LSX, BGRAToUVRow_LSX, 0, 4, 15)
+#endif
 #ifdef HAS_ABGRTOUVROW_NEON
 ANY12S(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, 0, 4, 15)
 #endif
+#ifdef HAS_ABGRTOUVROW_MSA
+ANY12S(ABGRToUVRow_Any_MSA, ABGRToUVRow_MSA, 0, 4, 15)
+#endif
+#ifdef HAS_ABGRTOUVROW_LSX
+ANY12S(ABGRToUVRow_Any_LSX, ABGRToUVRow_LSX, 0, 4, 15)
+#endif
 #ifdef HAS_RGBATOUVROW_NEON
 ANY12S(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, 0, 4, 15)
 #endif
+#ifdef HAS_RGBATOUVROW_MSA
+ANY12S(RGBAToUVRow_Any_MSA, RGBAToUVRow_MSA, 0, 4, 15)
+#endif
+#ifdef HAS_RGBATOUVROW_LSX
+ANY12S(RGBAToUVRow_Any_LSX, RGBAToUVRow_LSX, 0, 4, 15)
+#endif
 #ifdef HAS_RGB24TOUVROW_NEON
 ANY12S(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, 0, 3, 15)
 #endif
+#ifdef HAS_RGB24TOUVJROW_NEON
+ANY12S(RGB24ToUVJRow_Any_NEON, RGB24ToUVJRow_NEON, 0, 3, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_MSA
+ANY12S(RGB24ToUVRow_Any_MSA, RGB24ToUVRow_MSA, 0, 3, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_LSX
+ANY12S(RGB24ToUVRow_Any_LSX, RGB24ToUVRow_LSX, 0, 3, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_LASX
+ANY12S(RGB24ToUVRow_Any_LASX, RGB24ToUVRow_LASX, 0, 3, 31)
+#endif
 #ifdef HAS_RAWTOUVROW_NEON
 ANY12S(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, 0, 3, 15)
 #endif
+#ifdef HAS_RAWTOUVJROW_NEON
+ANY12S(RAWToUVJRow_Any_NEON, RAWToUVJRow_NEON, 0, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_MSA
+ANY12S(RAWToUVRow_Any_MSA, RAWToUVRow_MSA, 0, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_LSX
+ANY12S(RAWToUVRow_Any_LSX, RAWToUVRow_LSX, 0, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_LASX
+ANY12S(RAWToUVRow_Any_LASX, RAWToUVRow_LASX, 0, 3, 31)
+#endif
 #ifdef HAS_RGB565TOUVROW_NEON
 ANY12S(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, 0, 2, 15)
 #endif
+#ifdef HAS_RGB565TOUVROW_MSA
+ANY12S(RGB565ToUVRow_Any_MSA, RGB565ToUVRow_MSA, 0, 2, 15)
+#endif
+#ifdef HAS_RGB565TOUVROW_LSX
+ANY12S(RGB565ToUVRow_Any_LSX, RGB565ToUVRow_LSX, 0, 2, 15)
+#endif
+#ifdef HAS_RGB565TOUVROW_LASX
+ANY12S(RGB565ToUVRow_Any_LASX, RGB565ToUVRow_LASX, 0, 2, 31)
+#endif
 #ifdef HAS_ARGB1555TOUVROW_NEON
 ANY12S(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, 0, 2, 15)
 #endif
+#ifdef HAS_ARGB1555TOUVROW_MSA
+ANY12S(ARGB1555ToUVRow_Any_MSA, ARGB1555ToUVRow_MSA, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB1555TOUVROW_LSX
+ANY12S(ARGB1555ToUVRow_Any_LSX, ARGB1555ToUVRow_LSX, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB1555TOUVROW_LASX
+ANY12S(ARGB1555ToUVRow_Any_LASX, ARGB1555ToUVRow_LASX, 0, 2, 31)
+#endif
 #ifdef HAS_ARGB4444TOUVROW_NEON
 ANY12S(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, 0, 2, 15)
 #endif
@@ -816,8 +2236,131 @@ ANY12S(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, 1, 4, 15)
 #ifdef HAS_UYVYTOUVROW_NEON
 ANY12S(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, 1, 4, 15)
 #endif
+#ifdef HAS_YUY2TOUVROW_MSA
+ANY12S(YUY2ToUVRow_Any_MSA, YUY2ToUVRow_MSA, 1, 4, 31)
+#endif
+#ifdef HAS_YUY2TOUVROW_LASX
+ANY12S(YUY2ToUVRow_Any_LASX, YUY2ToUVRow_LASX, 1, 4, 31)
+#endif
+#ifdef HAS_UYVYTOUVROW_MSA
+ANY12S(UYVYToUVRow_Any_MSA, UYVYToUVRow_MSA, 1, 4, 31)
+#endif
+#ifdef HAS_UYVYTOUVROW_LASX
+ANY12S(UYVYToUVRow_Any_LASX, UYVYToUVRow_LASX, 1, 4, 31)
+#endif
 #undef ANY12S
 
+// Any 1 to 1 with source stride (2 rows of source).  Outputs UV plane.
+// 128 byte row allows for 32 avx ARGB pixels.
+#define ANY11S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                        \
+  void NAMEANY(const uint8_t* src_ptr, int src_stride, uint8_t* dst_vu,      \
+               int width) {                                                  \
+    SIMD_ALIGNED(uint8_t temp[128 * 3]);                                     \
+    memset(temp, 0, 128 * 2); /* for msan */                                 \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(src_ptr, src_stride, dst_vu, n);                              \
+    }                                                                        \
+    memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);      \
+    memcpy(temp + 128, src_ptr + src_stride + (n >> UVSHIFT) * BPP,          \
+           SS(r, UVSHIFT) * BPP);                                            \
+    if ((width & 1) && UVSHIFT == 0) { /* repeat last pixel for subsample */ \
+      memcpy(temp + SS(r, UVSHIFT) * BPP, temp + SS(r, UVSHIFT) * BPP - BPP, \
+             BPP);                                                           \
+      memcpy(temp + 128 + SS(r, UVSHIFT) * BPP,                              \
+             temp + 128 + SS(r, UVSHIFT) * BPP - BPP, BPP);                  \
+    }                                                                        \
+    ANY_SIMD(temp, 128, temp + 256, MASK + 1);                               \
+    memcpy(dst_vu + (n >> 1) * 2, temp + 256, SS(r, 1) * 2);                 \
+  }
+
+#ifdef HAS_AYUVTOVUROW_NEON
+ANY11S(AYUVToUVRow_Any_NEON, AYUVToUVRow_NEON, 0, 4, 15)
+ANY11S(AYUVToVURow_Any_NEON, AYUVToVURow_NEON, 0, 4, 15)
+#endif
+#undef ANY11S
+
+#define ANYDETILE(NAMEANY, ANY_SIMD, T, BPP, MASK)                           \
+  void NAMEANY(const T* src, ptrdiff_t src_tile_stride, T* dst, int width) { \
+    SIMD_ALIGNED(T temp[16 * 2]);                                            \
+    memset(temp, 0, 16 * BPP); /* for msan */                                \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(src, src_tile_stride, dst, n);                                \
+    }                                                                        \
+    memcpy(temp, src + (n / 16) * src_tile_stride, r * BPP);                 \
+    ANY_SIMD(temp, src_tile_stride, temp + 16, MASK + 1);                    \
+    memcpy(dst + n, temp + 16, r * BPP);                                     \
+  }
+
+#ifdef HAS_DETILEROW_NEON
+ANYDETILE(DetileRow_Any_NEON, DetileRow_NEON, uint8_t, 1, 15)
+#endif
+#ifdef HAS_DETILEROW_SSE2
+ANYDETILE(DetileRow_Any_SSE2, DetileRow_SSE2, uint8_t, 1, 15)
+#endif
+#ifdef HAS_DETILEROW_16_NEON
+ANYDETILE(DetileRow_16_Any_NEON, DetileRow_16_NEON, uint16_t, 2, 15)
+#endif
+#ifdef HAS_DETILEROW_16_SSE2
+ANYDETILE(DetileRow_16_Any_SSE2, DetileRow_16_SSE2, uint16_t, 2, 15)
+#endif
+#ifdef HAS_DETILEROW_16_AVX
+ANYDETILE(DetileRow_16_Any_AVX, DetileRow_16_AVX, uint16_t, 2, 15)
+#endif
+
+#define ANYDETILESPLITUV(NAMEANY, ANY_SIMD, MASK)                \
+  void NAMEANY(const uint8_t* src_uv, ptrdiff_t src_tile_stride, \
+               uint8_t* dst_u, uint8_t* dst_v, int width) {      \
+    SIMD_ALIGNED(uint8_t temp[16 * 2]);                          \
+    memset(temp, 0, 16 * 2); /* for msan */                      \
+    int r = width & MASK;                                        \
+    int n = width & ~MASK;                                       \
+    if (n > 0) {                                                 \
+      ANY_SIMD(src_uv, src_tile_stride, dst_u, dst_v, n);        \
+    }                                                            \
+    memcpy(temp, src_uv + (n / 16) * src_tile_stride, r);        \
+    ANY_SIMD(temp, src_tile_stride, temp + 16, temp + 24, r);    \
+    memcpy(dst_u + n / 2, temp + 16, (r + 1) / 2);               \
+    memcpy(dst_v + n / 2, temp + 24, (r + 1) / 2);               \
+  }
+
+#ifdef HAS_DETILESPLITUVROW_NEON
+ANYDETILESPLITUV(DetileSplitUVRow_Any_NEON, DetileSplitUVRow_NEON, 15)
+#endif
+#ifdef HAS_DETILESPLITUVROW_SSSE3
+ANYDETILESPLITUV(DetileSplitUVRow_Any_SSSE3, DetileSplitUVRow_SSSE3, 15)
+#endif
+
+#define ANYDETILEMERGE(NAMEANY, ANY_SIMD, MASK)                                \
+  void NAMEANY(const uint8_t* src_y, ptrdiff_t src_y_tile_stride,              \
+               const uint8_t* src_uv, ptrdiff_t src_uv_tile_stride,            \
+               uint8_t* dst_yuy2, int width) {                                 \
+    SIMD_ALIGNED(uint8_t temp[16 * 4]);                                        \
+    memset(temp, 0, 16 * 4); /* for msan */                                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(src_y, src_y_tile_stride, src_uv, src_uv_tile_stride, dst_yuy2, \
+               n);                                                             \
+    }                                                                          \
+    memcpy(temp, src_y + (n / 16) * src_y_tile_stride, r);                     \
+    memcpy(temp + 16, src_uv + (n / 16) * src_uv_tile_stride, r);              \
+    ANY_SIMD(temp, src_y_tile_stride, temp + 16, src_uv_tile_stride,           \
+             temp + 32, r);                                                    \
+    memcpy(dst_yuy2 + 2 * n, temp + 32, 2 * r);                                \
+  }
+
+#ifdef HAS_DETILETOYUY2_NEON
+ANYDETILEMERGE(DetileToYUY2_Any_NEON, DetileToYUY2_NEON, 15)
+#endif
+
+#ifdef HAS_DETILETOYUY2_SSE2
+ANYDETILEMERGE(DetileToYUY2_Any_SSE2, DetileToYUY2_SSE2, 15)
+#endif
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/media/libyuv/source/row_common.cc b/media/libyuv/source/row_common.cc
index cdbf717162..84afd35ba4 100644
--- a/media/libyuv/source/row_common.cc
+++ b/media/libyuv/source/row_common.cc
@@ -10,72 +10,126 @@
 
 #include "libyuv/row.h"
 
+#include <assert.h>
 #include <string.h>  // For memcpy and memset.
 
 #include "libyuv/basic_types.h"
+#include "libyuv/convert_argb.h"  // For kYuvI601Constants
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
+#ifdef __cplusplus
+#define STATIC_CAST(type, expr) static_cast<type>(expr)
+#else
+#define STATIC_CAST(type, expr) (type)(expr)
+#endif
+
+// This macro controls YUV to RGB using unsigned math to extend range of
+// YUV to RGB coefficients to 0 to 4 instead of 0 to 2 for more accuracy on B:
+// LIBYUV_UNLIMITED_DATA
+
+// Macros to enable unlimited data for each colorspace
+// LIBYUV_UNLIMITED_BT601
+// LIBYUV_UNLIMITED_BT709
+// LIBYUV_UNLIMITED_BT2020
+
+// The following macro from row_win makes the C code match the row_win code,
+// which is 7 bit fixed point for ARGBToI420:
+#if !defined(LIBYUV_BIT_EXACT) && !defined(LIBYUV_DISABLE_X86) && \
+    defined(_MSC_VER) && !defined(__clang__) &&                   \
+    (defined(_M_IX86) || defined(_M_X64))
+#define LIBYUV_RGB7 1
+#endif
+
+#if !defined(LIBYUV_BIT_EXACT) && (defined(__x86_64__) || defined(_M_X64) || \
+                                   defined(__i386__) || defined(_M_IX86))
+#define LIBYUV_ARGBTOUV_PAVGB 1
+#define LIBYUV_RGBTOU_TRUNCATE 1
+#define LIBYUV_ATTENUATE_DUP 1
+#endif
+#if defined(LIBYUV_BIT_EXACT)
+#define LIBYUV_UNATTENUATE_DUP 1
+#endif
+
 // llvm x86 is poor at ternary operator, so use branchless min/max.
 
 #define USE_BRANCHLESS 1
 #if USE_BRANCHLESS
-static __inline int32 clamp0(int32 v) {
-  return ((-(v) >> 31) & (v));
+static __inline int32_t clamp0(int32_t v) {
+  return -(v >= 0) & v;
+}
+// TODO(fbarchard): make clamp255 preserve negative values.
+static __inline int32_t clamp255(int32_t v) {
+  return (-(v >= 255) | v) & 255;
 }
 
-static __inline int32 clamp255(int32 v) {
-  return (((255 - (v)) >> 31) | (v)) & 255;
+static __inline int32_t clamp1023(int32_t v) {
+  return (-(v >= 1023) | v) & 1023;
 }
 
-static __inline uint32 Clamp(int32 val) {
-  int v = clamp0(val);
-  return (uint32)(clamp255(v));
+// clamp to max
+static __inline int32_t ClampMax(int32_t v, int32_t max) {
+  return (-(v >= max) | v) & max;
 }
 
-static __inline uint32 Abs(int32 v) {
-  int m = v >> 31;
+static __inline uint32_t Abs(int32_t v) {
+  int m = -(v < 0);
   return (v + m) ^ m;
 }
-#else  // USE_BRANCHLESS
-static __inline int32 clamp0(int32 v) {
+#else   // USE_BRANCHLESS
+static __inline int32_t clamp0(int32_t v) {
   return (v < 0) ? 0 : v;
 }
 
-static __inline int32 clamp255(int32 v) {
+static __inline int32_t clamp255(int32_t v) {
   return (v > 255) ? 255 : v;
 }
 
-static __inline uint32 Clamp(int32 val) {
-  int v = clamp0(val);
-  return (uint32)(clamp255(v));
+static __inline int32_t clamp1023(int32_t v) {
+  return (v > 1023) ? 1023 : v;
 }
 
-static __inline uint32 Abs(int32 v) {
+static __inline int32_t ClampMax(int32_t v, int32_t max) {
+  return (v > max) ? max : v;
+}
+
+static __inline uint32_t Abs(int32_t v) {
   return (v < 0) ? -v : v;
 }
 #endif  // USE_BRANCHLESS
+static __inline uint32_t Clamp(int32_t val) {
+  int v = clamp0(val);
+  return (uint32_t)(clamp255(v));
+}
+
+static __inline uint32_t Clamp10(int32_t val) {
+  int v = clamp0(val);
+  return (uint32_t)(clamp1023(v));
+}
 
-#ifdef LIBYUV_LITTLE_ENDIAN
-#define WRITEWORD(p, v) *(uint32*)(p) = v
+// Little Endian
+#if defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \
+    defined(_M_IX86) || defined(__arm__) || defined(_M_ARM) ||     \
+    (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#define WRITEWORD(p, v) *(uint32_t*)(p) = v
 #else
-static inline void WRITEWORD(uint8* p, uint32 v) {
-  p[0] = (uint8)(v & 255);
-  p[1] = (uint8)((v >> 8) & 255);
-  p[2] = (uint8)((v >> 16) & 255);
-  p[3] = (uint8)((v >> 24) & 255);
+static inline void WRITEWORD(uint8_t* p, uint32_t v) {
+  p[0] = (uint8_t)(v & 255);
+  p[1] = (uint8_t)((v >> 8) & 255);
+  p[2] = (uint8_t)((v >> 16) & 255);
+  p[3] = (uint8_t)((v >> 24) & 255);
 }
 #endif
 
-void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
+void RGB24ToARGBRow_C(const uint8_t* src_rgb24, uint8_t* dst_argb, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_rgb24[0];
-    uint8 g = src_rgb24[1];
-    uint8 r = src_rgb24[2];
+    uint8_t b = src_rgb24[0];
+    uint8_t g = src_rgb24[1];
+    uint8_t r = src_rgb24[2];
     dst_argb[0] = b;
     dst_argb[1] = g;
     dst_argb[2] = r;
@@ -85,12 +139,12 @@ void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
   }
 }
 
-void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
+void RAWToARGBRow_C(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 r = src_raw[0];
-    uint8 g = src_raw[1];
-    uint8 b = src_raw[2];
+    uint8_t r = src_raw[0];
+    uint8_t g = src_raw[1];
+    uint8_t b = src_raw[2];
     dst_argb[0] = b;
     dst_argb[1] = g;
     dst_argb[2] = r;
@@ -100,12 +154,27 @@ void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
   }
 }
 
-void RAWToRGB24Row_C(const uint8* src_raw, uint8* dst_rgb24, int width) {
+void RAWToRGBARow_C(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8_t r = src_raw[0];
+    uint8_t g = src_raw[1];
+    uint8_t b = src_raw[2];
+    dst_rgba[0] = 255u;
+    dst_rgba[1] = b;
+    dst_rgba[2] = g;
+    dst_rgba[3] = r;
+    dst_rgba += 4;
+    src_raw += 3;
+  }
+}
+
+void RAWToRGB24Row_C(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 r = src_raw[0];
-    uint8 g = src_raw[1];
-    uint8 b = src_raw[2];
+    uint8_t r = src_raw[0];
+    uint8_t g = src_raw[1];
+    uint8_t b = src_raw[2];
     dst_rgb24[0] = b;
     dst_rgb24[1] = g;
     dst_rgb24[2] = r;
@@ -114,61 +183,111 @@ void RAWToRGB24Row_C(const uint8* src_raw, uint8* dst_rgb24, int width) {
   }
 }
 
-void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) {
+void RGB565ToARGBRow_C(const uint8_t* src_rgb565,
+                       uint8_t* dst_argb,
+                       int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_rgb565[0] & 0x1f;
-    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
-    uint8 r = src_rgb565[1] >> 3;
-    dst_argb[0] = (b << 3) | (b >> 2);
-    dst_argb[1] = (g << 2) | (g >> 4);
-    dst_argb[2] = (r << 3) | (r >> 2);
+    uint8_t b = STATIC_CAST(uint8_t, src_rgb565[0] & 0x1f);
+    uint8_t g = STATIC_CAST(
+        uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3));
+    uint8_t r = STATIC_CAST(uint8_t, src_rgb565[1] >> 3);
+    dst_argb[0] = STATIC_CAST(uint8_t, (b << 3) | (b >> 2));
+    dst_argb[1] = STATIC_CAST(uint8_t, (g << 2) | (g >> 4));
+    dst_argb[2] = STATIC_CAST(uint8_t, (r << 3) | (r >> 2));
     dst_argb[3] = 255u;
     dst_argb += 4;
     src_rgb565 += 2;
   }
 }
 
-void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb,
+void ARGB1555ToARGBRow_C(const uint8_t* src_argb1555,
+                         uint8_t* dst_argb,
                          int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_argb1555[0] & 0x1f;
-    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
-    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
-    uint8 a = src_argb1555[1] >> 7;
-    dst_argb[0] = (b << 3) | (b >> 2);
-    dst_argb[1] = (g << 3) | (g >> 2);
-    dst_argb[2] = (r << 3) | (r >> 2);
+    uint8_t b = STATIC_CAST(uint8_t, src_argb1555[0] & 0x1f);
+    uint8_t g = STATIC_CAST(
+        uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3));
+    uint8_t r = STATIC_CAST(uint8_t, (src_argb1555[1] & 0x7c) >> 2);
+    uint8_t a = STATIC_CAST(uint8_t, src_argb1555[1] >> 7);
+    dst_argb[0] = STATIC_CAST(uint8_t, (b << 3) | (b >> 2));
+    dst_argb[1] = STATIC_CAST(uint8_t, (g << 3) | (g >> 2));
+    dst_argb[2] = STATIC_CAST(uint8_t, (r << 3) | (r >> 2));
     dst_argb[3] = -a;
     dst_argb += 4;
     src_argb1555 += 2;
   }
 }
 
-void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb,
+void ARGB4444ToARGBRow_C(const uint8_t* src_argb4444,
+                         uint8_t* dst_argb,
                          int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_argb4444[0] & 0x0f;
-    uint8 g = src_argb4444[0] >> 4;
-    uint8 r = src_argb4444[1] & 0x0f;
-    uint8 a = src_argb4444[1] >> 4;
-    dst_argb[0] = (b << 4) | b;
-    dst_argb[1] = (g << 4) | g;
-    dst_argb[2] = (r << 4) | r;
-    dst_argb[3] = (a << 4) | a;
+    uint8_t b = STATIC_CAST(uint8_t, src_argb4444[0] & 0x0f);
+    uint8_t g = STATIC_CAST(uint8_t, src_argb4444[0] >> 4);
+    uint8_t r = STATIC_CAST(uint8_t, src_argb4444[1] & 0x0f);
+    uint8_t a = STATIC_CAST(uint8_t, src_argb4444[1] >> 4);
+    dst_argb[0] = STATIC_CAST(uint8_t, (b << 4) | b);
+    dst_argb[1] = STATIC_CAST(uint8_t, (g << 4) | g);
+    dst_argb[2] = STATIC_CAST(uint8_t, (r << 4) | r);
+    dst_argb[3] = STATIC_CAST(uint8_t, (a << 4) | a);
     dst_argb += 4;
     src_argb4444 += 2;
   }
 }
 
-void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+void AR30ToARGBRow_C(const uint8_t* src_ar30, uint8_t* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint32_t ar30;
+    memcpy(&ar30, src_ar30, sizeof ar30);
+    uint32_t b = (ar30 >> 2) & 0xff;
+    uint32_t g = (ar30 >> 12) & 0xff;
+    uint32_t r = (ar30 >> 22) & 0xff;
+    uint32_t a = (ar30 >> 30) * 0x55;  // Replicate 2 bits to 8 bits.
+    *(uint32_t*)(dst_argb) = b | (g << 8) | (r << 16) | (a << 24);
+    dst_argb += 4;
+    src_ar30 += 4;
+  }
+}
+
+void AR30ToABGRRow_C(const uint8_t* src_ar30, uint8_t* dst_abgr, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint32_t ar30;
+    memcpy(&ar30, src_ar30, sizeof ar30);
+    uint32_t b = (ar30 >> 2) & 0xff;
+    uint32_t g = (ar30 >> 12) & 0xff;
+    uint32_t r = (ar30 >> 22) & 0xff;
+    uint32_t a = (ar30 >> 30) * 0x55;  // Replicate 2 bits to 8 bits.
+    *(uint32_t*)(dst_abgr) = r | (g << 8) | (b << 16) | (a << 24);
+    dst_abgr += 4;
+    src_ar30 += 4;
+  }
+}
+
+void AR30ToAB30Row_C(const uint8_t* src_ar30, uint8_t* dst_ab30, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_argb[0];
-    uint8 g = src_argb[1];
-    uint8 r = src_argb[2];
+    uint32_t ar30;
+    memcpy(&ar30, src_ar30, sizeof ar30);
+    uint32_t b = ar30 & 0x3ff;
+    uint32_t ga = ar30 & 0xc00ffc00;
+    uint32_t r = (ar30 >> 20) & 0x3ff;
+    *(uint32_t*)(dst_ab30) = r | ga | (b << 20);
+    dst_ab30 += 4;
+    src_ar30 += 4;
+  }
+}
+
+void ARGBToRGB24Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8_t b = src_argb[0];
+    uint8_t g = src_argb[1];
+    uint8_t r = src_argb[2];
     dst_rgb[0] = b;
     dst_rgb[1] = g;
     dst_rgb[2] = r;
@@ -177,12 +296,12 @@ void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
-void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+void ARGBToRAWRow_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_argb[0];
-    uint8 g = src_argb[1];
-    uint8 r = src_argb[2];
+    uint8_t b = src_argb[0];
+    uint8_t g = src_argb[1];
+    uint8_t r = src_argb[2];
     dst_rgb[0] = r;
     dst_rgb[1] = g;
     dst_rgb[2] = b;
@@ -191,25 +310,25 @@ void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
-void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+void ARGBToRGB565Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 b0 = src_argb[0] >> 3;
-    uint8 g0 = src_argb[1] >> 2;
-    uint8 r0 = src_argb[2] >> 3;
-    uint8 b1 = src_argb[4] >> 3;
-    uint8 g1 = src_argb[5] >> 2;
-    uint8 r1 = src_argb[6] >> 3;
-    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
-              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    uint8_t b0 = src_argb[0] >> 3;
+    uint8_t g0 = src_argb[1] >> 2;
+    uint8_t r0 = src_argb[2] >> 3;
+    uint8_t b1 = src_argb[4] >> 3;
+    uint8_t g1 = src_argb[5] >> 2;
+    uint8_t r1 = src_argb[6] >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) | (b1 << 16) | (g1 << 21) |
+                           (r1 << 27));
     dst_rgb += 4;
     src_argb += 8;
   }
   if (width & 1) {
-    uint8 b0 = src_argb[0] >> 3;
-    uint8 g0 = src_argb[1] >> 2;
-    uint8 r0 = src_argb[2] >> 3;
-    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+    uint8_t b0 = src_argb[0] >> 3;
+    uint8_t g0 = src_argb[1] >> 2;
+    uint8_t r0 = src_argb[2] >> 3;
+    *(uint16_t*)(dst_rgb) = STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11));
   }
 }
 
@@ -221,133 +340,341 @@ void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
 // endian will not affect order of the original matrix.  But the dither4
 // will containing the first pixel in the lower byte for little endian
 // or the upper byte for big endian.
-void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
-                             const uint32 dither4, int width) {
+void ARGBToRGB565DitherRow_C(const uint8_t* src_argb,
+                             uint8_t* dst_rgb,
+                             const uint32_t dither4,
+                             int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
     int dither0 = ((const unsigned char*)(&dither4))[x & 3];
     int dither1 = ((const unsigned char*)(&dither4))[(x + 1) & 3];
-    uint8 b0 = clamp255(src_argb[0] + dither0) >> 3;
-    uint8 g0 = clamp255(src_argb[1] + dither0) >> 2;
-    uint8 r0 = clamp255(src_argb[2] + dither0) >> 3;
-    uint8 b1 = clamp255(src_argb[4] + dither1) >> 3;
-    uint8 g1 = clamp255(src_argb[5] + dither1) >> 2;
-    uint8 r1 = clamp255(src_argb[6] + dither1) >> 3;
-    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
-              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    uint8_t b0 = STATIC_CAST(uint8_t, clamp255(src_argb[0] + dither0) >> 3);
+    uint8_t g0 = STATIC_CAST(uint8_t, clamp255(src_argb[1] + dither0) >> 2);
+    uint8_t r0 = STATIC_CAST(uint8_t, clamp255(src_argb[2] + dither0) >> 3);
+    uint8_t b1 = STATIC_CAST(uint8_t, clamp255(src_argb[4] + dither1) >> 3);
+    uint8_t g1 = STATIC_CAST(uint8_t, clamp255(src_argb[5] + dither1) >> 2);
+    uint8_t r1 = STATIC_CAST(uint8_t, clamp255(src_argb[6] + dither1) >> 3);
+    *(uint16_t*)(dst_rgb + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11));
+    *(uint16_t*)(dst_rgb + 2) =
+        STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 11));
     dst_rgb += 4;
     src_argb += 8;
   }
   if (width & 1) {
     int dither0 = ((const unsigned char*)(&dither4))[(width - 1) & 3];
-    uint8 b0 = clamp255(src_argb[0] + dither0) >> 3;
-    uint8 g0 = clamp255(src_argb[1] + dither0) >> 2;
-    uint8 r0 = clamp255(src_argb[2] + dither0) >> 3;
-    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+    uint8_t b0 = STATIC_CAST(uint8_t, clamp255(src_argb[0] + dither0) >> 3);
+    uint8_t g0 = STATIC_CAST(uint8_t, clamp255(src_argb[1] + dither0) >> 2);
+    uint8_t r0 = STATIC_CAST(uint8_t, clamp255(src_argb[2] + dither0) >> 3);
+    *(uint16_t*)(dst_rgb) = STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11));
   }
 }
 
-void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+void ARGBToARGB1555Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 b0 = src_argb[0] >> 3;
-    uint8 g0 = src_argb[1] >> 3;
-    uint8 r0 = src_argb[2] >> 3;
-    uint8 a0 = src_argb[3] >> 7;
-    uint8 b1 = src_argb[4] >> 3;
-    uint8 g1 = src_argb[5] >> 3;
-    uint8 r1 = src_argb[6] >> 3;
-    uint8 a1 = src_argb[7] >> 7;
-    *(uint32*)(dst_rgb) =
-        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
-        (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    uint8_t b0 = src_argb[0] >> 3;
+    uint8_t g0 = src_argb[1] >> 3;
+    uint8_t r0 = src_argb[2] >> 3;
+    uint8_t a0 = src_argb[3] >> 7;
+    uint8_t b1 = src_argb[4] >> 3;
+    uint8_t g1 = src_argb[5] >> 3;
+    uint8_t r1 = src_argb[6] >> 3;
+    uint8_t a1 = src_argb[7] >> 7;
+    *(uint16_t*)(dst_rgb + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | (a0 << 15));
+    *(uint16_t*)(dst_rgb + 2) =
+        STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 10) | (a1 << 15));
     dst_rgb += 4;
     src_argb += 8;
   }
   if (width & 1) {
-    uint8 b0 = src_argb[0] >> 3;
-    uint8 g0 = src_argb[1] >> 3;
-    uint8 r0 = src_argb[2] >> 3;
-    uint8 a0 = src_argb[3] >> 7;
-    *(uint16*)(dst_rgb) =
-        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+    uint8_t b0 = src_argb[0] >> 3;
+    uint8_t g0 = src_argb[1] >> 3;
+    uint8_t r0 = src_argb[2] >> 3;
+    uint8_t a0 = src_argb[3] >> 7;
+    *(uint16_t*)(dst_rgb) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | (a0 << 15));
   }
 }
 
-void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+void ARGBToARGB4444Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 b0 = src_argb[0] >> 4;
-    uint8 g0 = src_argb[1] >> 4;
-    uint8 r0 = src_argb[2] >> 4;
-    uint8 a0 = src_argb[3] >> 4;
-    uint8 b1 = src_argb[4] >> 4;
-    uint8 g1 = src_argb[5] >> 4;
-    uint8 r1 = src_argb[6] >> 4;
-    uint8 a1 = src_argb[7] >> 4;
-    *(uint32*)(dst_rgb) =
-        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
-        (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    uint8_t b0 = src_argb[0] >> 4;
+    uint8_t g0 = src_argb[1] >> 4;
+    uint8_t r0 = src_argb[2] >> 4;
+    uint8_t a0 = src_argb[3] >> 4;
+    uint8_t b1 = src_argb[4] >> 4;
+    uint8_t g1 = src_argb[5] >> 4;
+    uint8_t r1 = src_argb[6] >> 4;
+    uint8_t a1 = src_argb[7] >> 4;
+    *(uint16_t*)(dst_rgb + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | (a0 << 12));
+    *(uint16_t*)(dst_rgb + 2) =
+        STATIC_CAST(uint16_t, b1 | (g1 << 4) | (r1 << 8) | (a1 << 12));
     dst_rgb += 4;
     src_argb += 8;
   }
   if (width & 1) {
-    uint8 b0 = src_argb[0] >> 4;
-    uint8 g0 = src_argb[1] >> 4;
-    uint8 r0 = src_argb[2] >> 4;
-    uint8 a0 = src_argb[3] >> 4;
-    *(uint16*)(dst_rgb) =
-        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
-  }
-}
-
-static __inline int RGBToY(uint8 r, uint8 g, uint8 b) {
-  return (66 * r + 129 * g +  25 * b + 0x1080) >> 8;
-}
-
-static __inline int RGBToU(uint8 r, uint8 g, uint8 b) {
-  return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
-}
-static __inline int RGBToV(uint8 r, uint8 g, uint8 b) {
-  return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
-}
-
-#define MAKEROWY(NAME, R, G, B, BPP) \
-void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {       \
-  int x;                                                                       \
-  for (x = 0; x < width; ++x) {                                                \
-    dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);               \
-    src_argb0 += BPP;                                                          \
-    dst_y += 1;                                                                \
-  }                                                                            \
-}                                                                              \
-void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb,              \
-                       uint8* dst_u, uint8* dst_v, int width) {                \
-  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
-  int x;                                                                       \
-  for (x = 0; x < width - 1; x += 2) {                                         \
-    uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] +                              \
-               src_rgb1[B] + src_rgb1[B + BPP]) >> 2;                          \
-    uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] +                              \
-               src_rgb1[G] + src_rgb1[G + BPP]) >> 2;                          \
-    uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] +                              \
-               src_rgb1[R] + src_rgb1[R + BPP]) >> 2;                          \
-    dst_u[0] = RGBToU(ar, ag, ab);                                             \
-    dst_v[0] = RGBToV(ar, ag, ab);                                             \
-    src_rgb0 += BPP * 2;                                                       \
-    src_rgb1 += BPP * 2;                                                       \
-    dst_u += 1;                                                                \
-    dst_v += 1;                                                                \
-  }                                                                            \
-  if (width & 1) {                                                             \
-    uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                               \
-    uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                               \
-    uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                               \
-    dst_u[0] = RGBToU(ar, ag, ab);                                             \
-    dst_v[0] = RGBToV(ar, ag, ab);                                             \
-  }                                                                            \
+    uint8_t b0 = src_argb[0] >> 4;
+    uint8_t g0 = src_argb[1] >> 4;
+    uint8_t r0 = src_argb[2] >> 4;
+    uint8_t a0 = src_argb[3] >> 4;
+    *(uint16_t*)(dst_rgb) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | (a0 << 12));
+  }
+}
+
+void ABGRToAR30Row_C(const uint8_t* src_abgr, uint8_t* dst_ar30, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint32_t b0 = (src_abgr[0] >> 6) | ((uint32_t)(src_abgr[0]) << 2);
+    uint32_t g0 = (src_abgr[1] >> 6) | ((uint32_t)(src_abgr[1]) << 2);
+    uint32_t r0 = (src_abgr[2] >> 6) | ((uint32_t)(src_abgr[2]) << 2);
+    uint32_t a0 = (src_abgr[3] >> 6);
+    *(uint32_t*)(dst_ar30) =
+        STATIC_CAST(uint32_t, r0 | (g0 << 10) | (b0 << 20) | (a0 << 30));
+    dst_ar30 += 4;
+    src_abgr += 4;
+  }
+}
+
+void ARGBToAR30Row_C(const uint8_t* src_argb, uint8_t* dst_ar30, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint32_t b0 = (src_argb[0] >> 6) | ((uint32_t)(src_argb[0]) << 2);
+    uint32_t g0 = (src_argb[1] >> 6) | ((uint32_t)(src_argb[1]) << 2);
+    uint32_t r0 = (src_argb[2] >> 6) | ((uint32_t)(src_argb[2]) << 2);
+    uint32_t a0 = (src_argb[3] >> 6);
+    *(uint32_t*)(dst_ar30) =
+        STATIC_CAST(uint32_t, b0 | (g0 << 10) | (r0 << 20) | (a0 << 30));
+    dst_ar30 += 4;
+    src_argb += 4;
+  }
+}
+
+void ARGBToAR64Row_C(const uint8_t* src_argb, uint16_t* dst_ar64, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint16_t b = src_argb[0] * 0x0101;
+    uint16_t g = src_argb[1] * 0x0101;
+    uint16_t r = src_argb[2] * 0x0101;
+    uint16_t a = src_argb[3] * 0x0101;
+    dst_ar64[0] = b;
+    dst_ar64[1] = g;
+    dst_ar64[2] = r;
+    dst_ar64[3] = a;
+    dst_ar64 += 4;
+    src_argb += 4;
+  }
+}
+
+void ARGBToAB64Row_C(const uint8_t* src_argb, uint16_t* dst_ab64, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint16_t b = src_argb[0] * 0x0101;
+    uint16_t g = src_argb[1] * 0x0101;
+    uint16_t r = src_argb[2] * 0x0101;
+    uint16_t a = src_argb[3] * 0x0101;
+    dst_ab64[0] = r;
+    dst_ab64[1] = g;
+    dst_ab64[2] = b;
+    dst_ab64[3] = a;
+    dst_ab64 += 4;
+    src_argb += 4;
+  }
+}
+
+void AR64ToARGBRow_C(const uint16_t* src_ar64, uint8_t* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8_t b = src_ar64[0] >> 8;
+    uint8_t g = src_ar64[1] >> 8;
+    uint8_t r = src_ar64[2] >> 8;
+    uint8_t a = src_ar64[3] >> 8;
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    dst_argb += 4;
+    src_ar64 += 4;
+  }
+}
+
+void AB64ToARGBRow_C(const uint16_t* src_ab64, uint8_t* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8_t r = src_ab64[0] >> 8;
+    uint8_t g = src_ab64[1] >> 8;
+    uint8_t b = src_ab64[2] >> 8;
+    uint8_t a = src_ab64[3] >> 8;
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    dst_argb += 4;
+    src_ab64 += 4;
+  }
 }
 
+// TODO(fbarchard): Make shuffle compatible with SIMD versions
+void AR64ShuffleRow_C(const uint8_t* src_ar64,
+                      uint8_t* dst_ar64,
+                      const uint8_t* shuffler,
+                      int width) {
+  const uint16_t* src_ar64_16 = (const uint16_t*)src_ar64;
+  uint16_t* dst_ar64_16 = (uint16_t*)dst_ar64;
+  int index0 = shuffler[0] / 2;
+  int index1 = shuffler[2] / 2;
+  int index2 = shuffler[4] / 2;
+  int index3 = shuffler[6] / 2;
+  // Shuffle a row of AR64.
+  int x;
+  for (x = 0; x < width / 2; ++x) {
+    // To support in-place conversion.
+    uint16_t b = src_ar64_16[index0];
+    uint16_t g = src_ar64_16[index1];
+    uint16_t r = src_ar64_16[index2];
+    uint16_t a = src_ar64_16[index3];
+    dst_ar64_16[0] = b;
+    dst_ar64_16[1] = g;
+    dst_ar64_16[2] = r;
+    dst_ar64_16[3] = a;
+    src_ar64_16 += 4;
+    dst_ar64_16 += 4;
+  }
+}
+
+#ifdef LIBYUV_RGB7
+// Old 7 bit math for compatibility on unsupported platforms.
+static __inline uint8_t RGBToY(uint8_t r, uint8_t g, uint8_t b) {
+  return STATIC_CAST(uint8_t, ((33 * r + 65 * g + 13 * b) >> 7) + 16);
+}
+#else
+// 8 bit
+// Intel SSE/AVX uses the following equivalent formula
+// 0x7e80 = (66 + 129 + 25) * -128 + 0x1000 (for +16) and 0x0080 for round.
+//  return (66 * ((int)r - 128) + 129 * ((int)g - 128) + 25 * ((int)b - 128) +
+//  0x7e80) >> 8;
+
+static __inline uint8_t RGBToY(uint8_t r, uint8_t g, uint8_t b) {
+  return STATIC_CAST(uint8_t, (66 * r + 129 * g + 25 * b + 0x1080) >> 8);
+}
+#endif
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+
+// LIBYUV_RGBTOU_TRUNCATE mimics x86 code that does not round.
+#ifdef LIBYUV_RGBTOU_TRUNCATE
+static __inline uint8_t RGBToU(uint8_t r, uint8_t g, uint8_t b) {
+  return STATIC_CAST(uint8_t, (112 * b - 74 * g - 38 * r + 0x8000) >> 8);
+}
+static __inline uint8_t RGBToV(uint8_t r, uint8_t g, uint8_t b) {
+  return STATIC_CAST(uint8_t, (112 * r - 94 * g - 18 * b + 0x8000) >> 8);
+}
+#else
+// TODO(fbarchard): Add rounding to x86 SIMD and use this
+static __inline uint8_t RGBToU(uint8_t r, uint8_t g, uint8_t b) {
+  return STATIC_CAST(uint8_t, (112 * b - 74 * g - 38 * r + 0x8080) >> 8);
+}
+static __inline uint8_t RGBToV(uint8_t r, uint8_t g, uint8_t b) {
+  return STATIC_CAST(uint8_t, (112 * r - 94 * g - 18 * b + 0x8080) >> 8);
+}
+#endif
+
+// LIBYUV_ARGBTOUV_PAVGB mimics x86 code that subsamples with 2 pavgb.
+#if !defined(LIBYUV_ARGBTOUV_PAVGB)
+static __inline int RGB2xToU(uint16_t r, uint16_t g, uint16_t b) {
+  return STATIC_CAST(
+      uint8_t, ((112 / 2) * b - (74 / 2) * g - (38 / 2) * r + 0x8080) >> 8);
+}
+static __inline int RGB2xToV(uint16_t r, uint16_t g, uint16_t b) {
+  return STATIC_CAST(
+      uint8_t, ((112 / 2) * r - (94 / 2) * g - (18 / 2) * b + 0x8080) >> 8);
+}
+#endif
+
+// ARGBToY_C and ARGBToUV_C
+// Intel version mimic SSE/AVX which does 2 pavgb
+#if LIBYUV_ARGBTOUV_PAVGB
+#define MAKEROWY(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                 \
+    for (x = 0; x < width; ++x) {                                          \
+      dst_y[0] = RGBToY(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                      \
+      dst_y += 1;                                                          \
+    }                                                                      \
+  }                                                                        \
+  void NAME##ToUVRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                       uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                    \
+    int x;                                                                 \
+    for (x = 0; x < width - 1; x += 2) {                                   \
+      uint8_t ab = AVGB(AVGB(src_rgb[B], src_rgb1[B]),                     \
+                        AVGB(src_rgb[B + BPP], src_rgb1[B + BPP]));        \
+      uint8_t ag = AVGB(AVGB(src_rgb[G], src_rgb1[G]),                     \
+                        AVGB(src_rgb[G + BPP], src_rgb1[G + BPP]));        \
+      uint8_t ar = AVGB(AVGB(src_rgb[R], src_rgb1[R]),                     \
+                        AVGB(src_rgb[R + BPP], src_rgb1[R + BPP]));        \
+      dst_u[0] = RGBToU(ar, ag, ab);                                       \
+      dst_v[0] = RGBToV(ar, ag, ab);                                       \
+      src_rgb += BPP * 2;                                                  \
+      src_rgb1 += BPP * 2;                                                 \
+      dst_u += 1;                                                          \
+      dst_v += 1;                                                          \
+    }                                                                      \
+    if (width & 1) {                                                       \
+      uint8_t ab = AVGB(src_rgb[B], src_rgb1[B]);                          \
+      uint8_t ag = AVGB(src_rgb[G], src_rgb1[G]);                          \
+      uint8_t ar = AVGB(src_rgb[R], src_rgb1[R]);                          \
+      dst_u[0] = RGBToU(ar, ag, ab);                                       \
+      dst_v[0] = RGBToV(ar, ag, ab);                                       \
+    }                                                                      \
+  }
+#else
+// ARM version does sum / 2 then multiply by 2x smaller coefficients
+#define MAKEROWY(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                 \
+    for (x = 0; x < width; ++x) {                                          \
+      dst_y[0] = RGBToY(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                      \
+      dst_y += 1;                                                          \
+    }                                                                      \
+  }                                                                        \
+  void NAME##ToUVRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                       uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                    \
+    int x;                                                                 \
+    for (x = 0; x < width - 1; x += 2) {                                   \
+      uint16_t ab = (src_rgb[B] + src_rgb[B + BPP] + src_rgb1[B] +         \
+                     src_rgb1[B + BPP] + 1) >>                             \
+                    1;                                                     \
+      uint16_t ag = (src_rgb[G] + src_rgb[G + BPP] + src_rgb1[G] +         \
+                     src_rgb1[G + BPP] + 1) >>                             \
+                    1;                                                     \
+      uint16_t ar = (src_rgb[R] + src_rgb[R + BPP] + src_rgb1[R] +         \
+                     src_rgb1[R + BPP] + 1) >>                             \
+                    1;                                                     \
+      dst_u[0] = RGB2xToU(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToV(ar, ag, ab);                                     \
+      src_rgb += BPP * 2;                                                  \
+      src_rgb1 += BPP * 2;                                                 \
+      dst_u += 1;                                                          \
+      dst_v += 1;                                                          \
+    }                                                                      \
+    if (width & 1) {                                                       \
+      uint16_t ab = src_rgb[B] + src_rgb1[B];                              \
+      uint16_t ag = src_rgb[G] + src_rgb1[G];                              \
+      uint16_t ar = src_rgb[R] + src_rgb1[R];                              \
+      dst_u[0] = RGB2xToU(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToV(ar, ag, ab);                                     \
+    }                                                                      \
+  }
+#endif
+
 MAKEROWY(ARGB, 2, 1, 0, 4)
 MAKEROWY(BGRA, 1, 2, 3, 4)
 MAKEROWY(ABGR, 0, 1, 2, 4)
@@ -364,14 +691,14 @@ MAKEROWY(RAW, 0, 1, 2, 3)
 // b 0.1016 * 255 = 25.908 = 25
 // g 0.5078 * 255 = 129.489 = 129
 // r 0.2578 * 255 = 65.739 = 66
-// JPeg 8 bit Y (not used):
-// b 0.11400 * 256 = 29.184 = 29
-// g 0.58700 * 256 = 150.272 = 150
-// r 0.29900 * 256 = 76.544 = 77
-// JPeg 7 bit Y:
+// JPeg 7 bit Y (deprecated)
 // b 0.11400 * 128 = 14.592 = 15
 // g 0.58700 * 128 = 75.136 = 75
 // r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit Y:
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
 // JPeg 8 bit U:
 // b  0.50000 * 255 = 127.5 = 127
 // g -0.33126 * 255 = -84.4713 = -84
@@ -381,300 +708,453 @@ MAKEROWY(RAW, 0, 1, 2, 3)
 // g -0.41869 * 255 = -106.76595 = -107
 // r  0.50000 * 255 = 127.5 = 127
 
-static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) {
-  return (38 * r + 75 * g +  15 * b + 64) >> 7;
+#ifdef LIBYUV_RGB7
+// Old 7 bit math for compatibility on unsupported platforms.
+static __inline uint8_t RGBToYJ(uint8_t r, uint8_t g, uint8_t b) {
+  return (38 * r + 75 * g + 15 * b + 64) >> 7;
+}
+#else
+// 8 bit
+static __inline uint8_t RGBToYJ(uint8_t r, uint8_t g, uint8_t b) {
+  return (77 * r + 150 * g + 29 * b + 128) >> 8;
 }
+#endif
 
-static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) {
+#if defined(LIBYUV_ARGBTOUV_PAVGB)
+static __inline uint8_t RGBToUJ(uint8_t r, uint8_t g, uint8_t b) {
   return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
 }
-static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) {
+static __inline uint8_t RGBToVJ(uint8_t r, uint8_t g, uint8_t b) {
   return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
 }
+#else
+static __inline uint8_t RGB2xToUJ(uint16_t r, uint16_t g, uint16_t b) {
+  return ((127 / 2) * b - (84 / 2) * g - (43 / 2) * r + 0x8080) >> 8;
+}
+static __inline uint8_t RGB2xToVJ(uint16_t r, uint16_t g, uint16_t b) {
+  return ((127 / 2) * r - (107 / 2) * g - (20 / 2) * b + 0x8080) >> 8;
+}
+#endif
 
-#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+// ARGBToYJ_C and ARGBToUVJ_C
+// Intel version mimic SSE/AVX which does 2 pavgb
+#if LIBYUV_ARGBTOUV_PAVGB
+#define MAKEROWYJ(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                  \
+    for (x = 0; x < width; ++x) {                                           \
+      dst_y[0] = RGBToYJ(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                       \
+      dst_y += 1;                                                           \
+    }                                                                       \
+  }                                                                         \
+  void NAME##ToUVJRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                        uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                     \
+    int x;                                                                  \
+    for (x = 0; x < width - 1; x += 2) {                                    \
+      uint8_t ab = AVGB(AVGB(src_rgb[B], src_rgb1[B]),                      \
+                        AVGB(src_rgb[B + BPP], src_rgb1[B + BPP]));         \
+      uint8_t ag = AVGB(AVGB(src_rgb[G], src_rgb1[G]),                      \
+                        AVGB(src_rgb[G + BPP], src_rgb1[G + BPP]));         \
+      uint8_t ar = AVGB(AVGB(src_rgb[R], src_rgb1[R]),                      \
+                        AVGB(src_rgb[R + BPP], src_rgb1[R + BPP]));         \
+      dst_u[0] = RGBToUJ(ar, ag, ab);                                       \
+      dst_v[0] = RGBToVJ(ar, ag, ab);                                       \
+      src_rgb += BPP * 2;                                                   \
+      src_rgb1 += BPP * 2;                                                  \
+      dst_u += 1;                                                           \
+      dst_v += 1;                                                           \
+    }                                                                       \
+    if (width & 1) {                                                        \
+      uint8_t ab = AVGB(src_rgb[B], src_rgb1[B]);                           \
+      uint8_t ag = AVGB(src_rgb[G], src_rgb1[G]);                           \
+      uint8_t ar = AVGB(src_rgb[R], src_rgb1[R]);                           \
+      dst_u[0] = RGBToUJ(ar, ag, ab);                                       \
+      dst_v[0] = RGBToVJ(ar, ag, ab);                                       \
+    }                                                                       \
+  }
+#else
+// ARM version does sum / 2 then multiply by 2x smaller coefficients
+#define MAKEROWYJ(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                  \
+    for (x = 0; x < width; ++x) {                                           \
+      dst_y[0] = RGBToYJ(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                       \
+      dst_y += 1;                                                           \
+    }                                                                       \
+  }                                                                         \
+  void NAME##ToUVJRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                        uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                     \
+    int x;                                                                  \
+    for (x = 0; x < width - 1; x += 2) {                                    \
+      uint16_t ab = (src_rgb[B] + src_rgb[B + BPP] + src_rgb1[B] +          \
+                     src_rgb1[B + BPP] + 1) >>                              \
+                    1;                                                      \
+      uint16_t ag = (src_rgb[G] + src_rgb[G + BPP] + src_rgb1[G] +          \
+                     src_rgb1[G + BPP] + 1) >>                              \
+                    1;                                                      \
+      uint16_t ar = (src_rgb[R] + src_rgb[R + BPP] + src_rgb1[R] +          \
+                     src_rgb1[R + BPP] + 1) >>                              \
+                    1;                                                      \
+      dst_u[0] = RGB2xToUJ(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToVJ(ar, ag, ab);                                     \
+      src_rgb += BPP * 2;                                                   \
+      src_rgb1 += BPP * 2;                                                  \
+      dst_u += 1;                                                           \
+      dst_v += 1;                                                           \
+    }                                                                       \
+    if (width & 1) {                                                        \
+      uint16_t ab = (src_rgb[B] + src_rgb1[B]);                             \
+      uint16_t ag = (src_rgb[G] + src_rgb1[G]);                             \
+      uint16_t ar = (src_rgb[R] + src_rgb1[R]);                             \
+      dst_u[0] = RGB2xToUJ(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToVJ(ar, ag, ab);                                     \
+    }                                                                       \
+  }
 
-#define MAKEROWYJ(NAME, R, G, B, BPP) \
-void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) {      \
-  int x;                                                                       \
-  for (x = 0; x < width; ++x) {                                                \
-    dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);              \
-    src_argb0 += BPP;                                                          \
-    dst_y += 1;                                                                \
-  }                                                                            \
-}                                                                              \
-void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
-                        uint8* dst_u, uint8* dst_v, int width) {               \
-  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
-  int x;                                                                       \
-  for (x = 0; x < width - 1; x += 2) {                                         \
-    uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                            \
-                    AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));               \
-    uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                            \
-                    AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));               \
-    uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                            \
-                    AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));               \
-    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
-    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
-    src_rgb0 += BPP * 2;                                                       \
-    src_rgb1 += BPP * 2;                                                       \
-    dst_u += 1;                                                                \
-    dst_v += 1;                                                                \
-  }                                                                            \
-  if (width & 1) {                                                             \
-    uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]);                                 \
-    uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]);                                 \
-    uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]);                                 \
-    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
-    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
-  }                                                                            \
-}
+#endif
 
 MAKEROWYJ(ARGB, 2, 1, 0, 4)
+MAKEROWYJ(ABGR, 0, 1, 2, 4)
+MAKEROWYJ(RGBA, 3, 2, 1, 4)
+MAKEROWYJ(RGB24, 2, 1, 0, 3)
+MAKEROWYJ(RAW, 0, 1, 2, 3)
 #undef MAKEROWYJ
 
-void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
+void RGB565ToYRow_C(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_rgb565[0] & 0x1f;
-    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
-    uint8 r = src_rgb565[1] >> 3;
-    b = (b << 3) | (b >> 2);
-    g = (g << 2) | (g >> 4);
-    r = (r << 3) | (r >> 2);
+    uint8_t b = src_rgb565[0] & 0x1f;
+    uint8_t g = STATIC_CAST(
+        uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3));
+    uint8_t r = src_rgb565[1] >> 3;
+    b = STATIC_CAST(uint8_t, (b << 3) | (b >> 2));
+    g = STATIC_CAST(uint8_t, (g << 2) | (g >> 4));
+    r = STATIC_CAST(uint8_t, (r << 3) | (r >> 2));
     dst_y[0] = RGBToY(r, g, b);
     src_rgb565 += 2;
     dst_y += 1;
   }
 }
 
-void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) {
+void ARGB1555ToYRow_C(const uint8_t* src_argb1555, uint8_t* dst_y, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_argb1555[0] & 0x1f;
-    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
-    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
-    b = (b << 3) | (b >> 2);
-    g = (g << 3) | (g >> 2);
-    r = (r << 3) | (r >> 2);
+    uint8_t b = src_argb1555[0] & 0x1f;
+    uint8_t g = STATIC_CAST(
+        uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3));
+    uint8_t r = (src_argb1555[1] & 0x7c) >> 2;
+    b = STATIC_CAST(uint8_t, (b << 3) | (b >> 2));
+    g = STATIC_CAST(uint8_t, (g << 3) | (g >> 2));
+    r = STATIC_CAST(uint8_t, (r << 3) | (r >> 2));
     dst_y[0] = RGBToY(r, g, b);
     src_argb1555 += 2;
     dst_y += 1;
   }
 }
 
-void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) {
+void ARGB4444ToYRow_C(const uint8_t* src_argb4444, uint8_t* dst_y, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 b = src_argb4444[0] & 0x0f;
-    uint8 g = src_argb4444[0] >> 4;
-    uint8 r = src_argb4444[1] & 0x0f;
-    b = (b << 4) | b;
-    g = (g << 4) | g;
-    r = (r << 4) | r;
+    uint8_t b = src_argb4444[0] & 0x0f;
+    uint8_t g = src_argb4444[0] >> 4;
+    uint8_t r = src_argb4444[1] & 0x0f;
+    b = STATIC_CAST(uint8_t, (b << 4) | b);
+    g = STATIC_CAST(uint8_t, (g << 4) | g);
+    r = STATIC_CAST(uint8_t, (r << 4) | r);
     dst_y[0] = RGBToY(r, g, b);
     src_argb4444 += 2;
     dst_y += 1;
   }
 }
 
-void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
-                     uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565;
+void RGB565ToUVRow_C(const uint8_t* src_rgb565,
+                     int src_stride_rgb565,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  const uint8_t* next_rgb565 = src_rgb565 + src_stride_rgb565;
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 b0 = src_rgb565[0] & 0x1f;
-    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
-    uint8 r0 = src_rgb565[1] >> 3;
-    uint8 b1 = src_rgb565[2] & 0x1f;
-    uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3);
-    uint8 r1 = src_rgb565[3] >> 3;
-    uint8 b2 = next_rgb565[0] & 0x1f;
-    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
-    uint8 r2 = next_rgb565[1] >> 3;
-    uint8 b3 = next_rgb565[2] & 0x1f;
-    uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
-    uint8 r3 = next_rgb565[3] >> 3;
-    uint8 b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
-    uint8 g = (g0 + g1 + g2 + g3);
-    uint8 r = (r0 + r1 + r2 + r3);
-    b = (b << 1) | (b >> 6);  // 787 -> 888.
-    r = (r << 1) | (r >> 6);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+    uint8_t b0 = STATIC_CAST(uint8_t, src_rgb565[0] & 0x1f);
+    uint8_t g0 = STATIC_CAST(
+        uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3));
+    uint8_t r0 = STATIC_CAST(uint8_t, src_rgb565[1] >> 3);
+    uint8_t b1 = STATIC_CAST(uint8_t, src_rgb565[2] & 0x1f);
+    uint8_t g1 = STATIC_CAST(
+        uint8_t, (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3));
+    uint8_t r1 = STATIC_CAST(uint8_t, src_rgb565[3] >> 3);
+    uint8_t b2 = STATIC_CAST(uint8_t, next_rgb565[0] & 0x1f);
+    uint8_t g2 = STATIC_CAST(
+        uint8_t, (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3));
+    uint8_t r2 = STATIC_CAST(uint8_t, next_rgb565[1] >> 3);
+    uint8_t b3 = STATIC_CAST(uint8_t, next_rgb565[2] & 0x1f);
+    uint8_t g3 = STATIC_CAST(
+        uint8_t, (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3));
+    uint8_t r3 = STATIC_CAST(uint8_t, next_rgb565[3] >> 3);
+
+    b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2));
+    g0 = STATIC_CAST(uint8_t, (g0 << 2) | (g0 >> 4));
+    r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2));
+    b1 = STATIC_CAST(uint8_t, (b1 << 3) | (b1 >> 2));
+    g1 = STATIC_CAST(uint8_t, (g1 << 2) | (g1 >> 4));
+    r1 = STATIC_CAST(uint8_t, (r1 << 3) | (r1 >> 2));
+    b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2));
+    g2 = STATIC_CAST(uint8_t, (g2 << 2) | (g2 >> 4));
+    r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2));
+    b3 = STATIC_CAST(uint8_t, (b3 << 3) | (b3 >> 2));
+    g3 = STATIC_CAST(uint8_t, (g3 << 2) | (g3 >> 4));
+    r3 = STATIC_CAST(uint8_t, (r3 << 3) | (r3 >> 2));
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
+    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
+    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = (b0 + b1 + b2 + b3 + 1) >> 1;
+    uint16_t g = (g0 + g1 + g2 + g3 + 1) >> 1;
+    uint16_t r = (r0 + r1 + r2 + r3 + 1) >> 1;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+
     src_rgb565 += 4;
     next_rgb565 += 4;
     dst_u += 1;
     dst_v += 1;
   }
   if (width & 1) {
-    uint8 b0 = src_rgb565[0] & 0x1f;
-    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
-    uint8 r0 = src_rgb565[1] >> 3;
-    uint8 b2 = next_rgb565[0] & 0x1f;
-    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
-    uint8 r2 = next_rgb565[1] >> 3;
-    uint8 b = (b0 + b2);  // 565 * 2 = 676.
-    uint8 g = (g0 + g2);
-    uint8 r = (r0 + r2);
-    b = (b << 2) | (b >> 4);  // 676 -> 888
-    g = (g << 1) | (g >> 6);
-    r = (r << 2) | (r >> 4);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
-  }
-}
-
-void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555;
+    uint8_t b0 = STATIC_CAST(uint8_t, src_rgb565[0] & 0x1f);
+    uint8_t g0 = STATIC_CAST(
+        uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3));
+    uint8_t r0 = STATIC_CAST(uint8_t, src_rgb565[1] >> 3);
+    uint8_t b2 = STATIC_CAST(uint8_t, next_rgb565[0] & 0x1f);
+    uint8_t g2 = STATIC_CAST(
+        uint8_t, (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3));
+    uint8_t r2 = STATIC_CAST(uint8_t, next_rgb565[1] >> 3);
+    b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2));
+    g0 = STATIC_CAST(uint8_t, (g0 << 2) | (g0 >> 4));
+    r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2));
+    b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2));
+    g2 = STATIC_CAST(uint8_t, (g2 << 2) | (g2 >> 4));
+    r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2));
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(b0, b2);
+    uint8_t ag = AVGB(g0, g2);
+    uint8_t ar = AVGB(r0, r2);
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = b0 + b2;
+    uint16_t g = g0 + g2;
+    uint16_t r = r0 + r2;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+  }
+}
+
+void ARGB1555ToUVRow_C(const uint8_t* src_argb1555,
+                       int src_stride_argb1555,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  const uint8_t* next_argb1555 = src_argb1555 + src_stride_argb1555;
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 b0 = src_argb1555[0] & 0x1f;
-    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
-    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
-    uint8 b1 = src_argb1555[2] & 0x1f;
-    uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3);
-    uint8 r1 = (src_argb1555[3] & 0x7c) >> 2;
-    uint8 b2 = next_argb1555[0] & 0x1f;
-    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
-    uint8 r2 = (next_argb1555[1] & 0x7c) >> 2;
-    uint8 b3 = next_argb1555[2] & 0x1f;
-    uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
-    uint8 r3 = (next_argb1555[3] & 0x7c) >> 2;
-    uint8 b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
-    uint8 g = (g0 + g1 + g2 + g3);
-    uint8 r = (r0 + r1 + r2 + r3);
-    b = (b << 1) | (b >> 6);  // 777 -> 888.
-    g = (g << 1) | (g >> 6);
-    r = (r << 1) | (r >> 6);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+    uint8_t b0 = STATIC_CAST(uint8_t, src_argb1555[0] & 0x1f);
+    uint8_t g0 = STATIC_CAST(
+        uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3));
+    uint8_t r0 = STATIC_CAST(uint8_t, (src_argb1555[1] & 0x7c) >> 2);
+    uint8_t b1 = STATIC_CAST(uint8_t, src_argb1555[2] & 0x1f);
+    uint8_t g1 = STATIC_CAST(
+        uint8_t, (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3));
+    uint8_t r1 = STATIC_CAST(uint8_t, (src_argb1555[3] & 0x7c) >> 2);
+    uint8_t b2 = STATIC_CAST(uint8_t, next_argb1555[0] & 0x1f);
+    uint8_t g2 = STATIC_CAST(
+        uint8_t, (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3));
+    uint8_t r2 = STATIC_CAST(uint8_t, (next_argb1555[1] & 0x7c) >> 2);
+    uint8_t b3 = STATIC_CAST(uint8_t, next_argb1555[2] & 0x1f);
+    uint8_t g3 = STATIC_CAST(
+        uint8_t, (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3));
+    uint8_t r3 = STATIC_CAST(uint8_t, (next_argb1555[3] & 0x7c) >> 2);
+
+    b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2));
+    g0 = STATIC_CAST(uint8_t, (g0 << 3) | (g0 >> 2));
+    r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2));
+    b1 = STATIC_CAST(uint8_t, (b1 << 3) | (b1 >> 2));
+    g1 = STATIC_CAST(uint8_t, (g1 << 3) | (g1 >> 2));
+    r1 = STATIC_CAST(uint8_t, (r1 << 3) | (r1 >> 2));
+    b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2));
+    g2 = STATIC_CAST(uint8_t, (g2 << 3) | (g2 >> 2));
+    r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2));
+    b3 = STATIC_CAST(uint8_t, (b3 << 3) | (b3 >> 2));
+    g3 = STATIC_CAST(uint8_t, (g3 << 3) | (g3 >> 2));
+    r3 = STATIC_CAST(uint8_t, (r3 << 3) | (r3 >> 2));
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
+    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
+    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = (b0 + b1 + b2 + b3 + 1) >> 1;
+    uint16_t g = (g0 + g1 + g2 + g3 + 1) >> 1;
+    uint16_t r = (r0 + r1 + r2 + r3 + 1) >> 1;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+
     src_argb1555 += 4;
     next_argb1555 += 4;
     dst_u += 1;
     dst_v += 1;
   }
   if (width & 1) {
-    uint8 b0 = src_argb1555[0] & 0x1f;
-    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
-    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
-    uint8 b2 = next_argb1555[0] & 0x1f;
-    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
-    uint8 r2 = next_argb1555[1] >> 3;
-    uint8 b = (b0 + b2);  // 555 * 2 = 666.
-    uint8 g = (g0 + g2);
-    uint8 r = (r0 + r2);
-    b = (b << 2) | (b >> 4);  // 666 -> 888.
-    g = (g << 2) | (g >> 4);
-    r = (r << 2) | (r >> 4);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
-  }
-}
-
-void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444;
+    uint8_t b0 = STATIC_CAST(uint8_t, src_argb1555[0] & 0x1f);
+    uint8_t g0 = STATIC_CAST(
+        uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3));
+    uint8_t r0 = STATIC_CAST(uint8_t, (src_argb1555[1] & 0x7c) >> 2);
+    uint8_t b2 = STATIC_CAST(uint8_t, next_argb1555[0] & 0x1f);
+    uint8_t g2 = STATIC_CAST(
+        uint8_t, (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3));
+    uint8_t r2 = STATIC_CAST(uint8_t, (next_argb1555[1] & 0x7c) >> 2);
+
+    b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2));
+    g0 = STATIC_CAST(uint8_t, (g0 << 3) | (g0 >> 2));
+    r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2));
+    b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2));
+    g2 = STATIC_CAST(uint8_t, (g2 << 3) | (g2 >> 2));
+    r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2));
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(b0, b2);
+    uint8_t ag = AVGB(g0, g2);
+    uint8_t ar = AVGB(r0, r2);
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = b0 + b2;
+    uint16_t g = g0 + g2;
+    uint16_t r = r0 + r2;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+  }
+}
+
+void ARGB4444ToUVRow_C(const uint8_t* src_argb4444,
+                       int src_stride_argb4444,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  const uint8_t* next_argb4444 = src_argb4444 + src_stride_argb4444;
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 b0 = src_argb4444[0] & 0x0f;
-    uint8 g0 = src_argb4444[0] >> 4;
-    uint8 r0 = src_argb4444[1] & 0x0f;
-    uint8 b1 = src_argb4444[2] & 0x0f;
-    uint8 g1 = src_argb4444[2] >> 4;
-    uint8 r1 = src_argb4444[3] & 0x0f;
-    uint8 b2 = next_argb4444[0] & 0x0f;
-    uint8 g2 = next_argb4444[0] >> 4;
-    uint8 r2 = next_argb4444[1] & 0x0f;
-    uint8 b3 = next_argb4444[2] & 0x0f;
-    uint8 g3 = next_argb4444[2] >> 4;
-    uint8 r3 = next_argb4444[3] & 0x0f;
-    uint8 b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
-    uint8 g = (g0 + g1 + g2 + g3);
-    uint8 r = (r0 + r1 + r2 + r3);
-    b = (b << 2) | (b >> 4);  // 666 -> 888.
-    g = (g << 2) | (g >> 4);
-    r = (r << 2) | (r >> 4);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+    uint8_t b0 = src_argb4444[0] & 0x0f;
+    uint8_t g0 = src_argb4444[0] >> 4;
+    uint8_t r0 = src_argb4444[1] & 0x0f;
+    uint8_t b1 = src_argb4444[2] & 0x0f;
+    uint8_t g1 = src_argb4444[2] >> 4;
+    uint8_t r1 = src_argb4444[3] & 0x0f;
+    uint8_t b2 = next_argb4444[0] & 0x0f;
+    uint8_t g2 = next_argb4444[0] >> 4;
+    uint8_t r2 = next_argb4444[1] & 0x0f;
+    uint8_t b3 = next_argb4444[2] & 0x0f;
+    uint8_t g3 = next_argb4444[2] >> 4;
+    uint8_t r3 = next_argb4444[3] & 0x0f;
+
+    b0 = STATIC_CAST(uint8_t, (b0 << 4) | b0);
+    g0 = STATIC_CAST(uint8_t, (g0 << 4) | g0);
+    r0 = STATIC_CAST(uint8_t, (r0 << 4) | r0);
+    b1 = STATIC_CAST(uint8_t, (b1 << 4) | b1);
+    g1 = STATIC_CAST(uint8_t, (g1 << 4) | g1);
+    r1 = STATIC_CAST(uint8_t, (r1 << 4) | r1);
+    b2 = STATIC_CAST(uint8_t, (b2 << 4) | b2);
+    g2 = STATIC_CAST(uint8_t, (g2 << 4) | g2);
+    r2 = STATIC_CAST(uint8_t, (r2 << 4) | r2);
+    b3 = STATIC_CAST(uint8_t, (b3 << 4) | b3);
+    g3 = STATIC_CAST(uint8_t, (g3 << 4) | g3);
+    r3 = STATIC_CAST(uint8_t, (r3 << 4) | r3);
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
+    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
+    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = (b0 + b1 + b2 + b3 + 1) >> 1;
+    uint16_t g = (g0 + g1 + g2 + g3 + 1) >> 1;
+    uint16_t r = (r0 + r1 + r2 + r3 + 1) >> 1;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+
     src_argb4444 += 4;
     next_argb4444 += 4;
     dst_u += 1;
     dst_v += 1;
   }
   if (width & 1) {
-    uint8 b0 = src_argb4444[0] & 0x0f;
-    uint8 g0 = src_argb4444[0] >> 4;
-    uint8 r0 = src_argb4444[1] & 0x0f;
-    uint8 b2 = next_argb4444[0] & 0x0f;
-    uint8 g2 = next_argb4444[0] >> 4;
-    uint8 r2 = next_argb4444[1] & 0x0f;
-    uint8 b = (b0 + b2);  // 444 * 2 = 555.
-    uint8 g = (g0 + g2);
-    uint8 r = (r0 + r2);
-    b = (b << 3) | (b >> 2);  // 555 -> 888.
-    g = (g << 3) | (g >> 2);
-    r = (r << 3) | (r >> 2);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
-  }
-}
-
-void ARGBToUV444Row_C(const uint8* src_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  int x;
-  for (x = 0; x < width; ++x) {
-    uint8 ab = src_argb[0];
-    uint8 ag = src_argb[1];
-    uint8 ar = src_argb[2];
+    uint8_t b0 = src_argb4444[0] & 0x0f;
+    uint8_t g0 = src_argb4444[0] >> 4;
+    uint8_t r0 = src_argb4444[1] & 0x0f;
+    uint8_t b2 = next_argb4444[0] & 0x0f;
+    uint8_t g2 = next_argb4444[0] >> 4;
+    uint8_t r2 = next_argb4444[1] & 0x0f;
+
+    b0 = STATIC_CAST(uint8_t, (b0 << 4) | b0);
+    g0 = STATIC_CAST(uint8_t, (g0 << 4) | g0);
+    r0 = STATIC_CAST(uint8_t, (r0 << 4) | r0);
+    b2 = STATIC_CAST(uint8_t, (b2 << 4) | b2);
+    g2 = STATIC_CAST(uint8_t, (g2 << 4) | g2);
+    r2 = STATIC_CAST(uint8_t, (r2 << 4) | r2);
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(b0, b2);
+    uint8_t ag = AVGB(g0, g2);
+    uint8_t ar = AVGB(r0, r2);
     dst_u[0] = RGBToU(ar, ag, ab);
     dst_v[0] = RGBToV(ar, ag, ab);
-    src_argb += 4;
-    dst_u += 1;
-    dst_v += 1;
+#else
+    uint16_t b = b0 + b2;
+    uint16_t g = g0 + g2;
+    uint16_t r = r0 + r2;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
   }
 }
 
-void ARGBToUV411Row_C(const uint8* src_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
+void ARGBToUV444Row_C(const uint8_t* src_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   int x;
-  for (x = 0; x < width - 3; x += 4) {
-    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2;
-    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2;
-    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2;
+  for (x = 0; x < width; ++x) {
+    uint8_t ab = src_argb[0];
+    uint8_t ag = src_argb[1];
+    uint8_t ar = src_argb[2];
     dst_u[0] = RGBToU(ar, ag, ab);
     dst_v[0] = RGBToV(ar, ag, ab);
-    src_argb += 16;
+    src_argb += 4;
     dst_u += 1;
     dst_v += 1;
   }
-  // Odd width handling mimics 'any' function which replicates last pixel.
-  if ((width & 3) == 3) {
-    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[8]) >> 2;
-    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[9]) >> 2;
-    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[10]) >> 2;
-    dst_u[0] = RGBToU(ar, ag, ab);
-    dst_v[0] = RGBToV(ar, ag, ab);
-  } else if ((width & 3) == 2) {
-    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
-    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
-    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
-    dst_u[0] = RGBToU(ar, ag, ab);
-    dst_v[0] = RGBToV(ar, ag, ab);
-  } else if ((width & 3) == 1) {
-    uint8 ab = src_argb[0];
-    uint8 ag = src_argb[1];
-    uint8 ar = src_argb[2];
-    dst_u[0] = RGBToU(ar, ag, ab);
-    dst_v[0] = RGBToV(ar, ag, ab);
-  }
 }
 
-void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+void ARGBGrayRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
+    uint8_t y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
     dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
     dst_argb[3] = src_argb[3];
     dst_argb += 4;
@@ -683,7 +1163,7 @@ void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
 }
 
 // Convert a row of image to Sepia tone.
-void ARGBSepiaRow_C(uint8* dst_argb, int width) {
+void ARGBSepiaRow_C(uint8_t* dst_argb, int width) {
   int x;
   for (x = 0; x < width; ++x) {
     int b = dst_argb[0];
@@ -693,42 +1173,50 @@ void ARGBSepiaRow_C(uint8* dst_argb, int width) {
     int sg = (b * 22 + g * 88 + r * 45) >> 7;
     int sr = (b * 24 + g * 98 + r * 50) >> 7;
     // b does not over flow. a is preserved from original.
-    dst_argb[0] = sb;
-    dst_argb[1] = clamp255(sg);
-    dst_argb[2] = clamp255(sr);
+    dst_argb[0] = STATIC_CAST(uint8_t, sb);
+    dst_argb[1] = STATIC_CAST(uint8_t, clamp255(sg));
+    dst_argb[2] = STATIC_CAST(uint8_t, clamp255(sr));
     dst_argb += 4;
   }
 }
 
 // Apply color matrix to a row of image. Matrix is signed.
 // TODO(fbarchard): Consider adding rounding (+32).
-void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
-                          const int8* matrix_argb, int width) {
+void ARGBColorMatrixRow_C(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          const int8_t* matrix_argb,
+                          int width) {
   int x;
   for (x = 0; x < width; ++x) {
     int b = src_argb[0];
     int g = src_argb[1];
     int r = src_argb[2];
     int a = src_argb[3];
-    int sb = (b * matrix_argb[0] + g * matrix_argb[1] +
-              r * matrix_argb[2] + a * matrix_argb[3]) >> 6;
-    int sg = (b * matrix_argb[4] + g * matrix_argb[5] +
-              r * matrix_argb[6] + a * matrix_argb[7]) >> 6;
-    int sr = (b * matrix_argb[8] + g * matrix_argb[9] +
-              r * matrix_argb[10] + a * matrix_argb[11]) >> 6;
-    int sa = (b * matrix_argb[12] + g * matrix_argb[13] +
-              r * matrix_argb[14] + a * matrix_argb[15]) >> 6;
-    dst_argb[0] = Clamp(sb);
-    dst_argb[1] = Clamp(sg);
-    dst_argb[2] = Clamp(sr);
-    dst_argb[3] = Clamp(sa);
+    int sb = (b * matrix_argb[0] + g * matrix_argb[1] + r * matrix_argb[2] +
+              a * matrix_argb[3]) >>
+             6;
+    int sg = (b * matrix_argb[4] + g * matrix_argb[5] + r * matrix_argb[6] +
+              a * matrix_argb[7]) >>
+             6;
+    int sr = (b * matrix_argb[8] + g * matrix_argb[9] + r * matrix_argb[10] +
+              a * matrix_argb[11]) >>
+             6;
+    int sa = (b * matrix_argb[12] + g * matrix_argb[13] + r * matrix_argb[14] +
+              a * matrix_argb[15]) >>
+             6;
+    dst_argb[0] = STATIC_CAST(uint8_t, Clamp(sb));
+    dst_argb[1] = STATIC_CAST(uint8_t, Clamp(sg));
+    dst_argb[2] = STATIC_CAST(uint8_t, Clamp(sr));
+    dst_argb[3] = STATIC_CAST(uint8_t, Clamp(sa));
     src_argb += 4;
     dst_argb += 4;
   }
 }
 
 // Apply color table to a row of image.
-void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+void ARGBColorTableRow_C(uint8_t* dst_argb,
+                         const uint8_t* table_argb,
+                         int width) {
   int x;
   for (x = 0; x < width; ++x) {
     int b = dst_argb[0];
@@ -744,7 +1232,9 @@ void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
 }
 
 // Apply color table to a row of image.
-void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+void RGBColorTableRow_C(uint8_t* dst_argb,
+                        const uint8_t* table_argb,
+                        int width) {
   int x;
   for (x = 0; x < width; ++x) {
     int b = dst_argb[0];
@@ -757,36 +1247,44 @@ void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
   }
 }
 
-void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
-                       int interval_offset, int width) {
+void ARGBQuantizeRow_C(uint8_t* dst_argb,
+                       int scale,
+                       int interval_size,
+                       int interval_offset,
+                       int width) {
   int x;
   for (x = 0; x < width; ++x) {
     int b = dst_argb[0];
     int g = dst_argb[1];
     int r = dst_argb[2];
-    dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
-    dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset;
-    dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset;
+    dst_argb[0] = STATIC_CAST(
+        uint8_t, (b * scale >> 16) * interval_size + interval_offset);
+    dst_argb[1] = STATIC_CAST(
+        uint8_t, (g * scale >> 16) * interval_size + interval_offset);
+    dst_argb[2] = STATIC_CAST(
+        uint8_t, (r * scale >> 16) * interval_size + interval_offset);
     dst_argb += 4;
   }
 }
 
 #define REPEAT8(v) (v) | ((v) << 8)
-#define SHADE(f, v) v * f >> 24
+#define SHADE(f, v) v* f >> 24
 
-void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
-                    uint32 value) {
-  const uint32 b_scale = REPEAT8(value & 0xff);
-  const uint32 g_scale = REPEAT8((value >> 8) & 0xff);
-  const uint32 r_scale = REPEAT8((value >> 16) & 0xff);
-  const uint32 a_scale = REPEAT8(value >> 24);
+void ARGBShadeRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_argb,
+                    int width,
+                    uint32_t value) {
+  const uint32_t b_scale = REPEAT8(value & 0xff);
+  const uint32_t g_scale = REPEAT8((value >> 8) & 0xff);
+  const uint32_t r_scale = REPEAT8((value >> 16) & 0xff);
+  const uint32_t a_scale = REPEAT8(value >> 24);
 
   int i;
   for (i = 0; i < width; ++i) {
-    const uint32 b = REPEAT8(src_argb[0]);
-    const uint32 g = REPEAT8(src_argb[1]);
-    const uint32 r = REPEAT8(src_argb[2]);
-    const uint32 a = REPEAT8(src_argb[3]);
+    const uint32_t b = REPEAT8(src_argb[0]);
+    const uint32_t g = REPEAT8(src_argb[1]);
+    const uint32_t r = REPEAT8(src_argb[2]);
+    const uint32_t a = REPEAT8(src_argb[3]);
     dst_argb[0] = SHADE(b, b_scale);
     dst_argb[1] = SHADE(g, g_scale);
     dst_argb[2] = SHADE(r, r_scale);
@@ -799,25 +1297,27 @@ void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
 #undef SHADE
 
 #define REPEAT8(v) (v) | ((v) << 8)
-#define SHADE(f, v) v * f >> 16
+#define SHADE(f, v) v* f >> 16
 
-void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
-                       uint8* dst_argb, int width) {
+void ARGBMultiplyRow_C(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    const uint32 b = REPEAT8(src_argb0[0]);
-    const uint32 g = REPEAT8(src_argb0[1]);
-    const uint32 r = REPEAT8(src_argb0[2]);
-    const uint32 a = REPEAT8(src_argb0[3]);
-    const uint32 b_scale = src_argb1[0];
-    const uint32 g_scale = src_argb1[1];
-    const uint32 r_scale = src_argb1[2];
-    const uint32 a_scale = src_argb1[3];
-    dst_argb[0] = SHADE(b, b_scale);
-    dst_argb[1] = SHADE(g, g_scale);
-    dst_argb[2] = SHADE(r, r_scale);
-    dst_argb[3] = SHADE(a, a_scale);
-    src_argb0 += 4;
+    const uint32_t b = REPEAT8(src_argb[0]);
+    const uint32_t g = REPEAT8(src_argb[1]);
+    const uint32_t r = REPEAT8(src_argb[2]);
+    const uint32_t a = REPEAT8(src_argb[3]);
+    const uint32_t b_scale = src_argb1[0];
+    const uint32_t g_scale = src_argb1[1];
+    const uint32_t r_scale = src_argb1[2];
+    const uint32_t a_scale = src_argb1[3];
+    dst_argb[0] = STATIC_CAST(uint8_t, SHADE(b, b_scale));
+    dst_argb[1] = STATIC_CAST(uint8_t, SHADE(g, g_scale));
+    dst_argb[2] = STATIC_CAST(uint8_t, SHADE(r, r_scale));
+    dst_argb[3] = STATIC_CAST(uint8_t, SHADE(a, a_scale));
+    src_argb += 4;
     src_argb1 += 4;
     dst_argb += 4;
   }
@@ -827,23 +1327,25 @@ void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
 
 #define SHADE(f, v) clamp255(v + f)
 
-void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
-                  uint8* dst_argb, int width) {
+void ARGBAddRow_C(const uint8_t* src_argb,
+                  const uint8_t* src_argb1,
+                  uint8_t* dst_argb,
+                  int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    const int b = src_argb0[0];
-    const int g = src_argb0[1];
-    const int r = src_argb0[2];
-    const int a = src_argb0[3];
+    const int b = src_argb[0];
+    const int g = src_argb[1];
+    const int r = src_argb[2];
+    const int a = src_argb[3];
     const int b_add = src_argb1[0];
     const int g_add = src_argb1[1];
     const int r_add = src_argb1[2];
     const int a_add = src_argb1[3];
-    dst_argb[0] = SHADE(b, b_add);
-    dst_argb[1] = SHADE(g, g_add);
-    dst_argb[2] = SHADE(r, r_add);
-    dst_argb[3] = SHADE(a, a_add);
-    src_argb0 += 4;
+    dst_argb[0] = STATIC_CAST(uint8_t, SHADE(b, b_add));
+    dst_argb[1] = STATIC_CAST(uint8_t, SHADE(g, g_add));
+    dst_argb[2] = STATIC_CAST(uint8_t, SHADE(r, r_add));
+    dst_argb[3] = STATIC_CAST(uint8_t, SHADE(a, a_add));
+    src_argb += 4;
     src_argb1 += 4;
     dst_argb += 4;
   }
@@ -852,23 +1354,25 @@ void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
 
 #define SHADE(f, v) clamp0(f - v)
 
-void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
-                       uint8* dst_argb, int width) {
+void ARGBSubtractRow_C(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    const int b = src_argb0[0];
-    const int g = src_argb0[1];
-    const int r = src_argb0[2];
-    const int a = src_argb0[3];
+    const int b = src_argb[0];
+    const int g = src_argb[1];
+    const int r = src_argb[2];
+    const int a = src_argb[3];
     const int b_sub = src_argb1[0];
     const int g_sub = src_argb1[1];
     const int r_sub = src_argb1[2];
     const int a_sub = src_argb1[3];
-    dst_argb[0] = SHADE(b, b_sub);
-    dst_argb[1] = SHADE(g, g_sub);
-    dst_argb[2] = SHADE(r, r_sub);
-    dst_argb[3] = SHADE(a, a_sub);
-    src_argb0 += 4;
+    dst_argb[0] = STATIC_CAST(uint8_t, SHADE(b, b_sub));
+    dst_argb[1] = STATIC_CAST(uint8_t, SHADE(g, g_sub));
+    dst_argb[2] = STATIC_CAST(uint8_t, SHADE(r, r_sub));
+    dst_argb[3] = STATIC_CAST(uint8_t, SHADE(a, a_sub));
+    src_argb += 4;
     src_argb1 += 4;
     dst_argb += 4;
   }
@@ -876,8 +1380,11 @@ void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
 #undef SHADE
 
 // Sobel functions which mimics SSSE3.
-void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
-                 uint8* dst_sobelx, int width) {
+void SobelXRow_C(const uint8_t* src_y0,
+                 const uint8_t* src_y1,
+                 const uint8_t* src_y2,
+                 uint8_t* dst_sobelx,
+                 int width) {
   int i;
   for (i = 0; i < width; ++i) {
     int a = src_y0[i];
@@ -890,12 +1397,14 @@ void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
     int b_diff = b - b_sub;
     int c_diff = c - c_sub;
     int sobel = Abs(a_diff + b_diff * 2 + c_diff);
-    dst_sobelx[i] = (uint8)(clamp255(sobel));
+    dst_sobelx[i] = (uint8_t)(clamp255(sobel));
   }
 }
 
-void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
-                 uint8* dst_sobely, int width) {
+void SobelYRow_C(const uint8_t* src_y0,
+                 const uint8_t* src_y1,
+                 uint8_t* dst_sobely,
+                 int width) {
   int i;
   for (i = 0; i < width; ++i) {
     int a = src_y0[i + 0];
@@ -908,56 +1417,62 @@ void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
     int b_diff = b - b_sub;
     int c_diff = c - c_sub;
     int sobel = Abs(a_diff + b_diff * 2 + c_diff);
-    dst_sobely[i] = (uint8)(clamp255(sobel));
+    dst_sobely[i] = (uint8_t)(clamp255(sobel));
   }
 }
 
-void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
-                uint8* dst_argb, int width) {
+void SobelRow_C(const uint8_t* src_sobelx,
+                const uint8_t* src_sobely,
+                uint8_t* dst_argb,
+                int width) {
   int i;
   for (i = 0; i < width; ++i) {
     int r = src_sobelx[i];
     int b = src_sobely[i];
     int s = clamp255(r + b);
-    dst_argb[0] = (uint8)(s);
-    dst_argb[1] = (uint8)(s);
-    dst_argb[2] = (uint8)(s);
-    dst_argb[3] = (uint8)(255u);
+    dst_argb[0] = (uint8_t)(s);
+    dst_argb[1] = (uint8_t)(s);
+    dst_argb[2] = (uint8_t)(s);
+    dst_argb[3] = (uint8_t)(255u);
     dst_argb += 4;
   }
 }
 
-void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
-                       uint8* dst_y, int width) {
+void SobelToPlaneRow_C(const uint8_t* src_sobelx,
+                       const uint8_t* src_sobely,
+                       uint8_t* dst_y,
+                       int width) {
   int i;
   for (i = 0; i < width; ++i) {
     int r = src_sobelx[i];
     int b = src_sobely[i];
     int s = clamp255(r + b);
-    dst_y[i] = (uint8)(s);
+    dst_y[i] = (uint8_t)(s);
   }
 }
 
-void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
-                  uint8* dst_argb, int width) {
+void SobelXYRow_C(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width) {
   int i;
   for (i = 0; i < width; ++i) {
     int r = src_sobelx[i];
     int b = src_sobely[i];
     int g = clamp255(r + b);
-    dst_argb[0] = (uint8)(b);
-    dst_argb[1] = (uint8)(g);
-    dst_argb[2] = (uint8)(r);
-    dst_argb[3] = (uint8)(255u);
+    dst_argb[0] = (uint8_t)(b);
+    dst_argb[1] = (uint8_t)(g);
+    dst_argb[2] = (uint8_t)(r);
+    dst_argb[3] = (uint8_t)(255u);
     dst_argb += 4;
   }
 }
 
-void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
+void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width) {
   // Copy a Y to RGB.
   int x;
   for (x = 0; x < width; ++x) {
-    uint8 y = src_y[0];
+    uint8_t y = src_y[0];
     dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
     dst_argb[3] = 255u;
     dst_argb += 4;
@@ -965,390 +1480,576 @@ void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
   }
 }
 
-// TODO(fbarchard): Unify these structures to be platform independent.
-// TODO(fbarchard): Generate SIMD structures from float matrix.
+// Macros to create SIMD specific yuv to rgb conversion constants.
 
-// BT.601 YUV to RGB reference
-//  R = (Y - 16) * 1.164              - V * -1.596
-//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
-//  B = (Y - 16) * 1.164 - U * -2.018
+// clang-format off
 
-// Y contribution to R,G,B.  Scale and bias.
-#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
-#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+#if defined(__aarch64__) || defined(__arm__)
+// Bias values include subtract 128 from U and V, bias from Y and rounding.
+// For B and R bias is negative. For G bias is positive.
+#define YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR)                             \
+  {{UB, VR, UG, VG, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},                     \
+   {YG, (UB * 128 - YB), (UG * 128 + VG * 128 + YB), (VR * 128 - YB), YB, 0, \
+    0, 0}}
+#else
+#define YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR)                     \
+  {{UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,          \
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0},         \
+   {UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,  \
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG}, \
+   {0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,          \
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR},         \
+   {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}, \
+   {YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB}}
+#endif
+
+// clang-format on
+
+#define MAKEYUVCONSTANTS(name, YG, YB, UB, UG, VG, VR)            \
+  const struct YuvConstants SIMD_ALIGNED(kYuv##name##Constants) = \
+      YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR);                   \
+  const struct YuvConstants SIMD_ALIGNED(kYvu##name##Constants) = \
+      YUVCONSTANTSBODY(YG, YB, VR, VG, UG, UB);
+
+// TODO(fbarchard): Generate SIMD structures from float matrix.
+
+// BT.601 limited range YUV to RGB reference
+//  R = (Y - 16) * 1.164             + V * 1.596
+//  G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813
+//  B = (Y - 16) * 1.164 + U * 2.018
+// KR = 0.299; KB = 0.114
 
 // U and V contributions to R,G,B.
-#define UB -128 /* max(-128, round(-2.018 * 64)) */
-#define UG 25 /* round(0.391 * 64) */
-#define VG 52 /* round(0.813 * 64) */
-#define VR -102 /* round(-1.596 * 64) */
-
-// Bias values to subtract 16 from Y and 128 from U and V.
-#define BB (UB * 128            + YGB)
-#define BG (UG * 128 + VG * 128 + YGB)
-#define BR            (VR * 128 + YGB)
-
-#if defined(__aarch64__)
-const YuvConstants SIMD_ALIGNED(kYuvI601Constants) = {
-  { -UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR },
-  { -UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR },
-  { UG, VG, UG, VG, UG, VG, UG, VG },
-  { UG, VG, UG, VG, UG, VG, UG, VG },
-  { BB, BG, BR, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-const YuvConstants SIMD_ALIGNED(kYvuI601Constants) = {
-  { -VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB },
-  { -VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB },
-  { VG, UG, VG, UG, VG, UG, VG, UG },
-  { VG, UG, VG, UG, VG, UG, VG, UG },
-  { BR, BG, BB, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-#elif defined(__arm__)
-const YuvConstants SIMD_ALIGNED(kYuvI601Constants) = {
-  { -UB, -UB, -UB, -UB, -VR, -VR, -VR, -VR, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { UG, UG, UG, UG, VG, VG, VG, VG, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { BB, BG, BR, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-const YuvConstants SIMD_ALIGNED(kYvuI601Constants) = {
-  { -VR, -VR, -VR, -VR, -UB, -UB, -UB, -UB, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { VG, VG, VG, VG, UG, UG, UG, UG, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { BR, BG, BB, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
+#if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT601)
+#define UB 129 /* round(2.018 * 64) */
 #else
-const YuvConstants SIMD_ALIGNED(kYuvI601Constants) = {
-  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
-    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
-  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
-  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
-    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
-  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
-  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
-  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
-  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
-};
-const YuvConstants SIMD_ALIGNED(kYvuI601Constants) = {
-  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
-    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
-  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
-  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
-    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
-  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
-  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
-  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
-  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
-};
+#define UB 128 /* max(128, round(2.018 * 64)) */
 #endif
+#define UG 25  /* round(0.391 * 64) */
+#define VG 52  /* round(0.813 * 64) */
+#define VR 102 /* round(1.596 * 64) */
 
-#undef BB
-#undef BG
-#undef BR
-#undef YGB
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+MAKEYUVCONSTANTS(I601, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
 #undef UB
 #undef UG
 #undef VG
 #undef VR
-#undef YG
 
-// JPEG YUV to RGB reference
-// *  R = Y                - V * -1.40200
-// *  G = Y - U *  0.34414 - V *  0.71414
-// *  B = Y - U * -1.77200
+// BT.601 full range YUV to RGB reference (aka JPEG)
+// *  R = Y               + V * 1.40200
+// *  G = Y - U * 0.34414 - V * 0.71414
+// *  B = Y + U * 1.77200
+// KR = 0.299; KB = 0.114
+
+// U and V contributions to R,G,B.
+#define UB 113 /* round(1.77200 * 64) */
+#define UG 22  /* round(0.34414 * 64) */
+#define VG 46  /* round(0.71414 * 64) */
+#define VR 90  /* round(1.40200 * 64) */
 
 // Y contribution to R,G,B.  Scale and bias.
 #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
-#define YGB 32  /* 64 / 2 */
+#define YB 32    /* 64 / 2 */
+
+MAKEYUVCONSTANTS(JPEG, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+
+// BT.709 limited range YUV to RGB reference
+//  R = (Y - 16) * 1.164             + V * 1.793
+//  G = (Y - 16) * 1.164 - U * 0.213 - V * 0.533
+//  B = (Y - 16) * 1.164 + U * 2.112
+//  KR = 0.2126, KB = 0.0722
 
 // U and V contributions to R,G,B.
-#define UB -113 /* round(-1.77200 * 64) */
-#define UG 22 /* round(0.34414 * 64) */
-#define VG 46 /* round(0.71414  * 64) */
-#define VR -90 /* round(-1.40200 * 64) */
-
-// Bias values to round, and subtract 128 from U and V.
-#define BB (UB * 128            + YGB)
-#define BG (UG * 128 + VG * 128 + YGB)
-#define BR            (VR * 128 + YGB)
-
-#if defined(__aarch64__)
-const YuvConstants SIMD_ALIGNED(kYuvJPEGConstants) = {
-  { -UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR },
-  { -UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR },
-  { UG, VG, UG, VG, UG, VG, UG, VG },
-  { UG, VG, UG, VG, UG, VG, UG, VG },
-  { BB, BG, BR, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-const YuvConstants SIMD_ALIGNED(kYvuJPEGConstants) = {
-  { -VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB },
-  { -VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB },
-  { VG, UG, VG, UG, VG, UG, VG, UG },
-  { VG, UG, VG, UG, VG, UG, VG, UG },
-  { BR, BG, BB, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-#elif defined(__arm__)
-const YuvConstants SIMD_ALIGNED(kYuvJPEGConstants) = {
-  { -UB, -UB, -UB, -UB, -VR, -VR, -VR, -VR, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { UG, UG, UG, UG, VG, VG, VG, VG, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { BB, BG, BR, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-const YuvConstants SIMD_ALIGNED(kYvuJPEGConstants) = {
-  { -VR, -VR, -VR, -VR, -UB, -UB, -UB, -UB, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { VG, VG, VG, VG, UG, UG, UG, UG, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { BR, BG, BB, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
+#if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT709)
+#define UB 135 /* round(2.112 * 64) */
 #else
-const YuvConstants SIMD_ALIGNED(kYuvJPEGConstants) = {
-  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
-    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
-  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
-  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
-    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
-  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
-  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
-  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
-  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
-};
-const YuvConstants SIMD_ALIGNED(kYvuJPEGConstants) = {
-  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
-    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
-  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
-  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
-    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
-  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
-  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
-  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
-  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
-};
+#define UB 128 /* max(128, round(2.112 * 64)) */
 #endif
+#define UG 14  /* round(0.213 * 64) */
+#define VG 34  /* round(0.533 * 64) */
+#define VR 115 /* round(1.793 * 64) */
 
-#undef BB
-#undef BG
-#undef BR
-#undef YGB
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+MAKEYUVCONSTANTS(H709, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
 #undef UB
 #undef UG
 #undef VG
 #undef VR
-#undef YG
 
-// BT.709 YUV to RGB reference
-//  R = (Y - 16) * 1.164              - V * -1.793
-//  G = (Y - 16) * 1.164 - U *  0.213 - V *  0.533
-//  B = (Y - 16) * 1.164 - U * -2.112
+// BT.709 full range YUV to RGB reference
+//  R = Y               + V * 1.5748
+//  G = Y - U * 0.18732 - V * 0.46812
+//  B = Y + U * 1.8556
+//  KR = 0.2126, KB = 0.0722
 
-// Y contribution to R,G,B.  Scale and bias.
-#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
-#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+// U and V contributions to R,G,B.
+#define UB 119 /* round(1.8556 * 64) */
+#define UG 12  /* round(0.18732 * 64) */
+#define VG 30  /* round(0.46812 * 64) */
+#define VR 101 /* round(1.5748 * 64) */
+
+// Y contribution to R,G,B.  Scale and bias.  (same as jpeg)
+#define YG 16320 /* round(1 * 64 * 256 * 256 / 257) */
+#define YB 32    /* 64 / 2 */
+
+MAKEYUVCONSTANTS(F709, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+
+// BT.2020 limited range YUV to RGB reference
+//  R = (Y - 16) * 1.164384                + V * 1.67867
+//  G = (Y - 16) * 1.164384 - U * 0.187326 - V * 0.65042
+//  B = (Y - 16) * 1.164384 + U * 2.14177
+// KR = 0.2627; KB = 0.0593
 
-// TODO(fbarchard): Find way to express 2.112 instead of 2.0.
 // U and V contributions to R,G,B.
-#define UB -128 /* max(-128, round(-2.112 * 64)) */
-#define UG 14 /* round(0.213 * 64) */
-#define VG 34 /* round(0.533  * 64) */
-#define VR -115 /* round(-1.793 * 64) */
-
-// Bias values to round, and subtract 128 from U and V.
-#define BB (UB * 128            + YGB)
-#define BG (UG * 128 + VG * 128 + YGB)
-#define BR            (VR * 128 + YGB)
-
-#if defined(__aarch64__)
-const YuvConstants SIMD_ALIGNED(kYuvH709Constants) = {
-  { -UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR },
-  { -UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR },
-  { UG, VG, UG, VG, UG, VG, UG, VG },
-  { UG, VG, UG, VG, UG, VG, UG, VG },
-  { BB, BG, BR, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-const YuvConstants SIMD_ALIGNED(kYvuH709Constants) = {
-  { -VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB },
-  { -VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB },
-  { VG, UG, VG, UG, VG, UG, VG, UG },
-  { VG, UG, VG, UG, VG, UG, VG, UG },
-  { BR, BG, BB, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-#elif defined(__arm__)
-const YuvConstants SIMD_ALIGNED(kYuvH709Constants) = {
-  { -UB, -UB, -UB, -UB, -VR, -VR, -VR, -VR, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { UG, UG, UG, UG, VG, VG, VG, VG, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { BB, BG, BR, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
-const YuvConstants SIMD_ALIGNED(kYvuH709Constants) = {
-  { -VR, -VR, -VR, -VR, -UB, -UB, -UB, -UB, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { VG, VG, VG, VG, UG, UG, UG, UG, 0, 0, 0, 0, 0, 0, 0, 0 },
-  { BR, BG, BB, 0, 0, 0, 0, 0 },
-  { 0x0101 * YG, 0, 0, 0 }
-};
+#if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT2020)
+#define UB 137 /* round(2.142 * 64) */
 #else
-const YuvConstants SIMD_ALIGNED(kYuvH709Constants) = {
-  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
-    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
-  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
-  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
-    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
-  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
-  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
-  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
-  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
-};
-const YuvConstants SIMD_ALIGNED(kYvuH709Constants) = {
-  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
-    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
-  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
-  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
-    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
-  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
-  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
-  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
-  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
-};
+#define UB 128 /* max(128, round(2.142 * 64)) */
 #endif
+#define UG 12  /* round(0.187326 * 64) */
+#define VG 42  /* round(0.65042 * 64) */
+#define VR 107 /* round(1.67867 * 64) */
 
-#undef BB
-#undef BG
-#undef BR
-#undef YGB
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 19003 /* round(1.164384 * 64 * 256 * 256 / 257) */
+#define YB -1160 /* 1.164384 * 64 * -16 + 64 / 2 */
+
+MAKEYUVCONSTANTS(2020, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
 #undef UB
 #undef UG
 #undef VG
 #undef VR
+
+// BT.2020 full range YUV to RGB reference
+//  R = Y                + V * 1.474600
+//  G = Y - U * 0.164553 - V * 0.571353
+//  B = Y + U * 1.881400
+// KR = 0.2627; KB = 0.0593
+
+#define UB 120 /* round(1.881400 * 64) */
+#define UG 11  /* round(0.164553 * 64) */
+#define VG 37  /* round(0.571353 * 64) */
+#define VR 94  /* round(1.474600 * 64) */
+
+// Y contribution to R,G,B.  Scale and bias.  (same as jpeg)
+#define YG 16320 /* round(1 * 64 * 256 * 256 / 257) */
+#define YB 32    /* 64 / 2 */
+
+MAKEYUVCONSTANTS(V2020, YG, YB, UB, UG, VG, VR)
+
 #undef YG
+#undef YB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+
+#undef BB
+#undef BG
+#undef BR
+
+#undef MAKEYUVCONSTANTS
+
+#if defined(__aarch64__) || defined(__arm__)
+#define LOAD_YUV_CONSTANTS                 \
+  int ub = yuvconstants->kUVCoeff[0];      \
+  int vr = yuvconstants->kUVCoeff[1];      \
+  int ug = yuvconstants->kUVCoeff[2];      \
+  int vg = yuvconstants->kUVCoeff[3];      \
+  int yg = yuvconstants->kRGBCoeffBias[0]; \
+  int bb = yuvconstants->kRGBCoeffBias[1]; \
+  int bg = yuvconstants->kRGBCoeffBias[2]; \
+  int br = yuvconstants->kRGBCoeffBias[3]
+
+#define CALC_RGB16                         \
+  int32_t y1 = (uint32_t)(y32 * yg) >> 16; \
+  int b16 = y1 + (u * ub) - bb;            \
+  int g16 = y1 + bg - (u * ug + v * vg);   \
+  int r16 = y1 + (v * vr) - br
+#else
+#define LOAD_YUV_CONSTANTS           \
+  int ub = yuvconstants->kUVToB[0];  \
+  int ug = yuvconstants->kUVToG[0];  \
+  int vg = yuvconstants->kUVToG[1];  \
+  int vr = yuvconstants->kUVToR[1];  \
+  int yg = yuvconstants->kYToRgb[0]; \
+  int yb = yuvconstants->kYBiasToRgb[0]
+
+#define CALC_RGB16                                \
+  int32_t y1 = ((uint32_t)(y32 * yg) >> 16) + yb; \
+  int8_t ui = (int8_t)u;                          \
+  int8_t vi = (int8_t)v;                          \
+  ui -= 0x80;                                     \
+  vi -= 0x80;                                     \
+  int b16 = y1 + (ui * ub);                       \
+  int g16 = y1 - (ui * ug + vi * vg);             \
+  int r16 = y1 + (vi * vr)
+#endif
 
 // C reference code that mimics the YUV assembly.
-static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
-                              uint8* b, uint8* g, uint8* r,
+// Reads 8 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel(uint8_t y,
+                              uint8_t u,
+                              uint8_t v,
+                              uint8_t* b,
+                              uint8_t* g,
+                              uint8_t* r,
                               const struct YuvConstants* yuvconstants) {
-#if defined(__aarch64__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = -yuvconstants->kUVToRB[1];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#elif defined(__arm__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[4];
-  int vr = -yuvconstants->kUVToRB[4];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y * 0x0101;
+  CALC_RGB16;
+  *b = STATIC_CAST(uint8_t, Clamp((int32_t)(b16) >> 6));
+  *g = STATIC_CAST(uint8_t, Clamp((int32_t)(g16) >> 6));
+  *r = STATIC_CAST(uint8_t, Clamp((int32_t)(r16) >> 6));
+}
+
+// Reads 8 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel8_16(uint8_t y,
+                                  uint8_t u,
+                                  uint8_t v,
+                                  int* b,
+                                  int* g,
+                                  int* r,
+                                  const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y * 0x0101;
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
+}
+
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 10 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel10_16(uint16_t y,
+                                   uint16_t u,
+                                   uint16_t v,
+                                   int* b,
+                                   int* g,
+                                   int* r,
+                                   const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = (y << 6) | (y >> 4);
+  u = STATIC_CAST(uint8_t, clamp255(u >> 2));
+  v = STATIC_CAST(uint8_t, clamp255(v >> 2));
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
+}
+
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 12 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel12_16(int16_t y,
+                                   int16_t u,
+                                   int16_t v,
+                                   int* b,
+                                   int* g,
+                                   int* r,
+                                   const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = (y << 4) | (y >> 8);
+  u = STATIC_CAST(uint8_t, clamp255(u >> 4));
+  v = STATIC_CAST(uint8_t, clamp255(v >> 4));
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
+}
+
+// C reference code that mimics the YUV 10 bit assembly.
+// Reads 10 bit YUV and clamps down to 8 bit RGB.
+static __inline void YuvPixel10(uint16_t y,
+                                uint16_t u,
+                                uint16_t v,
+                                uint8_t* b,
+                                uint8_t* g,
+                                uint8_t* r,
+                                const struct YuvConstants* yuvconstants) {
+  int b16;
+  int g16;
+  int r16;
+  YuvPixel10_16(y, u, v, &b16, &g16, &r16, yuvconstants);
+  *b = STATIC_CAST(uint8_t, Clamp(b16 >> 6));
+  *g = STATIC_CAST(uint8_t, Clamp(g16 >> 6));
+  *r = STATIC_CAST(uint8_t, Clamp(r16 >> 6));
+}
+
+// C reference code that mimics the YUV 12 bit assembly.
+// Reads 12 bit YUV and clamps down to 8 bit RGB.
+static __inline void YuvPixel12(uint16_t y,
+                                uint16_t u,
+                                uint16_t v,
+                                uint8_t* b,
+                                uint8_t* g,
+                                uint8_t* r,
+                                const struct YuvConstants* yuvconstants) {
+  int b16;
+  int g16;
+  int r16;
+  YuvPixel12_16(y, u, v, &b16, &g16, &r16, yuvconstants);
+  *b = STATIC_CAST(uint8_t, Clamp(b16 >> 6));
+  *g = STATIC_CAST(uint8_t, Clamp(g16 >> 6));
+  *r = STATIC_CAST(uint8_t, Clamp(r16 >> 6));
+}
+
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 16 bit YUV and leaves result as 8 bit.
+static __inline void YuvPixel16_8(uint16_t y,
+                                  uint16_t u,
+                                  uint16_t v,
+                                  uint8_t* b,
+                                  uint8_t* g,
+                                  uint8_t* r,
+                                  const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y;
+  u = STATIC_CAST(uint16_t, clamp255(u >> 8));
+  v = STATIC_CAST(uint16_t, clamp255(v >> 8));
+  CALC_RGB16;
+  *b = STATIC_CAST(uint8_t, Clamp((int32_t)(b16) >> 6));
+  *g = STATIC_CAST(uint8_t, Clamp((int32_t)(g16) >> 6));
+  *r = STATIC_CAST(uint8_t, Clamp((int32_t)(r16) >> 6));
+}
+
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 16 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel16_16(uint16_t y,
+                                   uint16_t u,
+                                   uint16_t v,
+                                   int* b,
+                                   int* g,
+                                   int* r,
+                                   const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y;
+  u = STATIC_CAST(uint16_t, clamp255(u >> 8));
+  v = STATIC_CAST(uint16_t, clamp255(v >> 8));
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
+}
+
+// C reference code that mimics the YUV assembly.
+// Reads 8 bit YUV and leaves result as 8 bit.
+static __inline void YPixel(uint8_t y,
+                            uint8_t* b,
+                            uint8_t* g,
+                            uint8_t* r,
+                            const struct YuvConstants* yuvconstants) {
+#if defined(__aarch64__) || defined(__arm__)
+  int yg = yuvconstants->kRGBCoeffBias[0];
+  int ygb = yuvconstants->kRGBCoeffBias[4];
 #else
-  int ub = yuvconstants->kUVToB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = yuvconstants->kUVToR[1];
-  int bb = yuvconstants->kUVBiasB[0];
-  int bg = yuvconstants->kUVBiasG[0];
-  int br = yuvconstants->kUVBiasR[0];
+  int ygb = yuvconstants->kYBiasToRgb[0];
   int yg = yuvconstants->kYToRgb[0];
 #endif
+  uint32_t y1 = (uint32_t)(y * 0x0101 * yg) >> 16;
+  *b = STATIC_CAST(uint8_t, Clamp(((int32_t)(y1) + ygb) >> 6));
+  *g = STATIC_CAST(uint8_t, Clamp(((int32_t)(y1) + ygb) >> 6));
+  *r = STATIC_CAST(uint8_t, Clamp(((int32_t)(y1) + ygb) >> 6));
+}
 
-  uint32 y1 = (uint32)(y * 0x0101 * yg) >> 16;
-  *b = Clamp((int32)(-(u * ub         ) + y1 + bb) >> 6);
-  *g = Clamp((int32)(-(u * ug + v * vg) + y1 + bg) >> 6);
-  *r = Clamp((int32)(-(         v * vr) + y1 + br) >> 6);
+void I444ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
 }
 
-// Y contribution to R,G,B.  Scale and bias.
-#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
-#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+void I444ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_u,
+                      const uint8_t* src_v,
+                      uint8_t* rgb_buf,
+                      const struct YuvConstants* yuvconstants,
+                      int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 3;  // Advance 1 pixel.
+  }
+}
 
-// C reference code that mimics the YUV assembly.
-static __inline void YPixel(uint8 y, uint8* b, uint8* g, uint8* r) {
-  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
-  *b = Clamp((int32)(y1 + YGB) >> 6);
-  *g = Clamp((int32)(y1 + YGB) >> 6);
-  *r = Clamp((int32)(y1 + YGB) >> 6);
+// Also used for 420
+void I422ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+             rgb_buf + 6, yuvconstants);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = 255;
+  }
 }
 
-#undef YG
-#undef YGB
-
-#if !defined(LIBYUV_DISABLE_NEON) && \
-    (defined(__ARM_NEON__) || defined(__aarch64__) || defined(LIBYUV_NEON))
-// C mimic assembly.
-// TODO(fbarchard): Remove subsampling from Neon.
-void I444ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* rgb_buf,
+// 10 bit YUV to ARGB
+void I210ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint8 u = (src_u[0] + src_u[1] + 1) >> 1;
-    uint8 v = (src_v[0] + src_v[1] + 1) >> 1;
-    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2,
-             yuvconstants);
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6,
-             yuvconstants);
+    YuvPixel10(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+               rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_y += 2;
-    src_u += 2;
-    src_v += 2;
+    src_u += 1;
+    src_v += 1;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
-#else
-void I444ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* rgb_buf,
+
+void I410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
     src_y += 1;
     src_u += 1;
     src_v += 1;
-    rgb_buf += 4;  // Advance 1 pixel.
+    rgb_buf += 4;  // Advance 1 pixels.
   }
 }
-#endif
 
-// Also used for 420
-void I422ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* rgb_buf,
+void I210AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = STATIC_CAST(uint8_t, clamp255(src_a[0] >> 2));
+    YuvPixel10(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+               rgb_buf + 6, yuvconstants);
+    rgb_buf[7] = STATIC_CAST(uint8_t, clamp255(src_a[1] >> 2));
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    src_a += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = STATIC_CAST(uint8_t, clamp255(src_a[0] >> 2));
+  }
+}
+
+void I410AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = STATIC_CAST(uint8_t, clamp255(src_a[0] >> 2));
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    src_a += 1;
+    rgb_buf += 4;  // Advance 1 pixels.
+  }
+}
+
+// 12 bit YUV to ARGB
+void I212ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel12(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YuvPixel(src_y[1], src_u[0], src_v[0],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
+    YuvPixel12(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+               rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_y += 2;
     src_u += 1;
@@ -1356,26 +2057,243 @@ void I422ToARGBRow_C(const uint8* src_y,
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel12(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
 
-void I422AlphaToARGBRow_C(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          const uint8* src_a,
-                          uint8* rgb_buf,
+static void StoreAR30(uint8_t* rgb_buf, int b, int g, int r) {
+  uint32_t ar30;
+  b = b >> 4;  // convert 8 bit 10.6 to 10 bit.
+  g = g >> 4;
+  r = r >> 4;
+  b = Clamp10(b);
+  g = Clamp10(g);
+  r = Clamp10(r);
+  ar30 = b | ((uint32_t)g << 10) | ((uint32_t)r << 20) | 0xc0000000;
+  (*(uint32_t*)rgb_buf) = ar30;
+}
+
+// 10 bit YUV to 10 bit AR30
+void I210ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+    YuvPixel10_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf + 4, b, g, r);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+  }
+}
+
+// 12 bit YUV to 10 bit AR30
+void I212ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel12_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+    YuvPixel12_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf + 4, b, g, r);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel12_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+  }
+}
+
+void I410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width; ++x) {
+    YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+
+// P210 has 10 bits in msb of 16 bit NV12 style layout.
+void P210ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1,
+                 dst_argb + 2, yuvconstants);
+    dst_argb[3] = 255;
+    YuvPixel16_8(src_y[1], src_uv[0], src_uv[1], dst_argb + 4, dst_argb + 5,
+                 dst_argb + 6, yuvconstants);
+    dst_argb[7] = 255;
+    src_y += 2;
+    src_uv += 2;
+    dst_argb += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1,
+                 dst_argb + 2, yuvconstants);
+    dst_argb[3] = 255;
+  }
+}
+
+void P410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1,
+                 dst_argb + 2, yuvconstants);
+    dst_argb[3] = 255;
+    src_y += 1;
+    src_uv += 2;
+    dst_argb += 4;  // Advance 1 pixels.
+  }
+}
+
+void P210ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30, b, g, r);
+    YuvPixel16_16(src_y[1], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30 + 4, b, g, r);
+    src_y += 2;
+    src_uv += 2;
+    dst_ar30 += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30, b, g, r);
+  }
+}
+
+void P410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width; ++x) {
+    YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30, b, g, r);
+    src_y += 1;
+    src_uv += 2;
+    dst_ar30 += 4;  // Advance 1 pixel.
+  }
+}
+
+// 8 bit YUV to 10 bit AR30
+// Uses same code as 10 bit YUV bit shifts the 8 bit values up to 10 bits.
+void I422ToAR30Row_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel8_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+    YuvPixel8_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf + 4, b, g, r);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel8_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+  }
+}
+
+void I444AlphaToARGBRow_C(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          const uint8_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = src_a[0];
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    src_a += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+
+void I422AlphaToARGBRow_C(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          const uint8_t* src_a,
+                          uint8_t* rgb_buf,
                           const struct YuvConstants* yuvconstants,
                           int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = src_a[0];
-    YuvPixel(src_y[1], src_u[0], src_v[0],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
+    YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+             rgb_buf + 6, yuvconstants);
     rgb_buf[7] = src_a[1];
     src_y += 2;
     src_u += 1;
@@ -1384,47 +2302,47 @@ void I422AlphaToARGBRow_C(const uint8* src_y,
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = src_a[0];
   }
 }
 
-void I422ToRGB24Row_C(const uint8* src_y,
-                      const uint8* src_u,
-                      const uint8* src_v,
-                      uint8* rgb_buf,
+void I422ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_u,
+                      const uint8_t* src_v,
+                      uint8_t* rgb_buf,
                       const struct YuvConstants* yuvconstants,
                       int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
-    YuvPixel(src_y[1], src_u[0], src_v[0],
-             rgb_buf + 3, rgb_buf + 4, rgb_buf + 5, yuvconstants);
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 3, rgb_buf + 4,
+             rgb_buf + 5, yuvconstants);
     src_y += 2;
     src_u += 1;
     src_v += 1;
     rgb_buf += 6;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
   }
 }
 
-void I422ToARGB4444Row_C(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb4444,
+void I422ToARGB4444Row_C(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb4444,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  uint8 b0;
-  uint8 g0;
-  uint8 r0;
-  uint8 b1;
-  uint8 g1;
-  uint8 r1;
+  uint8_t b0;
+  uint8_t g0;
+  uint8_t r0;
+  uint8_t b1;
+  uint8_t g1;
+  uint8_t r1;
   int x;
   for (x = 0; x < width - 1; x += 2) {
     YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants);
@@ -1435,8 +2353,10 @@ void I422ToARGB4444Row_C(const uint8* src_y,
     b1 = b1 >> 4;
     g1 = g1 >> 4;
     r1 = r1 >> 4;
-    *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
-        (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000;
+    *(uint16_t*)(dst_argb4444 + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | 0xf000);
+    *(uint16_t*)(dst_argb4444 + 2) =
+        STATIC_CAST(uint16_t, b1 | (g1 << 4) | (r1 << 8) | 0xf000);
     src_y += 2;
     src_u += 1;
     src_v += 1;
@@ -1447,23 +2367,23 @@ void I422ToARGB4444Row_C(const uint8* src_y,
     b0 = b0 >> 4;
     g0 = g0 >> 4;
     r0 = r0 >> 4;
-    *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
-        0xf000;
+    *(uint16_t*)(dst_argb4444) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | 0xf000);
   }
 }
 
-void I422ToARGB1555Row_C(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_argb1555,
+void I422ToARGB1555Row_C(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb1555,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  uint8 b0;
-  uint8 g0;
-  uint8 r0;
-  uint8 b1;
-  uint8 g1;
-  uint8 r1;
+  uint8_t b0;
+  uint8_t g0;
+  uint8_t r0;
+  uint8_t b1;
+  uint8_t g1;
+  uint8_t r1;
   int x;
   for (x = 0; x < width - 1; x += 2) {
     YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants);
@@ -1474,8 +2394,10 @@ void I422ToARGB1555Row_C(const uint8* src_y,
     b1 = b1 >> 3;
     g1 = g1 >> 3;
     r1 = r1 >> 3;
-    *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
-        (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000;
+    *(uint16_t*)(dst_argb1555 + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | 0x8000);
+    *(uint16_t*)(dst_argb1555 + 2) =
+        STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 10) | 0x8000);
     src_y += 2;
     src_u += 1;
     src_v += 1;
@@ -1486,23 +2408,23 @@ void I422ToARGB1555Row_C(const uint8* src_y,
     b0 = b0 >> 3;
     g0 = g0 >> 3;
     r0 = r0 >> 3;
-    *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
-        0x8000;
+    *(uint16_t*)(dst_argb1555) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | 0x8000);
   }
 }
 
-void I422ToRGB565Row_C(const uint8* src_y,
-                       const uint8* src_u,
-                       const uint8* src_v,
-                       uint8* dst_rgb565,
+void I422ToRGB565Row_C(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_rgb565,
                        const struct YuvConstants* yuvconstants,
                        int width) {
-  uint8 b0;
-  uint8 g0;
-  uint8 r0;
-  uint8 b1;
-  uint8 g1;
-  uint8 r1;
+  uint8_t b0;
+  uint8_t g0;
+  uint8_t r0;
+  uint8_t b1;
+  uint8_t g1;
+  uint8_t r1;
   int x;
   for (x = 0; x < width - 1; x += 2) {
     YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants);
@@ -1513,8 +2435,10 @@ void I422ToRGB565Row_C(const uint8* src_y,
     b1 = b1 >> 3;
     g1 = g1 >> 2;
     r1 = r1 >> 3;
-    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
-        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    *(uint16_t*)(dst_rgb565 + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11));
+    *(uint16_t*)(dst_rgb565 + 2) =
+        STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 11));
     src_y += 2;
     src_u += 1;
     src_v += 1;
@@ -1525,111 +2449,112 @@ void I422ToRGB565Row_C(const uint8* src_y,
     b0 = b0 >> 3;
     g0 = g0 >> 2;
     r0 = r0 >> 3;
-    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+    *(uint16_t*)(dst_rgb565 + 0) =
+        STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11));
   }
 }
 
-void I411ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* rgb_buf,
+void NV12ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_uv,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
-  for (x = 0; x < width - 3; x += 4) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
-    rgb_buf[3] = 255;
-    YuvPixel(src_y[1], src_u[0], src_v[0],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
-    rgb_buf[7] = 255;
-    YuvPixel(src_y[2], src_u[0], src_v[0],
-             rgb_buf + 8, rgb_buf + 9, rgb_buf + 10, yuvconstants);
-    rgb_buf[11] = 255;
-    YuvPixel(src_y[3], src_u[0], src_v[0],
-             rgb_buf + 12, rgb_buf + 13, rgb_buf + 14, yuvconstants);
-    rgb_buf[15] = 255;
-    src_y += 4;
-    src_u += 1;
-    src_v += 1;
-    rgb_buf += 16;  // Advance 4 pixels.
-  }
-  if (width & 2) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YuvPixel(src_y[1], src_u[0], src_v[0],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
+    YuvPixel(src_y[1], src_uv[0], src_uv[1], rgb_buf + 4, rgb_buf + 5,
+             rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_y += 2;
+    src_uv += 2;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
 
-void NV12ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_uv,
-                     uint8* rgb_buf,
+void NV21ToARGBRow_C(const uint8_t* src_y,
+                     const uint8_t* src_vu,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], src_uv[0], src_uv[1],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YuvPixel(src_y[1], src_uv[0], src_uv[1],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
+    YuvPixel(src_y[1], src_vu[1], src_vu[0], rgb_buf + 4, rgb_buf + 5,
+             rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_y += 2;
-    src_uv += 2;
+    src_vu += 2;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_uv[0], src_uv[1],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
 
-void NV21ToARGBRow_C(const uint8* src_y,
-                     const uint8* src_vu,
-                     uint8* rgb_buf,
-                     const struct YuvConstants* yuvconstants,
-                     int width) {
+void NV12ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_uv,
+                      uint8_t* rgb_buf,
+                      const struct YuvConstants* yuvconstants,
+                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], src_vu[1], src_vu[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
-    rgb_buf[3] = 255;
-    YuvPixel(src_y[1], src_vu[1], src_vu[0],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
-    rgb_buf[7] = 255;
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[1], src_uv[0], src_uv[1], rgb_buf + 3, rgb_buf + 4,
+             rgb_buf + 5, yuvconstants);
+    src_y += 2;
+    src_uv += 2;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+  }
+}
+
+void NV21ToRGB24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_vu,
+                      uint8_t* rgb_buf,
+                      const struct YuvConstants* yuvconstants,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    YuvPixel(src_y[1], src_vu[1], src_vu[0], rgb_buf + 3, rgb_buf + 4,
+             rgb_buf + 5, yuvconstants);
     src_y += 2;
     src_vu += 2;
-    rgb_buf += 8;  // Advance 2 pixels.
+    rgb_buf += 6;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_vu[1], src_vu[0],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
-    rgb_buf[3] = 255;
+    YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
   }
 }
 
-void NV12ToRGB565Row_C(const uint8* src_y,
-                       const uint8* src_uv,
-                       uint8* dst_rgb565,
+void NV12ToRGB565Row_C(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_rgb565,
                        const struct YuvConstants* yuvconstants,
                        int width) {
-  uint8 b0;
-  uint8 g0;
-  uint8 r0;
-  uint8 b1;
-  uint8 g1;
-  uint8 r1;
+  uint8_t b0;
+  uint8_t g0;
+  uint8_t r0;
+  uint8_t b1;
+  uint8_t g1;
+  uint8_t r1;
   int x;
   for (x = 0; x < width - 1; x += 2) {
     YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0, yuvconstants);
@@ -1640,8 +2565,12 @@ void NV12ToRGB565Row_C(const uint8* src_y,
     b1 = b1 >> 3;
     g1 = g1 >> 2;
     r1 = r1 >> 3;
-    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
-        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    *(uint16_t*)(dst_rgb565 + 0) = STATIC_CAST(uint16_t, b0) |
+                                   STATIC_CAST(uint16_t, g0 << 5) |
+                                   STATIC_CAST(uint16_t, r0 << 11);
+    *(uint16_t*)(dst_rgb565 + 2) = STATIC_CAST(uint16_t, b1) |
+                                   STATIC_CAST(uint16_t, g1 << 5) |
+                                   STATIC_CAST(uint16_t, r1 << 11);
     src_y += 2;
     src_uv += 2;
     dst_rgb565 += 4;  // Advance 2 pixels.
@@ -1651,67 +2580,69 @@ void NV12ToRGB565Row_C(const uint8* src_y,
     b0 = b0 >> 3;
     g0 = g0 >> 2;
     r0 = r0 >> 3;
-    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+    *(uint16_t*)(dst_rgb565) = STATIC_CAST(uint16_t, b0) |
+                               STATIC_CAST(uint16_t, g0 << 5) |
+                               STATIC_CAST(uint16_t, r0 << 11);
   }
 }
 
-void YUY2ToARGBRow_C(const uint8* src_yuy2,
-                     uint8* rgb_buf,
+void YUY2ToARGBRow_C(const uint8_t* src_yuy2,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
+    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3], rgb_buf + 4, rgb_buf + 5,
+             rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_yuy2 += 4;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
 
-void UYVYToARGBRow_C(const uint8* src_uyvy,
-                     uint8* rgb_buf,
+void UYVYToARGBRow_C(const uint8_t* src_uyvy,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2],
-             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
+    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2], rgb_buf + 4, rgb_buf + 5,
+             rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_uyvy += 4;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
-             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
 
-void I422ToRGBARow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* rgb_buf,
+void I422ToRGBARow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3, yuvconstants);
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 1, rgb_buf + 2,
+             rgb_buf + 3, yuvconstants);
     rgb_buf[0] = 255;
-    YuvPixel(src_y[1], src_u[0], src_v[0],
-             rgb_buf + 5, rgb_buf + 6, rgb_buf + 7, yuvconstants);
+    YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 5, rgb_buf + 6,
+             rgb_buf + 7, yuvconstants);
     rgb_buf[4] = 255;
     src_y += 2;
     src_u += 1;
@@ -1719,29 +2650,32 @@ void I422ToRGBARow_C(const uint8* src_y,
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0],
-             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3, yuvconstants);
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 1, rgb_buf + 2,
+             rgb_buf + 3, yuvconstants);
     rgb_buf[0] = 255;
   }
 }
 
-void I400ToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
+void I400ToARGBRow_C(const uint8_t* src_y,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_y += 2;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
 
-void MirrorRow_C(const uint8* src, uint8* dst, int width) {
+void MirrorRow_C(const uint8_t* src, uint8_t* dst, int width) {
   int x;
   src += width - 1;
   for (x = 0; x < width - 1; x += 2) {
@@ -1754,7 +2688,34 @@ void MirrorRow_C(const uint8* src, uint8* dst, int width) {
   }
 }
 
-void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+void MirrorRow_16_C(const uint16_t* src, uint16_t* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
+void MirrorUVRow_C(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width; ++x) {
+    dst_uv[0] = src_uv[0];
+    dst_uv[1] = src_uv[1];
+    src_uv -= 2;
+    dst_uv += 2;
+  }
+}
+
+void MirrorSplitUVRow_C(const uint8_t* src_uv,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
   int x;
   src_uv += (width - 1) << 1;
   for (x = 0; x < width - 1; x += 2) {
@@ -1770,10 +2731,10 @@ void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
   }
 }
 
-void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
+void ARGBMirrorRow_C(const uint8_t* src, uint8_t* dst, int width) {
   int x;
-  const uint32* src32 = (const uint32*)(src);
-  uint32* dst32 = (uint32*)(dst);
+  const uint32_t* src32 = (const uint32_t*)(src);
+  uint32_t* dst32 = (uint32_t*)(dst);
   src32 += width - 1;
   for (x = 0; x < width - 1; x += 2) {
     dst32[x] = src32[0];
@@ -1785,7 +2746,25 @@ void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
   }
 }
 
-void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+void RGB24MirrorRow_C(const uint8_t* src_rgb24, uint8_t* dst_rgb24, int width) {
+  int x;
+  src_rgb24 += width * 3 - 3;
+  for (x = 0; x < width; ++x) {
+    uint8_t b = src_rgb24[0];
+    uint8_t g = src_rgb24[1];
+    uint8_t r = src_rgb24[2];
+    dst_rgb24[0] = b;
+    dst_rgb24[1] = g;
+    dst_rgb24[2] = r;
+    src_rgb24 -= 3;
+    dst_rgb24 += 3;
+  }
+}
+
+void SplitUVRow_C(const uint8_t* src_uv,
+                  uint8_t* dst_u,
+                  uint8_t* dst_v,
+                  int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
     dst_u[x] = src_uv[0];
@@ -1800,7 +2779,9 @@ void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
   }
 }
 
-void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_C(const uint8_t* src_u,
+                  const uint8_t* src_v,
+                  uint8_t* dst_uv,
                   int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
@@ -1816,29 +2797,402 @@ void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
   }
 }
 
-void CopyRow_C(const uint8* src, uint8* dst, int count) {
+void DetileRow_C(const uint8_t* src,
+                 ptrdiff_t src_tile_stride,
+                 uint8_t* dst,
+                 int width) {
+  int x;
+  for (x = 0; x < width - 15; x += 16) {
+    memcpy(dst, src, 16);
+    dst += 16;
+    src += src_tile_stride;
+  }
+  if (width & 15) {
+    memcpy(dst, src, width & 15);
+  }
+}
+
+void DetileRow_16_C(const uint16_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint16_t* dst,
+                    int width) {
+  int x;
+  for (x = 0; x < width - 15; x += 16) {
+    memcpy(dst, src, 16 * sizeof(uint16_t));
+    dst += 16;
+    src += src_tile_stride;
+  }
+  if (width & 15) {
+    memcpy(dst, src, (width & 15) * sizeof(uint16_t));
+  }
+}
+
+void DetileSplitUVRow_C(const uint8_t* src_uv,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  int x;
+  for (x = 0; x < width - 15; x += 16) {
+    SplitUVRow_C(src_uv, dst_u, dst_v, 8);
+    dst_u += 8;
+    dst_v += 8;
+    src_uv += src_tile_stride;
+  }
+  if (width & 15) {
+    SplitUVRow_C(src_uv, dst_u, dst_v, ((width & 15) + 1) / 2);
+  }
+}
+
+void DetileToYUY2_C(const uint8_t* src_y,
+                    ptrdiff_t src_y_tile_stride,
+                    const uint8_t* src_uv,
+                    ptrdiff_t src_uv_tile_stride,
+                    uint8_t* dst_yuy2,
+                    int width) {
+  for (int x = 0; x < width - 15; x += 16) {
+    for (int i = 0; i < 8; i++) {
+      dst_yuy2[0] = src_y[0];
+      dst_yuy2[1] = src_uv[0];
+      dst_yuy2[2] = src_y[1];
+      dst_yuy2[3] = src_uv[1];
+      dst_yuy2 += 4;
+      src_y += 2;
+      src_uv += 2;
+    }
+    src_y += src_y_tile_stride - 16;
+    src_uv += src_uv_tile_stride - 16;
+  }
+}
+
+// Unpack MT2T into tiled P010 64 pixels at a time. MT2T's bitstream is encoded
+// in 80 byte blocks representing 64 pixels each. The first 16 bytes of the
+// block contain all of the lower 2 bits of each pixel packed together, and the
+// next 64 bytes represent all the upper 8 bits of the pixel.
+void UnpackMT2T_C(const uint8_t* src, uint16_t* dst, size_t size) {
+  for (size_t i = 0; i < size; i += 80) {
+    const uint8_t* src_lower_bits = src;
+    const uint8_t* src_upper_bits = src + 16;
+
+    for (int j = 0; j < 16; j++) {
+      uint8_t lower_bits = src_lower_bits[j];
+      *dst++ = (lower_bits & 0x03) << 6 | (uint16_t)src_upper_bits[j * 4] << 8 |
+               (uint16_t)src_upper_bits[j * 4] >> 2;
+      *dst++ = (lower_bits & 0x0C) << 4 |
+               (uint16_t)src_upper_bits[j * 4 + 1] << 8 |
+               (uint16_t)src_upper_bits[j * 4 + 1] >> 2;
+      *dst++ = (lower_bits & 0x30) << 2 |
+               (uint16_t)src_upper_bits[j * 4 + 2] << 8 |
+               (uint16_t)src_upper_bits[j * 4 + 2] >> 2;
+      *dst++ = (lower_bits & 0xC0) | (uint16_t)src_upper_bits[j * 4 + 3] << 8 |
+               (uint16_t)src_upper_bits[j * 4 + 3] >> 2;
+    }
+
+    src += 80;
+  }
+}
+
+void SplitRGBRow_C(const uint8_t* src_rgb,
+                   uint8_t* dst_r,
+                   uint8_t* dst_g,
+                   uint8_t* dst_b,
+                   int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_r[x] = src_rgb[0];
+    dst_g[x] = src_rgb[1];
+    dst_b[x] = src_rgb[2];
+    src_rgb += 3;
+  }
+}
+
+void MergeRGBRow_C(const uint8_t* src_r,
+                   const uint8_t* src_g,
+                   const uint8_t* src_b,
+                   uint8_t* dst_rgb,
+                   int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_rgb[0] = src_r[x];
+    dst_rgb[1] = src_g[x];
+    dst_rgb[2] = src_b[x];
+    dst_rgb += 3;
+  }
+}
+
+void SplitARGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    uint8_t* dst_a,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_b[x] = src_argb[0];
+    dst_g[x] = src_argb[1];
+    dst_r[x] = src_argb[2];
+    dst_a[x] = src_argb[3];
+    src_argb += 4;
+  }
+}
+
+void MergeARGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    const uint8_t* src_a,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = src_b[x];
+    dst_argb[1] = src_g[x];
+    dst_argb[2] = src_r[x];
+    dst_argb[3] = src_a[x];
+    dst_argb += 4;
+  }
+}
+
+void MergeXR30Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint8_t* dst_ar30,
+                    int depth,
+                    int width) {
+  assert(depth >= 10);
+  assert(depth <= 16);
+  int x;
+  int shift = depth - 10;
+  uint32_t* dst_ar30_32 = (uint32_t*)dst_ar30;
+  for (x = 0; x < width; ++x) {
+    uint32_t r = clamp1023(src_r[x] >> shift);
+    uint32_t g = clamp1023(src_g[x] >> shift);
+    uint32_t b = clamp1023(src_b[x] >> shift);
+    dst_ar30_32[x] = b | (g << 10) | (r << 20) | 0xc0000000;
+  }
+}
+
+void MergeAR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    const uint16_t* src_a,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width) {
+  assert(depth >= 1);
+  assert(depth <= 16);
+  int x;
+  int shift = 16 - depth;
+  int max = (1 << depth) - 1;
+  for (x = 0; x < width; ++x) {
+    dst_ar64[0] = STATIC_CAST(uint16_t, ClampMax(src_b[x], max) << shift);
+    dst_ar64[1] = STATIC_CAST(uint16_t, ClampMax(src_g[x], max) << shift);
+    dst_ar64[2] = STATIC_CAST(uint16_t, ClampMax(src_r[x], max) << shift);
+    dst_ar64[3] = STATIC_CAST(uint16_t, ClampMax(src_a[x], max) << shift);
+    dst_ar64 += 4;
+  }
+}
+
+void MergeARGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         const uint16_t* src_a,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width) {
+  assert(depth >= 8);
+  assert(depth <= 16);
+  int x;
+  int shift = depth - 8;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = STATIC_CAST(uint8_t, clamp255(src_b[x] >> shift));
+    dst_argb[1] = STATIC_CAST(uint8_t, clamp255(src_g[x] >> shift));
+    dst_argb[2] = STATIC_CAST(uint8_t, clamp255(src_r[x] >> shift));
+    dst_argb[3] = STATIC_CAST(uint8_t, clamp255(src_a[x] >> shift));
+    dst_argb += 4;
+  }
+}
+
+void MergeXR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width) {
+  assert(depth >= 1);
+  assert(depth <= 16);
+  int x;
+  int shift = 16 - depth;
+  int max = (1 << depth) - 1;
+  for (x = 0; x < width; ++x) {
+    dst_ar64[0] = STATIC_CAST(uint16_t, ClampMax(src_b[x], max) << shift);
+    dst_ar64[1] = STATIC_CAST(uint16_t, ClampMax(src_g[x], max) << shift);
+    dst_ar64[2] = STATIC_CAST(uint16_t, ClampMax(src_r[x], max) << shift);
+    dst_ar64[3] = 0xffff;
+    dst_ar64 += 4;
+  }
+}
+
+void MergeXRGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width) {
+  assert(depth >= 8);
+  assert(depth <= 16);
+  int x;
+  int shift = depth - 8;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = STATIC_CAST(uint8_t, clamp255(src_b[x] >> shift));
+    dst_argb[1] = STATIC_CAST(uint8_t, clamp255(src_g[x] >> shift));
+    dst_argb[2] = STATIC_CAST(uint8_t, clamp255(src_r[x] >> shift));
+    dst_argb[3] = 0xff;
+    dst_argb += 4;
+  }
+}
+
+void SplitXRGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_b[x] = src_argb[0];
+    dst_g[x] = src_argb[1];
+    dst_r[x] = src_argb[2];
+    src_argb += 4;
+  }
+}
+
+void MergeXRGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = src_b[x];
+    dst_argb[1] = src_g[x];
+    dst_argb[2] = src_r[x];
+    dst_argb[3] = 255;
+    dst_argb += 4;
+  }
+}
+
+// Convert lsb formats to msb, depending on sample depth.
+void MergeUVRow_16_C(const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint16_t* dst_uv,
+                     int depth,
+                     int width) {
+  int shift = 16 - depth;
+  assert(depth >= 8);
+  assert(depth <= 16);
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_uv[0] = STATIC_CAST(uint16_t, src_u[x] << shift);
+    dst_uv[1] = STATIC_CAST(uint16_t, src_v[x] << shift);
+    dst_uv += 2;
+  }
+}
+
+// Convert msb formats to lsb, depending on sample depth.
+void SplitUVRow_16_C(const uint16_t* src_uv,
+                     uint16_t* dst_u,
+                     uint16_t* dst_v,
+                     int depth,
+                     int width) {
+  int shift = 16 - depth;
+  int x;
+  assert(depth >= 8);
+  assert(depth <= 16);
+  for (x = 0; x < width; ++x) {
+    dst_u[x] = src_uv[0] >> shift;
+    dst_v[x] = src_uv[1] >> shift;
+    src_uv += 2;
+  }
+}
+
+void MultiplyRow_16_C(const uint16_t* src_y,
+                      uint16_t* dst_y,
+                      int scale,
+                      int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_y[x] = STATIC_CAST(uint16_t, src_y[x] * scale);
+  }
+}
+
+void DivideRow_16_C(const uint16_t* src_y,
+                    uint16_t* dst_y,
+                    int scale,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_y[x] = (src_y[x] * scale) >> 16;
+  }
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+#define C16TO8(v, scale) clamp255(((v) * (scale)) >> 16)
+
+void Convert16To8Row_C(const uint16_t* src_y,
+                       uint8_t* dst_y,
+                       int scale,
+                       int width) {
+  int x;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+
+  for (x = 0; x < width; ++x) {
+    dst_y[x] = STATIC_CAST(uint8_t, C16TO8(src_y[x], scale));
+  }
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 1024 = 10 bits
+void Convert8To16Row_C(const uint8_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  int x;
+  scale *= 0x0101;  // replicates the byte.
+  for (x = 0; x < width; ++x) {
+    dst_y[x] = (src_y[x] * scale) >> 16;
+  }
+}
+
+void CopyRow_C(const uint8_t* src, uint8_t* dst, int count) {
   memcpy(dst, src, count);
 }
 
-void CopyRow_16_C(const uint16* src, uint16* dst, int count) {
+void CopyRow_16_C(const uint16_t* src, uint16_t* dst, int count) {
   memcpy(dst, src, count * 2);
 }
 
-void SetRow_C(uint8* dst, uint8 v8, int width) {
+void SetRow_C(uint8_t* dst, uint8_t v8, int width) {
   memset(dst, v8, width);
 }
 
-void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int width) {
-  uint32* d = (uint32*)(dst_argb);
+void ARGBSetRow_C(uint8_t* dst_argb, uint32_t v32, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    d[x] = v32;
+    memcpy(dst_argb + x * sizeof v32, &v32, sizeof v32);
   }
 }
 
 // Filter 2 rows of YUY2 UV's (422) into U and V (420).
-void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
-                   uint8* dst_u, uint8* dst_v, int width) {
+void YUY2ToUVRow_C(const uint8_t* src_yuy2,
+                   int src_stride_yuy2,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width) {
   // Output a row of UV values, filtering 2 rows of YUY2.
   int x;
   for (x = 0; x < width; x += 2) {
@@ -1850,9 +3204,26 @@ void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
   }
 }
 
+// Filter 2 rows of YUY2 UV's (422) into UV (NV12).
+void YUY2ToNVUVRow_C(const uint8_t* src_yuy2,
+                     int src_stride_yuy2,
+                     uint8_t* dst_uv,
+                     int width) {
+  // Output a row of UV values, filtering 2 rows of YUY2.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_uv[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1;
+    dst_uv[1] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1;
+    src_yuy2 += 4;
+    dst_uv += 2;
+  }
+}
+
 // Copy row of YUY2 UV's (422) into U and V (422).
-void YUY2ToUV422Row_C(const uint8* src_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
+void YUY2ToUV422Row_C(const uint8_t* src_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   // Output a row of UV values.
   int x;
   for (x = 0; x < width; x += 2) {
@@ -1865,7 +3236,7 @@ void YUY2ToUV422Row_C(const uint8* src_yuy2,
 }
 
 // Copy row of YUY2 Y's (422) into Y (420/422).
-void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
+void YUY2ToYRow_C(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
   // Output a row of Y values.
   int x;
   for (x = 0; x < width - 1; x += 2) {
@@ -1879,8 +3250,11 @@ void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
 }
 
 // Filter 2 rows of UYVY UV's (422) into U and V (420).
-void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
-                   uint8* dst_u, uint8* dst_v, int width) {
+void UYVYToUVRow_C(const uint8_t* src_uyvy,
+                   int src_stride_uyvy,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width) {
   // Output a row of UV values.
   int x;
   for (x = 0; x < width; x += 2) {
@@ -1893,8 +3267,10 @@ void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
 }
 
 // Copy row of UYVY UV's (422) into U and V (422).
-void UYVYToUV422Row_C(const uint8* src_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
+void UYVYToUV422Row_C(const uint8_t* src_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   // Output a row of UV values.
   int x;
   for (x = 0; x < width; x += 2) {
@@ -1907,7 +3283,7 @@ void UYVYToUV422Row_C(const uint8* src_uyvy,
 }
 
 // Copy row of UYVY Y's (422) into Y (420/422).
-void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
+void UYVYToYRow_C(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
   // Output a row of Y values.
   int x;
   for (x = 0; x < width - 1; x += 2) {
@@ -1920,62 +3296,67 @@ void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
   }
 }
 
-#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+#define BLEND(f, b, a) clamp255((((256 - a) * b) >> 8) + f)
 
-// Blend src_argb0 over src_argb1 and store to dst_argb.
-// dst_argb may be src_argb0 or src_argb1.
+// Blend src_argb over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb or src_argb1.
 // This code mimics the SSSE3 version for better testability.
-void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1,
-                    uint8* dst_argb, int width) {
+void ARGBBlendRow_C(const uint8_t* src_argb,
+                    const uint8_t* src_argb1,
+                    uint8_t* dst_argb,
+                    int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint32 fb = src_argb0[0];
-    uint32 fg = src_argb0[1];
-    uint32 fr = src_argb0[2];
-    uint32 a = src_argb0[3];
-    uint32 bb = src_argb1[0];
-    uint32 bg = src_argb1[1];
-    uint32 br = src_argb1[2];
-    dst_argb[0] = BLEND(fb, bb, a);
-    dst_argb[1] = BLEND(fg, bg, a);
-    dst_argb[2] = BLEND(fr, br, a);
+    uint32_t fb = src_argb[0];
+    uint32_t fg = src_argb[1];
+    uint32_t fr = src_argb[2];
+    uint32_t a = src_argb[3];
+    uint32_t bb = src_argb1[0];
+    uint32_t bg = src_argb1[1];
+    uint32_t br = src_argb1[2];
+    dst_argb[0] = STATIC_CAST(uint8_t, BLEND(fb, bb, a));
+    dst_argb[1] = STATIC_CAST(uint8_t, BLEND(fg, bg, a));
+    dst_argb[2] = STATIC_CAST(uint8_t, BLEND(fr, br, a));
     dst_argb[3] = 255u;
 
-    fb = src_argb0[4 + 0];
-    fg = src_argb0[4 + 1];
-    fr = src_argb0[4 + 2];
-    a = src_argb0[4 + 3];
+    fb = src_argb[4 + 0];
+    fg = src_argb[4 + 1];
+    fr = src_argb[4 + 2];
+    a = src_argb[4 + 3];
     bb = src_argb1[4 + 0];
     bg = src_argb1[4 + 1];
     br = src_argb1[4 + 2];
-    dst_argb[4 + 0] = BLEND(fb, bb, a);
-    dst_argb[4 + 1] = BLEND(fg, bg, a);
-    dst_argb[4 + 2] = BLEND(fr, br, a);
+    dst_argb[4 + 0] = STATIC_CAST(uint8_t, BLEND(fb, bb, a));
+    dst_argb[4 + 1] = STATIC_CAST(uint8_t, BLEND(fg, bg, a));
+    dst_argb[4 + 2] = STATIC_CAST(uint8_t, BLEND(fr, br, a));
     dst_argb[4 + 3] = 255u;
-    src_argb0 += 8;
+    src_argb += 8;
     src_argb1 += 8;
     dst_argb += 8;
   }
 
   if (width & 1) {
-    uint32 fb = src_argb0[0];
-    uint32 fg = src_argb0[1];
-    uint32 fr = src_argb0[2];
-    uint32 a = src_argb0[3];
-    uint32 bb = src_argb1[0];
-    uint32 bg = src_argb1[1];
-    uint32 br = src_argb1[2];
-    dst_argb[0] = BLEND(fb, bb, a);
-    dst_argb[1] = BLEND(fg, bg, a);
-    dst_argb[2] = BLEND(fr, br, a);
+    uint32_t fb = src_argb[0];
+    uint32_t fg = src_argb[1];
+    uint32_t fr = src_argb[2];
+    uint32_t a = src_argb[3];
+    uint32_t bb = src_argb1[0];
+    uint32_t bg = src_argb1[1];
+    uint32_t br = src_argb1[2];
+    dst_argb[0] = STATIC_CAST(uint8_t, BLEND(fb, bb, a));
+    dst_argb[1] = STATIC_CAST(uint8_t, BLEND(fg, bg, a));
+    dst_argb[2] = STATIC_CAST(uint8_t, BLEND(fr, br, a));
     dst_argb[3] = 255u;
   }
 }
 #undef BLEND
 
-#define UBLEND(f, b, a) (((a) * f) + ((255 - a) * b) + 255) >> 8
-void BlendPlaneRow_C(const uint8* src0, const uint8* src1,
-                     const uint8* alpha, uint8* dst, int width) {
+#define UBLEND(f, b, a) (((a)*f) + ((255 - a) * b) + 255) >> 8
+void BlendPlaneRow_C(const uint8_t* src0,
+                     const uint8_t* src1,
+                     const uint8_t* alpha,
+                     uint8_t* dst,
+                     int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
     dst[0] = UBLEND(src0[0], src1[0], alpha[0]);
@@ -1991,21 +3372,25 @@ void BlendPlaneRow_C(const uint8* src0, const uint8* src1,
 }
 #undef UBLEND
 
+#if LIBYUV_ATTENUATE_DUP
+// This code mimics the SSSE3 version for better testability.
 #define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+#else
+#define ATTENUATE(f, a) (f * a + 128) >> 8
+#endif
 
 // Multiply source RGB by alpha and store to destination.
-// This code mimics the SSSE3 version for better testability.
-void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+void ARGBAttenuateRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   int i;
   for (i = 0; i < width - 1; i += 2) {
-    uint32 b = src_argb[0];
-    uint32 g = src_argb[1];
-    uint32 r = src_argb[2];
-    uint32 a = src_argb[3];
+    uint32_t b = src_argb[0];
+    uint32_t g = src_argb[1];
+    uint32_t r = src_argb[2];
+    uint32_t a = src_argb[3];
     dst_argb[0] = ATTENUATE(b, a);
     dst_argb[1] = ATTENUATE(g, a);
     dst_argb[2] = ATTENUATE(r, a);
-    dst_argb[3] = a;
+    dst_argb[3] = STATIC_CAST(uint8_t, a);
     b = src_argb[4];
     g = src_argb[5];
     r = src_argb[6];
@@ -2013,20 +3398,20 @@ void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
     dst_argb[4] = ATTENUATE(b, a);
     dst_argb[5] = ATTENUATE(g, a);
     dst_argb[6] = ATTENUATE(r, a);
-    dst_argb[7] = a;
+    dst_argb[7] = STATIC_CAST(uint8_t, a);
     src_argb += 8;
     dst_argb += 8;
   }
 
   if (width & 1) {
-    const uint32 b = src_argb[0];
-    const uint32 g = src_argb[1];
-    const uint32 r = src_argb[2];
-    const uint32 a = src_argb[3];
+    const uint32_t b = src_argb[0];
+    const uint32_t g = src_argb[1];
+    const uint32_t r = src_argb[2];
+    const uint32_t a = src_argb[3];
     dst_argb[0] = ATTENUATE(b, a);
     dst_argb[1] = ATTENUATE(g, a);
     dst_argb[2] = ATTENUATE(r, a);
-    dst_argb[3] = a;
+    dst_argb[3] = STATIC_CAST(uint8_t, a);
   }
 }
 #undef ATTENUATE
@@ -2038,87 +3423,117 @@ void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
 // Reciprocal method is off by 1 on some values. ie 125
 // 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower.
 #define T(a) 0x01000000 + (0x10000 / a)
-const uint32 fixed_invtbl8[256] = {
-  0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07),
-  T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f),
-  T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17),
-  T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f),
-  T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27),
-  T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f),
-  T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37),
-  T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f),
-  T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47),
-  T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f),
-  T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57),
-  T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f),
-  T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67),
-  T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f),
-  T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77),
-  T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f),
-  T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87),
-  T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f),
-  T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97),
-  T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f),
-  T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7),
-  T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf),
-  T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7),
-  T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf),
-  T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7),
-  T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf),
-  T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7),
-  T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf),
-  T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7),
-  T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef),
-  T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7),
-  T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 };
+const uint32_t fixed_invtbl8[256] = {
+    0x01000000, 0x0100ffff, T(0x02), T(0x03),   T(0x04), T(0x05), T(0x06),
+    T(0x07),    T(0x08),    T(0x09), T(0x0a),   T(0x0b), T(0x0c), T(0x0d),
+    T(0x0e),    T(0x0f),    T(0x10), T(0x11),   T(0x12), T(0x13), T(0x14),
+    T(0x15),    T(0x16),    T(0x17), T(0x18),   T(0x19), T(0x1a), T(0x1b),
+    T(0x1c),    T(0x1d),    T(0x1e), T(0x1f),   T(0x20), T(0x21), T(0x22),
+    T(0x23),    T(0x24),    T(0x25), T(0x26),   T(0x27), T(0x28), T(0x29),
+    T(0x2a),    T(0x2b),    T(0x2c), T(0x2d),   T(0x2e), T(0x2f), T(0x30),
+    T(0x31),    T(0x32),    T(0x33), T(0x34),   T(0x35), T(0x36), T(0x37),
+    T(0x38),    T(0x39),    T(0x3a), T(0x3b),   T(0x3c), T(0x3d), T(0x3e),
+    T(0x3f),    T(0x40),    T(0x41), T(0x42),   T(0x43), T(0x44), T(0x45),
+    T(0x46),    T(0x47),    T(0x48), T(0x49),   T(0x4a), T(0x4b), T(0x4c),
+    T(0x4d),    T(0x4e),    T(0x4f), T(0x50),   T(0x51), T(0x52), T(0x53),
+    T(0x54),    T(0x55),    T(0x56), T(0x57),   T(0x58), T(0x59), T(0x5a),
+    T(0x5b),    T(0x5c),    T(0x5d), T(0x5e),   T(0x5f), T(0x60), T(0x61),
+    T(0x62),    T(0x63),    T(0x64), T(0x65),   T(0x66), T(0x67), T(0x68),
+    T(0x69),    T(0x6a),    T(0x6b), T(0x6c),   T(0x6d), T(0x6e), T(0x6f),
+    T(0x70),    T(0x71),    T(0x72), T(0x73),   T(0x74), T(0x75), T(0x76),
+    T(0x77),    T(0x78),    T(0x79), T(0x7a),   T(0x7b), T(0x7c), T(0x7d),
+    T(0x7e),    T(0x7f),    T(0x80), T(0x81),   T(0x82), T(0x83), T(0x84),
+    T(0x85),    T(0x86),    T(0x87), T(0x88),   T(0x89), T(0x8a), T(0x8b),
+    T(0x8c),    T(0x8d),    T(0x8e), T(0x8f),   T(0x90), T(0x91), T(0x92),
+    T(0x93),    T(0x94),    T(0x95), T(0x96),   T(0x97), T(0x98), T(0x99),
+    T(0x9a),    T(0x9b),    T(0x9c), T(0x9d),   T(0x9e), T(0x9f), T(0xa0),
+    T(0xa1),    T(0xa2),    T(0xa3), T(0xa4),   T(0xa5), T(0xa6), T(0xa7),
+    T(0xa8),    T(0xa9),    T(0xaa), T(0xab),   T(0xac), T(0xad), T(0xae),
+    T(0xaf),    T(0xb0),    T(0xb1), T(0xb2),   T(0xb3), T(0xb4), T(0xb5),
+    T(0xb6),    T(0xb7),    T(0xb8), T(0xb9),   T(0xba), T(0xbb), T(0xbc),
+    T(0xbd),    T(0xbe),    T(0xbf), T(0xc0),   T(0xc1), T(0xc2), T(0xc3),
+    T(0xc4),    T(0xc5),    T(0xc6), T(0xc7),   T(0xc8), T(0xc9), T(0xca),
+    T(0xcb),    T(0xcc),    T(0xcd), T(0xce),   T(0xcf), T(0xd0), T(0xd1),
+    T(0xd2),    T(0xd3),    T(0xd4), T(0xd5),   T(0xd6), T(0xd7), T(0xd8),
+    T(0xd9),    T(0xda),    T(0xdb), T(0xdc),   T(0xdd), T(0xde), T(0xdf),
+    T(0xe0),    T(0xe1),    T(0xe2), T(0xe3),   T(0xe4), T(0xe5), T(0xe6),
+    T(0xe7),    T(0xe8),    T(0xe9), T(0xea),   T(0xeb), T(0xec), T(0xed),
+    T(0xee),    T(0xef),    T(0xf0), T(0xf1),   T(0xf2), T(0xf3), T(0xf4),
+    T(0xf5),    T(0xf6),    T(0xf7), T(0xf8),   T(0xf9), T(0xfa), T(0xfb),
+    T(0xfc),    T(0xfd),    T(0xfe), 0x01000100};
 #undef T
 
-void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+#if LIBYUV_UNATTENUATE_DUP
+// This code mimics the Intel SIMD version for better testability.
+#define UNATTENUATE(f, ia) clamp255(((f | (f << 8)) * ia) >> 16)
+#else
+#define UNATTENUATE(f, ia) clamp255((f * ia) >> 8)
+#endif
+
+// mimics the Intel SIMD code for exactness.
+void ARGBUnattenuateRow_C(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    uint32 b = src_argb[0];
-    uint32 g = src_argb[1];
-    uint32 r = src_argb[2];
-    const uint32 a = src_argb[3];
-    const uint32 ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
-    b = (b * ia) >> 8;
-    g = (g * ia) >> 8;
-    r = (r * ia) >> 8;
+    uint32_t b = src_argb[0];
+    uint32_t g = src_argb[1];
+    uint32_t r = src_argb[2];
+    const uint32_t a = src_argb[3];
+    const uint32_t ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
+
     // Clamping should not be necessary but is free in assembly.
-    dst_argb[0] = clamp255(b);
-    dst_argb[1] = clamp255(g);
-    dst_argb[2] = clamp255(r);
-    dst_argb[3] = a;
+    dst_argb[0] = STATIC_CAST(uint8_t, UNATTENUATE(b, ia));
+    dst_argb[1] = STATIC_CAST(uint8_t, UNATTENUATE(g, ia));
+    dst_argb[2] = STATIC_CAST(uint8_t, UNATTENUATE(r, ia));
+    dst_argb[3] = STATIC_CAST(uint8_t, a);
     src_argb += 4;
     dst_argb += 4;
   }
 }
 
-void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
-                               const int32* previous_cumsum, int width) {
-  int32 row_sum[4] = {0, 0, 0, 0};
+void ComputeCumulativeSumRow_C(const uint8_t* row,
+                               int32_t* cumsum,
+                               const int32_t* previous_cumsum,
+                               int width) {
+  int32_t row_sum[4] = {0, 0, 0, 0};
   int x;
   for (x = 0; x < width; ++x) {
     row_sum[0] += row[x * 4 + 0];
     row_sum[1] += row[x * 4 + 1];
     row_sum[2] += row[x * 4 + 2];
     row_sum[3] += row[x * 4 + 3];
-    cumsum[x * 4 + 0] = row_sum[0]  + previous_cumsum[x * 4 + 0];
-    cumsum[x * 4 + 1] = row_sum[1]  + previous_cumsum[x * 4 + 1];
-    cumsum[x * 4 + 2] = row_sum[2]  + previous_cumsum[x * 4 + 2];
-    cumsum[x * 4 + 3] = row_sum[3]  + previous_cumsum[x * 4 + 3];
+    cumsum[x * 4 + 0] = row_sum[0] + previous_cumsum[x * 4 + 0];
+    cumsum[x * 4 + 1] = row_sum[1] + previous_cumsum[x * 4 + 1];
+    cumsum[x * 4 + 2] = row_sum[2] + previous_cumsum[x * 4 + 2];
+    cumsum[x * 4 + 3] = row_sum[3] + previous_cumsum[x * 4 + 3];
   }
 }
 
-void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
-                                int w, int area, uint8* dst, int count) {
-  float ooa = 1.0f / area;
+void CumulativeSumToAverageRow_C(const int32_t* tl,
+                                 const int32_t* bl,
+                                 int w,
+                                 int area,
+                                 uint8_t* dst,
+                                 int count) {
+  float ooa;
   int i;
+  assert(area != 0);
+
+  ooa = 1.0f / STATIC_CAST(float, area);
   for (i = 0; i < count; ++i) {
-    dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
-    dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
-    dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa);
-    dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa);
+    dst[0] =
+        (uint8_t)(STATIC_CAST(float, bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) *
+                  ooa);
+    dst[1] =
+        (uint8_t)(STATIC_CAST(float, bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) *
+                  ooa);
+    dst[2] =
+        (uint8_t)(STATIC_CAST(float, bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) *
+                  ooa);
+    dst[3] =
+        (uint8_t)(STATIC_CAST(float, bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) *
+                  ooa);
     dst += 4;
     tl += 4;
     bl += 4;
@@ -2127,8 +3542,11 @@ void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
 
 // Copy pixels from rotated source to destination row with a slope.
 LIBYUV_API
-void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
-                     uint8* dst_argb, const float* uv_dudv, int width) {
+void ARGBAffineRow_C(const uint8_t* src_argb,
+                     int src_argb_stride,
+                     uint8_t* dst_argb,
+                     const float* uv_dudv,
+                     int width) {
   int i;
   // Render a row of pixels from source into a buffer.
   float uv[2];
@@ -2137,9 +3555,8 @@ void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
   for (i = 0; i < width; ++i) {
     int x = (int)(uv[0]);
     int y = (int)(uv[1]);
-    *(uint32*)(dst_argb) =
-        *(const uint32*)(src_argb + y * src_argb_stride +
-                                         x * 4);
+    *(uint32_t*)(dst_argb) =
+        *(const uint32_t*)(src_argb + y * src_argb_stride + x * 4);
     dst_argb += 4;
     uv[0] += uv_dudv[2];
     uv[1] += uv_dudv[3];
@@ -2147,30 +3564,52 @@ void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
 }
 
 // Blend 2 rows into 1.
-static void HalfRow_C(const uint8* src_uv, ptrdiff_t src_uv_stride,
-                      uint8* dst_uv, int width) {
+static void HalfRow_C(const uint8_t* src_uv,
+                      ptrdiff_t src_uv_stride,
+                      uint8_t* dst_uv,
+                      int width) {
   int x;
   for (x = 0; x < width; ++x) {
     dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
   }
 }
 
-static void HalfRow_16_C(const uint16* src_uv, ptrdiff_t src_uv_stride,
-                         uint16* dst_uv, int width) {
+static void HalfRow_16_C(const uint16_t* src_uv,
+                         ptrdiff_t src_uv_stride,
+                         uint16_t* dst_uv,
+                         int width) {
   int x;
   for (x = 0; x < width; ++x) {
     dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
   }
 }
 
+static void HalfRow_16To8_C(const uint16_t* src_uv,
+                            ptrdiff_t src_uv_stride,
+                            uint8_t* dst_uv,
+                            int scale,
+                            int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_uv[x] = STATIC_CAST(
+        uint8_t,
+        C16TO8((src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1, scale));
+  }
+}
+
 // C version 2x2 -> 2x1.
-void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+void InterpolateRow_C(uint8_t* dst_ptr,
+                      const uint8_t* src_ptr,
                       ptrdiff_t src_stride,
-                      int width, int source_y_fraction) {
-  int y1_fraction = source_y_fraction ;
+                      int width,
+                      int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
   int y0_fraction = 256 - y1_fraction;
-  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
   int x;
+  assert(source_y_fraction >= 0);
+  assert(source_y_fraction < 256);
+
   if (y1_fraction == 0) {
     memcpy(dst_ptr, src_ptr, width);
     return;
@@ -2179,51 +3618,93 @@ void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
     HalfRow_C(src_ptr, src_stride, dst_ptr, width);
     return;
   }
-  for (x = 0; x < width - 1; x += 2) {
-    dst_ptr[0] =
-        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8;
-    dst_ptr[1] =
-        (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction + 128) >> 8;
-    src_ptr += 2;
-    src_ptr1 += 2;
-    dst_ptr += 2;
-  }
-  if (width & 1) {
-    dst_ptr[0] =
-        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8;
+  for (x = 0; x < width; ++x) {
+    dst_ptr[0] = STATIC_CAST(
+        uint8_t,
+        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8);
+    ++src_ptr;
+    ++src_ptr1;
+    ++dst_ptr;
   }
 }
 
-void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+// C version 2x2 -> 2x1.
+void InterpolateRow_16_C(uint16_t* dst_ptr,
+                         const uint16_t* src_ptr,
                          ptrdiff_t src_stride,
-                         int width, int source_y_fraction) {
+                         int width,
+                         int source_y_fraction) {
   int y1_fraction = source_y_fraction;
   int y0_fraction = 256 - y1_fraction;
-  const uint16* src_ptr1 = src_ptr + src_stride;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
   int x;
-  if (source_y_fraction == 0) {
+  assert(source_y_fraction >= 0);
+  assert(source_y_fraction < 256);
+
+  if (y1_fraction == 0) {
     memcpy(dst_ptr, src_ptr, width * 2);
     return;
   }
-  if (source_y_fraction == 128) {
+  if (y1_fraction == 128) {
     HalfRow_16_C(src_ptr, src_stride, dst_ptr, width);
     return;
   }
-  for (x = 0; x < width - 1; x += 2) {
-    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
-    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
-    src_ptr += 2;
-    src_ptr1 += 2;
-    dst_ptr += 2;
+  for (x = 0; x < width; ++x) {
+    dst_ptr[0] = STATIC_CAST(
+        uint16_t,
+        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8);
+    ++src_ptr;
+    ++src_ptr1;
+    ++dst_ptr;
   }
-  if (width & 1) {
-    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+}
+
+// C version 2x2 16 bit-> 2x1 8 bit.
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+
+void InterpolateRow_16To8_C(uint8_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int scale,
+                            int width,
+                            int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  int x;
+  assert(source_y_fraction >= 0);
+  assert(source_y_fraction < 256);
+
+  if (source_y_fraction == 0) {
+    Convert16To8Row_C(src_ptr, dst_ptr, scale, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_16To8_C(src_ptr, src_stride, dst_ptr, scale, width);
+    return;
+  }
+  for (x = 0; x < width; ++x) {
+    dst_ptr[0] = STATIC_CAST(
+        uint8_t,
+        C16TO8(
+            (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8,
+            scale));
+    src_ptr += 1;
+    src_ptr1 += 1;
+    dst_ptr += 1;
   }
 }
 
 // Use first 4 shuffler values to reorder ARGB channels.
-void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
-                      const uint8* shuffler, int width) {
+void ARGBShuffleRow_C(const uint8_t* src_argb,
+                      uint8_t* dst_argb,
+                      const uint8_t* shuffler,
+                      int width) {
   int index0 = shuffler[0];
   int index1 = shuffler[1];
   int index2 = shuffler[2];
@@ -2232,10 +3713,10 @@ void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
   int x;
   for (x = 0; x < width; ++x) {
     // To support in-place conversion.
-    uint8 b = src_argb[index0];
-    uint8 g = src_argb[index1];
-    uint8 r = src_argb[index2];
-    uint8 a = src_argb[index3];
+    uint8_t b = src_argb[index0];
+    uint8_t g = src_argb[index1];
+    uint8_t r = src_argb[index2];
+    uint8_t a = src_argb[index3];
     dst_argb[0] = b;
     dst_argb[1] = g;
     dst_argb[2] = r;
@@ -2245,10 +3726,11 @@ void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
   }
 }
 
-void I422ToYUY2Row_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_frame, int width) {
+void I422ToYUY2Row_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_frame,
+                     int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
     dst_frame[0] = src_y[0];
@@ -2268,10 +3750,11 @@ void I422ToYUY2Row_C(const uint8* src_y,
   }
 }
 
-void I422ToUYVYRow_C(const uint8* src_y,
-                     const uint8* src_u,
-                     const uint8* src_v,
-                     uint8* dst_frame, int width) {
+void I422ToUYVYRow_C(const uint8_t* src_y,
+                     const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_frame,
+                     int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
     dst_frame[0] = src_u[0];
@@ -2291,9 +3774,8 @@ void I422ToUYVYRow_C(const uint8* src_y,
   }
 }
 
-
-void ARGBPolynomialRow_C(const uint8* src_argb,
-                         uint8* dst_argb,
+void ARGBPolynomialRow_C(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
                          const float* poly,
                          int width) {
   int i;
@@ -2323,33 +3805,75 @@ void ARGBPolynomialRow_C(const uint8* src_argb,
     dr += poly[14] * r3;
     da += poly[15] * a3;
 
-    dst_argb[0] = Clamp((int32)(db));
-    dst_argb[1] = Clamp((int32)(dg));
-    dst_argb[2] = Clamp((int32)(dr));
-    dst_argb[3] = Clamp((int32)(da));
+    dst_argb[0] = STATIC_CAST(uint8_t, Clamp((int32_t)(db)));
+    dst_argb[1] = STATIC_CAST(uint8_t, Clamp((int32_t)(dg)));
+    dst_argb[2] = STATIC_CAST(uint8_t, Clamp((int32_t)(dr)));
+    dst_argb[3] = STATIC_CAST(uint8_t, Clamp((int32_t)(da)));
     src_argb += 4;
     dst_argb += 4;
   }
 }
 
-void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
-                             const uint8* luma, uint32 lumacoeff) {
-  uint32 bc = lumacoeff & 0xff;
-  uint32 gc = (lumacoeff >> 8) & 0xff;
-  uint32 rc = (lumacoeff >> 16) & 0xff;
+// Samples assumed to be unsigned in low 9, 10 or 12 bits.  Scale factor
+// adjust the source integer range to the half float range desired.
+
+// This magic constant is 2^-112. Multiplying by this
+// is the same as subtracting 112 from the exponent, which
+// is the difference in exponent bias between 32-bit and
+// 16-bit floats. Once we've done this subtraction, we can
+// simply extract the low bits of the exponent and the high
+// bits of the mantissa from our float and we're done.
+
+// Work around GCC 7 punning warning -Wstrict-aliasing
+#if defined(__GNUC__)
+typedef uint32_t __attribute__((__may_alias__)) uint32_alias_t;
+#else
+typedef uint32_t uint32_alias_t;
+#endif
+
+void HalfFloatRow_C(const uint16_t* src,
+                    uint16_t* dst,
+                    float scale,
+                    int width) {
+  int i;
+  float mult = 1.9259299444e-34f * scale;
+  for (i = 0; i < width; ++i) {
+    float value = src[i] * mult;
+    dst[i] = (uint16_t)((*(const uint32_alias_t*)&value) >> 13);
+  }
+}
+
+void ByteToFloatRow_C(const uint8_t* src, float* dst, float scale, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    float value = src[i] * scale;
+    dst[i] = value;
+  }
+}
+
+void ARGBLumaColorTableRow_C(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             int width,
+                             const uint8_t* luma,
+                             uint32_t lumacoeff) {
+  uint32_t bc = lumacoeff & 0xff;
+  uint32_t gc = (lumacoeff >> 8) & 0xff;
+  uint32_t rc = (lumacoeff >> 16) & 0xff;
 
   int i;
   for (i = 0; i < width - 1; i += 2) {
     // Luminance in rows, color values in columns.
-    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
-                           src_argb[2] * rc) & 0x7F00u) + luma;
-    const uint8* luma1;
+    const uint8_t* luma0 =
+        ((src_argb[0] * bc + src_argb[1] * gc + src_argb[2] * rc) & 0x7F00u) +
+        luma;
+    const uint8_t* luma1;
     dst_argb[0] = luma0[src_argb[0]];
     dst_argb[1] = luma0[src_argb[1]];
     dst_argb[2] = luma0[src_argb[2]];
     dst_argb[3] = src_argb[3];
-    luma1 = ((src_argb[4] * bc + src_argb[5] * gc +
-              src_argb[6] * rc) & 0x7F00u) + luma;
+    luma1 =
+        ((src_argb[4] * bc + src_argb[5] * gc + src_argb[6] * rc) & 0x7F00u) +
+        luma;
     dst_argb[4] = luma1[src_argb[4]];
     dst_argb[5] = luma1[src_argb[5]];
     dst_argb[6] = luma1[src_argb[6]];
@@ -2359,8 +3883,9 @@ void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
   }
   if (width & 1) {
     // Luminance in rows, color values in columns.
-    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
-                           src_argb[2] * rc) & 0x7F00u) + luma;
+    const uint8_t* luma0 =
+        ((src_argb[0] * bc + src_argb[1] * gc + src_argb[2] * rc) & 0x7F00u) +
+        luma;
     dst_argb[0] = luma0[src_argb[0]];
     dst_argb[1] = luma0[src_argb[1]];
     dst_argb[2] = luma0[src_argb[2]];
@@ -2368,7 +3893,7 @@ void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
   }
 }
 
-void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+void ARGBCopyAlphaRow_C(const uint8_t* src, uint8_t* dst, int width) {
   int i;
   for (i = 0; i < width - 1; i += 2) {
     dst[3] = src[3];
@@ -2381,7 +3906,7 @@ void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
   }
 }
 
-void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width) {
+void ARGBExtractAlphaRow_C(const uint8_t* src_argb, uint8_t* dst_a, int width) {
   int i;
   for (i = 0; i < width - 1; i += 2) {
     dst_a[0] = src_argb[3];
@@ -2394,7 +3919,7 @@ void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width) {
   }
 }
 
-void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
+void ARGBCopyYToAlphaRow_C(const uint8_t* src, uint8_t* dst, int width) {
   int i;
   for (i = 0; i < width - 1; i += 2) {
     dst[3] = src[0];
@@ -2410,16 +3935,16 @@ void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
 // Maximum temporary width for wrappers to process at a time, in pixels.
 #define MAXTWIDTH 2048
 
-#if !(defined(_MSC_VER) && defined(_M_IX86)) && \
+#if !(defined(_MSC_VER) && !defined(__clang__) && defined(_M_IX86)) && \
     defined(HAS_I422TORGB565ROW_SSSE3)
 // row_win.cc has asm version, but GCC uses 2 step wrapper.
-void I422ToRGB565Row_SSSE3(const uint8* src_y,
-                           const uint8* src_u,
-                           const uint8* src_v,
-                           uint8* dst_rgb565,
+void I422ToRGB565Row_SSSE3(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_rgb565,
                            const struct YuvConstants* yuvconstants,
                            int width) {
-  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, yuvconstants, twidth);
@@ -2434,14 +3959,14 @@ void I422ToRGB565Row_SSSE3(const uint8* src_y,
 #endif
 
 #if defined(HAS_I422TOARGB1555ROW_SSSE3)
-void I422ToARGB1555Row_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb1555,
+void I422ToARGB1555Row_SSSE3(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             uint8_t* dst_argb1555,
                              const struct YuvConstants* yuvconstants,
                              int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, yuvconstants, twidth);
@@ -2456,14 +3981,14 @@ void I422ToARGB1555Row_SSSE3(const uint8* src_y,
 #endif
 
 #if defined(HAS_I422TOARGB4444ROW_SSSE3)
-void I422ToARGB4444Row_SSSE3(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb4444,
+void I422ToARGB4444Row_SSSE3(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             uint8_t* dst_argb4444,
                              const struct YuvConstants* yuvconstants,
                              int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, yuvconstants, twidth);
@@ -2478,13 +4003,13 @@ void I422ToARGB4444Row_SSSE3(const uint8* src_y,
 #endif
 
 #if defined(HAS_NV12TORGB565ROW_SSSE3)
-void NV12ToRGB565Row_SSSE3(const uint8* src_y,
-                           const uint8* src_uv,
-                           uint8* dst_rgb565,
+void NV12ToRGB565Row_SSSE3(const uint8_t* src_y,
+                           const uint8_t* src_uv,
+                           uint8_t* dst_rgb565,
                            const struct YuvConstants* yuvconstants,
                            int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     NV12ToARGBRow_SSSE3(src_y, src_uv, row, yuvconstants, twidth);
@@ -2497,14 +4022,102 @@ void NV12ToRGB565Row_SSSE3(const uint8* src_y,
 }
 #endif
 
+#if defined(HAS_NV12TORGB24ROW_SSSE3)
+void NV12ToRGB24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb24,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_SSSE3(src_y, src_uv, row, yuvconstants, twidth);
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB24ROW_SSSE3)
+void NV21ToRGB24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_vu,
+                          uint8_t* dst_rgb24,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_SSSE3(src_y, src_vu, row, yuvconstants, twidth);
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV12TORGB24ROW_AVX2)
+void NV12ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_AVX2(src_y, src_uv, row, yuvconstants, twidth);
+#if defined(HAS_ARGBTORGB24ROW_AVX2)
+    ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth);
+#else
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+#endif
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB24ROW_AVX2)
+void NV21ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_AVX2(src_y, src_vu, row, yuvconstants, twidth);
+#if defined(HAS_ARGBTORGB24ROW_AVX2)
+    ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth);
+#else
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+#endif
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
 #if defined(HAS_I422TORGB565ROW_AVX2)
-void I422ToRGB565Row_AVX2(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          uint8* dst_rgb565,
+void I422ToRGB565Row_AVX2(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width) {
-  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth);
@@ -2523,14 +4136,14 @@ void I422ToRGB565Row_AVX2(const uint8* src_y,
 #endif
 
 #if defined(HAS_I422TOARGB1555ROW_AVX2)
-void I422ToARGB1555Row_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb1555,
+void I422ToARGB1555Row_AVX2(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
                             const struct YuvConstants* yuvconstants,
                             int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth);
@@ -2549,14 +4162,14 @@ void I422ToARGB1555Row_AVX2(const uint8* src_y,
 #endif
 
 #if defined(HAS_I422TOARGB4444ROW_AVX2)
-void I422ToARGB4444Row_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb4444,
+void I422ToARGB4444Row_AVX2(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
                             const struct YuvConstants* yuvconstants,
                             int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth);
@@ -2575,19 +4188,22 @@ void I422ToARGB4444Row_AVX2(const uint8* src_y,
 #endif
 
 #if defined(HAS_I422TORGB24ROW_AVX2)
-void I422ToRGB24Row_AVX2(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_rgb24,
-                            const struct YuvConstants* yuvconstants,
-                            int width) {
+void I422ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth);
-    // TODO(fbarchard): ARGBToRGB24Row_AVX2
+#if defined(HAS_ARGBTORGB24ROW_AVX2)
+    ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth);
+#else
     ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+#endif
     src_y += twidth;
     src_u += twidth / 2;
     src_v += twidth / 2;
@@ -2597,14 +4213,40 @@ void I422ToRGB24Row_AVX2(const uint8* src_y,
 }
 #endif
 
+#if defined(HAS_I444TORGB24ROW_AVX2)
+void I444ToRGB24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I444ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth);
+#if defined(HAS_ARGBTORGB24ROW_AVX2)
+    ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth);
+#else
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+#endif
+    src_y += twidth;
+    src_u += twidth;
+    src_v += twidth;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
 #if defined(HAS_NV12TORGB565ROW_AVX2)
-void NV12ToRGB565Row_AVX2(const uint8* src_y,
-                          const uint8* src_uv,
-                          uint8* dst_rgb565,
+void NV12ToRGB565Row_AVX2(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width) {
   // Row buffer for intermediate ARGB pixels.
-  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
   while (width > 0) {
     int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
     NV12ToARGBRow_AVX2(src_y, src_uv, row, yuvconstants, twidth);
@@ -2621,7 +4263,288 @@ void NV12ToRGB565Row_AVX2(const uint8* src_y,
 }
 #endif
 
+#ifdef HAS_RGB24TOYJROW_AVX2
+// Convert 16 RGB24 pixels (64 bytes) to 16 YJ values.
+void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth);
+    ARGBToYJRow_AVX2(row, dst_yj, twidth);
+    src_rgb24 += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RGB24TOYJROW_AVX2
+
+#ifdef HAS_RAWTOYJROW_AVX2
+// Convert 16 RAW pixels (64 bytes) to 16 YJ values.
+void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RAWToARGBRow_SSSE3(src_raw, row, twidth);
+    ARGBToYJRow_AVX2(row, dst_yj, twidth);
+    src_raw += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RAWTOYJROW_AVX2
+
+#ifdef HAS_RGB24TOYJROW_SSSE3
+// Convert 16 RGB24 pixels (64 bytes) to 16 YJ values.
+void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth);
+    ARGBToYJRow_SSSE3(row, dst_yj, twidth);
+    src_rgb24 += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RGB24TOYJROW_SSSE3
+
+#ifdef HAS_RAWTOYJROW_SSSE3
+// Convert 16 RAW pixels (64 bytes) to 16 YJ values.
+void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RAWToARGBRow_SSSE3(src_raw, row, twidth);
+    ARGBToYJRow_SSSE3(row, dst_yj, twidth);
+    src_raw += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RAWTOYJROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_16TO8_AVX2
+void InterpolateRow_16To8_AVX2(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int width,
+                               int source_y_fraction) {
+  // Row buffer for intermediate 16 bit pixels.
+  SIMD_ALIGNED(uint16_t row[MAXTWIDTH]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    InterpolateRow_16_C(row, src_ptr, src_stride, twidth, source_y_fraction);
+    Convert16To8Row_AVX2(row, dst_ptr, scale, twidth);
+    src_ptr += twidth;
+    dst_ptr += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_INTERPOLATEROW_16TO8_AVX2
+
+float ScaleSumSamples_C(const float* src, float* dst, float scale, int width) {
+  float fsum = 0.f;
+  int i;
+  for (i = 0; i < width; ++i) {
+    float v = *src++;
+    fsum += v * v;
+    *dst++ = v * scale;
+  }
+  return fsum;
+}
+
+float ScaleMaxSamples_C(const float* src, float* dst, float scale, int width) {
+  float fmax = 0.f;
+  int i;
+  for (i = 0; i < width; ++i) {
+    float v = *src++;
+    float vs = v * scale;
+    fmax = (v > fmax) ? v : fmax;
+    *dst++ = vs;
+  }
+  return fmax;
+}
+
+void ScaleSamples_C(const float* src, float* dst, float scale, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = *src++ * scale;
+  }
+}
+
+void GaussRow_C(const uint32_t* src, uint16_t* dst, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = STATIC_CAST(
+        uint16_t,
+        (src[0] + src[1] * 4 + src[2] * 6 + src[3] * 4 + src[4] + 128) >> 8);
+    ++src;
+  }
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_C(const uint16_t* src0,
+                const uint16_t* src1,
+                const uint16_t* src2,
+                const uint16_t* src3,
+                const uint16_t* src4,
+                uint32_t* dst,
+                int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = *src0++ + *src1++ * 4 + *src2++ * 6 + *src3++ * 4 + *src4++;
+  }
+}
+
+void GaussRow_F32_C(const float* src, float* dst, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = (src[0] + src[1] * 4 + src[2] * 6 + src[3] * 4 + src[4]) *
+             (1.0f / 256.0f);
+    ++src;
+  }
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_F32_C(const float* src0,
+                    const float* src1,
+                    const float* src2,
+                    const float* src3,
+                    const float* src4,
+                    float* dst,
+                    int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = *src0++ + *src1++ * 4 + *src2++ * 6 + *src3++ * 4 + *src4++;
+  }
+}
+
+// Convert biplanar NV21 to packed YUV24
+void NV21ToYUV24Row_C(const uint8_t* src_y,
+                      const uint8_t* src_vu,
+                      uint8_t* dst_yuv24,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_yuv24[0] = src_vu[0];  // V
+    dst_yuv24[1] = src_vu[1];  // U
+    dst_yuv24[2] = src_y[0];   // Y0
+    dst_yuv24[3] = src_vu[0];  // V
+    dst_yuv24[4] = src_vu[1];  // U
+    dst_yuv24[5] = src_y[1];   // Y1
+    src_y += 2;
+    src_vu += 2;
+    dst_yuv24 += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    dst_yuv24[0] = src_vu[0];  // V
+    dst_yuv24[1] = src_vu[1];  // U
+    dst_yuv24[2] = src_y[0];   // Y0
+  }
+}
+
+// Filter 2 rows of AYUV UV's (444) into UV (420).
+// AYUV is VUYA in memory.  UV for NV12 is UV order in memory.
+void AYUVToUVRow_C(const uint8_t* src_ayuv,
+                   int src_stride_ayuv,
+                   uint8_t* dst_uv,
+                   int width) {
+  // Output a row of UV values, filtering 2x2 rows of AYUV.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = (src_ayuv[1] + src_ayuv[5] + src_ayuv[src_stride_ayuv + 1] +
+                 src_ayuv[src_stride_ayuv + 5] + 2) >>
+                2;
+    dst_uv[1] = (src_ayuv[0] + src_ayuv[4] + src_ayuv[src_stride_ayuv + 0] +
+                 src_ayuv[src_stride_ayuv + 4] + 2) >>
+                2;
+    src_ayuv += 8;
+    dst_uv += 2;
+  }
+  if (width & 1) {
+    dst_uv[0] = (src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] + 1) >> 1;
+    dst_uv[1] = (src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] + 1) >> 1;
+  }
+}
+
+// Filter 2 rows of AYUV UV's (444) into VU (420).
+void AYUVToVURow_C(const uint8_t* src_ayuv,
+                   int src_stride_ayuv,
+                   uint8_t* dst_vu,
+                   int width) {
+  // Output a row of VU values, filtering 2x2 rows of AYUV.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_vu[0] = (src_ayuv[0] + src_ayuv[4] + src_ayuv[src_stride_ayuv + 0] +
+                 src_ayuv[src_stride_ayuv + 4] + 2) >>
+                2;
+    dst_vu[1] = (src_ayuv[1] + src_ayuv[5] + src_ayuv[src_stride_ayuv + 1] +
+                 src_ayuv[src_stride_ayuv + 5] + 2) >>
+                2;
+    src_ayuv += 8;
+    dst_vu += 2;
+  }
+  if (width & 1) {
+    dst_vu[0] = (src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] + 1) >> 1;
+    dst_vu[1] = (src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] + 1) >> 1;
+  }
+}
+
+// Copy row of AYUV Y's into Y
+void AYUVToYRow_C(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_y[x] = src_ayuv[2];  // v,u,y,a
+    src_ayuv += 4;
+  }
+}
+
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8_t u = src_uv[0];
+    uint8_t v = src_uv[1];
+    dst_vu[0] = v;
+    dst_vu[1] = u;
+    src_uv += 2;
+    dst_vu += 2;
+  }
+}
+
+void HalfMergeUVRow_C(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = (src_u[0] + src_u[1] + src_u[src_stride_u] +
+                 src_u[src_stride_u + 1] + 2) >>
+                2;
+    dst_uv[1] = (src_v[0] + src_v[1] + src_v[src_stride_v] +
+                 src_v[src_stride_v + 1] + 2) >>
+                2;
+    src_u += 2;
+    src_v += 2;
+    dst_uv += 2;
+  }
+  if (width & 1) {
+    dst_uv[0] = (src_u[0] + src_u[src_stride_u] + 1) >> 1;
+    dst_uv[1] = (src_v[0] + src_v[src_stride_v] + 1) >> 1;
+  }
+}
+
+#undef STATIC_CAST
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
-#endif
+#endif
\ No newline at end of file
diff --git a/media/libyuv/source/row_gcc.cc b/media/libyuv/source/row_gcc.cc
index 1ac7ef1aa3..aa4c0d11ec 100644
--- a/media/libyuv/source/row_gcc.cc
+++ b/media/libyuv/source/row_gcc.cc
@@ -1,4 +1,3 @@
-// VERSION 2
 /*
  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
@@ -10,1725 +9,2720 @@
  */
 
 #include "libyuv/row.h"
-
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
 
 // Constants for ARGB
-static vec8 kARGBToY = {
-  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
-};
+static const uvec8 kARGBToY = {25u, 129u, 66u, 0u, 25u, 129u, 66u, 0u,
+                               25u, 129u, 66u, 0u, 25u, 129u, 66u, 0u};
 
 // JPeg full range.
-static vec8 kARGBToYJ = {
-  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
-};
+static const uvec8 kARGBToYJ = {29u, 150u, 77u, 0u, 29u, 150u, 77u, 0u,
+                                29u, 150u, 77u, 0u, 29u, 150u, 77u, 0u};
+
+static const uvec8 kABGRToYJ = {77u, 150u, 29u, 0u, 77u, 150u, 29u, 0u,
+                                77u, 150u, 29u, 0u, 77u, 150u, 29u, 0u};
+
+static const uvec8 kRGBAToYJ = {0u, 29u, 150u, 77u, 0u, 29u, 150u, 77u,
+                                0u, 29u, 150u, 77u, 0u, 29u, 150u, 77u};
 #endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
 
 #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
 
-static vec8 kARGBToU = {
-  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
-};
+static const vec8 kARGBToU = {112, -74, -38, 0, 112, -74, -38, 0,
+                              112, -74, -38, 0, 112, -74, -38, 0};
 
-static vec8 kARGBToUJ = {
-  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
-};
+static const vec8 kARGBToUJ = {127, -84, -43, 0, 127, -84, -43, 0,
+                               127, -84, -43, 0, 127, -84, -43, 0};
 
-static vec8 kARGBToV = {
-  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
-};
+static const vec8 kABGRToUJ = {-43, -84, 127, 0, -43, -84, 127, 0,
+                               -43, -84, 127, 0, -43, -84, 127, 0};
 
-static vec8 kARGBToVJ = {
-  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
-};
+static const vec8 kARGBToV = {-18, -94, 112, 0, -18, -94, 112, 0,
+                              -18, -94, 112, 0, -18, -94, 112, 0};
+
+static const vec8 kARGBToVJ = {-20, -107, 127, 0, -20, -107, 127, 0,
+                               -20, -107, 127, 0, -20, -107, 127, 0};
+
+static const vec8 kABGRToVJ = {127, -107, -20, 0, 127, -107, -20, 0,
+                               127, -107, -20, 0, 127, -107, -20, 0};
 
 // Constants for BGRA
-static vec8 kBGRAToY = {
-  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
-};
+static const uvec8 kBGRAToY = {0u, 66u, 129u, 25u, 0u, 66u, 129u, 25u,
+                               0u, 66u, 129u, 25u, 0u, 66u, 129u, 25u};
 
-static vec8 kBGRAToU = {
-  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
-};
+static const vec8 kBGRAToU = {0, -38, -74, 112, 0, -38, -74, 112,
+                              0, -38, -74, 112, 0, -38, -74, 112};
 
-static vec8 kBGRAToV = {
-  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
-};
+static const vec8 kBGRAToV = {0, 112, -94, -18, 0, 112, -94, -18,
+                              0, 112, -94, -18, 0, 112, -94, -18};
 
 // Constants for ABGR
-static vec8 kABGRToY = {
-  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
-};
+static const uvec8 kABGRToY = {66u, 129u, 25u, 0u, 66u, 129u, 25u, 0u,
+                               66u, 129u, 25u, 0u, 66u, 129u, 25u, 0u};
 
-static vec8 kABGRToU = {
-  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
-};
+static const vec8 kABGRToU = {-38, -74, 112, 0, -38, -74, 112, 0,
+                              -38, -74, 112, 0, -38, -74, 112, 0};
 
-static vec8 kABGRToV = {
-  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
-};
+static const vec8 kABGRToV = {112, -94, -18, 0, 112, -94, -18, 0,
+                              112, -94, -18, 0, 112, -94, -18, 0};
 
 // Constants for RGBA.
-static vec8 kRGBAToY = {
-  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
-};
+static const uvec8 kRGBAToY = {0u, 25u, 129u, 66u, 0u, 25u, 129u, 66u,
+                               0u, 25u, 129u, 66u, 0u, 25u, 129u, 66u};
 
-static vec8 kRGBAToU = {
-  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
-};
+static const vec8 kRGBAToU = {0, 112, -74, -38, 0, 112, -74, -38,
+                              0, 112, -74, -38, 0, 112, -74, -38};
 
-static vec8 kRGBAToV = {
-  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
-};
+static const vec8 kRGBAToV = {0, -18, -94, 112, 0, -18, -94, 112,
+                              0, -18, -94, 112, 0, -18, -94, 112};
 
-static uvec8 kAddY16 = {
-  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
-};
+static const uvec16 kAddY16 = {0x7e80u, 0x7e80u, 0x7e80u, 0x7e80u,
+                               0x7e80u, 0x7e80u, 0x7e80u, 0x7e80u};
 
-// 7 bit fixed point 0.5.
-static vec16 kAddYJ64 = {
-  64, 64, 64, 64, 64, 64, 64, 64
-};
+static const uvec8 kAddUV128 = {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+                                128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
-static uvec8 kAddUV128 = {
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+static const uvec16 kSub128 = {0x8080u, 0x8080u, 0x8080u, 0x8080u,
+                               0x8080u, 0x8080u, 0x8080u, 0x8080u};
 
-static uvec16 kAddUVJ128 = {
-  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
-};
 #endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
 
 #ifdef HAS_RGB24TOARGBROW_SSSE3
 
 // Shuffle table for converting RGB24 to ARGB.
-static uvec8 kShuffleMaskRGB24ToARGB = {
-  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
-};
+static const uvec8 kShuffleMaskRGB24ToARGB = {
+    0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u};
 
 // Shuffle table for converting RAW to ARGB.
-static uvec8 kShuffleMaskRAWToARGB = {
-  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
-};
+static const uvec8 kShuffleMaskRAWToARGB = {2u, 1u, 0u, 12u, 5u,  4u,  3u, 13u,
+                                            8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u};
+
+// Shuffle table for converting RAW to RGBA.
+static const uvec8 kShuffleMaskRAWToRGBA = {12u, 2u, 1u, 0u, 13u, 5u,  4u,  3u,
+                                            14u, 8u, 7u, 6u, 15u, 11u, 10u, 9u};
 
 // Shuffle table for converting RAW to RGB24.  First 8.
 static const uvec8 kShuffleMaskRAWToRGB24_0 = {
-  2u, 1u, 0u, 5u, 4u, 3u, 8u, 7u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+    2u,   1u,   0u,   5u,   4u,   3u,   8u,   7u,
+    128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting RAW to RGB24.  Middle 8.
 static const uvec8 kShuffleMaskRAWToRGB24_1 = {
-  2u, 7u, 6u, 5u, 10u, 9u, 8u, 13u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+    2u,   7u,   6u,   5u,   10u,  9u,   8u,   13u,
+    128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting RAW to RGB24.  Last 8.
 static const uvec8 kShuffleMaskRAWToRGB24_2 = {
-  8u, 7u, 12u, 11u, 10u, 15u, 14u, 13u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+    8u,   7u,   12u,  11u,  10u,  15u,  14u,  13u,
+    128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting ARGB to RGB24.
-static uvec8 kShuffleMaskARGBToRGB24 = {
-  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
-};
+static const uvec8 kShuffleMaskARGBToRGB24 = {
+    0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting ARGB to RAW.
-static uvec8 kShuffleMaskARGBToRAW = {
-  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
-};
+static const uvec8 kShuffleMaskARGBToRAW = {
+    2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
-static uvec8 kShuffleMaskARGBToRGB24_0 = {
-  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
-};
+static const uvec8 kShuffleMaskARGBToRGB24_0 = {
+    0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u};
 
 // YUY2 shuf 16 Y to 32 Y.
-static const lvec8 kShuffleYUY2Y = {
-  0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12, 14, 14,
-  0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12, 14, 14
-};
+static const lvec8 kShuffleYUY2Y = {0,  0,  2,  2,  4,  4,  6,  6,  8,  8, 10,
+                                    10, 12, 12, 14, 14, 0,  0,  2,  2,  4, 4,
+                                    6,  6,  8,  8,  10, 10, 12, 12, 14, 14};
 
 // YUY2 shuf 8 UV to 16 UV.
-static const lvec8 kShuffleYUY2UV = {
-  1, 3, 1, 3, 5, 7, 5, 7, 9, 11, 9, 11, 13, 15, 13, 15,
-  1, 3, 1, 3, 5, 7, 5, 7, 9, 11, 9, 11, 13, 15, 13, 15
-};
+static const lvec8 kShuffleYUY2UV = {1,  3,  1,  3,  5,  7,  5,  7,  9,  11, 9,
+                                     11, 13, 15, 13, 15, 1,  3,  1,  3,  5,  7,
+                                     5,  7,  9,  11, 9,  11, 13, 15, 13, 15};
 
 // UYVY shuf 16 Y to 32 Y.
-static const lvec8 kShuffleUYVYY = {
-  1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15,
-  1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15
-};
+static const lvec8 kShuffleUYVYY = {1,  1,  3,  3,  5,  5,  7,  7,  9,  9, 11,
+                                    11, 13, 13, 15, 15, 1,  1,  3,  3,  5, 5,
+                                    7,  7,  9,  9,  11, 11, 13, 13, 15, 15};
 
 // UYVY shuf 8 UV to 16 UV.
-static const lvec8 kShuffleUYVYUV = {
-  0, 2, 0, 2, 4, 6, 4, 6, 8, 10, 8, 10, 12, 14, 12, 14,
-  0, 2, 0, 2, 4, 6, 4, 6, 8, 10, 8, 10, 12, 14, 12, 14
-};
+static const lvec8 kShuffleUYVYUV = {0,  2,  0,  2,  4,  6,  4,  6,  8,  10, 8,
+                                     10, 12, 14, 12, 14, 0,  2,  0,  2,  4,  6,
+                                     4,  6,  8,  10, 8,  10, 12, 14, 12, 14};
 
 // NV21 shuf 8 VU to 16 UV.
 static const lvec8 kShuffleNV21 = {
-  1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
-  1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
+    1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
+    1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
 };
 #endif  // HAS_RGB24TOARGBROW_SSSE3
 
 #ifdef HAS_J400TOARGBROW_SSE2
-void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pslld     $0x18,%%xmm5                    \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm0,%%xmm0                   \n"
-    "punpckhwd %%xmm1,%%xmm1                   \n"
-    "por       %%xmm5,%%xmm0                   \n"
-    "por       %%xmm5,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_y),     // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
-  );
+void J400ToARGBRow_SSE2(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0x18,%%xmm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklwd   %%xmm0,%%xmm0                 \n"
+      "punpckhwd   %%xmm1,%%xmm1                 \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm5");
 }
 #endif  // HAS_J400TOARGBROW_SSE2
 
 #ifdef HAS_RGB24TOARGBROW_SSSE3
-void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
-    "pslld     $0x18,%%xmm5                    \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
-    "lea       " MEMLEA(0x30,0) ",%0           \n"
-    "movdqa    %%xmm3,%%xmm2                   \n"
-    "palignr   $0x8,%%xmm1,%%xmm2              \n"
-    "pshufb    %%xmm4,%%xmm2                   \n"
-    "por       %%xmm5,%%xmm2                   \n"
-    "palignr   $0xc,%%xmm0,%%xmm1              \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
-    "por       %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm4,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "por       %%xmm5,%%xmm1                   \n"
-    "palignr   $0x4,%%xmm3,%%xmm3              \n"
-    "pshufb    %%xmm4,%%xmm3                   \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "por       %%xmm5,%%xmm3                   \n"
-    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_rgb24),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  : "m"(kShuffleMaskRGB24ToARGB)  // %3
-  : "memory", "cc" , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // 0xff000000
+      "pslld       $0x18,%%xmm5                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm3               \n"
+      "lea         0x30(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm2            \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "palignr     $0x4,%%xmm3,%%xmm3            \n"
+      "pshufb      %%xmm4,%%xmm3                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm3,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_rgb24),              // %0
+        "+r"(dst_argb),               // %1
+        "+r"(width)                   // %2
+      : "m"(kShuffleMaskRGB24ToARGB)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
-    "pslld     $0x18,%%xmm5                    \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
-    "lea       " MEMLEA(0x30,0) ",%0           \n"
-    "movdqa    %%xmm3,%%xmm2                   \n"
-    "palignr   $0x8,%%xmm1,%%xmm2              \n"
-    "pshufb    %%xmm4,%%xmm2                   \n"
-    "por       %%xmm5,%%xmm2                   \n"
-    "palignr   $0xc,%%xmm0,%%xmm1              \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
-    "por       %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm4,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "por       %%xmm5,%%xmm1                   \n"
-    "palignr   $0x4,%%xmm3,%%xmm3              \n"
-    "pshufb    %%xmm4,%%xmm3                   \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "por       %%xmm5,%%xmm3                   \n"
-    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_raw),   // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  : "m"(kShuffleMaskRAWToARGB)  // %3
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void RAWToARGBRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // 0xff000000
+      "pslld       $0x18,%%xmm5                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm3               \n"
+      "lea         0x30(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm2            \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "palignr     $0x4,%%xmm3,%%xmm3            \n"
+      "pshufb      %%xmm4,%%xmm3                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm3,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_raw),              // %0
+        "+r"(dst_argb),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleMaskRAWToARGB)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-void RAWToRGB24Row_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width) {
-  asm volatile (
-   "movdqa     %3,%%xmm3                       \n"
-   "movdqa     %4,%%xmm4                       \n"
-   "movdqa     %5,%%xmm5                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x4,0) ",%%xmm1    \n"
-    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm2    \n"
-    "lea       " MEMLEA(0x18,0) ",%0           \n"
-    "pshufb    %%xmm3,%%xmm0                   \n"
-    "pshufb    %%xmm4,%%xmm1                   \n"
-    "pshufb    %%xmm5,%%xmm2                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
-    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x18,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_raw),    // %0
-    "+r"(dst_rgb24),  // %1
-    "+r"(width)       // %2
-  : "m"(kShuffleMaskRAWToRGB24_0),  // %3
-    "m"(kShuffleMaskRAWToRGB24_1),  // %4
-    "m"(kShuffleMaskRAWToRGB24_2)   // %5
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+// Same code as RAWToARGB with different shuffler and A in low bits
+void RAWToRGBARow_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // 0x000000ff
+      "psrld       $0x18,%%xmm5                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm3               \n"
+      "lea         0x30(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm2            \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "palignr     $0x4,%%xmm3,%%xmm3            \n"
+      "pshufb      %%xmm4,%%xmm3                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm3,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_raw),              // %0
+        "+r"(dst_rgba),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleMaskRAWToRGBA)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "mov       $0x1080108,%%eax                \n"
-    "movd      %%eax,%%xmm5                    \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    "mov       $0x20802080,%%eax               \n"
-    "movd      %%eax,%%xmm6                    \n"
-    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
-    "pcmpeqb   %%xmm3,%%xmm3                   \n"
-    "psllw     $0xb,%%xmm3                     \n"
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "psllw     $0xa,%%xmm4                     \n"
-    "psrlw     $0x5,%%xmm4                     \n"
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"
-    "psllw     $0x8,%%xmm7                     \n"
-    "sub       %0,%1                           \n"
-    "sub       %0,%1                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "pand      %%xmm3,%%xmm1                   \n"
-    "psllw     $0xb,%%xmm2                     \n"
-    "pmulhuw   %%xmm5,%%xmm1                   \n"
-    "pmulhuw   %%xmm5,%%xmm2                   \n"
-    "psllw     $0x8,%%xmm1                     \n"
-    "por       %%xmm2,%%xmm1                   \n"
-    "pand      %%xmm4,%%xmm0                   \n"
-    "pmulhuw   %%xmm6,%%xmm0                   \n"
-    "por       %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm0,%%xmm1                   \n"
-    "punpckhbw %%xmm0,%%xmm2                   \n"
-    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
-    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  :
-  : "memory", "cc", "eax", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void RAWToRGB24Row_SSSE3(const uint8_t* src_raw,
+                         uint8_t* dst_rgb24,
+                         int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm3                     \n"
+      "movdqa      %4,%%xmm4                     \n"
+      "movdqa      %5,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x4(%0),%%xmm1                \n"
+      "movdqu      0x8(%0),%%xmm2                \n"
+      "lea         0x18(%0),%0                   \n"
+      "pshufb      %%xmm3,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x8(%1)                \n"
+      "movq        %%xmm2,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_raw),                  // %0
+        "+r"(dst_rgb24),                // %1
+        "+r"(width)                     // %2
+      : "m"(kShuffleMaskRAWToRGB24_0),  // %3
+        "m"(kShuffleMaskRAWToRGB24_1),  // %4
+        "m"(kShuffleMaskRAWToRGB24_2)   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "mov       $0x1080108,%%eax                \n"
-    "movd      %%eax,%%xmm5                    \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    "mov       $0x42004200,%%eax               \n"
-    "movd      %%eax,%%xmm6                    \n"
-    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
-    "pcmpeqb   %%xmm3,%%xmm3                   \n"
-    "psllw     $0xb,%%xmm3                     \n"
-    "movdqa    %%xmm3,%%xmm4                   \n"
-    "psrlw     $0x6,%%xmm4                     \n"
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"
-    "psllw     $0x8,%%xmm7                     \n"
-    "sub       %0,%1                           \n"
-    "sub       %0,%1                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "psllw     $0x1,%%xmm1                     \n"
-    "psllw     $0xb,%%xmm2                     \n"
-    "pand      %%xmm3,%%xmm1                   \n"
-    "pmulhuw   %%xmm5,%%xmm2                   \n"
-    "pmulhuw   %%xmm5,%%xmm1                   \n"
-    "psllw     $0x8,%%xmm1                     \n"
-    "por       %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "pand      %%xmm4,%%xmm0                   \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "pmulhuw   %%xmm6,%%xmm0                   \n"
-    "pand      %%xmm7,%%xmm2                   \n"
-    "por       %%xmm2,%%xmm0                   \n"
-    "movdqa    %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm0,%%xmm1                   \n"
-    "punpckhbw %%xmm0,%%xmm2                   \n"
-    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
-    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  :
-  : "memory", "cc", "eax", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void RGB565ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "mov         $0x1080108,%%eax              \n"
+      "movd        %%eax,%%xmm5                  \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "mov         $0x20802080,%%eax             \n"
+      "movd        %%eax,%%xmm6                  \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psllw       $0xb,%%xmm3                   \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psllw       $0xa,%%xmm4                   \n"
+      "psrlw       $0x5,%%xmm4                   \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psllw       $0x8,%%xmm7                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%1                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "psllw       $0xb,%%xmm2                   \n"
+      "pmulhuw     %%xmm5,%%xmm1                 \n"
+      "pmulhuw     %%xmm5,%%xmm2                 \n"
+      "psllw       $0x8,%%xmm1                   \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "pmulhuw     %%xmm6,%%xmm0                 \n"
+      "por         %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm0,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x00(%1,%0,2)          \n"
+      "movdqu      %%xmm2,0x10(%1,%0,2)          \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+        "xmm6", "xmm7");
 }
 
-void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "mov       $0xf0f0f0f,%%eax                \n"
-    "movd      %%eax,%%xmm4                    \n"
-    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
-    "movdqa    %%xmm4,%%xmm5                   \n"
-    "pslld     $0x4,%%xmm5                     \n"
-    "sub       %0,%1                           \n"
-    "sub       %0,%1                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "pand      %%xmm4,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm3                   \n"
-    "psllw     $0x4,%%xmm1                     \n"
-    "psrlw     $0x4,%%xmm3                     \n"
-    "por       %%xmm1,%%xmm0                   \n"
-    "por       %%xmm3,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm2,%%xmm0                   \n"
-    "punpckhbw %%xmm2,%%xmm1                   \n"
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,2)           //  movdqu  %%xmm0,(%1,%0,2)
-    MEMOPMEM(movdqu,xmm1,0x10,1,0,2)           //  movdqu  %%xmm1,0x10(%1,%0,2)
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  :
-  : "memory", "cc", "eax", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGB1555ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "mov         $0x1080108,%%eax              \n"
+      "movd        %%eax,%%xmm5                  \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "mov         $0x42004200,%%eax             \n"
+      "movd        %%eax,%%xmm6                  \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psllw       $0xb,%%xmm3                   \n"
+      "movdqa      %%xmm3,%%xmm4                 \n"
+      "psrlw       $0x6,%%xmm4                   \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psllw       $0x8,%%xmm7                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%1                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "psllw       $0x1,%%xmm1                   \n"
+      "psllw       $0xb,%%xmm2                   \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "pmulhuw     %%xmm5,%%xmm2                 \n"
+      "pmulhuw     %%xmm5,%%xmm1                 \n"
+      "psllw       $0x8,%%xmm1                   \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "psraw       $0x8,%%xmm2                   \n"
+      "pmulhuw     %%xmm6,%%xmm0                 \n"
+      "pand        %%xmm7,%%xmm2                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm0,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x00(%1,%0,2)          \n"
+      "movdqu      %%xmm2,0x10(%1,%0,2)          \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+        "xmm6", "xmm7");
 }
 
-void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "movdqa    %3,%%xmm6                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "pshufb    %%xmm6,%%xmm0                   \n"
-    "pshufb    %%xmm6,%%xmm1                   \n"
-    "pshufb    %%xmm6,%%xmm2                   \n"
-    "pshufb    %%xmm6,%%xmm3                   \n"
-    "movdqa    %%xmm1,%%xmm4                   \n"
-    "psrldq    $0x4,%%xmm1                     \n"
-    "pslldq    $0xc,%%xmm4                     \n"
-    "movdqa    %%xmm2,%%xmm5                   \n"
-    "por       %%xmm4,%%xmm0                   \n"
-    "pslldq    $0x8,%%xmm5                     \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "por       %%xmm5,%%xmm1                   \n"
-    "psrldq    $0x8,%%xmm2                     \n"
-    "pslldq    $0x4,%%xmm3                     \n"
-    "por       %%xmm3,%%xmm2                   \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
-    "lea       " MEMLEA(0x30,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  : "m"(kShuffleMaskARGBToRGB24)  // %3
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+void ARGB4444ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "mov         $0xf0f0f0f,%%eax              \n"
+      "movd        %%eax,%%xmm4                  \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "pslld       $0x4,%%xmm5                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%1                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "psllw       $0x4,%%xmm1                   \n"
+      "psrlw       $0x4,%%xmm3                   \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "movdqu      %%xmm0,0x00(%1,%0,2)          \n"
+      "movdqu      %%xmm1,0x10(%1,%0,2)          \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "movdqa    %3,%%xmm6                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "pshufb    %%xmm6,%%xmm0                   \n"
-    "pshufb    %%xmm6,%%xmm1                   \n"
-    "pshufb    %%xmm6,%%xmm2                   \n"
-    "pshufb    %%xmm6,%%xmm3                   \n"
-    "movdqa    %%xmm1,%%xmm4                   \n"
-    "psrldq    $0x4,%%xmm1                     \n"
-    "pslldq    $0xc,%%xmm4                     \n"
-    "movdqa    %%xmm2,%%xmm5                   \n"
-    "por       %%xmm4,%%xmm0                   \n"
-    "pslldq    $0x8,%%xmm5                     \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "por       %%xmm5,%%xmm1                   \n"
-    "psrldq    $0x8,%%xmm2                     \n"
-    "pslldq    $0x4,%%xmm3                     \n"
-    "por       %%xmm3,%%xmm2                   \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
-    "lea       " MEMLEA(0x30,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  : "m"(kShuffleMaskARGBToRAW)  // %3
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+void ARGBToRGB24Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+
+      "movdqa      %3,%%xmm6                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "lea         0x40(%0),%0                   \n"
+      "pshufb      %%xmm6,%%xmm0                 \n"
+      "pshufb      %%xmm6,%%xmm1                 \n"
+      "pshufb      %%xmm6,%%xmm2                 \n"
+      "pshufb      %%xmm6,%%xmm3                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "psrldq      $0x4,%%xmm1                   \n"
+      "pslldq      $0xc,%%xmm4                   \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pslldq      $0x8,%%xmm5                   \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "psrldq      $0x8,%%xmm2                   \n"
+      "pslldq      $0x4,%%xmm3                   \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "lea         0x30(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src),                    // %0
+        "+r"(dst),                    // %1
+        "+r"(width)                   // %2
+      : "m"(kShuffleMaskARGBToRGB24)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 
-void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm3,%%xmm3                   \n"
-    "psrld     $0x1b,%%xmm3                    \n"
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "psrld     $0x1a,%%xmm4                    \n"
-    "pslld     $0x5,%%xmm4                     \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pslld     $0xb,%%xmm5                     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "pslld     $0x8,%%xmm0                     \n"
-    "psrld     $0x3,%%xmm1                     \n"
-    "psrld     $0x5,%%xmm2                     \n"
-    "psrad     $0x10,%%xmm0                    \n"
-    "pand      %%xmm3,%%xmm1                   \n"
-    "pand      %%xmm4,%%xmm2                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "por       %%xmm2,%%xmm1                   \n"
-    "por       %%xmm1,%%xmm0                   \n"
-    "packssdw  %%xmm0,%%xmm0                   \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGBToRAWRow_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+
+      "movdqa      %3,%%xmm6                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "lea         0x40(%0),%0                   \n"
+      "pshufb      %%xmm6,%%xmm0                 \n"
+      "pshufb      %%xmm6,%%xmm1                 \n"
+      "pshufb      %%xmm6,%%xmm2                 \n"
+      "pshufb      %%xmm6,%%xmm3                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "psrldq      $0x4,%%xmm1                   \n"
+      "pslldq      $0xc,%%xmm4                   \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pslldq      $0x8,%%xmm5                   \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "psrldq      $0x8,%%xmm2                   \n"
+      "pslldq      $0x4,%%xmm3                   \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "lea         0x30(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst),                  // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleMaskARGBToRAW)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 
-void ARGBToRGB565DitherRow_SSE2(const uint8* src, uint8* dst,
-                                const uint32 dither4, int width) {
-  asm volatile (
-    "movd       %3,%%xmm6                      \n"
-    "punpcklbw  %%xmm6,%%xmm6                  \n"
-    "movdqa     %%xmm6,%%xmm7                  \n"
-    "punpcklwd  %%xmm6,%%xmm6                  \n"
-    "punpckhwd  %%xmm7,%%xmm7                  \n"
-    "pcmpeqb    %%xmm3,%%xmm3                  \n"
-    "psrld      $0x1b,%%xmm3                   \n"
-    "pcmpeqb    %%xmm4,%%xmm4                  \n"
-    "psrld      $0x1a,%%xmm4                   \n"
-    "pslld      $0x5,%%xmm4                    \n"
-    "pcmpeqb    %%xmm5,%%xmm5                  \n"
-    "pslld      $0xb,%%xmm5                    \n"
+#ifdef HAS_ARGBTORGB24ROW_AVX2
+// vpermd for 12+12 to 24
+static const lvec32 kPermdRGB24_AVX = {0, 1, 2, 4, 5, 6, 3, 7};
+
+void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm6                  \n"
+      "vmovdqa     %4,%%ymm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "lea         0x80(%0),%0                   \n"
+      "vpshufb     %%ymm6,%%ymm0,%%ymm0          \n"  // xxx0yyy0
+      "vpshufb     %%ymm6,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm6,%%ymm3,%%ymm3          \n"
+      "vpermd      %%ymm0,%%ymm7,%%ymm0          \n"  // pack to 24 bytes
+      "vpermd      %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpermd      %%ymm3,%%ymm7,%%ymm3          \n"
+      "vpermq      $0x3f,%%ymm1,%%ymm4           \n"  // combine 24 + 8
+      "vpor        %%ymm4,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vpermq      $0xf9,%%ymm1,%%ymm1           \n"  // combine 16 + 16
+      "vpermq      $0x4f,%%ymm2,%%ymm4           \n"
+      "vpor        %%ymm4,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "vpermq      $0xfe,%%ymm2,%%ymm2           \n"  // combine 8 + 24
+      "vpermq      $0x93,%%ymm3,%%ymm3           \n"
+      "vpor        %%ymm3,%%ymm2,%%ymm2          \n"
+      "vmovdqu     %%ymm2,0x40(%1)               \n"
+      "lea         0x60(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),                     // %0
+        "+r"(dst),                     // %1
+        "+r"(width)                    // %2
+      : "m"(kShuffleMaskARGBToRGB24),  // %3
+        "m"(kPermdRGB24_AVX)           // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
 
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu     (%0),%%xmm0                    \n"
-    "paddusb    %%xmm6,%%xmm0                  \n"
-    "movdqa     %%xmm0,%%xmm1                  \n"
-    "movdqa     %%xmm0,%%xmm2                  \n"
-    "pslld      $0x8,%%xmm0                    \n"
-    "psrld      $0x3,%%xmm1                    \n"
-    "psrld      $0x5,%%xmm2                    \n"
-    "psrad      $0x10,%%xmm0                   \n"
-    "pand       %%xmm3,%%xmm1                  \n"
-    "pand       %%xmm4,%%xmm2                  \n"
-    "pand       %%xmm5,%%xmm0                  \n"
-    "por        %%xmm2,%%xmm1                  \n"
-    "por        %%xmm1,%%xmm0                  \n"
-    "packssdw   %%xmm0,%%xmm0                  \n"
-    "lea        0x10(%0),%0                    \n"
-    "movq       %%xmm0,(%1)                    \n"
-    "lea        0x8(%1),%1                     \n"
-    "sub        $0x4,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  : "m"(dither4) // %3
-  : "memory", "cc",
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+#ifdef HAS_ARGBTORGB24ROW_AVX512VBMI
+// Shuffle table for converting ARGBToRGB24
+static const ulvec8 kPermARGBToRGB24_0 = {
+    0u,  1u,  2u,  4u,  5u,  6u,  8u,  9u,  10u, 12u, 13u,
+    14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u, 25u, 26u, 28u,
+    29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u, 40u, 41u};
+static const ulvec8 kPermARGBToRGB24_1 = {
+    10u, 12u, 13u, 14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u,
+    25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u,
+    40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u, 50u, 52u};
+static const ulvec8 kPermARGBToRGB24_2 = {
+    21u, 22u, 24u, 25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u,
+    36u, 37u, 38u, 40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u,
+    50u, 52u, 53u, 54u, 56u, 57u, 58u, 60u, 61u, 62u};
+
+void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vmovdqa     %3,%%ymm5                     \n"
+      "vmovdqa     %4,%%ymm6                     \n"
+      "vmovdqa     %5,%%ymm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "lea         0x80(%0),%0                   \n"
+      "vpermt2b    %%ymm1,%%ymm5,%%ymm0          \n"
+      "vpermt2b    %%ymm2,%%ymm6,%%ymm1          \n"
+      "vpermt2b    %%ymm3,%%ymm7,%%ymm2          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "vmovdqu     %%ymm2,0x40(%1)               \n"
+      "lea         0x60(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),                // %0
+        "+r"(dst),                // %1
+        "+r"(width)               // %2
+      : "m"(kPermARGBToRGB24_0),  // %3
+        "m"(kPermARGBToRGB24_1),  // %4
+        "m"(kPermARGBToRGB24_2)   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6", "xmm7");
 }
+#endif
 
-#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
-void ARGBToRGB565DitherRow_AVX2(const uint8* src, uint8* dst,
-                                const uint32 dither4, int width) {
-  asm volatile (
-    "vbroadcastss %3,%%xmm6                    \n"
-    "vpunpcklbw %%xmm6,%%xmm6,%%xmm6           \n"
-    "vpermq     $0xd8,%%ymm6,%%ymm6            \n"
-    "vpunpcklwd %%ymm6,%%ymm6,%%ymm6           \n"
-    "vpcmpeqb   %%ymm3,%%ymm3,%%ymm3           \n"
-    "vpsrld     $0x1b,%%ymm3,%%ymm3            \n"
-    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpsrld     $0x1a,%%ymm4,%%ymm4            \n"
-    "vpslld     $0x5,%%ymm4,%%ymm4             \n"
-    "vpslld     $0xb,%%ymm3,%%ymm5             \n"
+#ifdef HAS_ARGBTORAWROW_AVX2
+void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm6                  \n"
+      "vmovdqa     %4,%%ymm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "lea         0x80(%0),%0                   \n"
+      "vpshufb     %%ymm6,%%ymm0,%%ymm0          \n"  // xxx0yyy0
+      "vpshufb     %%ymm6,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm6,%%ymm3,%%ymm3          \n"
+      "vpermd      %%ymm0,%%ymm7,%%ymm0          \n"  // pack to 24 bytes
+      "vpermd      %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpermd      %%ymm3,%%ymm7,%%ymm3          \n"
+      "vpermq      $0x3f,%%ymm1,%%ymm4           \n"  // combine 24 + 8
+      "vpor        %%ymm4,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vpermq      $0xf9,%%ymm1,%%ymm1           \n"  // combine 16 + 16
+      "vpermq      $0x4f,%%ymm2,%%ymm4           \n"
+      "vpor        %%ymm4,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "vpermq      $0xfe,%%ymm2,%%ymm2           \n"  // combine 8 + 24
+      "vpermq      $0x93,%%ymm3,%%ymm3           \n"
+      "vpor        %%ymm3,%%ymm2,%%ymm2          \n"
+      "vmovdqu     %%ymm2,0x40(%1)               \n"
+      "lea         0x60(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),                   // %0
+        "+r"(dst),                   // %1
+        "+r"(width)                  // %2
+      : "m"(kShuffleMaskARGBToRAW),  // %3
+        "m"(kPermdRGB24_AVX)         // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
 
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    (%0),%%ymm0                    \n"
-    "vpaddusb   %%ymm6,%%ymm0,%%ymm0           \n"
-    "vpsrld     $0x5,%%ymm0,%%ymm2             \n"
-    "vpsrld     $0x3,%%ymm0,%%ymm1             \n"
-    "vpsrld     $0x8,%%ymm0,%%ymm0             \n"
-    "vpand      %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpand      %%ymm3,%%ymm1,%%ymm1           \n"
-    "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpor       %%ymm2,%%ymm1,%%ymm1           \n"
-    "vpor       %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpackusdw  %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "lea        0x20(%0),%0                    \n"
-    "vmovdqu    %%xmm0,(%1)                    \n"
-    "lea        0x10(%1),%1                    \n"
-    "sub        $0x8,%2                        \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  : "m"(dither4) // %3
-  : "memory", "cc",
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void ARGBToRGB565Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psrld       $0x1b,%%xmm3                  \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrld       $0x1a,%%xmm4                  \n"
+      "pslld       $0x5,%%xmm4                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0xb,%%xmm5                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pslld       $0x8,%%xmm0                   \n"
+      "psrld       $0x3,%%xmm1                   \n"
+      "psrld       $0x5,%%xmm2                   \n"
+      "psrad       $0x10,%%xmm0                  \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm4,%%xmm2                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
-#endif  // HAS_ARGBTORGB565DITHERROW_AVX2
 
+void ARGBToRGB565DitherRow_SSE2(const uint8_t* src,
+                                uint8_t* dst,
+                                const uint32_t dither4,
+                                int width) {
+  asm volatile(
+      "movd        %3,%%xmm6                     \n"
+      "punpcklbw   %%xmm6,%%xmm6                 \n"
+      "movdqa      %%xmm6,%%xmm7                 \n"
+      "punpcklwd   %%xmm6,%%xmm6                 \n"
+      "punpckhwd   %%xmm7,%%xmm7                 \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psrld       $0x1b,%%xmm3                  \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrld       $0x1a,%%xmm4                  \n"
+      "pslld       $0x5,%%xmm4                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0xb,%%xmm5                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "paddusb     %%xmm6,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pslld       $0x8,%%xmm0                   \n"
+      "psrld       $0x3,%%xmm1                   \n"
+      "psrld       $0x5,%%xmm2                   \n"
+      "psrad       $0x10,%%xmm0                  \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm4,%%xmm2                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),    // %0
+        "+r"(dst),    // %1
+        "+r"(width)   // %2
+      : "m"(dither4)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
 
-void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "psrld     $0x1b,%%xmm4                    \n"
-    "movdqa    %%xmm4,%%xmm5                   \n"
-    "pslld     $0x5,%%xmm5                     \n"
-    "movdqa    %%xmm4,%%xmm6                   \n"
-    "pslld     $0xa,%%xmm6                     \n"
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"
-    "pslld     $0xf,%%xmm7                     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm3                   \n"
-    "psrad     $0x10,%%xmm0                    \n"
-    "psrld     $0x3,%%xmm1                     \n"
-    "psrld     $0x6,%%xmm2                     \n"
-    "psrld     $0x9,%%xmm3                     \n"
-    "pand      %%xmm7,%%xmm0                   \n"
-    "pand      %%xmm4,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm2                   \n"
-    "pand      %%xmm6,%%xmm3                   \n"
-    "por       %%xmm1,%%xmm0                   \n"
-    "por       %%xmm3,%%xmm2                   \n"
-    "por       %%xmm2,%%xmm0                   \n"
-    "packssdw  %%xmm0,%%xmm0                   \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  :: "memory", "cc",
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
+void ARGBToRGB565DitherRow_AVX2(const uint8_t* src,
+                                uint8_t* dst,
+                                const uint32_t dither4,
+                                int width) {
+  asm volatile(
+      "vbroadcastss %3,%%xmm6                    \n"
+      "vpunpcklbw  %%xmm6,%%xmm6,%%xmm6          \n"
+      "vpermq      $0xd8,%%ymm6,%%ymm6           \n"
+      "vpunpcklwd  %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpcmpeqb    %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpsrld      $0x1b,%%ymm3,%%ymm3           \n"
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrld      $0x1a,%%ymm4,%%ymm4           \n"
+      "vpslld      $0x5,%%ymm4,%%ymm4            \n"
+      "vpslld      $0xb,%%ymm3,%%ymm5            \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpaddusb    %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpsrld      $0x5,%%ymm0,%%ymm2            \n"
+      "vpsrld      $0x3,%%ymm0,%%ymm1            \n"
+      "vpsrld      $0x8,%%ymm0,%%ymm0            \n"
+      "vpand       %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpand       %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpor        %%ymm2,%%ymm1,%%ymm1          \n"
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "lea         0x20(%0),%0                   \n"
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),    // %0
+        "+r"(dst),    // %1
+        "+r"(width)   // %2
+      : "m"(dither4)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
+#endif  // HAS_ARGBTORGB565DITHERROW_AVX2
 
-void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "psllw     $0xc,%%xmm4                     \n"
-    "movdqa    %%xmm4,%%xmm3                   \n"
-    "psrlw     $0x8,%%xmm3                     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm3,%%xmm0                   \n"
-    "pand      %%xmm4,%%xmm1                   \n"
-    "psrlq     $0x4,%%xmm0                     \n"
-    "psrlq     $0x8,%%xmm1                     \n"
-    "por       %%xmm1,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
-  );
+void ARGBToARGB1555Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrld       $0x1b,%%xmm4                  \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "pslld       $0x5,%%xmm5                   \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "pslld       $0xa,%%xmm6                   \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "pslld       $0xf,%%xmm7                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm3                 \n"
+      "psrad       $0x10,%%xmm0                  \n"
+      "psrld       $0x3,%%xmm1                   \n"
+      "psrld       $0x6,%%xmm2                   \n"
+      "psrld       $0x9,%%xmm3                   \n"
+      "pand        %%xmm7,%%xmm0                 \n"
+      "pand        %%xmm4,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm2                 \n"
+      "pand        %%xmm6,%%xmm3                 \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7");
+}
+
+void ARGBToARGB4444Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psllw       $0xc,%%xmm4                   \n"
+      "movdqa      %%xmm4,%%xmm3                 \n"
+      "psrlw       $0x8,%%xmm3                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm3,%%xmm0                 \n"
+      "pand        %%xmm4,%%xmm1                 \n"
+      "psrlq       $0x4,%%xmm0                   \n"
+      "psrlq       $0x8,%%xmm1                   \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
 }
 #endif  // HAS_RGB24TOARGBROW_SSSE3
 
+/*
+
+ARGBToAR30Row:
+
+Red Blue
+With the 8 bit value in the upper bits of a short, vpmulhuw by (1024+4) will
+produce a 10 bit value in the low 10 bits of each 16 bit value. This is whats
+wanted for the blue channel. The red needs to be shifted 4 left, so multiply by
+(1024+4)*16 for red.
+
+Alpha Green
+Alpha and Green are already in the high bits so vpand can zero out the other
+bits, keeping just 2 upper bits of alpha and 8 bit green. The same multiplier
+could be used for Green - (1024+4) putting the 10 bit green in the lsb.  Alpha
+would be a simple multiplier to shift it into position.  It wants a gap of 10
+above the green.  Green is 10 bits, so there are 6 bits in the low short.  4
+more are needed, so a multiplier of 4 gets the 2 bits into the upper 16 bits,
+and then a shift of 4 is a multiply of 16, so (4*16) = 64.  Then shift the
+result left 10 to position the A and G channels.
+*/
+
+// Shuffle table for converting RAW to RGB24.  Last 8.
+static const uvec8 kShuffleRB30 = {128u, 0u, 128u, 2u,  128u, 4u,  128u, 6u,
+                                   128u, 8u, 128u, 10u, 128u, 12u, 128u, 14u};
+
+static const uvec8 kShuffleBR30 = {128u, 2u,  128u, 0u, 128u, 6u,  128u, 4u,
+                                   128u, 10u, 128u, 8u, 128u, 14u, 128u, 12u};
+
+static const uint32_t kMulRB10 = 1028 * 16 * 65536 + 1028;
+static const uint32_t kMaskRB10 = 0x3ff003ff;
+static const uint32_t kMaskAG10 = 0xc000ff00;
+static const uint32_t kMulAG10 = 64 * 65536 + 1028;
+
+void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm2                     \n"  // shuffler for RB
+      "movd        %4,%%xmm3                     \n"  // multipler for RB
+      "movd        %5,%%xmm4                     \n"  // mask for R10 B10
+      "movd        %6,%%xmm5                     \n"  // mask for AG
+      "movd        %7,%%xmm6                     \n"  // multipler for AG
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // fetch 4 ARGB pixels
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm1                 \n"  // R0B0
+      "pand        %%xmm5,%%xmm0                 \n"  // A0G0
+      "pmulhuw     %%xmm3,%%xmm1                 \n"  // X2 R16 X4  B10
+      "pmulhuw     %%xmm6,%%xmm0                 \n"  // X10 A2 X10 G10
+      "pand        %%xmm4,%%xmm1                 \n"  // X2 R10 X10 B10
+      "pslld       $10,%%xmm0                    \n"  // A2 x10 G10 x10
+      "por         %%xmm1,%%xmm0                 \n"  // A2 R10 G10 B10
+      "movdqu      %%xmm0,(%1,%0)                \n"  // store 4 AR30 pixels
+      "add         $0x10,%0                      \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+
+      : "+r"(src),          // %0
+        "+r"(dst),          // %1
+        "+r"(width)         // %2
+      : "m"(kShuffleRB30),  // %3
+        "m"(kMulRB10),      // %4
+        "m"(kMaskRB10),     // %5
+        "m"(kMaskAG10),     // %6
+        "m"(kMulAG10)       // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+void ABGRToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm2                     \n"  // shuffler for RB
+      "movd        %4,%%xmm3                     \n"  // multipler for RB
+      "movd        %5,%%xmm4                     \n"  // mask for R10 B10
+      "movd        %6,%%xmm5                     \n"  // mask for AG
+      "movd        %7,%%xmm6                     \n"  // multipler for AG
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // fetch 4 ABGR pixels
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm1                 \n"  // R0B0
+      "pand        %%xmm5,%%xmm0                 \n"  // A0G0
+      "pmulhuw     %%xmm3,%%xmm1                 \n"  // X2 R16 X4  B10
+      "pmulhuw     %%xmm6,%%xmm0                 \n"  // X10 A2 X10 G10
+      "pand        %%xmm4,%%xmm1                 \n"  // X2 R10 X10 B10
+      "pslld       $10,%%xmm0                    \n"  // A2 x10 G10 x10
+      "por         %%xmm1,%%xmm0                 \n"  // A2 R10 G10 B10
+      "movdqu      %%xmm0,(%1,%0)                \n"  // store 4 AR30 pixels
+      "add         $0x10,%0                      \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+
+      : "+r"(src),          // %0
+        "+r"(dst),          // %1
+        "+r"(width)         // %2
+      : "m"(kShuffleBR30),  // %3  reversed shuffler
+        "m"(kMulRB10),      // %4
+        "m"(kMaskRB10),     // %5
+        "m"(kMaskAG10),     // %6
+        "m"(kMulAG10)       // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+#ifdef HAS_ARGBTOAR30ROW_AVX2
+void ARGBToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm2                  \n"  // shuffler for RB
+      "vbroadcastss %4,%%ymm3                    \n"  // multipler for RB
+      "vbroadcastss %5,%%ymm4                    \n"  // mask for R10 B10
+      "vbroadcastss %6,%%ymm5                    \n"  // mask for AG
+      "vbroadcastss %7,%%ymm6                    \n"  // multipler for AG
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // fetch 8 ARGB pixels
+      "vpshufb     %%ymm2,%%ymm0,%%ymm1          \n"  // R0B0
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"  // A0G0
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"  // X2 R16 X4  B10
+      "vpmulhuw    %%ymm6,%%ymm0,%%ymm0          \n"  // X10 A2 X10 G10
+      "vpand       %%ymm4,%%ymm1,%%ymm1          \n"  // X2 R10 X10 B10
+      "vpslld      $10,%%ymm0,%%ymm0             \n"  // A2 x10 G10 x10
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"  // A2 R10 G10 B10
+      "vmovdqu     %%ymm0,(%1,%0)                \n"  // store 8 AR30 pixels
+      "add         $0x20,%0                      \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+      : "+r"(src),          // %0
+        "+r"(dst),          // %1
+        "+r"(width)         // %2
+      : "m"(kShuffleRB30),  // %3
+        "m"(kMulRB10),      // %4
+        "m"(kMaskRB10),     // %5
+        "m"(kMaskAG10),     // %6
+        "m"(kMulAG10)       // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_ABGRTOAR30ROW_AVX2
+void ABGRToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm2                  \n"  // shuffler for RB
+      "vbroadcastss %4,%%ymm3                    \n"  // multipler for RB
+      "vbroadcastss %5,%%ymm4                    \n"  // mask for R10 B10
+      "vbroadcastss %6,%%ymm5                    \n"  // mask for AG
+      "vbroadcastss %7,%%ymm6                    \n"  // multipler for AG
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // fetch 8 ABGR pixels
+      "vpshufb     %%ymm2,%%ymm0,%%ymm1          \n"  // R0B0
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"  // A0G0
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"  // X2 R16 X4  B10
+      "vpmulhuw    %%ymm6,%%ymm0,%%ymm0          \n"  // X10 A2 X10 G10
+      "vpand       %%ymm4,%%ymm1,%%ymm1          \n"  // X2 R10 X10 B10
+      "vpslld      $10,%%ymm0,%%ymm0             \n"  // A2 x10 G10 x10
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"  // A2 R10 G10 B10
+      "vmovdqu     %%ymm0,(%1,%0)                \n"  // store 8 AR30 pixels
+      "add         $0x20,%0                      \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+      : "+r"(src),          // %0
+        "+r"(dst),          // %1
+        "+r"(width)         // %2
+      : "m"(kShuffleBR30),  // %3  reversed shuffler
+        "m"(kMulRB10),      // %4
+        "m"(kMaskRB10),     // %5
+        "m"(kMaskAG10),     // %6
+        "m"(kMulAG10)       // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+static const uvec8 kShuffleARGBToABGR = {2,  1, 0, 3,  6,  5,  4,  7,
+                                         10, 9, 8, 11, 14, 13, 12, 15};
+
+static const uvec8 kShuffleARGBToAB64Lo = {2, 2, 1, 1, 0, 0, 3, 3,
+                                           6, 6, 5, 5, 4, 4, 7, 7};
+static const uvec8 kShuffleARGBToAB64Hi = {10, 10, 9,  9,  8,  8,  11, 11,
+                                           14, 14, 13, 13, 12, 12, 15, 15};
+
+void ARGBToAR64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ar64,
+                         int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+
+void ARGBToAB64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ab64,
+                         int width) {
+  asm volatile(
+
+      "movdqa      %3,%%xmm2                     \n"
+      "movdqa      %4,%%xmm3                     \n" LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm0                 \n"
+      "pshufb      %%xmm3,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),             // %0
+        "+r"(dst_ab64),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleARGBToAB64Lo),  // %3
+        "m"(kShuffleARGBToAB64Hi)   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+
+void AR64ToARGBRow_SSSE3(const uint16_t* src_ar64,
+                         uint8_t* dst_argb,
+                         int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psrlw       $8,%%xmm0                     \n"
+      "psrlw       $8,%%xmm1                     \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ar64),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+
+void AB64ToARGBRow_SSSE3(const uint16_t* src_ab64,
+                         uint8_t* dst_argb,
+                         int width) {
+  asm volatile(
+
+      "movdqa      %3,%%xmm2                     \n" LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psrlw       $8,%%xmm0                     \n"
+      "psrlw       $8,%%xmm1                     \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "pshufb      %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ab64),          // %0
+        "+r"(dst_argb),          // %1
+        "+r"(width)              // %2
+      : "m"(kShuffleARGBToABGR)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+
+#ifdef HAS_ARGBTOAR64ROW_AVX2
+void ARGBToAR64Row_AVX2(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpunpckhbw  %%ymm0,%%ymm0,%%ymm1          \n"
+      "vpunpcklbw  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_AVX2
+void ARGBToAB64Row_AVX2(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm2                  \n"
+      "vbroadcastf128 %4,%%ymm3                  \n" LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %%ymm3,%%ymm0,%%ymm1          \n"
+      "vpshufb     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),             // %0
+        "+r"(dst_ab64),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleARGBToAB64Lo),  // %3
+        "m"(kShuffleARGBToAB64Hi)   // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_AR64TOARGBROW_AVX2
+void AR64ToARGBRow_AVX2(const uint16_t* src_ar64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpsrlw      $8,%%ymm0,%%ymm0              \n"
+      "vpsrlw      $8,%%ymm1,%%ymm1              \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x40(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ar64),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+#endif
+
+#ifdef HAS_AB64TOARGBROW_AVX2
+void AB64ToARGBRow_AVX2(const uint16_t* src_ab64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm2                  \n" LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpsrlw      $8,%%ymm0,%%ymm0              \n"
+      "vpsrlw      $8,%%ymm1,%%ymm1              \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x40(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ab64),          // %0
+        "+r"(dst_argb),          // %1
+        "+r"(width)              // %2
+      : "m"(kShuffleARGBToABGR)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+// clang-format off
+
+// TODO(mraptis): Consider passing R, G, B multipliers as parameter.
+// round parameter is register containing value to add before shift.
+#define RGBTOY(round)                            \
+  "1:                                        \n" \
+  "movdqu    (%0),%%xmm0                     \n" \
+  "movdqu    0x10(%0),%%xmm1                 \n" \
+  "movdqu    0x20(%0),%%xmm2                 \n" \
+  "movdqu    0x30(%0),%%xmm3                 \n" \
+  "psubb     %%xmm5,%%xmm0                   \n" \
+  "psubb     %%xmm5,%%xmm1                   \n" \
+  "psubb     %%xmm5,%%xmm2                   \n" \
+  "psubb     %%xmm5,%%xmm3                   \n" \
+  "movdqu    %%xmm4,%%xmm6                   \n" \
+  "pmaddubsw %%xmm0,%%xmm6                   \n" \
+  "movdqu    %%xmm4,%%xmm0                   \n" \
+  "pmaddubsw %%xmm1,%%xmm0                   \n" \
+  "movdqu    %%xmm4,%%xmm1                   \n" \
+  "pmaddubsw %%xmm2,%%xmm1                   \n" \
+  "movdqu    %%xmm4,%%xmm2                   \n" \
+  "pmaddubsw %%xmm3,%%xmm2                   \n" \
+  "lea       0x40(%0),%0                     \n" \
+  "phaddw    %%xmm0,%%xmm6                   \n" \
+  "phaddw    %%xmm2,%%xmm1                   \n" \
+  "prefetcht0 1280(%0)                       \n" \
+  "paddw     %%" #round ",%%xmm6             \n" \
+  "paddw     %%" #round ",%%xmm1             \n" \
+  "psrlw     $0x8,%%xmm6                     \n" \
+  "psrlw     $0x8,%%xmm1                     \n" \
+  "packuswb  %%xmm1,%%xmm6                   \n" \
+  "movdqu    %%xmm6,(%1)                     \n" \
+  "lea       0x10(%1),%1                     \n" \
+  "sub       $0x10,%2                        \n" \
+  "jg        1b                              \n"
+
+#define RGBTOY_AVX2(round)                                       \
+  "1:                                        \n"                 \
+  "vmovdqu    (%0),%%ymm0                    \n"                 \
+  "vmovdqu    0x20(%0),%%ymm1                \n"                 \
+  "vmovdqu    0x40(%0),%%ymm2                \n"                 \
+  "vmovdqu    0x60(%0),%%ymm3                \n"                 \
+  "vpsubb     %%ymm5, %%ymm0, %%ymm0         \n"                 \
+  "vpsubb     %%ymm5, %%ymm1, %%ymm1         \n"                 \
+  "vpsubb     %%ymm5, %%ymm2, %%ymm2         \n"                 \
+  "vpsubb     %%ymm5, %%ymm3, %%ymm3         \n"                 \
+  "vpmaddubsw %%ymm0,%%ymm4,%%ymm0           \n"                 \
+  "vpmaddubsw %%ymm1,%%ymm4,%%ymm1           \n"                 \
+  "vpmaddubsw %%ymm2,%%ymm4,%%ymm2           \n"                 \
+  "vpmaddubsw %%ymm3,%%ymm4,%%ymm3           \n"                 \
+  "lea       0x80(%0),%0                     \n"                 \
+  "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n" /* mutates. */  \
+  "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"                 \
+  "prefetcht0 1280(%0)                       \n"                 \
+  "vpaddw     %%" #round ",%%ymm0,%%ymm0     \n" /* Add .5 for rounding. */             \
+  "vpaddw     %%" #round ",%%ymm2,%%ymm2     \n" \
+  "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"                 \
+  "vpsrlw     $0x8,%%ymm2,%%ymm2             \n"                 \
+  "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n" /* mutates. */  \
+  "vpermd     %%ymm0,%%ymm6,%%ymm0           \n" /* unmutate. */ \
+  "vmovdqu    %%ymm0,(%1)                    \n"                 \
+  "lea       0x20(%1),%1                     \n"                 \
+  "sub       $0x20,%2                        \n"                 \
+  "jg        1b                              \n"                 \
+  "vzeroupper                                \n"
+
+// clang-format on
+
 #ifdef HAS_ARGBTOYROW_SSSE3
 // Convert 16 ARGB pixels (64 bytes) to 16 Y values.
-void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "movdqa    %3,%%xmm4                       \n"
-    "movdqa    %4,%%xmm5                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kARGBToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGBToYRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
+      : "+r"(src_argb),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kARGBToY),   // %3
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOYROW_SSSE3
 
 #ifdef HAS_ARGBTOYJROW_SSSE3
 // Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
-// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
-void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "movdqa    %3,%%xmm4                       \n"
-    "movdqa    %4,%%xmm5                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "paddw     %%xmm5,%%xmm0                   \n"
-    "paddw     %%xmm5,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kARGBToYJ),  // %3
-    "m"(kAddYJ64)    // %4
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+// Same as ARGBToYRow but different coefficients, no add 16.
+void ARGBToYJRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN RGBTOY(xmm5)
+      : "+r"(src_argb),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kARGBToYJ),  // %3
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBTOYJROW_SSSE3
 
-#ifdef HAS_ARGBTOYROW_AVX2
+#ifdef HAS_ABGRTOYJROW_SSSE3
+// Convert 16 ABGR pixels (64 bytes) to 16 YJ values.
+// Same as ABGRToYRow but different coefficients, no add 16.
+void ABGRToYJRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN RGBTOY(xmm5)
+      : "+r"(src_abgr),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kABGRToYJ),  // %3
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif  // HAS_ABGRTOYJROW_SSSE3
+
+#ifdef HAS_RGBATOYJROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16.
+void RGBAToYJRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN RGBTOY(xmm5)
+      : "+r"(src_rgba),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kRGBAToYJ),  // %3
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif  // HAS_RGBATOYJROW_SSSE3
+
+#if defined(HAS_ARGBTOYROW_AVX2) || defined(HAS_ABGRTOYROW_AVX2) || \
+    defined(HAS_ARGBEXTRACTALPHAROW_AVX2)
 // vpermd for vphaddw + vpackuswb vpermd.
-static const lvec32 kPermdARGBToY_AVX = {
-  0, 4, 1, 5, 2, 6, 3, 7
-};
+static const lvec32 kPermdARGBToY_AVX = {0, 4, 1, 5, 2, 6, 3, 7};
+#endif
+
+#ifdef HAS_ARGBTOYROW_AVX2
 
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
-void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "vbroadcastf128 %3,%%ymm4                  \n"
-    "vbroadcastf128 %4,%%ymm5                  \n"
-    "vmovdqu    %5,%%ymm6                      \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
-    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
-    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "lea       " MEMLEA(0x80,0) ",%0           \n"
-    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
-    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
-    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
-    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
-    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
-    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
-    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"  // add 16 for Y
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kARGBToY),   // %3
-    "m"(kAddY16),    // %4
-    "m"(kPermdARGBToY_AVX)  // %5
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+void ARGBToYRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vbroadcastf128 %5,%%ymm7                  \n"
+      "vmovdqu     %6,%%ymm6                     \n" LABELALIGN RGBTOY_AVX2(
+      ymm7) "vzeroupper                                \n"
+      : "+r"(src_argb),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kARGBToY),          // %3
+        "m"(kSub128),           // %4
+        "m"(kAddY16),           // %5
+        "m"(kPermdARGBToY_AVX)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOYROW_AVX2
 
+#ifdef HAS_ABGRTOYROW_AVX2
+// Convert 32 ABGR pixels (128 bytes) to 32 Y values.
+void ABGRToYRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vbroadcastf128 %5,%%ymm7                  \n"
+      "vmovdqu     %6,%%ymm6                     \n" LABELALIGN RGBTOY_AVX2(
+      ymm7) "vzeroupper                                \n"
+      : "+r"(src_abgr),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kABGRToY),          // %3
+        "m"(kSub128),           // %4
+        "m"(kAddY16),           // %5
+        "m"(kPermdARGBToY_AVX)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // HAS_ABGRTOYROW_AVX2
+
 #ifdef HAS_ARGBTOYJROW_AVX2
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
-void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "vbroadcastf128 %3,%%ymm4                  \n"
-    "vbroadcastf128 %4,%%ymm5                  \n"
-    "vmovdqu    %5,%%ymm6                      \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
-    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
-    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "lea       " MEMLEA(0x80,0) ",%0           \n"
-    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
-    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
-    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"  // Add .5 for rounding.
-    "vpaddw     %%ymm5,%%ymm2,%%ymm2           \n"
-    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
-    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
-    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
-    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kARGBToYJ),   // %3
-    "m"(kAddYJ64),    // %4
-    "m"(kPermdARGBToY_AVX)  // %5
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+void ARGBToYJRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vmovdqu     %5,%%ymm6                     \n" LABELALIGN RGBTOY_AVX2(
+      ymm5) "vzeroupper                                \n"
+      : "+r"(src_argb),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kARGBToYJ),         // %3
+        "m"(kSub128),           // %4
+        "m"(kPermdARGBToY_AVX)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOYJROW_AVX2
 
+#ifdef HAS_ABGRTOYJROW_AVX2
+// Convert 32 ABGR pixels (128 bytes) to 32 Y values.
+void ABGRToYJRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vmovdqu     %5,%%ymm6                     \n" LABELALIGN RGBTOY_AVX2(
+      ymm5) "vzeroupper                                \n"
+      : "+r"(src_abgr),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kABGRToYJ),         // %3
+        "m"(kSub128),           // %4
+        "m"(kPermdARGBToY_AVX)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // HAS_ABGRTOYJROW_AVX2
+
+#ifdef HAS_RGBATOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void RGBAToYJRow_AVX2(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vmovdqu     %5,%%ymm6                     \n" LABELALIGN RGBTOY_AVX2(
+      ymm5) "vzeroupper                                \n"
+      : "+r"(src_rgba),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kRGBAToYJ),         // %3
+        "m"(kSub128),           // %4
+        "m"(kPermdARGBToY_AVX)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif  // HAS_RGBATOYJROW_AVX2
+
 #ifdef HAS_ARGBTOUVROW_SSSE3
-void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %5,%%xmm3                       \n"
-    "movdqa    %6,%%xmm4                       \n"
-    "movdqa    %7,%%xmm5                       \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm1                   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)), // %4
-    "m"(kARGBToV),  // %5
-    "m"(kARGBToU),  // %6
-    "m"(kAddUV128)  // %7
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-  );
+void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile(
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_argb)),  // %4
+        "m"(kARGBToV),                     // %5
+        "m"(kARGBToU),                     // %6
+        "m"(kAddUV128)                     // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 #endif  // HAS_ARGBTOUVROW_SSSE3
 
-#ifdef HAS_ARGBTOUVROW_AVX2
+#if defined(HAS_ARGBTOUVROW_AVX2) || defined(HAS_ABGRTOUVROW_AVX2) || \
+    defined(HAS_ARGBTOUVJROW_AVX2) || defined(HAS_ABGRTOUVJROW_AVX2)
 // vpshufb for vphaddw + vpackuswb packed to shorts.
 static const lvec8 kShufARGBToUV_AVX = {
-  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
-  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
-};
-void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "vbroadcastf128 %5,%%ymm5                  \n"
-    "vbroadcastf128 %6,%%ymm6                  \n"
-    "vbroadcastf128 %7,%%ymm7                  \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
-    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
-    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
-    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
-    VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2)
-    VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3)
-    "lea       " MEMLEA(0x80,0) ",%0           \n"
-    "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
-    "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
-    "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
-    "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
-    "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
-    "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
-
-    "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
-    "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
-    "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
-    "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
-    "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
-    "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vpshufb    %8,%%ymm0,%%ymm0               \n"
-    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"
-
-    "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n"
-    VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x20,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)), // %4
-    "m"(kAddUV128),  // %5
-    "m"(kARGBToV),   // %6
-    "m"(kARGBToU),   // %7
-    "m"(kShufARGBToUV_AVX)  // %8
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+    0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+    0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
+#endif
+
+#if defined(HAS_ARGBTOUVROW_AVX2)
+void ARGBToUVRow_AVX2(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vbroadcastf128 %6,%%ymm6                  \n"
+      "vbroadcastf128 %7,%%ymm7                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
+      "vpaddb      %%ymm5,%%ymm0,%%ymm0          \n"
+
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_argb)),  // %4
+        "m"(kAddUV128),                    // %5
+        "m"(kARGBToV),                     // %6
+        "m"(kARGBToU),                     // %7
+        "m"(kShufARGBToUV_AVX)             // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOUVROW_AVX2
 
+#ifdef HAS_ABGRTOUVROW_AVX2
+void ABGRToUVRow_AVX2(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vbroadcastf128 %6,%%ymm6                  \n"
+      "vbroadcastf128 %7,%%ymm7                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
+      "vpaddb      %%ymm5,%%ymm0,%%ymm0          \n"
+
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_abgr),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_abgr)),  // %4
+        "m"(kAddUV128),                    // %5
+        "m"(kABGRToV),                     // %6
+        "m"(kABGRToU),                     // %7
+        "m"(kShufARGBToUV_AVX)             // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // HAS_ABGRTOUVROW_AVX2
+
 #ifdef HAS_ARGBTOUVJROW_AVX2
-void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "vbroadcastf128 %5,%%ymm5                  \n"
-    "vbroadcastf128 %6,%%ymm6                  \n"
-    "vbroadcastf128 %7,%%ymm7                  \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
-    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
-    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
-    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
-    VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2)
-    VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3)
-    "lea       " MEMLEA(0x80,0) ",%0           \n"
-    "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
-    "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
-    "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
-    "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
-    "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
-    "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
-
-    "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
-    "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
-    "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
-    "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
-    "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm5,%%ymm1,%%ymm1           \n"
-    "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
-    "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vpshufb    %8,%%ymm0,%%ymm0               \n"
-
-    "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n"
-    VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x20,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)), // %4
-    "m"(kAddUVJ128),  // %5
-    "m"(kARGBToVJ),  // %6
-    "m"(kARGBToUJ),  // %7
-    "m"(kShufARGBToUV_AVX)  // %8
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vbroadcastf128 %6,%%ymm6                  \n"
+      "vbroadcastf128 %7,%%ymm7                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
+
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_argb)),  // %4
+        "m"(kSub128),                      // %5
+        "m"(kARGBToVJ),                    // %6
+        "m"(kARGBToUJ),                    // %7
+        "m"(kShufARGBToUV_AVX)             // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOUVJROW_AVX2
 
+// TODO(fbarchard): Pass kABGRToVJ / kABGRToUJ as matrix
+#ifdef HAS_ABGRTOUVJROW_AVX2
+void ABGRToUVJRow_AVX2(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vbroadcastf128 %6,%%ymm6                  \n"
+      "vbroadcastf128 %7,%%ymm7                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
+
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_abgr),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_abgr)),  // %4
+        "m"(kSub128),                      // %5
+        "m"(kABGRToVJ),                    // %6
+        "m"(kABGRToUJ),                    // %7
+        "m"(kShufARGBToUV_AVX)             // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // HAS_ABGRTOUVJROW_AVX2
+
 #ifdef HAS_ARGBTOUVJROW_SSSE3
-void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                        uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %5,%%xmm3                       \n"
-    "movdqa    %6,%%xmm4                       \n"
-    "movdqa    %7,%%xmm5                       \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm1                   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "paddw     %%xmm5,%%xmm0                   \n"
-    "paddw     %%xmm5,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_argb)), // %4
-    "m"(kARGBToVJ),  // %5
-    "m"(kARGBToUJ),  // %6
-    "m"(kAddUVJ128)  // %7
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-  );
+void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
+                        int src_stride_argb,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  asm volatile(
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "paddw       %%xmm5,%%xmm0                 \n"
+      "paddw       %%xmm5,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_argb)),  // %4
+        "m"(kARGBToVJ),                    // %5
+        "m"(kARGBToUJ),                    // %6
+        "m"(kSub128)                       // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 #endif  // HAS_ARGBTOUVJROW_SSSE3
 
+#ifdef HAS_ABGRTOUVJROW_SSSE3
+void ABGRToUVJRow_SSSE3(const uint8_t* src_abgr,
+                        int src_stride_abgr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  asm volatile(
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "paddw       %%xmm5,%%xmm0                 \n"
+      "paddw       %%xmm5,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_abgr),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_abgr)),  // %4
+        "m"(kABGRToVJ),                    // %5
+        "m"(kABGRToUJ),                    // %6
+        "m"(kSub128)                       // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
+}
+#endif  // HAS_ABGRTOUVJROW_SSSE3
+
 #ifdef HAS_ARGBTOUV444ROW_SSSE3
-void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
                           int width) {
-  asm volatile (
-    "movdqa    %4,%%xmm3                       \n"
-    "movdqa    %5,%%xmm4                       \n"
-    "movdqa    %6,%%xmm5                       \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm6                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm2                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "packsswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm3,%%xmm0                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm2                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm2                     \n"
-    "packsswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),        // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "m"(kARGBToV),  // %4
-    "m"(kARGBToU),  // %5
-    "m"(kAddUV128)  // %6
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm6"
-  );
+  asm volatile(
+      "movdqa      %4,%%xmm3                     \n"
+      "movdqa      %5,%%xmm4                     \n"
+      "movdqa      %6,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm6                 \n"
+      "phaddw      %%xmm1,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm2                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm2                   \n"
+      "packsswb    %%xmm2,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm1,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm2                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm2                   \n"
+      "packsswb    %%xmm2,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "lea         0x40(%0),%0                   \n"
+      "movdqu      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+rm"(width)     // %3
+      : "m"(kARGBToV),   // %4
+        "m"(kARGBToU),   // %5
+        "m"(kAddUV128)   // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6");
 }
 #endif  // HAS_ARGBTOUV444ROW_SSSE3
 
-void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int width) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_bgra),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kBGRAToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
+      : "+r"(src_bgra),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kBGRAToY),   // %3
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
-void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %5,%%xmm3                       \n"
-    "movdqa    %6,%%xmm4                       \n"
-    "movdqa    %7,%%xmm5                       \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm1                   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_bgra0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_bgra)), // %4
-    "m"(kBGRAToV),  // %5
-    "m"(kBGRAToU),  // %6
-    "m"(kAddUV128)  // %7
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-  );
+void BGRAToUVRow_SSSE3(const uint8_t* src_bgra,
+                       int src_stride_bgra,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile(
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_bgra),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_bgra)),  // %4
+        "m"(kBGRAToV),                     // %5
+        "m"(kBGRAToU),                     // %6
+        "m"(kAddUV128)                     // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 
-void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int width) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_abgr),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kABGRToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
+      : "+r"(src_abgr),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kABGRToY),   // %3
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
-void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int width) {
-  asm volatile (
-    "movdqa    %4,%%xmm5                       \n"
-    "movdqa    %3,%%xmm4                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm4,%%xmm3                   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "phaddw    %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_rgba),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  : "m"(kRGBAToY),   // %3
-    "m"(kAddY16)     // %4
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void RGBAToYRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
+      : "+r"(src_rgba),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kRGBAToY),   // %3
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
-void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %5,%%xmm3                       \n"
-    "movdqa    %6,%%xmm4                       \n"
-    "movdqa    %7,%%xmm5                       \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm1                   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_abgr0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_abgr)), // %4
-    "m"(kABGRToV),  // %5
-    "m"(kABGRToU),  // %6
-    "m"(kAddUV128)  // %7
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-  );
+void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile(
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_abgr),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_abgr)),  // %4
+        "m"(kABGRToV),                     // %5
+        "m"(kABGRToU),                     // %6
+        "m"(kAddUV128)                     // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 
-void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "movdqa    %5,%%xmm3                       \n"
-    "movdqa    %6,%%xmm4                       \n"
-    "movdqa    %7,%%xmm5                       \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm1                   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm7                   \n"
-    "shufps    $0x88,%%xmm6,%%xmm2             \n"
-    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-    "pavgb     %%xmm7,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm2                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "phaddw    %%xmm2,%%xmm0                   \n"
-    "phaddw    %%xmm6,%%xmm1                   \n"
-    "psraw     $0x8,%%xmm0                     \n"
-    "psraw     $0x8,%%xmm1                     \n"
-    "packsswb  %%xmm1,%%xmm0                   \n"
-    "paddb     %%xmm5,%%xmm0                   \n"
-    "movlps    %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_rgba0),       // %0
-    "+r"(dst_u),           // %1
-    "+r"(dst_v),           // %2
-    "+rm"(width)           // %3
-  : "r"((intptr_t)(src_stride_rgba)), // %4
-    "m"(kRGBAToV),  // %5
-    "m"(kRGBAToU),  // %6
-    "m"(kAddUV128)  // %7
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
-  );
+void RGBAToUVRow_SSSE3(const uint8_t* src_rgba,
+                       int src_stride_rgba,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile(
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_rgba),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_rgba)),  // %4
+        "m"(kRGBAToV),                     // %5
+        "m"(kRGBAToU),                     // %6
+        "m"(kAddUV128)                     // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 
 #if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2)
 
 // Read 8 UV from 444
-#define READYUV444                                                             \
-    "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
-    MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
-    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
-    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
-    "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"
+#define READYUV444                                                \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n"
 
 // Read 4 UV from 422, upsample to 8 UV
-#define READYUV422                                                             \
-    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
-    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
-    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
-    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
-    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
-    "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"
+#define READYUV422                                                \
+  "movd       (%[u_buf]),%%xmm3                               \n" \
+  "movd       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x4(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n"
+
+// Read 4 UV from 422 10 bit, upsample to 8 UV
+#define READYUV210                                                \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklwd  %%xmm1,%%xmm3                                   \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "movdqa     %%xmm4,%%xmm2                                   \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "psrlw      $4,%%xmm2                                       \n" \
+  "paddw      %%xmm2,%%xmm4                                   \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
+#define READYUVA210                                               \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklwd  %%xmm1,%%xmm3                                   \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "movdqa     %%xmm4,%%xmm2                                   \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "psrlw      $4,%%xmm2                                       \n" \
+  "paddw      %%xmm2,%%xmm4                                   \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n" \
+  "movdqu     (%[a_buf]),%%xmm5                               \n" \
+  "psraw      $2,%%xmm5                                       \n" \
+  "packuswb   %%xmm5,%%xmm5                                   \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                         \n"
+
+// Read 8 UV from 444 10 bit
+#define READYUV410                                                \
+  "movdqu     (%[u_buf]),%%xmm3                               \n" \
+  "movdqu     0x00(%[u_buf],%[v_buf],1),%%xmm2                \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                         \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "psraw      $2,%%xmm2                                       \n" \
+  "movdqa     %%xmm3,%%xmm1                                   \n" \
+  "punpcklwd  %%xmm2,%%xmm3                                   \n" \
+  "punpckhwd  %%xmm2,%%xmm1                                   \n" \
+  "packuswb   %%xmm1,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "movdqa     %%xmm4,%%xmm2                                   \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "psrlw      $4,%%xmm2                                       \n" \
+  "paddw      %%xmm2,%%xmm4                                   \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
+// Read 8 UV from 444 10 bit.  With 8 Alpha.
+#define READYUVA410                                               \
+  "movdqu     (%[u_buf]),%%xmm3                               \n" \
+  "movdqu     0x00(%[u_buf],%[v_buf],1),%%xmm2                \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                         \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "psraw      $2,%%xmm2                                       \n" \
+  "movdqa     %%xmm3,%%xmm1                                   \n" \
+  "punpcklwd  %%xmm2,%%xmm3                                   \n" \
+  "punpckhwd  %%xmm2,%%xmm1                                   \n" \
+  "packuswb   %%xmm1,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "movdqa     %%xmm4,%%xmm2                                   \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "psrlw      $4,%%xmm2                                       \n" \
+  "paddw      %%xmm2,%%xmm4                                   \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n" \
+  "movdqu     (%[a_buf]),%%xmm5                               \n" \
+  "psraw      $2,%%xmm5                                       \n" \
+  "packuswb   %%xmm5,%%xmm5                                   \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                         \n"
+
+// Read 4 UV from 422 12 bit, upsample to 8 UV
+#define READYUV212                                                \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklwd  %%xmm1,%%xmm3                                   \n" \
+  "psraw      $0x4,%%xmm3                                     \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "movdqa     %%xmm4,%%xmm2                                   \n" \
+  "psllw      $4,%%xmm4                                       \n" \
+  "psrlw      $8,%%xmm2                                       \n" \
+  "paddw      %%xmm2,%%xmm4                                   \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
-#define READYUVA422                                                            \
-    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
-    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
-    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
-    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
-    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
-    "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
-    "movq       " MEMACCESS([a_buf]) ",%%xmm5                   \n"            \
-    "lea        " MEMLEA(0x8, [a_buf]) ",%[a_buf]               \n"
-
-// Read 2 UV from 411, upsample to 8 UV.
-// reading 4 bytes is an msan violation.
-//    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"
-//    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)
-// pinsrw fails with drmemory
-//  __asm pinsrw     xmm0, [esi], 0        /* U */
-//  __asm pinsrw     xmm1, [esi + edi], 0  /* V */
-#define READYUV411_TEMP                                                        \
-    "movzwl     " MEMACCESS([u_buf]) ",%[temp]                  \n"            \
-    "movd       %[temp],%%xmm0                                  \n"            \
-    MEMOPARG(movzwl, 0x00, [u_buf], [v_buf], 1, [temp]) "       \n"            \
-    "movd       %[temp],%%xmm1                                  \n"            \
-    "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
-    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
-    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
-    "punpckldq  %%xmm0,%%xmm0                                   \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
-    "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"
+#define READYUVA422                                               \
+  "movd       (%[u_buf]),%%xmm3                               \n" \
+  "movd       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x4(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n" \
+  "movq       (%[a_buf]),%%xmm5                               \n" \
+  "lea        0x8(%[a_buf]),%[a_buf]                          \n"
+
+// Read 8 UV from 444.  With 8 Alpha.
+#define READYUVA444                                               \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n" \
+  "movq       (%[a_buf]),%%xmm5                               \n" \
+  "lea        0x8(%[a_buf]),%[a_buf]                          \n"
 
 // Read 4 UV from NV12, upsample to 8 UV
-#define READNV12                                                               \
-    "movq       " MEMACCESS([uv_buf]) ",%%xmm0                  \n"            \
-    "lea        " MEMLEA(0x8, [uv_buf]) ",%[uv_buf]             \n"            \
-    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
-    "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"
+#define READNV12                                                  \
+  "movq       (%[uv_buf]),%%xmm3                              \n" \
+  "lea        0x8(%[uv_buf]),%[uv_buf]                        \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n"
 
 // Read 4 VU from NV21, upsample to 8 UV
-#define READNV21                                                               \
-    "movq       " MEMACCESS([vu_buf]) ",%%xmm0                  \n"            \
-    "lea        " MEMLEA(0x8, [vu_buf]) ",%[vu_buf]             \n"            \
-    "pshufb     %[kShuffleNV21], %%xmm0                         \n"            \
-    "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
-    "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
-    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"
+#define READNV21                                                  \
+  "movq       (%[vu_buf]),%%xmm3                              \n" \
+  "lea        0x8(%[vu_buf]),%[vu_buf]                        \n" \
+  "pshufb     %[kShuffleNV21], %%xmm3                         \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n"
 
 // Read 4 YUY2 with 8 Y and update 4 UV to 8 UV.
-#define READYUY2                                                               \
-    "movdqu     " MEMACCESS([yuy2_buf]) ",%%xmm4                \n"            \
-    "pshufb     %[kShuffleYUY2Y], %%xmm4                        \n"            \
-    "movdqu     " MEMACCESS([yuy2_buf]) ",%%xmm0                \n"            \
-    "pshufb     %[kShuffleYUY2UV], %%xmm0                       \n"            \
-    "lea        " MEMLEA(0x10, [yuy2_buf]) ",%[yuy2_buf]        \n"
+#define READYUY2                                                  \
+  "movdqu     (%[yuy2_buf]),%%xmm4                            \n" \
+  "pshufb     %[kShuffleYUY2Y], %%xmm4                        \n" \
+  "movdqu     (%[yuy2_buf]),%%xmm3                            \n" \
+  "pshufb     %[kShuffleYUY2UV], %%xmm3                       \n" \
+  "lea        0x10(%[yuy2_buf]),%[yuy2_buf]                   \n"
 
 // Read 4 UYVY with 8 Y and update 4 UV to 8 UV.
-#define READUYVY                                                               \
-    "movdqu     " MEMACCESS([uyvy_buf]) ",%%xmm4                \n"            \
-    "pshufb     %[kShuffleUYVYY], %%xmm4                        \n"            \
-    "movdqu     " MEMACCESS([uyvy_buf]) ",%%xmm0                \n"            \
-    "pshufb     %[kShuffleUYVYUV], %%xmm0                       \n"            \
-    "lea        " MEMLEA(0x10, [uyvy_buf]) ",%[uyvy_buf]        \n"
+#define READUYVY                                                  \
+  "movdqu     (%[uyvy_buf]),%%xmm4                            \n" \
+  "pshufb     %[kShuffleUYVYY], %%xmm4                        \n" \
+  "movdqu     (%[uyvy_buf]),%%xmm3                            \n" \
+  "pshufb     %[kShuffleUYVYUV], %%xmm3                       \n" \
+  "lea        0x10(%[uyvy_buf]),%[uyvy_buf]                   \n"
+
+// Read 4 UV from P210, upsample to 8 UV
+#define READP210                                                  \
+  "movdqu     (%[uv_buf]),%%xmm3                              \n" \
+  "lea        0x10(%[uv_buf]),%[uv_buf]                       \n" \
+  "psrlw      $0x8,%%xmm3                                     \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
+// Read 8 UV from P410
+#define READP410                                                  \
+  "movdqu     (%[uv_buf]),%%xmm3                              \n" \
+  "movdqu     0x10(%[uv_buf]),%%xmm1                          \n" \
+  "lea        0x20(%[uv_buf]),%[uv_buf]                       \n" \
+  "psrlw      $0x8,%%xmm3                                     \n" \
+  "psrlw      $0x8,%%xmm1                                     \n" \
+  "packuswb   %%xmm1,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
 
 #if defined(__x86_64__)
-#define YUVTORGB_SETUP(yuvconstants)                                           \
-    "movdqa     " MEMACCESS([yuvconstants]) ",%%xmm8            \n"            \
-    "movdqa     " MEMACCESS2(32, [yuvconstants]) ",%%xmm9       \n"            \
-    "movdqa     " MEMACCESS2(64, [yuvconstants]) ",%%xmm10      \n"            \
-    "movdqa     " MEMACCESS2(96, [yuvconstants]) ",%%xmm11      \n"            \
-    "movdqa     " MEMACCESS2(128, [yuvconstants]) ",%%xmm12     \n"            \
-    "movdqa     " MEMACCESS2(160, [yuvconstants]) ",%%xmm13     \n"            \
-    "movdqa     " MEMACCESS2(192, [yuvconstants]) ",%%xmm14     \n"
+#define YUVTORGB_SETUP(yuvconstants)                              \
+  "pcmpeqb    %%xmm13,%%xmm13                                 \n" \
+  "movdqa     (%[yuvconstants]),%%xmm8                        \n" \
+  "pxor       %%xmm12,%%xmm12                                 \n" \
+  "movdqa     32(%[yuvconstants]),%%xmm9                      \n" \
+  "psllw      $7,%%xmm13                                      \n" \
+  "movdqa     64(%[yuvconstants]),%%xmm10                     \n" \
+  "pshufb     %%xmm12,%%xmm13                                 \n" \
+  "movdqa     96(%[yuvconstants]),%%xmm11                     \n" \
+  "movdqa     128(%[yuvconstants]),%%xmm12                    \n"
+
 // Convert 8 pixels: 8 UV and 8 Y
-#define YUVTORGB(yuvconstants)                                                 \
-    "movdqa     %%xmm0,%%xmm1                                   \n"            \
-    "movdqa     %%xmm0,%%xmm2                                   \n"            \
-    "movdqa     %%xmm0,%%xmm3                                   \n"            \
-    "movdqa     %%xmm11,%%xmm0                                  \n"            \
-    "pmaddubsw  %%xmm8,%%xmm1                                   \n"            \
-    "psubw      %%xmm1,%%xmm0                                   \n"            \
-    "movdqa     %%xmm12,%%xmm1                                  \n"            \
-    "pmaddubsw  %%xmm9,%%xmm2                                   \n"            \
-    "psubw      %%xmm2,%%xmm1                                   \n"            \
-    "movdqa     %%xmm13,%%xmm2                                  \n"            \
-    "pmaddubsw  %%xmm10,%%xmm3                                  \n"            \
-    "psubw      %%xmm3,%%xmm2                                   \n"            \
-    "pmulhuw    %%xmm14,%%xmm4                                  \n"            \
-    "paddsw     %%xmm4,%%xmm0                                   \n"            \
-    "paddsw     %%xmm4,%%xmm1                                   \n"            \
-    "paddsw     %%xmm4,%%xmm2                                   \n"            \
-    "psraw      $0x6,%%xmm0                                     \n"            \
-    "psraw      $0x6,%%xmm1                                     \n"            \
-    "psraw      $0x6,%%xmm2                                     \n"            \
-    "packuswb   %%xmm0,%%xmm0                                   \n"            \
-    "packuswb   %%xmm1,%%xmm1                                   \n"            \
-    "packuswb   %%xmm2,%%xmm2                                   \n"
-#define YUVTORGB_REGS \
-    "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14",
+#define YUVTORGB16(yuvconstants)                                  \
+  "psubb      %%xmm13,%%xmm3                                  \n" \
+  "pmulhuw    %%xmm11,%%xmm4                                  \n" \
+  "movdqa     %%xmm8,%%xmm0                                   \n" \
+  "movdqa     %%xmm9,%%xmm1                                   \n" \
+  "movdqa     %%xmm10,%%xmm2                                  \n" \
+  "paddw      %%xmm12,%%xmm4                                  \n" \
+  "pmaddubsw  %%xmm3,%%xmm0                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm1                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm2                                   \n" \
+  "paddsw     %%xmm4,%%xmm0                                   \n" \
+  "paddsw     %%xmm4,%%xmm2                                   \n" \
+  "psubsw     %%xmm1,%%xmm4                                   \n" \
+  "movdqa     %%xmm4,%%xmm1                                   \n"
+
+#define YUVTORGB_REGS "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",
 
 #else
 #define YUVTORGB_SETUP(yuvconstants)
 // Convert 8 pixels: 8 UV and 8 Y
-#define YUVTORGB(yuvconstants)                                                 \
-    "movdqa     %%xmm0,%%xmm1                                   \n"            \
-    "movdqa     %%xmm0,%%xmm2                                   \n"            \
-    "movdqa     %%xmm0,%%xmm3                                   \n"            \
-    "movdqa     " MEMACCESS2(96, [yuvconstants]) ",%%xmm0       \n"            \
-    "pmaddubsw  " MEMACCESS([yuvconstants]) ",%%xmm1            \n"            \
-    "psubw      %%xmm1,%%xmm0                                   \n"            \
-    "movdqa     " MEMACCESS2(128, [yuvconstants]) ",%%xmm1      \n"            \
-    "pmaddubsw  " MEMACCESS2(32, [yuvconstants]) ",%%xmm2       \n"            \
-    "psubw      %%xmm2,%%xmm1                                   \n"            \
-    "movdqa     " MEMACCESS2(160, [yuvconstants]) ",%%xmm2      \n"            \
-    "pmaddubsw  " MEMACCESS2(64, [yuvconstants]) ",%%xmm3       \n"            \
-    "psubw      %%xmm3,%%xmm2                                   \n"            \
-    "pmulhuw    " MEMACCESS2(192, [yuvconstants]) ",%%xmm4      \n"            \
-    "paddsw     %%xmm4,%%xmm0                                   \n"            \
-    "paddsw     %%xmm4,%%xmm1                                   \n"            \
-    "paddsw     %%xmm4,%%xmm2                                   \n"            \
-    "psraw      $0x6,%%xmm0                                     \n"            \
-    "psraw      $0x6,%%xmm1                                     \n"            \
-    "psraw      $0x6,%%xmm2                                     \n"            \
-    "packuswb   %%xmm0,%%xmm0                                   \n"            \
-    "packuswb   %%xmm1,%%xmm1                                   \n"            \
-    "packuswb   %%xmm2,%%xmm2                                   \n"
+#define YUVTORGB16(yuvconstants)                                  \
+  "pcmpeqb    %%xmm0,%%xmm0                                   \n" \
+  "pxor       %%xmm1,%%xmm1                                   \n" \
+  "psllw      $7,%%xmm0                                       \n" \
+  "pshufb     %%xmm1,%%xmm0                                   \n" \
+  "psubb      %%xmm0,%%xmm3                                   \n" \
+  "pmulhuw    96(%[yuvconstants]),%%xmm4                      \n" \
+  "movdqa     (%[yuvconstants]),%%xmm0                        \n" \
+  "movdqa     32(%[yuvconstants]),%%xmm1                      \n" \
+  "movdqa     64(%[yuvconstants]),%%xmm2                      \n" \
+  "pmaddubsw  %%xmm3,%%xmm0                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm1                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm2                                   \n" \
+  "movdqa     128(%[yuvconstants]),%%xmm3                     \n" \
+  "paddw      %%xmm3,%%xmm4                                   \n" \
+  "paddsw     %%xmm4,%%xmm0                                   \n" \
+  "paddsw     %%xmm4,%%xmm2                                   \n" \
+  "psubsw     %%xmm1,%%xmm4                                   \n" \
+  "movdqa     %%xmm4,%%xmm1                                   \n"
+
 #define YUVTORGB_REGS
 #endif
 
+#define YUVTORGB(yuvconstants)                                    \
+  YUVTORGB16(yuvconstants)                                        \
+  "psraw      $0x6,%%xmm0                                     \n" \
+  "psraw      $0x6,%%xmm1                                     \n" \
+  "psraw      $0x6,%%xmm2                                     \n" \
+  "packuswb   %%xmm0,%%xmm0                                   \n" \
+  "packuswb   %%xmm1,%%xmm1                                   \n" \
+  "packuswb   %%xmm2,%%xmm2                                   \n"
+
 // Store 8 ARGB values.
-#define STOREARGB                                                              \
-    "punpcklbw  %%xmm1,%%xmm0                                    \n"           \
-    "punpcklbw  %%xmm5,%%xmm2                                    \n"           \
-    "movdqa     %%xmm0,%%xmm1                                    \n"           \
-    "punpcklwd  %%xmm2,%%xmm0                                    \n"           \
-    "punpckhwd  %%xmm2,%%xmm1                                    \n"           \
-    "movdqu     %%xmm0," MEMACCESS([dst_argb]) "                 \n"           \
-    "movdqu     %%xmm1," MEMACCESS2(0x10, [dst_argb]) "          \n"           \
-    "lea        " MEMLEA(0x20, [dst_argb]) ", %[dst_argb]        \n"
+#define STOREARGB                                                  \
+  "punpcklbw  %%xmm1,%%xmm0                                    \n" \
+  "punpcklbw  %%xmm5,%%xmm2                                    \n" \
+  "movdqa     %%xmm0,%%xmm1                                    \n" \
+  "punpcklwd  %%xmm2,%%xmm0                                    \n" \
+  "punpckhwd  %%xmm2,%%xmm1                                    \n" \
+  "movdqu     %%xmm0,(%[dst_argb])                             \n" \
+  "movdqu     %%xmm1,0x10(%[dst_argb])                         \n" \
+  "lea        0x20(%[dst_argb]), %[dst_argb]                   \n"
 
 // Store 8 RGBA values.
-#define STORERGBA                                                              \
-    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
-    "punpcklbw %%xmm2,%%xmm1                                     \n"           \
-    "punpcklbw %%xmm0,%%xmm5                                     \n"           \
-    "movdqa    %%xmm5,%%xmm0                                     \n"           \
-    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
-    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
-    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "                  \n"           \
-    "movdqu    %%xmm0," MEMACCESS2(0x10, [dst_rgba]) "           \n"           \
-    "lea       " MEMLEA(0x20, [dst_rgba]) ",%[dst_rgba]          \n"
-
-void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
-                                const uint8* u_buf,
-                                const uint8* v_buf,
-                                uint8* dst_argb,
+#define STORERGBA                                                  \
+  "pcmpeqb   %%xmm5,%%xmm5                                     \n" \
+  "punpcklbw %%xmm2,%%xmm1                                     \n" \
+  "punpcklbw %%xmm0,%%xmm5                                     \n" \
+  "movdqa    %%xmm5,%%xmm0                                     \n" \
+  "punpcklwd %%xmm1,%%xmm5                                     \n" \
+  "punpckhwd %%xmm1,%%xmm0                                     \n" \
+  "movdqu    %%xmm5,(%[dst_rgba])                              \n" \
+  "movdqu    %%xmm0,0x10(%[dst_rgba])                          \n" \
+  "lea       0x20(%[dst_rgba]),%[dst_rgba]                     \n"
+
+// Store 8 RGB24 values.
+#define STORERGB24                                                      \
+  "punpcklbw   %%xmm1,%%xmm0                                        \n" \
+  "punpcklbw   %%xmm2,%%xmm2                                        \n" \
+  "movdqa      %%xmm0,%%xmm1                                        \n" \
+  "punpcklwd   %%xmm2,%%xmm0                                        \n" \
+  "punpckhwd   %%xmm2,%%xmm1                                        \n" \
+  "pshufb      %%xmm5,%%xmm0                                        \n" \
+  "pshufb      %%xmm6,%%xmm1                                        \n" \
+  "palignr     $0xc,%%xmm0,%%xmm1                                   \n" \
+  "movq        %%xmm0,(%[dst_rgb24])                                \n" \
+  "movdqu      %%xmm1,0x8(%[dst_rgb24])                             \n" \
+  "lea         0x18(%[dst_rgb24]),%[dst_rgb24]                      \n"
+
+// Store 8 AR30 values.
+#define STOREAR30                                                  \
+  "psraw      $0x4,%%xmm0                                      \n" \
+  "psraw      $0x4,%%xmm1                                      \n" \
+  "psraw      $0x4,%%xmm2                                      \n" \
+  "pminsw     %%xmm7,%%xmm0                                    \n" \
+  "pminsw     %%xmm7,%%xmm1                                    \n" \
+  "pminsw     %%xmm7,%%xmm2                                    \n" \
+  "pmaxsw     %%xmm6,%%xmm0                                    \n" \
+  "pmaxsw     %%xmm6,%%xmm1                                    \n" \
+  "pmaxsw     %%xmm6,%%xmm2                                    \n" \
+  "psllw      $0x4,%%xmm2                                      \n" \
+  "movdqa     %%xmm0,%%xmm3                                    \n" \
+  "punpcklwd  %%xmm2,%%xmm0                                    \n" \
+  "punpckhwd  %%xmm2,%%xmm3                                    \n" \
+  "movdqa     %%xmm1,%%xmm2                                    \n" \
+  "punpcklwd  %%xmm5,%%xmm1                                    \n" \
+  "punpckhwd  %%xmm5,%%xmm2                                    \n" \
+  "pslld      $0xa,%%xmm1                                      \n" \
+  "pslld      $0xa,%%xmm2                                      \n" \
+  "por        %%xmm1,%%xmm0                                    \n" \
+  "por        %%xmm2,%%xmm3                                    \n" \
+  "movdqu     %%xmm0,(%[dst_ar30])                             \n" \
+  "movdqu     %%xmm3,0x10(%[dst_ar30])                         \n" \
+  "lea        0x20(%[dst_ar30]), %[dst_ar30]                   \n"
+
+void OMITFP I444ToARGBRow_SSSE3(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV444
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS
+  : "memory", "cc", YUVTORGB_REGS
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
-void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
-                                 const uint8* u_buf,
-                                 const uint8* v_buf,
-                                 uint8* dst_rgb24,
+#ifdef HAS_I444ALPHATOARGBROW_SSSE3
+void OMITFP I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                                     const uint8_t* u_buf,
+                                     const uint8_t* v_buf,
+                                     const uint8_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  // clang-format off
+  asm volatile (
+  YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+  LABELALIGN
+      "1:                                        \n"
+  READYUVA444
+  YUVTORGB(yuvconstants)
+  STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [a_buf]"+r"(a_buf),    // %[a_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_I444ALPHATOARGBROW_SSSE3
+
+void OMITFP I422ToRGB24Row_SSSE3(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 uint8_t* dst_rgb24,
                                  const struct YuvConstants* yuvconstants,
                                  int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
-    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
-    "sub       %[u_buf],%[v_buf]               \n"
+      "movdqa      %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+      "movdqa      %[kShuffleMaskARGBToRGB24],%%xmm6 \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB(yuvconstants)
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm2,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "pshufb    %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm6,%%xmm1                   \n"
-    "palignr   $0xc,%%xmm0,%%xmm1              \n"
-    "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
-    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
-    "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
-    "subl      $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+    STORERGB24
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
-#if defined(__i386__) && defined(__pic__)
+#if defined(__i386__)
     [width]"+m"(width)     // %[width]
 #else
     [width]"+rm"(width)    // %[width]
@@ -1736,235 +2730,657 @@ void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
   : [yuvconstants]"r"(yuvconstants),  // %[yuvconstants]
     [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
     [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS
+  : "memory", "cc", YUVTORGB_REGS
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
   );
 }
 
-void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
-                                const uint8* u_buf,
-                                const uint8* v_buf,
-                                uint8* dst_argb,
+void OMITFP I444ToRGB24Row_SSSE3(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 uint8_t* dst_rgb24,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "movdqa      %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+      "movdqa      %[kShuffleMaskARGBToRGB24],%%xmm6 \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV444
+    YUVTORGB(yuvconstants)
+    STORERGB24
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+#if defined(__i386__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [yuvconstants]"r"(yuvconstants),  // %[yuvconstants]
+    [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void OMITFP I422ToARGBRow_SSSE3(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS
+  : "memory", "cc", YUVTORGB_REGS
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 
-#ifdef HAS_I422ALPHATOARGBROW_SSSE3
-void OMITFP I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
-                                     const uint8* u_buf,
-                                     const uint8* v_buf,
-                                     const uint8* a_buf,
-                                     uint8* dst_argb,
-                                     const struct YuvConstants* yuvconstants,
-                                     int width) {
+void OMITFP I422ToAR30Row_SSSE3(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // AR30 constants
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV422
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+// 10 bit YUV to ARGB
+void OMITFP I210ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
-    READYUVA422
+      "1:                                        \n"
+    READYUV210
     YUVTORGB(yuvconstants)
     STOREARGB
-    "subl      $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
-    [a_buf]"+r"(a_buf),    // %[a_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-#if defined(__i386__) && defined(__pic__)
-    [width]"+m"(width)     // %[width]
-#else
     [width]"+rm"(width)    // %[width]
-#endif
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS
+  : "memory", "cc", YUVTORGB_REGS
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
-#endif  // HAS_I422ALPHATOARGBROW_SSSE3
 
-#ifdef HAS_I411TOARGBROW_SSSE3
-void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
-                                const uint8* u_buf,
-                                const uint8* v_buf,
-                                uint8* dst_argb,
+// 12 bit YUV to ARGB
+void OMITFP I212ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212
+    YUVTORGB(yuvconstants)
+    STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+// 10 bit YUV to AR30
+void OMITFP I210ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV210
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+// 12 bit YUV to AR30
+void OMITFP I212ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+// 10 bit YUV to ARGB
+void OMITFP I410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
-  int temp;
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
-    READYUV411_TEMP
+      "1:                                        \n"
+    READYUV410
     YUVTORGB(yuvconstants)
     STOREARGB
-    "subl      $0x8,%[width]                   \n"
-    "jg        1b                              \n"
-  : [y_buf]"+r"(y_buf),        // %[y_buf]
-    [u_buf]"+r"(u_buf),        // %[u_buf]
-    [v_buf]"+r"(v_buf),        // %[v_buf]
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-    [temp]"=&r"(temp),         // %[temp]
-#if defined(__i386__) && defined(__pic__)
-    [width]"+m"(width)         // %[width]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+#ifdef HAS_I210ALPHATOARGBROW_SSSE3
+// 10 bit YUVA to ARGB
+void OMITFP I210AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                                     const uint16_t* u_buf,
+                                     const uint16_t* v_buf,
+                                     const uint16_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  asm volatile(
+      YUVTORGB_SETUP(
+      yuvconstants) "sub         %[u_buf],%[v_buf]             \n"
+
+      LABELALIGN "1:                                        \n" READYUVA210
+          YUVTORGB(yuvconstants) STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+      : [y_buf] "+r"(y_buf),  // %[y_buf]
+        [u_buf] "+r"(u_buf),  // %[u_buf]
+        [v_buf] "+r"(v_buf),  // %[v_buf]
+        [a_buf] "+r"(a_buf),
+        [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+        [width] "+m"(width)  // %[width]
+#else
+        [width] "+rm"(width)  // %[width]
+#endif
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+        "xmm5");
+}
+#endif
+
+#ifdef HAS_I410ALPHATOARGBROW_SSSE3
+// 10 bit YUVA to ARGB
+void OMITFP I410AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                                     const uint16_t* u_buf,
+                                     const uint16_t* v_buf,
+                                     const uint16_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  // clang-format off
+  asm volatile(
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUVA410
+    YUVTORGB(yuvconstants)
+    STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+    : [y_buf] "+r"(y_buf),  // %[y_buf]
+      [u_buf] "+r"(u_buf),  // %[u_buf]
+      [v_buf] "+r"(v_buf),  // %[v_buf]
+      [a_buf] "+r"(a_buf),
+      [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+      [width] "+m"(width)  // %[width]
 #else
-    [width]"+rm"(width)        // %[width]
+      [width] "+rm"(width)  // %[width]
+#endif
+    : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+    : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+      "xmm5");
+  // clang-format on
+}
+#endif
+
+// 10 bit YUV to AR30
+void OMITFP I410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV410
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)   // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+#ifdef HAS_I422ALPHATOARGBROW_SSSE3
+void OMITFP I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                                     const uint8_t* u_buf,
+                                     const uint8_t* v_buf,
+                                     const uint8_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  // clang-format off
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUVA422
+    YUVTORGB(yuvconstants)
+    STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [a_buf]"+r"(a_buf),    // %[a_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
 #endif
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS
+  : "memory", "cc", YUVTORGB_REGS
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
-#endif
+#endif  // HAS_I422ALPHATOARGBROW_SSSE3
 
-void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
-                                const uint8* uv_buf,
-                                uint8* dst_argb,
+void OMITFP NV12ToARGBRow_SSSE3(const uint8_t* y_buf,
+                                const uint8_t* uv_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READNV12
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [uv_buf]"+r"(uv_buf),    // %[uv_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-    : "memory", "cc", YUVTORGB_REGS  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS
       "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 
-void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
-                                const uint8* vu_buf,
-                                uint8* dst_argb,
+void OMITFP NV21ToARGBRow_SSSE3(const uint8_t* y_buf,
+                                const uint8_t* vu_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READNV21
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [vu_buf]"+r"(vu_buf),    // %[vu_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
     [kShuffleNV21]"m"(kShuffleNV21)
-    : "memory", "cc", YUVTORGB_REGS  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS
       "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 
-void OMITFP YUY2ToARGBRow_SSSE3(const uint8* yuy2_buf,
-                                uint8* dst_argb,
+void OMITFP YUY2ToARGBRow_SSSE3(const uint8_t* yuy2_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUY2
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [yuy2_buf]"+r"(yuy2_buf),    // %[yuy2_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
     [kShuffleYUY2Y]"m"(kShuffleYUY2Y),
     [kShuffleYUY2UV]"m"(kShuffleYUY2UV)
-    : "memory", "cc", YUVTORGB_REGS  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS
       "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 
-void OMITFP UYVYToARGBRow_SSSE3(const uint8* uyvy_buf,
-                                uint8* dst_argb,
+void OMITFP UYVYToARGBRow_SSSE3(const uint8_t* uyvy_buf,
+                                uint8_t* dst_argb,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READUYVY
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [uyvy_buf]"+r"(uyvy_buf),    // %[uyvy_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
     [kShuffleUYVYY]"m"(kShuffleUYVYY),
     [kShuffleUYVYUV]"m"(kShuffleUYVYUV)
-    : "memory", "cc", YUVTORGB_REGS  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS
       "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 
-void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
-                                const uint8* u_buf,
-                                const uint8* v_buf,
-                                uint8* dst_rgba,
+void OMITFP P210ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile(
+      YUVTORGB_SETUP(
+      yuvconstants) "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN "1:                                        \n" READP210
+          YUVTORGB(yuvconstants) STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+      : [y_buf] "+r"(y_buf),              // %[y_buf]
+        [uv_buf] "+r"(uv_buf),            // %[u_buf]
+        [dst_argb] "+r"(dst_argb),        // %[dst_argb]
+        [width] "+rm"(width)              // %[width]
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+        "xmm5");
+}
+
+void OMITFP P410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile(
+      YUVTORGB_SETUP(
+      yuvconstants) "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN "1:                                        \n" READP410
+          YUVTORGB(yuvconstants) STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+      : [y_buf] "+r"(y_buf),              // %[y_buf]
+        [uv_buf] "+r"(uv_buf),            // %[u_buf]
+        [dst_argb] "+r"(dst_argb),        // %[dst_argb]
+        [width] "+rm"(width)              // %[width]
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+        "xmm5");
+}
+
+void OMITFP P210ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READP210
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),              // %[y_buf]
+    [uv_buf]"+r"(uv_buf),            // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),        // %[dst_ar30]
+    [width]"+rm"(width)              // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void OMITFP P410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READP410
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),              // %[y_buf]
+    [uv_buf]"+r"(uv_buf),            // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),        // %[dst_ar30]
+    [width]"+rm"(width)              // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_rgba,
                                 const struct YuvConstants* yuvconstants,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB(yuvconstants)
     STORERGBA
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS
+  : "memory", "cc", YUVTORGB_REGS
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
@@ -1972,316 +3388,928 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
 #endif  // HAS_I422TOARGBROW_SSSE3
 
 // Read 16 UV from 444
-#define READYUV444_AVX2                                                        \
-    "vmovdqu    " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
-    MEMOPREG(vmovdqu, 0x00, [u_buf], [v_buf], 1, xmm1)                         \
-    "lea        " MEMLEA(0x10, [u_buf]) ",%[u_buf]                  \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n"        \
-    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
-    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
-    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
-    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
+#define READYUV444_AVX2                                               \
+  "vmovdqu    (%[u_buf]),%%xmm3                                   \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                             \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n"
 
 // Read 8 UV from 422, upsample to 16 UV.
-#define READYUV422_AVX2                                                        \
-    "vmovq      " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
-    MEMOPREG(vmovq, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
-    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]                   \n"        \
-    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"        \
-    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
-    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
-    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
+#define READYUV422_AVX2                                               \
+  "vmovq      (%[u_buf]),%%xmm3                                   \n" \
+  "vmovq      0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                              \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n"
+
+#define READYUV422_AVX512BW                                           \
+  "vmovdqu    (%[u_buf]),%%xmm3                                   \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "vpermq     %%zmm3,%%zmm16,%%zmm3                               \n" \
+  "vpermq     %%zmm1,%%zmm16,%%zmm1                               \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                             \n" \
+  "vpunpcklbw %%zmm1,%%zmm3,%%zmm3                                \n" \
+  "vpermq     $0xd8,%%zmm3,%%zmm3                                 \n" \
+  "vpunpcklwd %%zmm3,%%zmm3,%%zmm3                                \n" \
+  "vmovdqu8   (%[y_buf]),%%ymm4                                   \n" \
+  "vpermq     %%zmm4,%%zmm17,%%zmm4                               \n" \
+  "vpermq     $0xd8,%%zmm4,%%zmm4                                 \n" \
+  "vpunpcklbw %%zmm4,%%zmm4,%%zmm4                                \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                             \n"
+
+// Read 8 UV from 210, upsample to 16 UV
+// TODO(fbarchard): Consider vshufb to replace pack/unpack
+// TODO(fbarchard): Consider vunpcklpd to combine the 2 registers into 1.
+#define READYUV210_AVX2                                            \
+  "vmovdqu    (%[u_buf]),%%xmm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                 \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                          \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                              \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                              \n" \
+  "vpunpcklwd %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm2                                 \n" \
+  "vpsrlw     $4,%%ymm4,%%ymm4                                 \n" \
+  "vpaddw     %%ymm2,%%ymm4,%%ymm4                             \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n"
+
+// Read 8 UV from 210, upsample to 16 UV. With 16 Alpha.
+#define READYUVA210_AVX2                                           \
+  "vmovdqu    (%[u_buf]),%%xmm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                 \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                          \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                              \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                              \n" \
+  "vpunpcklwd %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm2                                 \n" \
+  "vpsrlw     $4,%%ymm4,%%ymm4                                 \n" \
+  "vpaddw     %%ymm2,%%ymm4,%%ymm4                             \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n" \
+  "vmovdqu    (%[a_buf]),%%ymm5                                \n" \
+  "vpsraw     $2,%%ymm5,%%ymm5                                 \n" \
+  "vpackuswb  %%ymm5,%%ymm5,%%ymm5                             \n" \
+  "lea        0x20(%[a_buf]),%[a_buf]                          \n"
+
+// Read 16 UV from 410
+#define READYUV410_AVX2                                            \
+  "vmovdqu    (%[u_buf]),%%ymm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%ymm2                 \n" \
+  "lea        0x20(%[u_buf]),%[u_buf]                          \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpsraw     $2,%%ymm2,%%ymm2                                 \n" \
+  "vpunpckhwd %%ymm2,%%ymm3,%%ymm1                             \n" \
+  "vpunpcklwd %%ymm2,%%ymm3,%%ymm3                             \n" \
+  "vpackuswb  %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm2                                 \n" \
+  "vpsrlw     $4,%%ymm4,%%ymm4                                 \n" \
+  "vpaddw     %%ymm2,%%ymm4,%%ymm4                             \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n"
+
+// Read 8 UV from 212 12 bit, upsample to 16 UV
+#define READYUV212_AVX2                                            \
+  "vmovdqu    (%[u_buf]),%%xmm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                 \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                          \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                              \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                              \n" \
+  "vpunpcklwd %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vpsraw     $0x4,%%ymm3,%%ymm3                               \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $4,%%ymm4,%%ymm2                                 \n" \
+  "vpsrlw     $8,%%ymm4,%%ymm4                                 \n" \
+  "vpaddw     %%ymm2,%%ymm4,%%ymm4                             \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n"
+
+// Read 16 UV from 410. With 16 Alpha.
+#define READYUVA410_AVX2                                           \
+  "vmovdqu    (%[u_buf]),%%ymm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%ymm2                 \n" \
+  "lea        0x20(%[u_buf]),%[u_buf]                          \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpsraw     $2,%%ymm2,%%ymm2                                 \n" \
+  "vpunpckhwd %%ymm2,%%ymm3,%%ymm1                             \n" \
+  "vpunpcklwd %%ymm2,%%ymm3,%%ymm3                             \n" \
+  "vpackuswb  %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm2                                 \n" \
+  "vpsrlw     $4,%%ymm4,%%ymm4                                 \n" \
+  "vpaddw     %%ymm2,%%ymm4,%%ymm4                             \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n" \
+  "vmovdqu    (%[a_buf]),%%ymm5                                \n" \
+  "vpsraw     $2,%%ymm5,%%ymm5                                 \n" \
+  "vpackuswb  %%ymm5,%%ymm5,%%ymm5                             \n" \
+  "lea        0x20(%[a_buf]),%[a_buf]                          \n"
+
+// Read 16 UV from 444.  With 16 Alpha.
+#define READYUVA444_AVX2                                              \
+  "vmovdqu    (%[u_buf]),%%xmm3                                   \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                             \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n" \
+  "vmovdqu    (%[a_buf]),%%xmm5                                   \n" \
+  "vpermq     $0xd8,%%ymm5,%%ymm5                                 \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                             \n"
 
 // Read 8 UV from 422, upsample to 16 UV.  With 16 Alpha.
-#define READYUVA422_AVX2                                                       \
-    "vmovq      " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
-    MEMOPREG(vmovq, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
-    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]                   \n"        \
-    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"        \
-    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
-    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
-    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"        \
-    "vmovdqu    " MEMACCESS([a_buf]) ",%%xmm5                       \n"        \
-    "vpermq     $0xd8,%%ymm5,%%ymm5                                 \n"        \
-    "lea        " MEMLEA(0x10, [a_buf]) ",%[a_buf]                  \n"
-
-// Read 4 UV from 411, upsample to 16 UV.
-#define READYUV411_AVX2                                                        \
-    "vmovd      " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
-    MEMOPREG(vmovd, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
-    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]                   \n"        \
-    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
-    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpunpckldq %%ymm0,%%ymm0,%%ymm0                                \n"        \
-    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
-    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
-    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
+#define READYUVA422_AVX2                                              \
+  "vmovq      (%[u_buf]),%%xmm3                                   \n" \
+  "vmovq      0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                              \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n" \
+  "vmovdqu    (%[a_buf]),%%xmm5                                   \n" \
+  "vpermq     $0xd8,%%ymm5,%%ymm5                                 \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                             \n"
 
 // Read 8 UV from NV12, upsample to 16 UV.
-#define READNV12_AVX2                                                          \
-    "vmovdqu    " MEMACCESS([uv_buf]) ",%%xmm0                      \n"        \
-    "lea        " MEMLEA(0x10, [uv_buf]) ",%[uv_buf]                \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"        \
-    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
-    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
-    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
+#define READNV12_AVX2                                                 \
+  "vmovdqu    (%[uv_buf]),%%xmm3                                  \n" \
+  "lea        0x10(%[uv_buf]),%[uv_buf]                           \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n"
 
 // Read 8 VU from NV21, upsample to 16 UV.
-#define READNV21_AVX2                                                          \
-    "vmovdqu    " MEMACCESS([vu_buf]) ",%%xmm0                      \n"        \
-    "lea        " MEMLEA(0x10, [vu_buf]) ",%[vu_buf]                \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpshufb     %[kShuffleNV21], %%ymm0, %%ymm0                    \n"        \
-    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
-    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
-    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
+#define READNV21_AVX2                                                 \
+  "vmovdqu    (%[vu_buf]),%%xmm3                                  \n" \
+  "lea        0x10(%[vu_buf]),%[vu_buf]                           \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpshufb     %[kShuffleNV21], %%ymm3, %%ymm3                    \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n"
+
+// Read 4 UV from P210, upsample to 8 UV
+#define READP210_AVX2                                                 \
+  "vmovdqu    (%[uv_buf]),%%ymm3                                  \n" \
+  "lea        0x20(%[uv_buf]),%[uv_buf]                           \n" \
+  "vpsrlw     $0x8,%%ymm3,%%ymm3                                  \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                   \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                             \n"
+
+// Read 8 UV from P410
+#define READP410_AVX2                                                 \
+  "vmovdqu    (%[uv_buf]),%%ymm3                                  \n" \
+  "vmovdqu    0x20(%[uv_buf]),%%ymm1                              \n" \
+  "lea        0x40(%[uv_buf]),%[uv_buf]                           \n" \
+  "vpsrlw     $0x8,%%ymm3,%%ymm3                                  \n" \
+  "vpsrlw     $0x8,%%ymm1,%%ymm1                                  \n" \
+  "vpackuswb  %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                   \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                             \n"
 
 // Read 8 YUY2 with 16 Y and upsample 8 UV to 16 UV.
-#define READYUY2_AVX2                                                          \
-    "vmovdqu    " MEMACCESS([yuy2_buf]) ",%%ymm4                    \n"        \
-    "vpshufb    %[kShuffleYUY2Y], %%ymm4, %%ymm4                    \n"        \
-    "vmovdqu    " MEMACCESS([yuy2_buf]) ",%%ymm0                    \n"        \
-    "vpshufb    %[kShuffleYUY2UV], %%ymm0, %%ymm0                   \n"        \
-    "lea        " MEMLEA(0x20, [yuy2_buf]) ",%[yuy2_buf]            \n"
+#define READYUY2_AVX2                                                 \
+  "vmovdqu    (%[yuy2_buf]),%%ymm4                                \n" \
+  "vpshufb    %[kShuffleYUY2Y], %%ymm4, %%ymm4                    \n" \
+  "vmovdqu    (%[yuy2_buf]),%%ymm3                                \n" \
+  "vpshufb    %[kShuffleYUY2UV], %%ymm3, %%ymm3                   \n" \
+  "lea        0x20(%[yuy2_buf]),%[yuy2_buf]                       \n"
 
 // Read 8 UYVY with 16 Y and upsample 8 UV to 16 UV.
-#define READUYVY_AVX2                                                          \
-    "vmovdqu     " MEMACCESS([uyvy_buf]) ",%%ymm4                   \n"        \
-    "vpshufb     %[kShuffleUYVYY], %%ymm4, %%ymm4                   \n"        \
-    "vmovdqu     " MEMACCESS([uyvy_buf]) ",%%ymm0                   \n"        \
-    "vpshufb     %[kShuffleUYVYUV], %%ymm0, %%ymm0                  \n"        \
-    "lea        " MEMLEA(0x20, [uyvy_buf]) ",%[uyvy_buf]            \n"
-
+#define READUYVY_AVX2                                                 \
+  "vmovdqu    (%[uyvy_buf]),%%ymm4                                \n" \
+  "vpshufb    %[kShuffleUYVYY], %%ymm4, %%ymm4                    \n" \
+  "vmovdqu    (%[uyvy_buf]),%%ymm3                                \n" \
+  "vpshufb    %[kShuffleUYVYUV], %%ymm3, %%ymm3                   \n" \
+  "lea        0x20(%[uyvy_buf]),%[uyvy_buf]                       \n"
+
+// TODO(fbarchard): Remove broadcastb
 #if defined(__x86_64__)
-#define YUVTORGB_SETUP_AVX2(yuvconstants)                                      \
-    "vmovdqa     " MEMACCESS([yuvconstants]) ",%%ymm8            \n"           \
-    "vmovdqa     " MEMACCESS2(32, [yuvconstants]) ",%%ymm9       \n"           \
-    "vmovdqa     " MEMACCESS2(64, [yuvconstants]) ",%%ymm10      \n"           \
-    "vmovdqa     " MEMACCESS2(96, [yuvconstants]) ",%%ymm11      \n"           \
-    "vmovdqa     " MEMACCESS2(128, [yuvconstants]) ",%%ymm12     \n"           \
-    "vmovdqa     " MEMACCESS2(160, [yuvconstants]) ",%%ymm13     \n"           \
-    "vmovdqa     " MEMACCESS2(192, [yuvconstants]) ",%%ymm14     \n"
-#define YUVTORGB_AVX2(yuvconstants)                                            \
-    "vpmaddubsw  %%ymm10,%%ymm0,%%ymm2                              \n"        \
-    "vpmaddubsw  %%ymm9,%%ymm0,%%ymm1                               \n"        \
-    "vpmaddubsw  %%ymm8,%%ymm0,%%ymm0                               \n"        \
-    "vpsubw      %%ymm2,%%ymm13,%%ymm2                              \n"        \
-    "vpsubw      %%ymm1,%%ymm12,%%ymm1                              \n"        \
-    "vpsubw      %%ymm0,%%ymm11,%%ymm0                              \n"        \
-    "vpmulhuw    %%ymm14,%%ymm4,%%ymm4                              \n"        \
-    "vpaddsw     %%ymm4,%%ymm0,%%ymm0                               \n"        \
-    "vpaddsw     %%ymm4,%%ymm1,%%ymm1                               \n"        \
-    "vpaddsw     %%ymm4,%%ymm2,%%ymm2                               \n"        \
-    "vpsraw      $0x6,%%ymm0,%%ymm0                                 \n"        \
-    "vpsraw      $0x6,%%ymm1,%%ymm1                                 \n"        \
-    "vpsraw      $0x6,%%ymm2,%%ymm2                                 \n"        \
-    "vpackuswb   %%ymm0,%%ymm0,%%ymm0                               \n"        \
-    "vpackuswb   %%ymm1,%%ymm1,%%ymm1                               \n"        \
-    "vpackuswb   %%ymm2,%%ymm2,%%ymm2                               \n"
-#define YUVTORGB_REGS_AVX2 \
-    "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14",
+#define YUVTORGB_SETUP_AVX2(yuvconstants)                             \
+  "vpcmpeqb    %%xmm13,%%xmm13,%%xmm13                            \n" \
+  "vmovdqa     (%[yuvconstants]),%%ymm8                           \n" \
+  "vpsllw      $7,%%xmm13,%%xmm13                                 \n" \
+  "vmovdqa     32(%[yuvconstants]),%%ymm9                         \n" \
+  "vpbroadcastb %%xmm13,%%ymm13                                   \n" \
+  "vmovdqa     64(%[yuvconstants]),%%ymm10                        \n" \
+  "vmovdqa     96(%[yuvconstants]),%%ymm11                        \n" \
+  "vmovdqa     128(%[yuvconstants]),%%ymm12                       \n"
+
+#define YUVTORGB_SETUP_AVX512BW(yuvconstants)                         \
+  "vpcmpeqb   %%xmm13,%%xmm13,%%xmm13                             \n" \
+  "movdqa     (%[yuvconstants]),%%xmm8                            \n" \
+  "vpbroadcastq %%xmm8, %%zmm8                                    \n" \
+  "vpsllw     $7,%%xmm13,%%xmm13                                  \n" \
+  "vpbroadcastb %%xmm13,%%zmm13                                   \n" \
+  "movq     32(%[yuvconstants]),%%xmm9                            \n" \
+  "vpbroadcastq %%xmm9,%%zmm9                                     \n" \
+  "movq     64(%[yuvconstants]),%%xmm10                           \n" \
+  "vpbroadcastq %%xmm10,%%zmm10                                   \n" \
+  "movq     96(%[yuvconstants]),%%xmm11                           \n" \
+  "vpbroadcastq %%xmm11,%%zmm11                                   \n" \
+  "movq     128(%[yuvconstants]),%%xmm12                          \n" \
+  "vpbroadcastq %%xmm12,%%zmm12                                   \n" \
+  "vmovdqu8 (%[quadsplitperm]),%%zmm16                            \n" \
+  "vmovdqu8 (%[dquadsplitperm]),%%zmm17                           \n" \
+  "vmovdqu8 (%[unperm]),%%zmm18                                   \n"
+
+#define YUVTORGB16_AVX2(yuvconstants)                                 \
+  "vpsubb      %%ymm13,%%ymm3,%%ymm3                              \n" \
+  "vpmulhuw    %%ymm11,%%ymm4,%%ymm4                              \n" \
+  "vpmaddubsw  %%ymm3,%%ymm8,%%ymm0                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm9,%%ymm1                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm10,%%ymm2                              \n" \
+  "vpaddw      %%ymm4,%%ymm12,%%ymm4                              \n" \
+  "vpaddsw     %%ymm4,%%ymm0,%%ymm0                               \n" \
+  "vpsubsw     %%ymm1,%%ymm4,%%ymm1                               \n" \
+  "vpaddsw     %%ymm4,%%ymm2,%%ymm2                               \n"
+
+#define YUVTORGB16_AVX512BW(yuvconstants)                             \
+  "vpsubb      %%zmm13,%%zmm3,%%zmm3                              \n" \
+  "vpmulhuw    %%zmm11,%%zmm4,%%zmm4                              \n" \
+  "vpmaddubsw  %%zmm3,%%zmm8,%%zmm0                               \n" \
+  "vpmaddubsw  %%zmm3,%%zmm9,%%zmm1                               \n" \
+  "vpmaddubsw  %%zmm3,%%zmm10,%%zmm2                              \n" \
+  "vpaddw      %%zmm4,%%zmm12,%%zmm4                              \n" \
+  "vpaddsw     %%zmm4,%%zmm0,%%zmm0                               \n" \
+  "vpsubsw     %%zmm1,%%zmm4,%%zmm1                               \n" \
+  "vpaddsw     %%zmm4,%%zmm2,%%zmm2                               \n"
+
+#define YUVTORGB_REGS_AVX2 "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",
+#define YUVTORGB_REGS_AVX512BW \
+  "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm16", "xmm17", "xmm18",
+
 #else  // Convert 16 pixels: 16 UV and 16 Y.
+
 #define YUVTORGB_SETUP_AVX2(yuvconstants)
-#define YUVTORGB_AVX2(yuvconstants)                                            \
-    "vpmaddubsw  " MEMACCESS2(64, [yuvconstants]) ",%%ymm0,%%ymm2   \n"        \
-    "vpmaddubsw  " MEMACCESS2(32, [yuvconstants]) ",%%ymm0,%%ymm1   \n"        \
-    "vpmaddubsw  " MEMACCESS([yuvconstants]) ",%%ymm0,%%ymm0        \n"        \
-    "vmovdqu     " MEMACCESS2(160, [yuvconstants]) ",%%ymm3         \n"        \
-    "vpsubw      %%ymm2,%%ymm3,%%ymm2                               \n"        \
-    "vmovdqu     " MEMACCESS2(128, [yuvconstants]) ",%%ymm3         \n"        \
-    "vpsubw      %%ymm1,%%ymm3,%%ymm1                               \n"        \
-    "vmovdqu     " MEMACCESS2(96, [yuvconstants]) ",%%ymm3          \n"        \
-    "vpsubw      %%ymm0,%%ymm3,%%ymm0                               \n"        \
-    "vpmulhuw    " MEMACCESS2(192, [yuvconstants]) ",%%ymm4,%%ymm4  \n"        \
-    "vpaddsw     %%ymm4,%%ymm0,%%ymm0                               \n"        \
-    "vpaddsw     %%ymm4,%%ymm1,%%ymm1                               \n"        \
-    "vpaddsw     %%ymm4,%%ymm2,%%ymm2                               \n"        \
-    "vpsraw      $0x6,%%ymm0,%%ymm0                                 \n"        \
-    "vpsraw      $0x6,%%ymm1,%%ymm1                                 \n"        \
-    "vpsraw      $0x6,%%ymm2,%%ymm2                                 \n"        \
-    "vpackuswb   %%ymm0,%%ymm0,%%ymm0                               \n"        \
-    "vpackuswb   %%ymm1,%%ymm1,%%ymm1                               \n"        \
-    "vpackuswb   %%ymm2,%%ymm2,%%ymm2                               \n"
+#define YUVTORGB16_AVX2(yuvconstants)                                 \
+  "vpcmpeqb    %%xmm0,%%xmm0,%%xmm0                               \n" \
+  "vpsllw      $7,%%xmm0,%%xmm0                                   \n" \
+  "vpbroadcastb %%xmm0,%%ymm0                                     \n" \
+  "vpsubb      %%ymm0,%%ymm3,%%ymm3                               \n" \
+  "vpmulhuw    96(%[yuvconstants]),%%ymm4,%%ymm4                  \n" \
+  "vmovdqa     (%[yuvconstants]),%%ymm0                           \n" \
+  "vmovdqa     32(%[yuvconstants]),%%ymm1                         \n" \
+  "vmovdqa     64(%[yuvconstants]),%%ymm2                         \n" \
+  "vpmaddubsw  %%ymm3,%%ymm0,%%ymm0                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm1,%%ymm1                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm2,%%ymm2                               \n" \
+  "vmovdqa     128(%[yuvconstants]),%%ymm3                        \n" \
+  "vpaddw      %%ymm4,%%ymm3,%%ymm4                               \n" \
+  "vpaddsw     %%ymm4,%%ymm0,%%ymm0                               \n" \
+  "vpsubsw     %%ymm1,%%ymm4,%%ymm1                               \n" \
+  "vpaddsw     %%ymm4,%%ymm2,%%ymm2                               \n"
+
 #define YUVTORGB_REGS_AVX2
 #endif
 
+#define YUVTORGB_AVX2(yuvconstants)                                   \
+  YUVTORGB16_AVX2(yuvconstants)                                       \
+  "vpsraw      $0x6,%%ymm0,%%ymm0                                 \n" \
+  "vpsraw      $0x6,%%ymm1,%%ymm1                                 \n" \
+  "vpsraw      $0x6,%%ymm2,%%ymm2                                 \n" \
+  "vpackuswb   %%ymm0,%%ymm0,%%ymm0                               \n" \
+  "vpackuswb   %%ymm1,%%ymm1,%%ymm1                               \n" \
+  "vpackuswb   %%ymm2,%%ymm2,%%ymm2                               \n"
+
+#define YUVTORGB_AVX512BW(yuvconstants)                               \
+  YUVTORGB16_AVX512BW(yuvconstants)                                   \
+  "vpsraw     $0x6,%%zmm0,%%zmm0                                  \n" \
+  "vpsraw     $0x6,%%zmm1,%%zmm1                                  \n" \
+  "vpsraw     $0x6,%%zmm2,%%zmm2                                  \n" \
+  "vpackuswb  %%zmm0,%%zmm0,%%zmm0                                \n" \
+  "vpackuswb  %%zmm1,%%zmm1,%%zmm1                                \n" \
+  "vpackuswb  %%zmm2,%%zmm2,%%zmm2                                \n"
+
 // Store 16 ARGB values.
-#define STOREARGB_AVX2                                                         \
-    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
-    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
-    "vpunpcklbw %%ymm5,%%ymm2,%%ymm2                                \n"        \
-    "vpermq     $0xd8,%%ymm2,%%ymm2                                 \n"        \
-    "vpunpcklwd %%ymm2,%%ymm0,%%ymm1                                \n"        \
-    "vpunpckhwd %%ymm2,%%ymm0,%%ymm0                                \n"        \
-    "vmovdqu    %%ymm1," MEMACCESS([dst_argb]) "                    \n"        \
-    "vmovdqu    %%ymm0," MEMACCESS2(0x20, [dst_argb]) "             \n"        \
-    "lea       " MEMLEA(0x40, [dst_argb]) ", %[dst_argb]            \n"
+#define STOREARGB_AVX2                                                \
+  "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n" \
+  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
+  "vpunpcklbw %%ymm5,%%ymm2,%%ymm2                                \n" \
+  "vpermq     $0xd8,%%ymm2,%%ymm2                                 \n" \
+  "vpunpcklwd %%ymm2,%%ymm0,%%ymm1                                \n" \
+  "vpunpckhwd %%ymm2,%%ymm0,%%ymm0                                \n" \
+  "vmovdqu    %%ymm1,(%[dst_argb])                                \n" \
+  "vmovdqu    %%ymm0,0x20(%[dst_argb])                            \n" \
+  "lea        0x40(%[dst_argb]), %[dst_argb]                      \n"
+
+// Store 32 ARGB values.
+#define STOREARGB_AVX512BW                                            \
+  "vpunpcklbw %%zmm1,%%zmm0,%%zmm0                                \n" \
+  "vpermq     %%zmm0,%%zmm18,%%zmm0                               \n" \
+  "vpunpcklbw %%zmm5,%%zmm2,%%zmm2                                \n" \
+  "vpermq     %%zmm2,%%zmm18,%%zmm2                               \n" \
+  "vpunpcklwd %%zmm2,%%zmm0,%%zmm1                                \n" \
+  "vpunpckhwd %%zmm2,%%zmm0,%%zmm0                                \n" \
+  "vmovdqu8   %%zmm1,(%[dst_argb])                                \n" \
+  "vmovdqu8   %%zmm0,0x40(%[dst_argb])                            \n" \
+  "lea        0x80(%[dst_argb]), %[dst_argb]                      \n"
+
+// Store 16 AR30 values.
+#define STOREAR30_AVX2                                                \
+  "vpsraw     $0x4,%%ymm0,%%ymm0                                  \n" \
+  "vpsraw     $0x4,%%ymm1,%%ymm1                                  \n" \
+  "vpsraw     $0x4,%%ymm2,%%ymm2                                  \n" \
+  "vpminsw    %%ymm7,%%ymm0,%%ymm0                                \n" \
+  "vpminsw    %%ymm7,%%ymm1,%%ymm1                                \n" \
+  "vpminsw    %%ymm7,%%ymm2,%%ymm2                                \n" \
+  "vpmaxsw    %%ymm6,%%ymm0,%%ymm0                                \n" \
+  "vpmaxsw    %%ymm6,%%ymm1,%%ymm1                                \n" \
+  "vpmaxsw    %%ymm6,%%ymm2,%%ymm2                                \n" \
+  "vpsllw     $0x4,%%ymm2,%%ymm2                                  \n" \
+  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n" \
+  "vpermq     $0xd8,%%ymm2,%%ymm2                                 \n" \
+  "vpunpckhwd %%ymm2,%%ymm0,%%ymm3                                \n" \
+  "vpunpcklwd %%ymm2,%%ymm0,%%ymm0                                \n" \
+  "vpunpckhwd %%ymm5,%%ymm1,%%ymm2                                \n" \
+  "vpunpcklwd %%ymm5,%%ymm1,%%ymm1                                \n" \
+  "vpslld     $0xa,%%ymm1,%%ymm1                                  \n" \
+  "vpslld     $0xa,%%ymm2,%%ymm2                                  \n" \
+  "vpor       %%ymm1,%%ymm0,%%ymm0                                \n" \
+  "vpor       %%ymm2,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    %%ymm0,(%[dst_ar30])                                \n" \
+  "vmovdqu    %%ymm3,0x20(%[dst_ar30])                            \n" \
+  "lea        0x40(%[dst_ar30]), %[dst_ar30]                      \n"
 
 #ifdef HAS_I444TOARGBROW_AVX2
 // 16 pixels
 // 16 UV values with 16 Y producing 16 ARGB (64 bytes).
-void OMITFP I444ToARGBRow_AVX2(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* dst_argb,
+void OMITFP I444ToARGBRow_AVX2(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV444_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS_AVX2
+  : "memory", "cc", YUVTORGB_REGS_AVX2
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
 #endif  // HAS_I444TOARGBROW_AVX2
 
-#ifdef HAS_I411TOARGBROW_AVX2
+#if defined(HAS_I422TOARGBROW_AVX2)
 // 16 pixels
-// 4 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-void OMITFP I411ToARGBRow_AVX2(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* dst_argb,
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I422ToARGBRow_AVX2(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
-    READYUV411_AVX2
+      "1:                                        \n"
+    READYUV422_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS_AVX2
+  : "memory", "cc", YUVTORGB_REGS_AVX2
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
-#endif  // HAS_I411TOARGBROW_AVX2
+#endif  // HAS_I422TOARGBROW_AVX2
 
-#if defined(HAS_I422TOARGBROW_AVX2)
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+static const uint64_t kSplitQuadWords[8] = {0, 2, 2, 2, 1, 2, 2, 2};
+static const uint64_t kSplitDoubleQuadWords[8] = {0, 1, 4, 4, 2, 3, 4, 4};
+static const uint64_t kUnpermuteAVX512[8] = {0, 4, 1, 5, 2, 6, 3, 7};
+
+// 32 pixels
+// 16 UV values upsampled to 32 UV, mixed with 32 Y producing 32 ARGB (128
+// bytes).
+void OMITFP I422ToARGBRow_AVX512BW(const uint8_t* y_buf,
+                                   const uint8_t* u_buf,
+                                   const uint8_t* v_buf,
+                                   uint8_t* dst_argb,
+                                   const struct YuvConstants* yuvconstants,
+                                   int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX512BW(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%xmm5,%%xmm5,%%xmm5          \n"
+      "vpbroadcastq %%xmm5,%%zmm5                \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV422_AVX512BW
+    YUVTORGB_AVX512BW(yuvconstants)
+    STOREARGB_AVX512BW
+      "sub         $0x20,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),                         // %[y_buf]
+    [u_buf]"+r"(u_buf),                         // %[u_buf]
+    [v_buf]"+r"(v_buf),                         // %[v_buf]
+    [dst_argb]"+r"(dst_argb),                   // %[dst_argb]
+    [width]"+rm"(width)                         // %[width]
+  : [yuvconstants]"r"(yuvconstants),            // %[yuvconstants]
+    [quadsplitperm]"r"(kSplitQuadWords),        // %[quadsplitperm]
+    [dquadsplitperm]"r"(kSplitDoubleQuadWords), // %[dquadsplitperm]
+    [unperm]"r"(kUnpermuteAVX512)               // %[unperm]
+  : "memory", "cc", YUVTORGB_REGS_AVX512BW
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I422TOARGBROW_AVX512BW
+
+#if defined(HAS_I422TOAR30ROW_AVX2)
 // 16 pixels
-// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* dst_argb,
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I422ToAR30Row_AVX2(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_ar30,
                                const struct YuvConstants* yuvconstants,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV422_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_I422TOAR30ROW_AVX2
+
+#if defined(HAS_I210TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I210ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV210_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS_AVX2
+  : "memory", "cc", YUVTORGB_REGS_AVX2
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
-#endif  // HAS_I422TOARGBROW_AVX2
+#endif  // HAS_I210TOARGBROW_AVX2
 
-#if defined(HAS_I422ALPHATOARGBROW_AVX2)
+#if defined(HAS_I212TOARGBROW_AVX2)
 // 16 pixels
-// 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB.
-void OMITFP I422AlphaToARGBRow_AVX2(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               const uint8* a_buf,
-                               uint8* dst_argb,
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I212ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I212TOARGBROW_AVX2
+
+#if defined(HAS_I210TOAR30ROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I210ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV210_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_I210TOAR30ROW_AVX2
+
+#if defined(HAS_I212TOAR30ROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I212ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_I212TOAR30ROW_AVX2
+
+#if defined(HAS_I410TOARGBROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I410ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
+    READYUV410_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I410TOARGBROW_AVX2
+
+#if defined(HAS_I210ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 8 UV, 16 Y and 16 A producing 16 ARGB (64 bytes).
+void OMITFP I210AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                                    const uint16_t* u_buf,
+                                    const uint16_t* v_buf,
+                                    const uint16_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  asm volatile(
+      YUVTORGB_SETUP_AVX2(
+      yuvconstants) "sub         %[u_buf],%[v_buf]             \n"
+
+      LABELALIGN "1:                                        \n" READYUVA210_AVX2
+          YUVTORGB_AVX2(yuvconstants) STOREARGB_AVX2
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+      : [y_buf] "+r"(y_buf),        // %[y_buf]
+        [u_buf] "+r"(u_buf),        // %[u_buf]
+        [v_buf] "+r"(v_buf),        // %[v_buf]
+        [a_buf] "+r"(a_buf),        // %[a_buf]
+        [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+        [width] "+m"(width)  // %[width]
+#else
+        [width] "+rm"(width)  // %[width]
+#endif
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", "xmm3",
+        "xmm4", "xmm5");
+}
+#endif  // HAS_I210TOARGBROW_AVX2
+
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 16 UV, 16 Y and 16 A producing 16 ARGB (64 bytes).
+void OMITFP I410AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                                    const uint16_t* u_buf,
+                                    const uint16_t* v_buf,
+                                    const uint16_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  asm volatile(
+      YUVTORGB_SETUP_AVX2(
+      yuvconstants) "sub         %[u_buf],%[v_buf]             \n"
+
+      LABELALIGN "1:                                        \n" READYUVA410_AVX2
+          YUVTORGB_AVX2(yuvconstants) STOREARGB_AVX2
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+      : [y_buf] "+r"(y_buf),        // %[y_buf]
+        [u_buf] "+r"(u_buf),        // %[u_buf]
+        [v_buf] "+r"(v_buf),        // %[v_buf]
+        [a_buf] "+r"(a_buf),        // %[a_buf]
+        [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+        [width] "+m"(width)  // %[width]
+#else
+        [width] "+rm"(width)  // %[width]
+#endif
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", "xmm3",
+        "xmm4", "xmm5");
+}
+#endif  // HAS_I410TOARGBROW_AVX2
+
+#if defined(HAS_I410TOAR30ROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I410ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV410_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_I410TOAR30ROW_AVX2
+
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y and 16 A producing 16 ARGB.
+void OMITFP I444AlphaToARGBRow_AVX2(const uint8_t* y_buf,
+                                    const uint8_t* u_buf,
+                                    const uint8_t* v_buf,
+                                    const uint8_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  // clang-format off
+  asm volatile (
+  YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+  LABELALIGN
+      "1:                                        \n"
+  READYUVA444_AVX2
+  YUVTORGB_AVX2(yuvconstants)
+  STOREARGB_AVX2
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [a_buf]"+r"(a_buf),    // %[a_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_I444ALPHATOARGBROW_AVX2
+
+#if defined(HAS_I422ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB.
+void OMITFP I422AlphaToARGBRow_AVX2(const uint8_t* y_buf,
+                                    const uint8_t* u_buf,
+                                    const uint8_t* v_buf,
+                                    const uint8_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  // clang-format off
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+    LABELALIGN
+      "1:                                        \n"
     READYUVA422_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "subl      $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [a_buf]"+r"(a_buf),    // %[a_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
-#if defined(__i386__) && defined(__pic__)
+#if defined(__i386__)
     [width]"+m"(width)     // %[width]
 #else
     [width]"+rm"(width)    // %[width]
 #endif
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS_AVX2
+  : "memory", "cc", YUVTORGB_REGS_AVX2
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 #endif  // HAS_I422ALPHATOARGBROW_AVX2
 
 #if defined(HAS_I422TORGBAROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
-void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* dst_argb,
+void OMITFP I422ToRGBARow_AVX2(const uint8_t* y_buf,
+                               const uint8_t* u_buf,
+                               const uint8_t* v_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUV422_AVX2
     YUVTORGB_AVX2(yuvconstants)
 
@@ -2292,11 +4320,11 @@ void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
     "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
     "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"
     "vpunpckhwd %%ymm1,%%ymm2,%%ymm1           \n"
-    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
-    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
-    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
+    "vmovdqu    %%ymm0,(%[dst_argb])           \n"
+    "vmovdqu    %%ymm1,0x20(%[dst_argb])       \n"
+    "lea        0x40(%[dst_argb]),%[dst_argb]  \n"
+    "sub        $0x10,%[width]                 \n"
+    "jg         1b                             \n"
     "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
@@ -2304,7 +4332,7 @@ void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-  : "memory", "cc", NACL_R14 YUVTORGB_REGS_AVX2
+  : "memory", "cc", YUVTORGB_REGS_AVX2
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
 }
@@ -2313,1243 +4341,2988 @@ void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
 #if defined(HAS_NV12TOARGBROW_AVX2)
 // 16 pixels.
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-void OMITFP NV12ToARGBRow_AVX2(const uint8* y_buf,
-                               const uint8* uv_buf,
-                               uint8* dst_argb,
+void OMITFP NV12ToARGBRow_AVX2(const uint8_t* y_buf,
+                               const uint8_t* uv_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READNV12_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [uv_buf]"+r"(uv_buf),    // %[uv_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-    : "memory", "cc", YUVTORGB_REGS_AVX2  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS_AVX2
     "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 #endif  // HAS_NV12TOARGBROW_AVX2
 
 #if defined(HAS_NV21TOARGBROW_AVX2)
 // 16 pixels.
 // 8 VU values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-void OMITFP NV21ToARGBRow_AVX2(const uint8* y_buf,
-                               const uint8* vu_buf,
-                               uint8* dst_argb,
+void OMITFP NV21ToARGBRow_AVX2(const uint8_t* y_buf,
+                               const uint8_t* vu_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READNV21_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [vu_buf]"+r"(vu_buf),    // %[vu_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
     [kShuffleNV21]"m"(kShuffleNV21)
-    : "memory", "cc", YUVTORGB_REGS_AVX2  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS_AVX2
       "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 #endif  // HAS_NV21TOARGBROW_AVX2
 
 #if defined(HAS_YUY2TOARGBROW_AVX2)
 // 16 pixels.
 // 8 YUY2 values with 16 Y and 8 UV producing 16 ARGB (64 bytes).
-void OMITFP YUY2ToARGBRow_AVX2(const uint8* yuy2_buf,
-                               uint8* dst_argb,
+void OMITFP YUY2ToARGBRow_AVX2(const uint8_t* yuy2_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READYUY2_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [yuy2_buf]"+r"(yuy2_buf),    // %[yuy2_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
     [kShuffleYUY2Y]"m"(kShuffleYUY2Y),
     [kShuffleYUY2UV]"m"(kShuffleYUY2UV)
-    : "memory", "cc", YUVTORGB_REGS_AVX2  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS_AVX2
       "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 #endif  // HAS_YUY2TOARGBROW_AVX2
 
 #if defined(HAS_UYVYTOARGBROW_AVX2)
 // 16 pixels.
 // 8 UYVY values with 16 Y and 8 UV producing 16 ARGB (64 bytes).
-void OMITFP UYVYToARGBRow_AVX2(const uint8* uyvy_buf,
-                               uint8* dst_argb,
+void OMITFP UYVYToARGBRow_AVX2(const uint8_t* uyvy_buf,
+                               uint8_t* dst_argb,
                                const struct YuvConstants* yuvconstants,
                                int width) {
+  // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
+      "1:                                        \n"
     READUYVY_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [uyvy_buf]"+r"(uyvy_buf),    // %[uyvy_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
     [kShuffleUYVYY]"m"(kShuffleUYVYY),
     [kShuffleUYVYUV]"m"(kShuffleUYVYUV)
-    : "memory", "cc", YUVTORGB_REGS_AVX2  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS_AVX2
       "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
 #endif  // HAS_UYVYTOARGBROW_AVX2
 
-#ifdef HAS_I400TOARGBROW_SSE2
-void I400ToARGBRow_SSE2(const uint8* y_buf, uint8* dst_argb, int width) {
+#if defined(HAS_P210TOARGBROW_AVX2)
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP P210ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  // clang-format off
   asm volatile (
-    "mov       $0x4a354a35,%%eax               \n"  // 4a35 = 18997 = 1.164
-    "movd      %%eax,%%xmm2                    \n"
-    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
-    "mov       $0x04880488,%%eax               \n"  // 0488 = 1160 = 1.164 * 16
-    "movd      %%eax,%%xmm3                    \n"
-    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "pslld     $0x18,%%xmm4                    \n"
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
     LABELALIGN
-  "1:                                          \n"
-    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "psubusw   %%xmm3,%%xmm0                   \n"
-    "psrlw     $6, %%xmm0                      \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-
-    // Step 2: Weave into ARGB
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm0,%%xmm0                   \n"
-    "punpckhwd %%xmm1,%%xmm1                   \n"
-    "por       %%xmm4,%%xmm0                   \n"
-    "por       %%xmm4,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(y_buf),     // %0
-    "+r"(dst_argb),  // %1
-    "+rm"(width)     // %2
-  :
-  : "memory", "cc", "eax"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+      "1:                                        \n"
+    READP210_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
+  // clang-format on
 }
-#endif  // HAS_I400TOARGBROW_SSE2
+#endif  // HAS_P210TOARGBROW_AVX2
 
-#ifdef HAS_I400TOARGBROW_AVX2
-// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
-// note: vpunpcklbw mutates and vpackuswb unmutates.
-void I400ToARGBRow_AVX2(const uint8* y_buf, uint8* dst_argb, int width) {
+#if defined(HAS_P410TOARGBROW_AVX2)
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP P410ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  // clang-format off
   asm volatile (
-    "mov        $0x4a354a35,%%eax              \n" // 0488 = 1160 = 1.164 * 16
-    "vmovd      %%eax,%%xmm2                   \n"
-    "vbroadcastss %%xmm2,%%ymm2                \n"
-    "mov        $0x4880488,%%eax               \n" // 4a35 = 18997 = 1.164
-    "vmovd      %%eax,%%xmm3                   \n"
-    "vbroadcastss %%xmm3,%%ymm3                \n"
-    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpslld     $0x18,%%ymm4,%%ymm4            \n"
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-  "1:                                          \n"
-    // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
-    "vmovdqu    " MEMACCESS(0) ",%%xmm0        \n"
-    "lea        " MEMLEA(0x10,0) ",%0          \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpsubusw   %%ymm3,%%ymm0,%%ymm0           \n"
-    "vpsrlw     $0x6,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpunpcklbw %%ymm0,%%ymm0,%%ymm1           \n"
-    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
-    "vpunpcklwd %%ymm1,%%ymm1,%%ymm0           \n"
-    "vpunpckhwd %%ymm1,%%ymm1,%%ymm1           \n"
-    "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "vmovdqu    %%ymm1," MEMACCESS2(0x20,1) "  \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub        $0x10,%2                       \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(y_buf),     // %0
-    "+r"(dst_argb),  // %1
-    "+rm"(width)     // %2
-  :
-  : "memory", "cc", "eax"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+      "1:                                        \n"
+    READP410_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_P410TOARGBROW_AVX2
+
+#if defined(HAS_P210TOAR30ROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP P210ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READP210_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
+#endif  // HAS_P210TOAR30ROW_AVX2
+
+#if defined(HAS_P410TOAR30ROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP P410ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READP410_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_P410TOAR30ROW_AVX2
+
+#ifdef HAS_I400TOARGBROW_SSE2
+void I400ToARGBRow_SSE2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  asm volatile(
+      "movdqa      96(%3),%%xmm2                 \n"  // yg = 18997 = 1.164
+      "movdqa      128(%3),%%xmm3                \n"  // ygb = 1160 = 1.164 * 16
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"  // 0xff000000
+      "pslld       $0x18,%%xmm4                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+      "movq      (%0),%%xmm0                     \n"
+      "lea       0x8(%0),%0                      \n"
+      "punpcklbw %%xmm0,%%xmm0                   \n"
+      "pmulhuw   %%xmm2,%%xmm0                   \n"
+      "paddsw    %%xmm3,%%xmm0                   \n"
+      "psraw     $6, %%xmm0                      \n"
+      "packuswb  %%xmm0,%%xmm0                   \n"
+
+      // Step 2: Weave into ARGB
+      "punpcklbw %%xmm0,%%xmm0                   \n"
+      "movdqa    %%xmm0,%%xmm1                   \n"
+      "punpcklwd %%xmm0,%%xmm0                   \n"
+      "punpckhwd %%xmm1,%%xmm1                   \n"
+      "por       %%xmm4,%%xmm0                   \n"
+      "por       %%xmm4,%%xmm1                   \n"
+      "movdqu    %%xmm0,(%1)                     \n"
+      "movdqu    %%xmm1,0x10(%1)                 \n"
+      "lea       0x20(%1),%1                     \n"
+
+      "sub       $0x8,%2                         \n"
+      "jg        1b                              \n"
+      : "+r"(y_buf),       // %0
+        "+r"(dst_argb),    // %1
+        "+rm"(width)       // %2
+      : "r"(yuvconstants)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_I400TOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+void I400ToARGBRow_AVX2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  asm volatile(
+      "vmovdqa     96(%3),%%ymm2                 \n"  // yg = 18997 = 1.164
+      "vmovdqa     128(%3),%%ymm3                \n"  // ygb = -1160 = 1.164*16
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"  // 0xff000000
+      "vpslld      $0x18,%%ymm4,%%ymm4           \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
+      "vmovdqu    (%0),%%xmm0                    \n"
+      "lea        0x10(%0),%0                    \n"
+      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+      "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
+      "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+      "vpaddsw    %%ymm3,%%ymm0,%%ymm0           \n"
+      "vpsraw     $0x6,%%ymm0,%%ymm0             \n"
+      "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
+      "vpunpcklbw %%ymm0,%%ymm0,%%ymm1           \n"
+      "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+      "vpunpcklwd %%ymm1,%%ymm1,%%ymm0           \n"
+      "vpunpckhwd %%ymm1,%%ymm1,%%ymm1           \n"
+      "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
+      "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
+      "vmovdqu    %%ymm0,(%1)                    \n"
+      "vmovdqu    %%ymm1,0x20(%1)                \n"
+      "lea        0x40(%1),%1                     \n"
+      "sub        $0x10,%2                       \n"
+      "jg        1b                              \n"
+      "vzeroupper                                \n"
+      : "+r"(y_buf),       // %0
+        "+r"(dst_argb),    // %1
+        "+rm"(width)       // %2
+      : "r"(yuvconstants)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
 #endif  // HAS_I400TOARGBROW_AVX2
 
 #ifdef HAS_MIRRORROW_SSSE3
 // Shuffle table for reversing the bytes.
-static uvec8 kShuffleMirror = {
-  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
-};
+static const uvec8 kShuffleMirror = {15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u,
+                                     7u,  6u,  5u,  4u,  3u,  2u,  1u, 0u};
 
-void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+void MirrorRow_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
-  asm volatile (
-    "movdqa    %3,%%xmm5                       \n"
-    LABELALIGN
-  "1:                                          \n"
-    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
-    "pshufb    %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(temp_width)  // %2
-  : "m"(kShuffleMirror) // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm5"
-  );
+  asm volatile(
+
+      "movdqa      %3,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      -0x10(%0,%2,1),%%xmm0         \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src),           // %0
+        "+r"(dst),           // %1
+        "+r"(temp_width)     // %2
+      : "m"(kShuffleMirror)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
 }
 #endif  // HAS_MIRRORROW_SSSE3
 
 #ifdef HAS_MIRRORROW_AVX2
-void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+void MirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
-  asm volatile (
-    "vbroadcastf128 %3,%%ymm5                  \n"
-    LABELALIGN
-  "1:                                          \n"
-    MEMOPREG(vmovdqu,-0x20,0,2,1,ymm0)         //  vmovdqu -0x20(%0,%2),%%ymm0
-    "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpermq     $0x4e,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(temp_width)  // %2
-  : "m"(kShuffleMirror) // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm5"
-  );
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     -0x20(%0,%2,1),%%ymm0         \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpermq      $0x4e,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),           // %0
+        "+r"(dst),           // %1
+        "+r"(temp_width)     // %2
+      : "m"(kShuffleMirror)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
 }
 #endif  // HAS_MIRRORROW_AVX2
 
 #ifdef HAS_MIRRORUVROW_SSSE3
-// Shuffle table for reversing the bytes of UV channels.
-static uvec8 kShuffleMirrorUV = {
-  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
-};
-void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
-                       int width) {
+// Shuffle table for reversing the UV.
+static const uvec8 kShuffleMirrorUV = {14u, 15u, 12u, 13u, 10u, 11u, 8u, 9u,
+                                       6u,  7u,  4u,  5u,  2u,  3u,  0u, 1u};
+
+void MirrorUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
   intptr_t temp_width = (intptr_t)(width);
-  asm volatile (
-    "movdqa    %4,%%xmm1                       \n"
-    "lea       " MEMLEA4(-0x10,0,3,2) ",%0     \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(-0x10,0) ",%0          \n"
-    "pshufb    %%xmm1,%%xmm0                   \n"
-    "movlpd    %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $8,%3                           \n"
-    "jg        1b                              \n"
-  : "+r"(src),      // %0
-    "+r"(dst_u),    // %1
-    "+r"(dst_v),    // %2
-    "+r"(temp_width)  // %3
-  : "m"(kShuffleMirrorUV)  // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1"
-  );
+  asm volatile(
+
+      "movdqa      %3,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      -0x10(%0,%2,2),%%xmm0         \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),          // %0
+        "+r"(dst_uv),          // %1
+        "+r"(temp_width)       // %2
+      : "m"(kShuffleMirrorUV)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
 }
 #endif  // HAS_MIRRORUVROW_SSSE3
 
+#ifdef HAS_MIRRORUVROW_AVX2
+void MirrorUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     -0x20(%0,%2,2),%%ymm0         \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpermq      $0x4e,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uv),          // %0
+        "+r"(dst_uv),          // %1
+        "+r"(temp_width)       // %2
+      : "m"(kShuffleMirrorUV)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
+}
+#endif  // HAS_MIRRORUVROW_AVX2
+
+#ifdef HAS_MIRRORSPLITUVROW_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static const uvec8 kShuffleMirrorSplitUV = {14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u,
+                                            15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u};
+void MirrorSplitUVRow_SSSE3(const uint8_t* src,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile(
+      "movdqa      %4,%%xmm1                     \n"
+      "lea         -0x10(%0,%3,2),%0             \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         -0x10(%0),%0                  \n"
+      "pshufb      %%xmm1,%%xmm0                 \n"
+      "movlpd      %%xmm0,(%1)                   \n"
+      "movhpd      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $8,%3                         \n"
+      "jg          1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(temp_width)            // %3
+      : "m"(kShuffleMirrorSplitUV)  // %4
+      : "memory", "cc", "xmm0", "xmm1");
+}
+#endif  // HAS_MIRRORSPLITUVROW_SSSE3
+
+#ifdef HAS_RGB24MIRRORROW_SSSE3
+
+// Shuffle first 5 pixels to last 5 mirrored.  first byte zero
+static const uvec8 kShuffleMirrorRGB0 = {128u, 12u, 13u, 14u, 9u, 10u, 11u, 6u,
+                                         7u,   8u,  3u,  4u,  5u, 0u,  1u,  2u};
+
+// Shuffle last 5 pixels to first 5 mirrored.  last byte zero
+static const uvec8 kShuffleMirrorRGB1 = {
+    13u, 14u, 15u, 10u, 11u, 12u, 7u, 8u, 9u, 4u, 5u, 6u, 1u, 2u, 3u, 128u};
+
+// Shuffle 5 pixels at a time (15 bytes)
+void RGB24MirrorRow_SSSE3(const uint8_t* src_rgb24,
+                          uint8_t* dst_rgb24,
+                          int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  src_rgb24 += width * 3 - 48;
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // first 5
+      "movdqu      15(%0),%%xmm1                 \n"  // next 5
+      "movdqu      30(%0),%%xmm2                 \n"  // next 5
+      "movdqu      32(%0),%%xmm3                 \n"  // last 1 special
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "pshufb      %%xmm5,%%xmm3                 \n"
+      "lea         -0x30(%0),%0                  \n"
+      "movdqu      %%xmm0,32(%1)                 \n"  // last 5
+      "movdqu      %%xmm1,17(%1)                 \n"  // next 5
+      "movdqu      %%xmm2,2(%1)                  \n"  // next 5
+      "movlpd      %%xmm3,0(%1)                  \n"  // first 1
+      "lea         0x30(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_rgb24),          // %0
+        "+r"(dst_rgb24),          // %1
+        "+r"(temp_width)          // %2
+      : "m"(kShuffleMirrorRGB0),  // %3
+        "m"(kShuffleMirrorRGB1)   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_RGB24MIRRORROW_SSSE3
+
 #ifdef HAS_ARGBMIRRORROW_SSE2
 
-void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+void ARGBMirrorRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
-  asm volatile (
-    "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "pshufd    $0x1b,%%xmm0,%%xmm0             \n"
-    "lea       " MEMLEA(-0x10,0) ",%0          \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(temp_width)  // %2
-  :
-  : "memory", "cc"
-    , "xmm0"
-  );
+  asm volatile(
+
+      "lea         -0x10(%0,%2,4),%0             \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "pshufd      $0x1b,%%xmm0,%%xmm0           \n"
+      "lea         -0x10(%0),%0                  \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),        // %0
+        "+r"(dst),        // %1
+        "+r"(temp_width)  // %2
+      :
+      : "memory", "cc", "xmm0");
 }
 #endif  // HAS_ARGBMIRRORROW_SSE2
 
 #ifdef HAS_ARGBMIRRORROW_AVX2
 // Shuffle table for reversing the bytes.
-static const ulvec32 kARGBShuffleMirror_AVX2 = {
-  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
-};
-void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+static const ulvec32 kARGBShuffleMirror_AVX2 = {7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u};
+void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
-  asm volatile (
-    "vmovdqu    %3,%%ymm5                      \n"
-    LABELALIGN
-  "1:                                          \n"
-    VMEMOPREG(vpermd,-0x20,0,2,4,ymm5,ymm0) // vpermd -0x20(%0,%2,4),ymm5,ymm0
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "sub        $0x8,%2                        \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src),  // %0
-    "+r"(dst),  // %1
-    "+r"(temp_width)  // %2
-  : "m"(kARGBShuffleMirror_AVX2) // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm5"
-  );
+  asm volatile(
+
+      "vmovdqu     %3,%%ymm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpermd      -0x20(%0,%2,4),%%ymm5,%%ymm0  \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),                    // %0
+        "+r"(dst),                    // %1
+        "+r"(temp_width)              // %2
+      : "m"(kARGBShuffleMirror_AVX2)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
 }
 #endif  // HAS_ARGBMIRRORROW_AVX2
 
 #ifdef HAS_SPLITUVROW_AVX2
-void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_AVX2(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5             \n"
-    "vpsrlw     $0x8,%%ymm5,%%ymm5               \n"
-    "sub        %1,%2                            \n"
-    LABELALIGN
-  "1:                                            \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0          \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1    \n"
-    "lea        " MEMLEA(0x40,0) ",%0            \n"
-    "vpsrlw     $0x8,%%ymm0,%%ymm2               \n"
-    "vpsrlw     $0x8,%%ymm1,%%ymm3               \n"
-    "vpand      %%ymm5,%%ymm0,%%ymm0             \n"
-    "vpand      %%ymm5,%%ymm1,%%ymm1             \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm3,%%ymm2,%%ymm2             \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0              \n"
-    "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "          \n"
-    MEMOPMEM(vmovdqu,ymm2,0x00,1,2,1)             //  vmovdqu %%ymm2,(%1,%2)
-    "lea        " MEMLEA(0x20,1) ",%1            \n"
-    "sub        $0x20,%3                         \n"
-    "jg         1b                               \n"
-    "vzeroupper                                  \n"
-  : "+r"(src_uv),     // %0
-    "+r"(dst_u),      // %1
-    "+r"(dst_v),      // %2
-    "+r"(width)         // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm2            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm3            \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpackuswb   %%ymm3,%%ymm2,%%ymm2          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpermq      $0xd8,%%ymm2,%%ymm2           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x00(%1,%2,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SPLITUVROW_AVX2
 
 #ifdef HAS_SPLITUVROW_SSE2
-void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_SSE2(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width) {
-  asm volatile (
-    "pcmpeqb    %%xmm5,%%xmm5                    \n"
-    "psrlw      $0x8,%%xmm5                      \n"
-    "sub        %1,%2                            \n"
-    LABELALIGN
-  "1:                                            \n"
-    "movdqu     " MEMACCESS(0) ",%%xmm0          \n"
-    "movdqu     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
-    "lea        " MEMLEA(0x20,0) ",%0            \n"
-    "movdqa     %%xmm0,%%xmm2                    \n"
-    "movdqa     %%xmm1,%%xmm3                    \n"
-    "pand       %%xmm5,%%xmm0                    \n"
-    "pand       %%xmm5,%%xmm1                    \n"
-    "packuswb   %%xmm1,%%xmm0                    \n"
-    "psrlw      $0x8,%%xmm2                      \n"
-    "psrlw      $0x8,%%xmm3                      \n"
-    "packuswb   %%xmm3,%%xmm2                    \n"
-    "movdqu     %%xmm0," MEMACCESS(1) "          \n"
-    MEMOPMEM(movdqu,xmm2,0x00,1,2,1)             //  movdqu     %%xmm2,(%1,%2)
-    "lea        " MEMLEA(0x10,1) ",%1            \n"
-    "sub        $0x10,%3                         \n"
-    "jg         1b                               \n"
-  : "+r"(src_uv),     // %0
-    "+r"(dst_u),      // %1
-    "+r"(dst_v),      // %2
-    "+r"(width)         // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "psrlw       $0x8,%%xmm3                   \n"
+      "packuswb    %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm2,0x00(%1,%2,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SPLITUVROW_SSE2
 
+#ifdef HAS_DETILEROW_SSE2
+void DetileRow_SSE2(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "sub         $0x10,%2                      \n"
+      "lea         (%0,%3),%0                    \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "jg          1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "xmm0");
+}
+#endif  // HAS_DETILEROW_SSE2
+
+#ifdef HAS_DETILEROW_16_SSE2
+void DetileRow_16_SSE2(const uint16_t* src,
+                       ptrdiff_t src_tile_stride,
+                       uint16_t* dst,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "xmm0", "xmm1");
+}
+#endif  // HAS_DETILEROW_SSE2
+
+#ifdef HAS_DETILEROW_16_AVX
+void DetileRow_16_AVX(const uint16_t* src,
+                      ptrdiff_t src_tile_stride,
+                      uint16_t* dst,
+                      int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "xmm0");
+}
+#endif  // HAS_DETILEROW_AVX
+
+#ifdef HAS_DETILETOYUY2_SSE2
+// Read 16 Y, 8 UV, and write 8 YUYV.
+void DetileToYUY2_SSE2(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // Load 16 Y
+      "sub         $0x10,%3                      \n"
+      "lea         (%0,%4),%0                    \n"
+      "movdqu      (%1),%%xmm1                   \n"  // Load 8 UV
+      "lea         (%1,%5),%1                    \n"
+      "movdqu      %%xmm0,%%xmm2                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "movdqu      %%xmm2,0x10(%2)               \n"
+      "lea         0x20(%2),%2                   \n"
+      "jg          1b                            \n"
+      : "+r"(src_y),                            // %0
+        "+r"(src_uv),                           // %1
+        "+r"(dst_yuy2),                         // %2
+        "+r"(width)                             // %3
+      : "r"(src_y_tile_stride),                 // %4
+        "r"(src_uv_tile_stride)                 // %5
+      : "cc", "memory", "xmm0", "xmm1", "xmm2"  // Clobber list
+  );
+}
+#endif
+
+#ifdef HAS_DETILESPLITUVROW_SSSE3
+// TODO(greenjustin): Look into generating these constants instead of loading
+// them since this can cause branch mispredicts for fPIC code on 32-bit
+// machines.
+static const uvec8 kDeinterlaceUV = {0, 2, 4, 6, 8, 10, 12, 14,
+                                     1, 3, 5, 7, 9, 11, 13, 15};
+
+// TODO(greenjustin): Research alternatives to pshufb, since pshufb can be very
+// slow on older SSE2 processors.
+void DetileSplitUVRow_SSSE3(const uint8_t* src_uv,
+                            ptrdiff_t src_tile_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width) {
+  asm volatile(
+      "movdqu      %4,%%xmm1                     \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         (%0, %5),%0                   \n"
+      "pshufb      %%xmm1,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "movhps      %%xmm0,(%2)                   \n"
+      "lea         0x8(%2),%2                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),         // %0
+        "+r"(dst_u),          // %1
+        "+r"(dst_v),          // %2
+        "+r"(width)           // %3
+      : "m"(kDeinterlaceUV),  // %4
+        "r"(src_tile_stride)  // %5
+      : "cc", "memory", "xmm0", "xmm1");
+}
+#endif  // HAS_DETILESPLITUVROW_SSSE3
+
+#ifdef HAS_MERGEUVROW_AVX512BW
+void MergeUVRow_AVX512BW(const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_uv,
+                         int width) {
+      asm volatile("sub         %0,%1                         \n"
+
+               LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbw   (%0),%%zmm0                   \n"
+      "vpmovzxbw   0x00(%0,%1,1),%%zmm1          \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpsllw      $0x8,%%zmm1,%%zmm1            \n"
+      "vporq       %%zmm0,%%zmm1,%%zmm2          \n"
+      "vmovdqu64   %%zmm2,(%2)                   \n"
+      "lea         0x40(%2),%2                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+               : "+r"(src_u),   // %0
+                 "+r"(src_v),   // %1
+                 "+r"(dst_uv),  // %2
+                 "+r"(width)    // %3
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_MERGEUVROW_AVX512BW
+
 #ifdef HAS_MERGEUVROW_AVX2
-void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_AVX2(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
                      int width) {
-  asm volatile (
-    "sub       %0,%1                             \n"
-    LABELALIGN
-  "1:                                            \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0           \n"
-    MEMOPREG(vmovdqu,0x00,0,1,1,ymm1)             //  vmovdqu (%0,%1,1),%%ymm1
-    "lea       " MEMLEA(0x20,0) ",%0             \n"
-    "vpunpcklbw %%ymm1,%%ymm0,%%ymm2             \n"
-    "vpunpckhbw %%ymm1,%%ymm0,%%ymm0             \n"
-    "vextractf128 $0x0,%%ymm2," MEMACCESS(2) "   \n"
-    "vextractf128 $0x0,%%ymm0," MEMACCESS2(0x10,2) "\n"
-    "vextractf128 $0x1,%%ymm2," MEMACCESS2(0x20,2) "\n"
-    "vextractf128 $0x1,%%ymm0," MEMACCESS2(0x30,2) "\n"
-    "lea       " MEMLEA(0x40,2) ",%2             \n"
-    "sub       $0x20,%3                          \n"
-    "jg        1b                                \n"
-    "vzeroupper                                  \n"
-  : "+r"(src_u),     // %0
-    "+r"(src_v),     // %1
-    "+r"(dst_uv),    // %2
-    "+r"(width)      // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2"
-  );
+      asm volatile("sub         %0,%1                         \n"
+
+               LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbw   (%0),%%ymm0                   \n"
+      "vpmovzxbw   0x00(%0,%1,1),%%ymm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "vpsllw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpor        %%ymm0,%%ymm1,%%ymm2          \n"
+      "vmovdqu     %%ymm2,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+               : "+r"(src_u),   // %0
+                 "+r"(src_v),   // %1
+                 "+r"(dst_uv),  // %2
+                 "+r"(width)    // %3
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_MERGEUVROW_AVX2
 
 #ifdef HAS_MERGEUVROW_SSE2
-void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_SSE2(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
                      int width) {
-  asm volatile (
-    "sub       %0,%1                             \n"
-    LABELALIGN
-  "1:                                            \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
-    MEMOPREG(movdqu,0x00,0,1,1,xmm1)             //  movdqu    (%0,%1,1),%%xmm1
-    "lea       " MEMLEA(0x10,0) ",%0             \n"
-    "movdqa    %%xmm0,%%xmm2                     \n"
-    "punpcklbw %%xmm1,%%xmm0                     \n"
-    "punpckhbw %%xmm1,%%xmm2                     \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "           \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "     \n"
-    "lea       " MEMLEA(0x20,2) ",%2             \n"
-    "sub       $0x10,%3                          \n"
-    "jg        1b                                \n"
-  : "+r"(src_u),     // %0
-    "+r"(src_v),     // %1
-    "+r"(dst_uv),    // %2
-    "+r"(width)      // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2"
-  );
+      asm volatile("sub         %0,%1                         \n"
+
+               LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "movdqu      %%xmm2,0x10(%2)               \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_u),   // %0
+                 "+r"(src_v),   // %1
+                 "+r"(dst_uv),  // %2
+                 "+r"(width)    // %3
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_MERGEUVROW_SSE2
 
-#ifdef HAS_COPYROW_SSE2
-void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+#ifdef HAS_MERGEUVROW_16_AVX2
+void MergeUVRow_16_AVX2(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  // clang-format off
   asm volatile (
-    "test       $0xf,%0                        \n"
-    "jne        2f                             \n"
-    "test       $0xf,%1                        \n"
-    "jne        2f                             \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "jmp       9f                              \n"
-    LABELALIGN
-  "2:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        2b                              \n"
-  "9:                                          \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(count)  // %2
+      "vmovd       %4,%%xmm3                     \n"
+      "vmovd       %5,%%xmm4                     \n"
+
+
+      "sub         %0,%1                         \n"
+      // 8 pixels per loop.
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxwd   (%0),%%ymm0                   \n"
+      "vpmovzxwd   0x00(%0,%1,1),%%ymm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "vpsllw      %%xmm3,%%ymm0,%%ymm0          \n"
+      "vpslld      %%xmm4,%%ymm1,%%ymm1          \n"
+      "vpor        %%ymm0,%%ymm1,%%ymm2          \n"
+      "vmovdqu     %%ymm2,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_u),      // %0
+    "+r"(src_v),      // %1
+    "+r"(dst_uv),     // %2
+    "+r"(width)       // %3
+  : "r"(16 - depth),  // %4
+    "r"(32 - depth)   // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+  // clang-format on
+}
+#endif  // HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_SPLITUVROW_16_AVX2
+const uvec8 kSplitUVShuffle16 = {0, 1, 4, 5, 8,  9,  12, 13,
+                                 2, 3, 6, 7, 10, 11, 14, 15};
+void SplitUVRow_16_AVX2(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width) {
+  depth = 16 - depth;
+  // clang-format off
+  asm volatile (
+      "vmovd       %4,%%xmm3                     \n"
+      "vbroadcastf128 %5,%%ymm4                  \n"
+      "sub         %1,%2                         \n"
+
+    // 16 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "add         $0x40,%0                      \n"
+
+      "vpsrlw      %%xmm3,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm3,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x0,%%ymm1,0x10(%1)         \n"
+      "vextractf128 $0x1,%%ymm0,(%1,%2)          \n"
+      "vextractf128 $0x1,%%ymm1,0x10(%1,%2)      \n"
+      "add         $0x20,%1                      \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_uv),   // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(width)     // %3
+  : "r"(depth),     // %4
+    "m"(kSplitUVShuffle16) // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+  // clang-format on
+}
+#endif  // HAS_SPLITUVROW_16_AVX2
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 128 = 9 bits
+// 64 = 10 bits
+// 16 = 12 bits
+// 1 = 16 bits
+#ifdef HAS_MULTIPLYROW_16_AVX2
+void MultiplyRow_16_AVX2(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width) {
+  // clang-format off
+  asm volatile (
+      "vmovd       %3,%%xmm3                     \n"
+      "vpunpcklwd  %%xmm3,%%xmm3,%%xmm3          \n"
+      "vbroadcastss %%xmm3,%%ymm3                \n"
+      "sub         %0,%1                         \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpmullw     %%ymm3,%%ymm0,%%ymm0          \n"
+      "vpmullw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%0,%1)                \n"
+      "vmovdqu     %%ymm1,0x20(%0,%1)            \n"
+      "add         $0x40,%0                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width)    // %2
+  : "r"(scale)     // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm3");
+  // clang-format on
+}
+#endif  // HAS_MULTIPLYROW_16_AVX2
+
+// Use scale to convert msb formats to lsb, depending how many bits there are:
+// 512 = 9 bits
+// 1024 = 10 bits
+// 4096 = 12 bits
+// 65536 = 16 bits
+#ifdef HAS_DIVIDEROW_16_AVX2
+void DivideRow_16_AVX2(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  // clang-format off
+  asm volatile (
+      "vmovd       %3,%%xmm3                     \n"
+      "vpunpcklwd  %%xmm3,%%xmm3,%%xmm3          \n"
+      "vbroadcastss %%xmm3,%%ymm3                \n"
+      "sub         %0,%1                         \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpmulhuw    %%ymm3,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%0,%1)                \n"
+      "vmovdqu     %%ymm1,0x20(%0,%1)            \n"
+      "add         $0x40,%0                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width),    // %2
+    "+r"(scale)     // %3
   :
-  : "memory", "cc"
-    , "xmm0", "xmm1"
-  );
+  : "memory", "cc", "xmm0", "xmm1", "xmm3");
+  // clang-format on
+}
+#endif  // HAS_MULTIPLYROW_16_AVX2
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+void Convert16To8Row_SSSE3(const uint16_t* src_y,
+                           uint8_t* dst_y,
+                           int scale,
+                           int width) {
+  // clang-format off
+  asm volatile (
+      "movd        %3,%%xmm2                     \n"
+      "punpcklwd   %%xmm2,%%xmm2                 \n"
+      "pshufd      $0x0,%%xmm2,%%xmm2            \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "add         $0x20,%0                      \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "add         $0x10,%1                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width)    // %2
+  : "r"(scale)     // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2");
+  // clang-format on
+}
+
+#ifdef HAS_CONVERT16TO8ROW_AVX2
+void Convert16To8Row_AVX2(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width) {
+  // clang-format off
+  asm volatile (
+      "vmovd       %3,%%xmm2                     \n"
+      "vpunpcklwd  %%xmm2,%%xmm2,%%xmm2          \n"
+      "vbroadcastss %%xmm2,%%ymm2                \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "add         $0x40,%0                      \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm2,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"  // mutates
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "add         $0x20,%1                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width)    // %2
+  : "r"(scale)     // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2");
+  // clang-format on
+}
+#endif  // HAS_CONVERT16TO8ROW_AVX2
+
+// Use scale to convert to lsb formats depending how many bits there are:
+// 512 = 9 bits
+// 1024 = 10 bits
+// 4096 = 12 bits
+void Convert8To16Row_SSE2(const uint8_t* src_y,
+                          uint16_t* dst_y,
+                          int scale,
+                          int width) {
+  // clang-format off
+  asm volatile (
+      "movd        %3,%%xmm2                     \n"
+      "punpcklwd   %%xmm2,%%xmm2                 \n"
+      "pshufd      $0x0,%%xmm2,%%xmm2            \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "add         $0x10,%0                      \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "add         $0x20,%1                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width)    // %2
+  : "r"(scale)     // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2");
+  // clang-format on
+}
+
+#ifdef HAS_CONVERT8TO16ROW_AVX2
+void Convert8To16Row_AVX2(const uint8_t* src_y,
+                          uint16_t* dst_y,
+                          int scale,
+                          int width) {
+  // clang-format off
+  asm volatile (
+      "vmovd       %3,%%xmm2                     \n"
+      "vpunpcklwd  %%xmm2,%%xmm2,%%xmm2          \n"
+      "vbroadcastss %%xmm2,%%ymm2                \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "add         $0x20,%0                      \n"
+      "vpunpckhbw  %%ymm0,%%ymm0,%%ymm1          \n"
+      "vpunpcklbw  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm2,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "add         $0x40,%1                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width)    // %2
+  : "r"(scale)     // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2");
+  // clang-format on
+}
+#endif  // HAS_CONVERT8TO16ROW_AVX2
+
+#ifdef HAS_SPLITRGBROW_SSSE3
+// Shuffle table for converting RGB to Planar.
+static const uvec8 kSplitRGBShuffle[9] = {
+    {0u, 3u, 6u, 9u, 12u, 15u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 2u, 5u, 8u, 11u, 14u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 1u, 4u,
+     7u, 10u, 13u},
+    {1u, 4u, 7u, 10u, 13u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 0u, 3u, 6u, 9u, 12u, 15u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 2u, 5u,
+     8u, 11u, 14u},
+    {2u, 5u, 8u, 11u, 14u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 1u, 4u, 7u, 10u, 13u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 0u, 3u, 6u, 9u,
+     12u, 15u}};
+
+void SplitRGBRow_SSSE3(const uint8_t* src_rgb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "pshufb      0(%5), %%xmm0                 \n"
+      "pshufb      16(%5), %%xmm1                \n"
+      "pshufb      32(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "pshufb      48(%5),%%xmm0                 \n"
+      "pshufb      64(%5),%%xmm1                 \n"
+      "pshufb      80(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "pshufb      96(%5), %%xmm0                \n"
+      "pshufb      112(%5), %%xmm1               \n"
+      "pshufb      128(%5), %%xmm2               \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%3)                   \n"
+      "lea         0x10(%3),%3                   \n"
+      "lea         0x30(%0),%0                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_rgb),             // %0
+        "+r"(dst_r),               // %1
+        "+r"(dst_g),               // %2
+        "+r"(dst_b),               // %3
+        "+r"(width)                // %4
+      : "r"(&kSplitRGBShuffle[0])  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_SPLITRGBROW_SSSE3
+
+#ifdef HAS_MERGERGBROW_SSSE3
+// Shuffle table for converting Planar to RGB.
+static const uvec8 kMergeRGBShuffle[9] = {
+    {0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, 4u, 128u,
+     128u, 5u},
+    {128u, 0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, 4u,
+     128u, 128u},
+    {128u, 128u, 0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u,
+     4u, 128u},
+    {128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, 128u, 128u,
+     10u, 128u},
+    {5u, 128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, 128u,
+     128u, 10u},
+    {128u, 5u, 128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u,
+     128u, 128u},
+    {128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, 128u, 128u,
+     15u, 128u, 128u},
+    {128u, 128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, 128u,
+     128u, 15u, 128u},
+    {10u, 128u, 128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u,
+     128u, 128u, 15u}};
+
+void MergeRGBRow_SSSE3(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_rgb,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "pshufb      (%5), %%xmm0                  \n"
+      "pshufb      16(%5), %%xmm1                \n"
+      "pshufb      32(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%3)                   \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "pshufb      48(%5), %%xmm0                \n"
+      "pshufb      64(%5), %%xmm1                \n"
+      "pshufb      80(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,16(%3)                 \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "pshufb      96(%5), %%xmm0                \n"
+      "pshufb      112(%5), %%xmm1               \n"
+      "pshufb      128(%5), %%xmm2               \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,32(%3)                 \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "lea         0x30(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_r),               // %0
+        "+r"(src_g),               // %1
+        "+r"(src_b),               // %2
+        "+r"(dst_rgb),             // %3
+        "+r"(width)                // %4
+      : "r"(&kMergeRGBShuffle[0])  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_MERGERGBROW_SSSE3
+
+#ifdef HAS_MERGEARGBROW_SSE2
+void MergeARGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movq        (%0,%2),%%xmm0                \n"  // B
+      "movq        (%0),%%xmm1                   \n"  // R
+      "movq        (%0,%1),%%xmm2                \n"  // G
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // BR
+      "movq        (%0,%3),%%xmm1                \n"  // A
+      "punpcklbw   %%xmm1,%%xmm2                 \n"  // GA
+      "movdqa      %%xmm0,%%xmm1                 \n"  // BR
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // BGRA (hi)
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // BGRA (lo)
+      "movdqu      %%xmm0,(%4)                   \n"
+      "movdqu      %%xmm1,16(%4)                 \n"
+
+      "lea         8(%0),%0                      \n"
+      "lea         32(%4),%4                     \n"
+      "sub         $0x8,%5                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_MERGEXRGBROW_SSE2
+void MergeXRGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movq        (%2),%%xmm0                   \n"  // B
+      "movq        (%0),%%xmm1                   \n"  // R
+      "movq        (%1),%%xmm2                   \n"  // G
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // BR
+      "pcmpeqd     %%xmm1,%%xmm1                 \n"  // A(255)
+      "punpcklbw   %%xmm1,%%xmm2                 \n"  // GA
+      "movdqa      %%xmm0,%%xmm1                 \n"  // BR
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // BGRA (hi)
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // BGRA (lo)
+      "movdqu      %%xmm0,(%3)                   \n"
+      "movdqu      %%xmm1,16(%3)                 \n"
+
+      "lea         8(%0),%0                      \n"
+      "lea         8(%1),%1                      \n"
+      "lea         8(%2),%2                      \n"
+      "lea         32(%3),%3                     \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+r"(width)      // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_MERGEARGBROW_SSE2
+
+#ifdef HAS_MERGEARGBROW_AVX2
+void MergeARGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0,%2),%%xmm0                \n"  // B
+      "vmovdqu     (%0,%1),%%xmm1                \n"  // R
+      "vinserti128 $1,(%0),%%ymm0,%%ymm0         \n"  // G
+      "vinserti128 $1,(%0,%3),%%ymm1,%%ymm1      \n"  // A
+      "vpunpckhbw  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vmovdqu     %%ymm0,(%4)                   \n"  // First 8
+      "vmovdqu     %%ymm1,32(%4)                 \n"  // Next 8
+
+      "lea         16(%0),%0                     \n"
+      "lea         64(%4),%4                     \n"
+      "sub         $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_MERGEXRGBROW_AVX2
+void MergeXRGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%2),%%xmm0                   \n"  // B
+      "vpcmpeqd    %%ymm1,%%ymm1,%%ymm1          \n"  // A(255)
+      "vinserti128 $0,(%1),%%ymm1,%%ymm1         \n"  // R
+      "vinserti128 $1,(%0),%%ymm0,%%ymm0         \n"  // G
+      "vpunpckhbw  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vmovdqu     %%ymm0,(%3)                   \n"  // First 8
+      "vmovdqu     %%ymm1,32(%3)                 \n"  // Next 8
+
+      "lea         16(%0),%0                     \n"
+      "lea         16(%1),%1                     \n"
+      "lea         16(%2),%2                     \n"
+      "lea         64(%3),%3                     \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_MERGEARGBROW_AVX2
+
+#ifdef HAS_SPLITARGBROW_SSE2
+void SplitARGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+      "sub         %1,%3                         \n"
+      "sub         %1,%4                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 00-07 10-17
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 08-0F 18-1F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 08192A3B4C5D6E7F (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 08192A3B4C5D6E7F (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 08192A3B08192A3B
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 4C5D6E7F4C5D6E7F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%1,%3)                \n"  // B
+      "movhps      %%xmm0,(%1,%2)                \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+      "movhps      %%xmm2,(%1,%4)                \n"  // A
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "sub         $0x8,%5                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+r"(dst_a),     // %4
+        "+rm"(width)     // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_SPLITXRGBROW_SSE2
+void SplitXRGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 00-07 10-17
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 08-0F 18-1F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 08192A3B4C5D6E7F (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 08192A3B4C5D6E7F (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 08192A3B08192A3B
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 4C5D6E7F4C5D6E7F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%3)                   \n"  // B
+      "movhps      %%xmm0,(%2)                   \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "lea         8(%2),%2                      \n"
+      "lea         8(%3),%3                      \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+static const uvec8 kShuffleMaskARGBSplit = {0, 4, 8,  12, 1, 5, 9,  13,
+                                            2, 6, 10, 14, 3, 7, 11, 15};
+#ifdef HAS_SPLITARGBROW_SSSE3
+void SplitARGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        uint8_t* dst_a,
+                        int width) {
+  asm volatile(
+
+      "movdqa      %6,%%xmm3                     \n"
+      "sub         %1,%2                         \n"
+      "sub         %1,%3                         \n"
+      "sub         %1,%4                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "pshufb      %%xmm3,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "pshufb      %%xmm3,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%1,%3)                \n"  // B
+      "movhps      %%xmm0,(%1,%2)                \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+      "movhps      %%xmm2,(%1,%4)                \n"  // A
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "subl        $0x8,%5                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+r"(dst_a),     // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(kShuffleMaskARGBSplit)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+}
+#endif
+
+#ifdef HAS_SPLITXRGBROW_SSSE3
+void SplitXRGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        int width) {
+  asm volatile(
+
+      "movdqa      %5,%%xmm3                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "pshufb      %%xmm3,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "pshufb      %%xmm3,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%3)                   \n"  // B
+      "movhps      %%xmm0,(%2)                   \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "lea         8(%2),%2                      \n"
+      "lea         8(%3),%3                      \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),             // %0
+        "+r"(dst_r),                // %1
+        "+r"(dst_g),                // %2
+        "+r"(dst_b),                // %3
+        "+r"(width)                 // %4
+      : "m"(kShuffleMaskARGBSplit)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+}
+#endif
+
+#ifdef HAS_SPLITARGBROW_AVX2
+static const ulvec32 kShuffleMaskARGBPermute = {0, 4, 1, 5, 2, 6, 3, 7};
+void SplitARGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+      "sub         %1,%3                         \n"
+      "sub         %1,%4                         \n"
+      "vmovdqa     %7,%%ymm3                     \n"
+      "vbroadcastf128 %6,%%ymm4                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00-0F
+      "vmovdqu     16(%0),%%xmm1                 \n"  // 10-1F
+      "vinserti128 $1,32(%0),%%ymm0,%%ymm0       \n"  // 00-0F 20-2F
+      "vinserti128 $1,48(%0),%%ymm1,%%ymm1       \n"  // 10-1F 30-3F
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpermd      %%ymm0,%%ymm3,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm3,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"  // GA
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"  // BR
+      "vmovdqu     %%xmm0,(%1,%3)                \n"  // B
+      "vextracti128 $1,%%ymm0,(%1)               \n"  // R
+      "vmovdqu     %%xmm2,(%1,%2)                \n"  // G
+      "vextracti128 $1,%%ymm2,(%1,%4)            \n"  // A
+      "lea         64(%0),%0                     \n"
+      "lea         16(%1),%1                     \n"
+      "subl        $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+r"(dst_a),     // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(kShuffleMaskARGBSplit),   // %6
+        "m"(kShuffleMaskARGBPermute)  // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SPLITXRGBROW_AVX2
+void SplitXRGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+
+      "vmovdqa     %6,%%ymm3                     \n"
+      "vbroadcastf128 %5,%%ymm4                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00-0F
+      "vmovdqu     16(%0),%%xmm1                 \n"  // 10-1F
+      "vinserti128 $1,32(%0),%%ymm0,%%ymm0       \n"  // 00-0F 20-2F
+      "vinserti128 $1,48(%0),%%ymm1,%%ymm1       \n"  // 10-1F 30-3F
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpermd      %%ymm0,%%ymm3,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm3,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"  // GA
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"  // BR
+      "vmovdqu     %%xmm0,(%3)                   \n"  // B
+      "vextracti128 $1,%%ymm0,(%1)               \n"  // R
+      "vmovdqu     %%xmm2,(%2)                   \n"  // G
+
+      "lea         64(%0),%0                     \n"
+      "lea         16(%1),%1                     \n"
+      "lea         16(%2),%2                     \n"
+      "lea         16(%3),%3                     \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),               // %0
+        "+r"(dst_r),                  // %1
+        "+r"(dst_g),                  // %2
+        "+r"(dst_b),                  // %3
+        "+r"(width)                   // %4
+      : "m"(kShuffleMaskARGBSplit),   // %5
+        "m"(kShuffleMaskARGBPermute)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_MERGEXR30ROW_AVX2
+void MergeXR30Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width) {
+  int shift = depth - 10;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpcmpeqb    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $6,%%ymm6,%%ymm6              \n"
+      "vmovd       %5,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     (%0,%1),%%ymm1                \n"
+      "vmovdqu     (%0,%2),%%ymm2                \n"
+      "vpsrlw      %%xmm4,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm4,%%ymm1,%%ymm1          \n"
+      "vpsrlw      %%xmm4,%%ymm2,%%ymm2          \n"
+      "vpminuw     %%ymm0,%%ymm6,%%ymm0          \n"
+      "vpminuw     %%ymm1,%%ymm6,%%ymm1          \n"
+      "vpminuw     %%ymm2,%%ymm6,%%ymm2          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm2,%%ymm2           \n"
+      "vpsllw      $0x4,%%ymm0,%%ymm0            \n"  // Shift R to target bit
+      "vpunpckhwd  %%ymm0,%%ymm2,%%ymm3          \n"  // RB
+      "vpunpcklwd  %%ymm0,%%ymm2,%%ymm0          \n"
+      "vpunpckhwd  %%ymm5,%%ymm1,%%ymm2          \n"  // AG
+      "vpunpcklwd  %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpslld      $0xa,%%ymm1,%%ymm1            \n"  // Shift AG to target bit
+      "vpslld      $0xa,%%ymm2,%%ymm2            \n"
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"  // Combine
+      "vpor        %%ymm2,%%ymm3,%%ymm3          \n"
+      "vmovdqu     %%ymm0,(%3)                   \n"
+      "vmovdqu     %%ymm3,0x20(%3)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+#if defined(__i386__)
+      : "m"(shift)  // %5
+#else
+      : "rm"(shift)           // %5
+#endif
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_MERGEAR64ROW_AVX2
+static const lvec32 MergeAR64Permute = {0, 4, 2, 6, 1, 5, 3, 7};
+void MergeAR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  mask = (mask << 16) + mask;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+      "vmovdqa     %8,%%ymm5                     \n"
+      "vmovd       %6,%%xmm6                     \n"
+      "vbroadcastss %7,%%ymm7                    \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vmovdqu     (%0,%3),%%ymm3                \n"  // A
+      "vpminuw     %%ymm0,%%ymm7,%%ymm0          \n"
+      "vpminuw     %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpminuw     %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpminuw     %%ymm3,%%ymm7,%%ymm3          \n"
+      "vpsllw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsllw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsllw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpsllw      %%xmm6,%%ymm3,%%ymm3          \n"
+      "vpermd      %%ymm0,%%ymm5,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm5,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm5,%%ymm2          \n"
+      "vpermd      %%ymm3,%%ymm5,%%ymm3          \n"
+      "vpunpcklwd  %%ymm1,%%ymm2,%%ymm4          \n"  // BG(low)
+      "vpunpckhwd  %%ymm1,%%ymm2,%%ymm1          \n"  // BG(hi)
+      "vpunpcklwd  %%ymm3,%%ymm0,%%ymm2          \n"  // RA(low)
+      "vpunpckhwd  %%ymm3,%%ymm0,%%ymm0          \n"  // RA(hi)
+      "vpunpckldq  %%ymm2,%%ymm4,%%ymm3          \n"  // BGRA(1)
+      "vpunpckhdq  %%ymm2,%%ymm4,%%ymm4          \n"  // BGRA(3)
+      "vpunpckldq  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA(2)
+      "vpunpckhdq  %%ymm0,%%ymm1,%%ymm1          \n"  // BGRA(4)
+      "vmovdqu     %%ymm3,(%4)                   \n"
+      "vmovdqu     %%ymm2,0x20(%4)               \n"
+      "vmovdqu     %%ymm4,0x40(%4)               \n"
+      "vmovdqu     %%ymm1,0x60(%4)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x80(%4),%4                   \n"
+      "subl        $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_ar64),  // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(shift),            // %6
+        "m"(mask),             // %7
+        "m"(MergeAR64Permute)  // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_MERGEXR64ROW_AVX2
+void MergeXR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  mask = (mask << 16) + mask;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "vmovdqa     %7,%%ymm5                     \n"
+      "vmovd       %5,%%xmm6                     \n"
+      "vbroadcastss %6,%%ymm7                    \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vpminuw     %%ymm0,%%ymm7,%%ymm0          \n"
+      "vpminuw     %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpminuw     %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpsllw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsllw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsllw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpermd      %%ymm0,%%ymm5,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm5,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm5,%%ymm2          \n"
+      "vpcmpeqb    %%ymm3,%%ymm3,%%ymm3          \n"  // A (0xffff)
+      "vpunpcklwd  %%ymm1,%%ymm2,%%ymm4          \n"  // BG(low)
+      "vpunpckhwd  %%ymm1,%%ymm2,%%ymm1          \n"  // BG(hi)
+      "vpunpcklwd  %%ymm3,%%ymm0,%%ymm2          \n"  // RA(low)
+      "vpunpckhwd  %%ymm3,%%ymm0,%%ymm0          \n"  // RA(hi)
+      "vpunpckldq  %%ymm2,%%ymm4,%%ymm3          \n"  // BGRA(1)
+      "vpunpckhdq  %%ymm2,%%ymm4,%%ymm4          \n"  // BGRA(3)
+      "vpunpckldq  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA(2)
+      "vpunpckhdq  %%ymm0,%%ymm1,%%ymm1          \n"  // BGRA(4)
+      "vmovdqu     %%ymm3,(%3)                   \n"
+      "vmovdqu     %%ymm2,0x20(%3)               \n"
+      "vmovdqu     %%ymm4,0x40(%3)               \n"
+      "vmovdqu     %%ymm1,0x60(%3)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x80(%3),%3                   \n"
+      "subl        $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),           // %0
+        "+r"(src_g),           // %1
+        "+r"(src_b),           // %2
+        "+r"(dst_ar64),        // %3
+        "+r"(width)            // %4
+      : "m"(shift),            // %5
+        "m"(mask),             // %6
+        "m"(MergeAR64Permute)  // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_MERGEARGB16TO8ROW_AVX2
+static const uvec8 MergeARGB16To8Shuffle = {0, 8,  1, 9,  2, 10, 3, 11,
+                                            4, 12, 5, 13, 6, 14, 7, 15};
+void MergeARGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = depth - 8;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+      "vbroadcastf128 %7,%%ymm5                  \n"
+      "vmovd       %6,%%xmm6                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vmovdqu     (%0,%3),%%ymm3                \n"  // A
+      "vpsrlw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsrlw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpsrlw      %%xmm6,%%ymm3,%%ymm3          \n"
+      "vpackuswb   %%ymm1,%%ymm2,%%ymm1          \n"  // BG (planar)
+      "vpackuswb   %%ymm3,%%ymm0,%%ymm0          \n"  // RA (planar)
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"  // BG (interleave)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"  // RA (interleave)
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpunpcklwd  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA (low)
+      "vpunpckhwd  %%ymm0,%%ymm1,%%ymm0          \n"  // BGRA (hi)
+      "vmovdqu     %%ymm2,(%4)                   \n"
+      "vmovdqu     %%ymm0,0x20(%4)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%4),%4                   \n"
+      "subl        $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(shift),                 // %6
+        "m"(MergeARGB16To8Shuffle)  // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_MERGEXRGB16TO8ROW_AVX2
+void MergeXRGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = depth - 8;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "vbroadcastf128 %6,%%ymm5                  \n"
+      "vmovd       %5,%%xmm6                     \n"
+      "vpcmpeqb    %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpsrlw      $8,%%ymm3,%%ymm3              \n"  // A (0xff)
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vpsrlw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsrlw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpackuswb   %%ymm1,%%ymm2,%%ymm1          \n"  // BG (planar)
+      "vpackuswb   %%ymm3,%%ymm0,%%ymm0          \n"  // RA (planar)
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"  // BG (interleave)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"  // RA (interleave)
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpunpcklwd  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA (low)
+      "vpunpckhwd  %%ymm0,%%ymm1,%%ymm0          \n"  // BGRA (hi)
+      "vmovdqu     %%ymm2,(%3)                   \n"
+      "vmovdqu     %%ymm0,0x20(%3)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%3),%3                   \n"
+      "subl        $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),                // %0
+        "+r"(src_g),                // %1
+        "+r"(src_b),                // %2
+        "+r"(dst_argb),             // %3
+        "+r"(width)                 // %4
+      : "m"(shift),                 // %5
+        "m"(MergeARGB16To8Shuffle)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_COPYROW_SSE2
+void CopyRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "test        $0xf,%0                       \n"
+      "jne         2f                            \n"
+      "test        $0xf,%1                       \n"
+      "jne         2f                            \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqa      (%0),%%xmm0                   \n"
+      "movdqa      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm0,(%1)                   \n"
+      "movdqa      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "jmp         9f                            \n"
+
+      LABELALIGN
+      "2:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          2b                            \n"
+
+      LABELALIGN "9:                                        \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 #endif  // HAS_COPYROW_SSE2
 
 #ifdef HAS_COPYROW_AVX
-void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
-    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x40,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(count)  // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1"
-  );
+void CopyRow_AVX(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x40,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 #endif  // HAS_COPYROW_AVX
 
 #ifdef HAS_COPYROW_ERMS
 // Multiple of 1.
-void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+void CopyRow_ERMS(const uint8_t* src, uint8_t* dst, int width) {
   size_t width_tmp = (size_t)(width);
-  asm volatile (
-    "rep movsb " MEMMOVESTRING(0,1) "          \n"
-  : "+S"(src),  // %0
-    "+D"(dst),  // %1
-    "+c"(width_tmp) // %2
-  :
-  : "memory", "cc"
-  );
+  asm volatile(
+
+      "rep         movsb                         \n"
+      : "+S"(src),       // %0
+        "+D"(dst),       // %1
+        "+c"(width_tmp)  // %2
+      :
+      : "memory", "cc");
 }
 #endif  // HAS_COPYROW_ERMS
 
 #ifdef HAS_ARGBCOPYALPHAROW_SSE2
 // width in pixels
-void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm0,%%xmm0                   \n"
-    "pslld     $0x18,%%xmm0                    \n"
-    "pcmpeqb   %%xmm1,%%xmm1                   \n"
-    "psrld     $0x8,%%xmm1                     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
-    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
-    "pand      %%xmm0,%%xmm2                   \n"
-    "pand      %%xmm0,%%xmm3                   \n"
-    "pand      %%xmm1,%%xmm4                   \n"
-    "pand      %%xmm1,%%xmm5                   \n"
-    "por       %%xmm4,%%xmm2                   \n"
-    "por       %%xmm5,%%xmm3                   \n"
-    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGBCopyAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm0,%%xmm0                 \n"
+      "pslld       $0x18,%%xmm0                  \n"
+      "pcmpeqb     %%xmm1,%%xmm1                 \n"
+      "psrld       $0x8,%%xmm1                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqu      0x10(%0),%%xmm3               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqu      (%1),%%xmm4                   \n"
+      "movdqu      0x10(%1),%%xmm5               \n"
+      "pand        %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm0,%%xmm3                 \n"
+      "pand        %%xmm1,%%xmm4                 \n"
+      "pand        %%xmm1,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "movdqu      %%xmm3,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_ARGBCOPYALPHAROW_SSE2
 
 #ifdef HAS_ARGBCOPYALPHAROW_AVX2
 // width in pixels
-void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
-    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm1         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm2   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
-    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
-    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
-    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2"
-  );
+void ARGBCopyAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpsrld      $0x8,%%ymm0,%%ymm0            \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm1                   \n"
+      "vmovdqu     0x20(%0),%%ymm2               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpblendvb   %%ymm0,(%1),%%ymm1,%%ymm1     \n"
+      "vpblendvb   %%ymm0,0x20(%1),%%ymm2,%%ymm2 \n"
+      "vmovdqu     %%ymm1,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_ARGBCOPYALPHAROW_AVX2
 
 #ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
 // width in pixels
-void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
- asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ", %%xmm0        \n"
-    "movdqu    " MEMACCESS2(0x10, 0) ", %%xmm1 \n"
-    "lea       " MEMLEA(0x20, 0) ", %0         \n"
-    "psrld     $0x18, %%xmm0                   \n"
-    "psrld     $0x18, %%xmm1                   \n"
-    "packssdw  %%xmm1, %%xmm0                  \n"
-    "packuswb  %%xmm0, %%xmm0                  \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8, 1) ", %1          \n"
-    "sub       $0x8, %2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_a),     // %1
-    "+rm"(width)     // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1"
-  );
+void ARGBExtractAlphaRow_SSE2(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0), %%xmm0                  \n"
+      "movdqu      0x10(%0), %%xmm1              \n"
+      "lea         0x20(%0), %0                  \n"
+      "psrld       $0x18, %%xmm0                 \n"
+      "psrld       $0x18, %%xmm1                 \n"
+      "packssdw    %%xmm1, %%xmm0                \n"
+      "packuswb    %%xmm0, %%xmm0                \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1), %1                   \n"
+      "sub         $0x8, %2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_a),     // %1
+        "+rm"(width)     // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 #endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
+static const uvec8 kShuffleAlphaShort_AVX2 = {
+    3u,  128u, 128u, 128u, 7u,  128u, 128u, 128u,
+    11u, 128u, 128u, 128u, 15u, 128u, 128u, 128u};
+
+void ARGBExtractAlphaRow_AVX2(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width) {
+  asm volatile(
+      "vmovdqa     %3,%%ymm4                     \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0), %%ymm0                  \n"
+      "vmovdqu     0x20(%0), %%ymm1              \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"  // vpsrld $0x18, %%ymm0
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vmovdqu     0x40(%0), %%ymm2              \n"
+      "vmovdqu     0x60(%0), %%ymm3              \n"
+      "lea         0x80(%0), %0                  \n"
+      "vpackssdw   %%ymm1, %%ymm0, %%ymm0        \n"  // mutates
+      "vpshufb     %%ymm5,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpackssdw   %%ymm3, %%ymm2, %%ymm2        \n"  // mutates
+      "vpackuswb   %%ymm2,%%ymm0,%%ymm0          \n"  // mutates.
+      "vpermd      %%ymm0,%%ymm4,%%ymm0          \n"  // unmutate.
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20, %2                     \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),               // %0
+        "+r"(dst_a),                  // %1
+        "+rm"(width)                  // %2
+      : "m"(kPermdARGBToY_AVX),       // %3
+        "m"(kShuffleAlphaShort_AVX2)  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_AVX2
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
-void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm0,%%xmm0                   \n"
-    "pslld     $0x18,%%xmm0                    \n"
-    "pcmpeqb   %%xmm1,%%xmm1                   \n"
-    "psrld     $0x8,%%xmm1                     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "punpcklbw %%xmm2,%%xmm2                   \n"
-    "punpckhwd %%xmm2,%%xmm3                   \n"
-    "punpcklwd %%xmm2,%%xmm2                   \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
-    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
-    "pand      %%xmm0,%%xmm2                   \n"
-    "pand      %%xmm0,%%xmm3                   \n"
-    "pand      %%xmm1,%%xmm4                   \n"
-    "pand      %%xmm1,%%xmm5                   \n"
-    "por       %%xmm4,%%xmm2                   \n"
-    "por       %%xmm5,%%xmm3                   \n"
-    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm0,%%xmm0                 \n"
+      "pslld       $0x18,%%xmm0                  \n"
+      "pcmpeqb     %%xmm1,%%xmm1                 \n"
+      "psrld       $0x8,%%xmm1                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm2                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "punpcklbw   %%xmm2,%%xmm2                 \n"
+      "punpckhwd   %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm2,%%xmm2                 \n"
+      "movdqu      (%1),%%xmm4                   \n"
+      "movdqu      0x10(%1),%%xmm5               \n"
+      "pand        %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm0,%%xmm3                 \n"
+      "pand        %%xmm1,%%xmm4                 \n"
+      "pand        %%xmm1,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "movdqu      %%xmm3,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
 
 #ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
 // width in pixels
-void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
-    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vpmovzxbd " MEMACCESS(0) ",%%ymm1         \n"
-    "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2    \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "vpslld    $0x18,%%ymm1,%%ymm1             \n"
-    "vpslld    $0x18,%%ymm2,%%ymm2             \n"
-    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
-    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
-    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
-    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2"
-  );
+void ARGBCopyYToAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpsrld      $0x8,%%ymm0,%%ymm0            \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbd   (%0),%%ymm1                   \n"
+      "vpmovzxbd   0x8(%0),%%ymm2                \n"
+      "lea         0x10(%0),%0                   \n"
+      "vpslld      $0x18,%%ymm1,%%ymm1           \n"
+      "vpslld      $0x18,%%ymm2,%%ymm2           \n"
+      "vpblendvb   %%ymm0,(%1),%%ymm1,%%ymm1     \n"
+      "vpblendvb   %%ymm0,0x20(%1),%%ymm2,%%ymm2 \n"
+      "vmovdqu     %%ymm1,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
 
 #ifdef HAS_SETROW_X86
-void SetRow_X86(uint8* dst, uint8 v8, int width) {
+void SetRow_X86(uint8_t* dst, uint8_t v8, int width) {
   size_t width_tmp = (size_t)(width >> 2);
-  const uint32 v32 = v8 * 0x01010101u;  // Duplicate byte to all bytes.
-  asm volatile (
-    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
-    : "+D"(dst),       // %0
-      "+c"(width_tmp)  // %1
-    : "a"(v32)         // %2
-    : "memory", "cc");
+  const uint32_t v32 = v8 * 0x01010101u;  // Duplicate byte to all bytes.
+  asm volatile(
+
+      "rep         stosl                         \n"
+      : "+D"(dst),       // %0
+        "+c"(width_tmp)  // %1
+      : "a"(v32)         // %2
+      : "memory", "cc");
 }
 
-void SetRow_ERMS(uint8* dst, uint8 v8, int width) {
+void SetRow_ERMS(uint8_t* dst, uint8_t v8, int width) {
   size_t width_tmp = (size_t)(width);
-  asm volatile (
-    "rep stosb " MEMSTORESTRING(al,0) "        \n"
-    : "+D"(dst),       // %0
-      "+c"(width_tmp)  // %1
-    : "a"(v8)          // %2
-    : "memory", "cc");
+  asm volatile(
+
+      "rep         stosb                         \n"
+      : "+D"(dst),       // %0
+        "+c"(width_tmp)  // %1
+      : "a"(v8)          // %2
+      : "memory", "cc");
 }
 
-void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int width) {
+void ARGBSetRow_X86(uint8_t* dst_argb, uint32_t v32, int width) {
   size_t width_tmp = (size_t)(width);
-  asm volatile (
-    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
-    : "+D"(dst_argb),  // %0
-      "+c"(width_tmp)  // %1
-    : "a"(v32)         // %2
-    : "memory", "cc");
+  asm volatile(
+
+      "rep         stosl                         \n"
+      : "+D"(dst_argb),  // %0
+        "+c"(width_tmp)  // %1
+      : "a"(v32)         // %2
+      : "memory", "cc");
 }
 #endif  // HAS_SETROW_X86
 
 #ifdef HAS_YUY2TOYROW_SSE2
-void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_yuy2),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm5"
-  );
+void YUY2ToYRow_SSE2(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_yuy2),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  : "r"((intptr_t)(stride_yuy2))  // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void YUY2ToNVUVRow_SSE2(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width) {
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_yuy2),               // %0
+                 "+r"(dst_uv),                 // %1
+                 "+r"(width)                   // %2
+               : "r"((intptr_t)(stride_yuy2))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
 }
 
-void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_yuy2),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
+void YUY2ToUVRow_SSE2(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_yuy2),               // %0
+        "+r"(dst_u),                  // %1
+        "+r"(dst_v),                  // %2
+        "+r"(width)                   // %3
+      : "r"((intptr_t)(stride_yuy2))  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 
-void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_uyvy),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1"
-  );
+void YUY2ToUV422Row_SSE2(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_uyvy),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  : "r"((intptr_t)(stride_uyvy))  // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void UYVYToYRow_SSE2(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 
-void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $0x8,%%xmm5                     \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_uyvy),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
+void UYVYToUVRow_SSE2(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uyvy),               // %0
+        "+r"(dst_u),                  // %1
+        "+r"(dst_v),                  // %2
+        "+r"(width)                   // %3
+      : "r"((intptr_t)(stride_uyvy))  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
+}
+
+void UYVYToUV422Row_SSE2(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 #endif  // HAS_YUY2TOYROW_SSE2
 
 #ifdef HAS_YUY2TOYROW_AVX2
-void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int width) {
-  asm volatile (
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
-    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
-    "lea      " MEMLEA(0x20,1) ",%1            \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_yuy2),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm5"
-  );
+void YUY2ToYRow_AVX2(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
-    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
-    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
-    "lea      " MEMLEA(0x10,1) ",%1            \n"
-    "sub       $0x20,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_yuy2),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  : "r"((intptr_t)(stride_yuy2))  // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
+void YUY2ToNVUVRow_AVX2(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpavgb      0x00(%0,%3,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%3,1),%%ymm1,%%ymm1   \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_yuy2),               // %0
+        "+r"(dst_uv),                 // %1
+        "+r"(width)                   // %2
+      : "r"((intptr_t)(stride_yuy2))  // %3
+      : "memory", "cc", "xmm0", "xmm1");
 }
 
-void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
-    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
-    "lea      " MEMLEA(0x10,1) ",%1            \n"
-    "sub       $0x20,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_yuy2),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
+void YUY2ToUVRow_AVX2(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vextractf128 $0x0,%%ymm1,(%1)             \n"
+      "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_yuy2),               // %0
+        "+r"(dst_u),                  // %1
+        "+r"(dst_v),                  // %2
+        "+r"(width)                   // %3
+      : "r"((intptr_t)(stride_yuy2))  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
-    "lea      " MEMLEA(0x20,1) ",%1            \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_uyvy),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm5"
-  );
+void YUY2ToUV422Row_AVX2(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vextractf128 $0x0,%%ymm1,(%1)             \n"
+      "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
-void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-    "sub       %1,%2                           \n"
 
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
-    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
-    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
-    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
-    "lea      " MEMLEA(0x10,1) ",%1            \n"
-    "sub       $0x20,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_uyvy),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  : "r"((intptr_t)(stride_uyvy))  // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
+void UYVYToYRow_AVX2(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
+}
+void UYVYToUVRow_AVX2(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vextractf128 $0x0,%%ymm1,(%1)             \n"
+      "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uyvy),               // %0
+        "+r"(dst_u),                  // %1
+        "+r"(dst_v),                  // %2
+        "+r"(width)                   // %3
+      : "r"((intptr_t)(stride_uyvy))  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-    "vpsrlw     $0x8,%%ymm5,%%ymm5             \n"
-    "sub       %1,%2                           \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
-    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
-    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
-    "lea      " MEMLEA(0x10,1) ",%1            \n"
-    "sub       $0x20,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_uyvy),    // %0
-    "+r"(dst_u),       // %1
-    "+r"(dst_v),       // %2
-    "+r"(width)          // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
+void UYVYToUV422Row_AVX2(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vextractf128 $0x0,%%ymm1,(%1)             \n"
+      "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 #endif  // HAS_YUY2TOYROW_AVX2
 
 #ifdef HAS_ARGBBLENDROW_SSSE3
 // Shuffle table for isolating alpha.
-static uvec8 kShuffleAlpha = {
-  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
-  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
-};
+static const uvec8 kShuffleAlpha = {3u,  0x80, 3u,  0x80, 7u,  0x80, 7u,  0x80,
+                                    11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80};
 
 // Blend 8 pixels at a time
-void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
-                        uint8* dst_argb, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"
-    "psrlw     $0xf,%%xmm7                     \n"
-    "pcmpeqb   %%xmm6,%%xmm6                   \n"
-    "psrlw     $0x8,%%xmm6                     \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psllw     $0x8,%%xmm5                     \n"
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "pslld     $0x18,%%xmm4                    \n"
-    "sub       $0x4,%3                         \n"
-    "jl        49f                             \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "40:                                         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm3,%%xmm0                   \n"
-    "pxor      %%xmm4,%%xmm3                   \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
-    "pshufb    %4,%%xmm3                       \n"
-    "pand      %%xmm6,%%xmm2                   \n"
-    "paddw     %%xmm7,%%xmm3                   \n"
-    "pmullw    %%xmm3,%%xmm2                   \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "por       %%xmm4,%%xmm0                   \n"
-    "pmullw    %%xmm3,%%xmm1                   \n"
-    "psrlw     $0x8,%%xmm2                     \n"
-    "paddusb   %%xmm2,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jge       40b                             \n"
-
-  "49:                                         \n"
-    "add       $0x3,%3                         \n"
-    "jl        99f                             \n"
-
-    // 1 pixel loop.
-  "91:                                         \n"
-    "movd      " MEMACCESS(0) ",%%xmm3         \n"
-    "lea       " MEMLEA(0x4,0) ",%0            \n"
-    "movdqa    %%xmm3,%%xmm0                   \n"
-    "pxor      %%xmm4,%%xmm3                   \n"
-    "movd      " MEMACCESS(1) ",%%xmm2         \n"
-    "pshufb    %4,%%xmm3                       \n"
-    "pand      %%xmm6,%%xmm2                   \n"
-    "paddw     %%xmm7,%%xmm3                   \n"
-    "pmullw    %%xmm3,%%xmm2                   \n"
-    "movd      " MEMACCESS(1) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x4,1) ",%1            \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "por       %%xmm4,%%xmm0                   \n"
-    "pmullw    %%xmm3,%%xmm1                   \n"
-    "psrlw     $0x8,%%xmm2                     \n"
-    "paddusb   %%xmm2,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "movd      %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x4,2) ",%2            \n"
-    "sub       $0x1,%3                         \n"
-    "jge       91b                             \n"
-  "99:                                         \n"
-  : "+r"(src_argb0),    // %0
-    "+r"(src_argb1),    // %1
-    "+r"(dst_argb),     // %2
-    "+r"(width)         // %3
-  : "m"(kShuffleAlpha)  // %4
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void ARGBBlendRow_SSSE3(const uint8_t* src_argb,
+                        const uint8_t* src_argb1,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $0xf,%%xmm7                   \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrlw       $0x8,%%xmm6                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psllw       $0x8,%%xmm5                   \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "pslld       $0x18,%%xmm4                  \n"
+      "sub         $0x4,%3                       \n"
+      "jl          49f                           \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "40:                                       \n"
+      "movdqu      (%0),%%xmm3                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm0                 \n"
+      "pxor        %%xmm4,%%xmm3                 \n"
+      "movdqu      (%1),%%xmm2                   \n"
+      "pshufb      %4,%%xmm3                     \n"
+      "pand        %%xmm6,%%xmm2                 \n"
+      "paddw       %%xmm7,%%xmm3                 \n"
+      "pmullw      %%xmm3,%%xmm2                 \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pmullw      %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "paddusb     %%xmm2,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         40b                           \n"
+
+      "49:                                       \n"
+      "add         $0x3,%3                       \n"
+      "jl          99f                           \n"
+
+      // 1 pixel loop.
+      "91:                                       \n"
+      "movd        (%0),%%xmm3                   \n"
+      "lea         0x4(%0),%0                    \n"
+      "movdqa      %%xmm3,%%xmm0                 \n"
+      "pxor        %%xmm4,%%xmm3                 \n"
+      "movd        (%1),%%xmm2                   \n"
+      "pshufb      %4,%%xmm3                     \n"
+      "pand        %%xmm6,%%xmm2                 \n"
+      "paddw       %%xmm7,%%xmm3                 \n"
+      "pmullw      %%xmm3,%%xmm2                 \n"
+      "movd        (%1),%%xmm1                   \n"
+      "lea         0x4(%1),%1                    \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pmullw      %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "paddusb     %%xmm2,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "lea         0x4(%2),%2                    \n"
+      "sub         $0x1,%3                       \n"
+      "jge         91b                           \n"
+      "99:                                       \n"
+      : "+r"(src_argb),     // %0
+        "+r"(src_argb1),    // %1
+        "+r"(dst_argb),     // %2
+        "+r"(width)         // %3
+      : "m"(kShuffleAlpha)  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBBLENDROW_SSSE3
 
@@ -3559,46 +7332,49 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
 // =((A2*C2)+(B2*(255-C2))+255)/256
 // signed version of math
 // =(((A2-128)*C2)+((B2-128)*(255-C2))+32768+127)/256
-void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
-                         const uint8* alpha, uint8* dst, int width) {
-  asm volatile (
-    "pcmpeqb    %%xmm5,%%xmm5                  \n"
-    "psllw      $0x8,%%xmm5                    \n"
-    "mov        $0x80808080,%%eax              \n"
-    "movd       %%eax,%%xmm6                   \n"
-    "pshufd     $0x0,%%xmm6,%%xmm6             \n"
-    "mov        $0x807f807f,%%eax              \n"
-    "movd       %%eax,%%xmm7                   \n"
-    "pshufd     $0x0,%%xmm7,%%xmm7             \n"
-    "sub        %2,%0                          \n"
-    "sub        %2,%1                          \n"
-    "sub        %2,%3                          \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movq       (%2),%%xmm0                    \n"
-    "punpcklbw  %%xmm0,%%xmm0                  \n"
-    "pxor       %%xmm5,%%xmm0                  \n"
-    "movq       (%0,%2,1),%%xmm1               \n"
-    "movq       (%1,%2,1),%%xmm2               \n"
-    "punpcklbw  %%xmm2,%%xmm1                  \n"
-    "psubb      %%xmm6,%%xmm1                  \n"
-    "pmaddubsw  %%xmm1,%%xmm0                  \n"
-    "paddw      %%xmm7,%%xmm0                  \n"
-    "psrlw      $0x8,%%xmm0                    \n"
-    "packuswb   %%xmm0,%%xmm0                  \n"
-    "movq       %%xmm0,(%3,%2,1)               \n"
-    "lea        0x8(%2),%2                     \n"
-    "sub        $0x8,%4                        \n"
-    "jg        1b                              \n"
-  : "+r"(src0),       // %0
-    "+r"(src1),       // %1
-    "+r"(alpha),      // %2
-    "+r"(dst),        // %3
-    "+rm"(width)      // %4
-  :: "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7"
-  );
+void BlendPlaneRow_SSSE3(const uint8_t* src0,
+                         const uint8_t* src1,
+                         const uint8_t* alpha,
+                         uint8_t* dst,
+                         int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psllw       $0x8,%%xmm5                   \n"
+      "mov         $0x80808080,%%eax             \n"
+      "movd        %%eax,%%xmm6                  \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "mov         $0x807f807f,%%eax             \n"
+      "movd        %%eax,%%xmm7                  \n"
+      "pshufd      $0x0,%%xmm7,%%xmm7            \n"
+      "sub         %2,%0                         \n"
+      "sub         %2,%1                         \n"
+      "sub         %2,%3                         \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%2),%%xmm0                   \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm5,%%xmm0                 \n"
+      "movq        (%0,%2,1),%%xmm1              \n"
+      "movq        (%1,%2,1),%%xmm2              \n"
+      "punpcklbw   %%xmm2,%%xmm1                 \n"
+      "psubb       %%xmm6,%%xmm1                 \n"
+      "pmaddubsw   %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm7,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%3,%2,1)              \n"
+      "lea         0x8(%2),%2                    \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src0),   // %0
+        "+r"(src1),   // %1
+        "+r"(alpha),  // %2
+        "+r"(dst),    // %3
+        "+rm"(width)  // %4
+        ::"memory",
+        "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7");
 }
 #endif  // HAS_BLENDPLANEROW_SSSE3
 
@@ -3608,312 +7384,312 @@ void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
 // =((A2*C2)+(B2*(255-C2))+255)/256
 // signed version of math
 // =(((A2-128)*C2)+((B2-128)*(255-C2))+32768+127)/256
-void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
-                        const uint8* alpha, uint8* dst, int width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-    "vpsllw     $0x8,%%ymm5,%%ymm5             \n"
-    "mov        $0x80808080,%%eax              \n"
-    "vmovd      %%eax,%%xmm6                   \n"
-    "vbroadcastss %%xmm6,%%ymm6                \n"
-    "mov        $0x807f807f,%%eax              \n"
-    "vmovd      %%eax,%%xmm7                   \n"
-    "vbroadcastss %%xmm7,%%ymm7                \n"
-    "sub        %2,%0                          \n"
-    "sub        %2,%1                          \n"
-    "sub        %2,%3                          \n"
-
-    // 32 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    (%2),%%ymm0                    \n"
-    "vpunpckhbw %%ymm0,%%ymm0,%%ymm3           \n"
-    "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpxor      %%ymm5,%%ymm3,%%ymm3           \n"
-    "vpxor      %%ymm5,%%ymm0,%%ymm0           \n"
-    "vmovdqu    (%0,%2,1),%%ymm1               \n"
-    "vmovdqu    (%1,%2,1),%%ymm2               \n"
-    "vpunpckhbw %%ymm2,%%ymm1,%%ymm4           \n"
-    "vpunpcklbw %%ymm2,%%ymm1,%%ymm1           \n"
-    "vpsubb     %%ymm6,%%ymm4,%%ymm4           \n"
-    "vpsubb     %%ymm6,%%ymm1,%%ymm1           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "vpmaddubsw %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm7,%%ymm3,%%ymm3           \n"
-    "vpaddw     %%ymm7,%%ymm0,%%ymm0           \n"
-    "vpsrlw     $0x8,%%ymm3,%%ymm3             \n"
-    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm3,%%ymm0,%%ymm0           \n"
-    "vmovdqu    %%ymm0,(%3,%2,1)               \n"
-    "lea        0x20(%2),%2                    \n"
-    "sub        $0x20,%4                       \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src0),       // %0
-    "+r"(src1),       // %1
-    "+r"(alpha),      // %2
-    "+r"(dst),        // %3
-    "+rm"(width)      // %4
-  :: "memory", "cc", "eax",
-     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void BlendPlaneRow_AVX2(const uint8_t* src0,
+                        const uint8_t* src1,
+                        const uint8_t* alpha,
+                        uint8_t* dst,
+                        int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsllw      $0x8,%%ymm5,%%ymm5            \n"
+      "mov         $0x80808080,%%eax             \n"
+      "vmovd       %%eax,%%xmm6                  \n"
+      "vbroadcastss %%xmm6,%%ymm6                \n"
+      "mov         $0x807f807f,%%eax             \n"
+      "vmovd       %%eax,%%xmm7                  \n"
+      "vbroadcastss %%xmm7,%%ymm7                \n"
+      "sub         %2,%0                         \n"
+      "sub         %2,%1                         \n"
+      "sub         %2,%3                         \n"
+
+      // 32 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%2),%%ymm0                   \n"
+      "vpunpckhbw  %%ymm0,%%ymm0,%%ymm3          \n"
+      "vpunpcklbw  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpxor       %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpxor       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vmovdqu     (%0,%2,1),%%ymm1              \n"
+      "vmovdqu     (%1,%2,1),%%ymm2              \n"
+      "vpunpckhbw  %%ymm2,%%ymm1,%%ymm4          \n"
+      "vpunpcklbw  %%ymm2,%%ymm1,%%ymm1          \n"
+      "vpsubb      %%ymm6,%%ymm4,%%ymm4          \n"
+      "vpsubb      %%ymm6,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpmaddubsw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm7,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm7,%%ymm0,%%ymm0          \n"
+      "vpsrlw      $0x8,%%ymm3,%%ymm3            \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm3,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%3,%2,1)              \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x20,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src0),   // %0
+        "+r"(src1),   // %1
+        "+r"(alpha),  // %2
+        "+r"(dst),    // %3
+        "+rm"(width)  // %4
+        ::"memory",
+        "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_BLENDPLANEROW_AVX2
 
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
-// Shuffle table duplicating alpha
-static uvec8 kShuffleAlpha0 = {
-  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u
-};
-static uvec8 kShuffleAlpha1 = {
-  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
-  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u
-};
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha0 = {3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u,
+                                     7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u};
+static const uvec8 kShuffleAlpha1 = {11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+                                     15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u};
 // Attenuate 4 pixels at a time.
-void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    "pcmpeqb   %%xmm3,%%xmm3                   \n"
-    "pslld     $0x18,%%xmm3                    \n"
-    "movdqa    %3,%%xmm4                       \n"
-    "movdqa    %4,%%xmm5                       \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "punpcklbw %%xmm1,%%xmm1                   \n"
-    "pmulhuw   %%xmm1,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
-    "punpckhbw %%xmm2,%%xmm2                   \n"
-    "pmulhuw   %%xmm2,%%xmm1                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "pand      %%xmm3,%%xmm2                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "por       %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),    // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)        // %2
-  : "m"(kShuffleAlpha0),  // %3
-    "m"(kShuffleAlpha1)  // %4
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGBAttenuateRow_SSSE3(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "pslld       $0x18,%%xmm3                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "punpcklbw   %%xmm1,%%xmm1                 \n"
+      "pmulhuw     %%xmm1,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "punpckhbw   %%xmm2,%%xmm2                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "pand        %%xmm3,%%xmm2                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),       // %0
+        "+r"(dst_argb),       // %1
+        "+r"(width)           // %2
+      : "m"(kShuffleAlpha0),  // %3
+        "m"(kShuffleAlpha1)   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_ARGBATTENUATEROW_SSSE3
 
 #ifdef HAS_ARGBATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
-static const uvec8 kShuffleAlpha_AVX2 = {
-  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
-};
+static const uvec8 kShuffleAlpha_AVX2 = {6u,   7u,   6u,   7u,  6u,  7u,
+                                         128u, 128u, 14u,  15u, 14u, 15u,
+                                         14u,  15u,  128u, 128u};
 // Attenuate 8 pixels at a time.
-void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    "vbroadcastf128 %3,%%ymm4                  \n"
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-    "vpslld     $0x18,%%ymm5,%%ymm5            \n"
-    "sub        %0,%1                          \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
-    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
-    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
-    "vpshufb    %%ymm4,%%ymm0,%%ymm2           \n"
-    "vpshufb    %%ymm4,%%ymm1,%%ymm3           \n"
-    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
-    "vpand      %%ymm5,%%ymm6,%%ymm6           \n"
-    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpor       %%ymm6,%%ymm0,%%ymm0           \n"
-    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "sub        $0x8,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb),    // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)        // %2
-  : "m"(kShuffleAlpha_AVX2)  // %3
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+void ARGBAttenuateRow_AVX2(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpslld      $0x18,%%ymm5,%%ymm5           \n"
+      "sub         %0,%1                         \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm6                   \n"
+      "vpunpcklbw  %%ymm6,%%ymm6,%%ymm0          \n"
+      "vpunpckhbw  %%ymm6,%%ymm6,%%ymm1          \n"
+      "vpshufb     %%ymm4,%%ymm0,%%ymm2          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm3          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpand       %%ymm5,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpor        %%ymm6,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,0x00(%0,%1,1)          \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),          // %0
+        "+r"(dst_argb),          // %1
+        "+r"(width)              // %2
+      : "m"(kShuffleAlpha_AVX2)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBATTENUATEROW_AVX2
 
 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
 // Unattenuate 4 pixels at a time.
-void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+void ARGBUnattenuateRow_SSE2(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
                              int width) {
   uintptr_t alpha;
-  asm volatile (
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
-    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
-    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
-    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
-    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
-    "movlhps   %%xmm3,%%xmm2                   \n"
-    "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
-    "punpckhbw %%xmm1,%%xmm1                   \n"
-    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
-    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
-    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
-    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
-    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
-    "movlhps   %%xmm3,%%xmm2                   \n"
-    "pmulhuw   %%xmm2,%%xmm1                   \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),     // %0
-    "+r"(dst_argb),     // %1
-    "+r"(width),        // %2
-    "=&r"(alpha)        // %3
-  : "r"(fixed_invtbl8)  // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+  asm volatile(
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movzb       0x03(%0),%3                   \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "movd        0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x07(%0),%3                   \n"
+      "movd        0x00(%4,%3,4),%%xmm3          \n"
+      "pshuflw     $0x40,%%xmm2,%%xmm2           \n"
+      "pshuflw     $0x40,%%xmm3,%%xmm3           \n"
+      "movlhps     %%xmm3,%%xmm2                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "movzb       0x0b(%0),%3                   \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "movd        0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x0f(%0),%3                   \n"
+      "movd        0x00(%4,%3,4),%%xmm3          \n"
+      "pshuflw     $0x40,%%xmm2,%%xmm2           \n"
+      "pshuflw     $0x40,%%xmm3,%%xmm3           \n"
+      "movlhps     %%xmm3,%%xmm2                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),     // %0
+        "+r"(dst_argb),     // %1
+        "+r"(width),        // %2
+        "=&r"(alpha)        // %3
+      : "r"(fixed_invtbl8)  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_ARGBUNATTENUATEROW_SSE2
 
 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
 static const uvec8 kUnattenShuffleAlpha_AVX2 = {
-  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
-};
+    0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u};
 // Unattenuate 8 pixels at a time.
-void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
                              int width) {
   uintptr_t alpha;
-  asm volatile (
-    "sub        %0,%1                          \n"
-    "vbroadcastf128 %5,%%ymm5                  \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    // replace VPGATHER
-    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
-    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
-    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
-    "vpunpckldq %%xmm1,%%xmm0,%%xmm6           \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
-    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
-    "movzb     " MEMACCESS2(0x13,0) ",%3       \n"
-    "vpunpckldq %%xmm3,%%xmm2,%%xmm7           \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
-    "movzb     " MEMACCESS2(0x17,0) ",%3       \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
-    "movzb     " MEMACCESS2(0x1b,0) ",%3       \n"
-    "vpunpckldq %%xmm1,%%xmm0,%%xmm0           \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
-    "movzb     " MEMACCESS2(0x1f,0) ",%3       \n"
-    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
-    "vpunpckldq %%xmm3,%%xmm2,%%xmm2           \n"
-    "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
-    "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
-    "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
-    // end of VPGATHER
-
-    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
-    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
-    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
-    "vpunpcklwd %%ymm3,%%ymm3,%%ymm2           \n"
-    "vpunpckhwd %%ymm3,%%ymm3,%%ymm3           \n"
-    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
-    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
-    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "sub        $0x8,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb),      // %0
-    "+r"(dst_argb),      // %1
-    "+r"(width),         // %2
-    "=&r"(alpha)         // %3
-  : "r"(fixed_invtbl8),  // %4
-    "m"(kUnattenShuffleAlpha_AVX2)  // %5
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "vbroadcastf128 %5,%%ymm5                  \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      // replace VPGATHER
+      "movzb       0x03(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm0          \n"
+      "movzb       0x07(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm1          \n"
+      "movzb       0x0b(%0),%3                   \n"
+      "vpunpckldq  %%xmm1,%%xmm0,%%xmm6          \n"
+      "vmovd       0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x0f(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm3          \n"
+      "movzb       0x13(%0),%3                   \n"
+      "vpunpckldq  %%xmm3,%%xmm2,%%xmm7          \n"
+      "vmovd       0x00(%4,%3,4),%%xmm0          \n"
+      "movzb       0x17(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm1          \n"
+      "movzb       0x1b(%0),%3                   \n"
+      "vpunpckldq  %%xmm1,%%xmm0,%%xmm0          \n"
+      "vmovd       0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x1f(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm3          \n"
+      "vpunpckldq  %%xmm3,%%xmm2,%%xmm2          \n"
+      "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
+      "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
+      "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
+      // end of VPGATHER
+
+      "vmovdqu     (%0),%%ymm6                   \n"
+      "vpunpcklbw  %%ymm6,%%ymm6,%%ymm0          \n"
+      "vpunpckhbw  %%ymm6,%%ymm6,%%ymm1          \n"
+      "vpunpcklwd  %%ymm3,%%ymm3,%%ymm2          \n"
+      "vpunpckhwd  %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpshufb     %%ymm5,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,0x00(%0,%1,1)          \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),                 // %0
+        "+r"(dst_argb),                 // %1
+        "+r"(width),                    // %2
+        "=&r"(alpha)                    // %3
+      : "r"(fixed_invtbl8),             // %4
+        "m"(kUnattenShuffleAlpha_AVX2)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBUNATTENUATEROW_AVX2
 
 #ifdef HAS_ARGBGRAYROW_SSSE3
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
-void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    "movdqa    %3,%%xmm4                       \n"
-    "movdqa    %4,%%xmm5                       \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "pmaddubsw %%xmm4,%%xmm0                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "phaddw    %%xmm1,%%xmm0                   \n"
-    "paddw     %%xmm5,%%xmm0                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrld     $0x18,%%xmm2                    \n"
-    "psrld     $0x18,%%xmm3                    \n"
-    "packuswb  %%xmm3,%%xmm2                   \n"
-    "packuswb  %%xmm2,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm3                   \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "punpcklbw %%xmm2,%%xmm3                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm3,%%xmm0                   \n"
-    "punpckhwd %%xmm3,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)       // %2
-  : "m"(kARGBToYJ),   // %3
-    "m"(kAddYJ64)     // %4
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ARGBGrayRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psubb       %%xmm5,%%xmm0                 \n"
+      "psubb       %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm4,%%xmm6                 \n"
+      "pmaddubsw   %%xmm0,%%xmm6                 \n"
+      "movdqu      %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm1,%%xmm0                 \n"
+      "phaddw      %%xmm0,%%xmm6                 \n"
+      "paddw       %%xmm5,%%xmm6                 \n"
+      "psrlw       $0x8,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqu      0x10(%0),%%xmm3               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrld       $0x18,%%xmm2                  \n"
+      "psrld       $0x18,%%xmm3                  \n"
+      "packuswb    %%xmm3,%%xmm2                 \n"
+      "packuswb    %%xmm2,%%xmm2                 \n"
+      "movdqa      %%xmm6,%%xmm3                 \n"
+      "punpcklbw   %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm6,%%xmm1                 \n"
+      "punpcklwd   %%xmm3,%%xmm6                 \n"
+      "punpckhwd   %%xmm3,%%xmm1                 \n"
+      "movdqu      %%xmm6,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "m"(kARGBToYJ),  // %3
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBGRAYROW_SSSE3
 
@@ -3922,412 +7698,410 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
 //    g = (r * 45 + g * 88 + b * 22) >> 7
 //    r = (r * 50 + g * 98 + b * 24) >> 7
 // Constant for ARGB color to sepia tone
-static vec8 kARGBToSepiaB = {
-  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
-};
+static const vec8 kARGBToSepiaB = {17, 68, 35, 0, 17, 68, 35, 0,
+                                   17, 68, 35, 0, 17, 68, 35, 0};
 
-static vec8 kARGBToSepiaG = {
-  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
-};
+static const vec8 kARGBToSepiaG = {22, 88, 45, 0, 22, 88, 45, 0,
+                                   22, 88, 45, 0, 22, 88, 45, 0};
 
-static vec8 kARGBToSepiaR = {
-  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
-};
+static const vec8 kARGBToSepiaR = {24, 98, 50, 0, 24, 98, 50, 0,
+                                   24, 98, 50, 0, 24, 98, 50, 0};
 
 // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
-void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
-  asm volatile (
-    "movdqa    %2,%%xmm2                       \n"
-    "movdqa    %3,%%xmm3                       \n"
-    "movdqa    %4,%%xmm4                       \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
-    "pmaddubsw %%xmm2,%%xmm0                   \n"
-    "pmaddubsw %%xmm2,%%xmm6                   \n"
-    "phaddw    %%xmm6,%%xmm0                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "pmaddubsw %%xmm3,%%xmm5                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "phaddw    %%xmm1,%%xmm5                   \n"
-    "psrlw     $0x7,%%xmm5                     \n"
-    "packuswb  %%xmm5,%%xmm5                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "pmaddubsw %%xmm4,%%xmm5                   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "phaddw    %%xmm1,%%xmm5                   \n"
-    "psrlw     $0x7,%%xmm5                     \n"
-    "packuswb  %%xmm5,%%xmm5                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "psrld     $0x18,%%xmm6                    \n"
-    "psrld     $0x18,%%xmm1                    \n"
-    "packuswb  %%xmm1,%%xmm6                   \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "punpcklbw %%xmm6,%%xmm5                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm5,%%xmm0                   \n"
-    "punpckhwd %%xmm5,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "sub       $0x8,%1                         \n"
-    "jg        1b                              \n"
-  : "+r"(dst_argb),      // %0
-    "+r"(width)          // %1
-  : "m"(kARGBToSepiaB),  // %2
-    "m"(kARGBToSepiaG),  // %3
-    "m"(kARGBToSepiaR)   // %4
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+void ARGBSepiaRow_SSSE3(uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movdqa      %2,%%xmm2                     \n"
+      "movdqa      %3,%%xmm3                     \n"
+      "movdqa      %4,%%xmm4                     \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm6               \n"
+      "pmaddubsw   %%xmm2,%%xmm0                 \n"
+      "pmaddubsw   %%xmm2,%%xmm6                 \n"
+      "phaddw      %%xmm6,%%xmm0                 \n"
+      "psrlw       $0x7,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm5                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "pmaddubsw   %%xmm3,%%xmm5                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "phaddw      %%xmm1,%%xmm5                 \n"
+      "psrlw       $0x7,%%xmm5                   \n"
+      "packuswb    %%xmm5,%%xmm5                 \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm5                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "pmaddubsw   %%xmm4,%%xmm5                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "phaddw      %%xmm1,%%xmm5                 \n"
+      "psrlw       $0x7,%%xmm5                   \n"
+      "packuswb    %%xmm5,%%xmm5                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psrld       $0x18,%%xmm6                  \n"
+      "psrld       $0x18,%%xmm1                  \n"
+      "packuswb    %%xmm1,%%xmm6                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm6,%%xmm5                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklwd   %%xmm5,%%xmm0                 \n"
+      "punpckhwd   %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "movdqu      %%xmm1,0x10(%0)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%1                       \n"
+      "jg          1b                            \n"
+      : "+r"(dst_argb),      // %0
+        "+r"(width)          // %1
+      : "m"(kARGBToSepiaB),  // %2
+        "m"(kARGBToSepiaG),  // %3
+        "m"(kARGBToSepiaR)   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBSEPIAROW_SSSE3
 
 #ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
 // Tranform 8 ARGB pixels (32 bytes) with color matrix.
 // Same as Sepia except matrix is provided.
-void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                              const int8* matrix_argb, int width) {
-  asm volatile (
-    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
-    "pshufd    $0x00,%%xmm5,%%xmm2             \n"
-    "pshufd    $0x55,%%xmm5,%%xmm3             \n"
-    "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
-    "pshufd    $0xff,%%xmm5,%%xmm5             \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
-    "pmaddubsw %%xmm2,%%xmm0                   \n"
-    "pmaddubsw %%xmm2,%%xmm7                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "pmaddubsw %%xmm3,%%xmm6                   \n"
-    "pmaddubsw %%xmm3,%%xmm1                   \n"
-    "phaddsw   %%xmm7,%%xmm0                   \n"
-    "phaddsw   %%xmm1,%%xmm6                   \n"
-    "psraw     $0x6,%%xmm0                     \n"
-    "psraw     $0x6,%%xmm6                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "punpcklbw %%xmm6,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
-    "pmaddubsw %%xmm4,%%xmm1                   \n"
-    "pmaddubsw %%xmm4,%%xmm7                   \n"
-    "phaddsw   %%xmm7,%%xmm1                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
-    "pmaddubsw %%xmm5,%%xmm6                   \n"
-    "pmaddubsw %%xmm5,%%xmm7                   \n"
-    "phaddsw   %%xmm7,%%xmm6                   \n"
-    "psraw     $0x6,%%xmm1                     \n"
-    "psraw     $0x6,%%xmm6                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "punpcklbw %%xmm6,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm6                   \n"
-    "punpcklwd %%xmm1,%%xmm0                   \n"
-    "punpckhwd %%xmm1,%%xmm6                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm6," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),      // %0
-    "+r"(dst_argb),      // %1
-    "+r"(width)          // %2
-  : "r"(matrix_argb)     // %3
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void ARGBColorMatrixRow_SSSE3(const uint8_t* src_argb,
+                              uint8_t* dst_argb,
+                              const int8_t* matrix_argb,
+                              int width) {
+  asm volatile(
+      "movdqu      (%3),%%xmm5                   \n"
+      "pshufd      $0x00,%%xmm5,%%xmm2           \n"
+      "pshufd      $0x55,%%xmm5,%%xmm3           \n"
+      "pshufd      $0xaa,%%xmm5,%%xmm4           \n"
+      "pshufd      $0xff,%%xmm5,%%xmm5           \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm7               \n"
+      "pmaddubsw   %%xmm2,%%xmm0                 \n"
+      "pmaddubsw   %%xmm2,%%xmm7                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "phaddsw     %%xmm7,%%xmm0                 \n"
+      "phaddsw     %%xmm1,%%xmm6                 \n"
+      "psraw       $0x6,%%xmm0                   \n"
+      "psraw       $0x6,%%xmm6                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm6,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "movdqu      0x10(%0),%%xmm7               \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm7                 \n"
+      "phaddsw     %%xmm7,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x10(%0),%%xmm7               \n"
+      "pmaddubsw   %%xmm5,%%xmm6                 \n"
+      "pmaddubsw   %%xmm5,%%xmm7                 \n"
+      "phaddsw     %%xmm7,%%xmm6                 \n"
+      "psraw       $0x6,%%xmm1                   \n"
+      "psraw       $0x6,%%xmm6                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm6,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm6                 \n"
+      "punpcklwd   %%xmm1,%%xmm0                 \n"
+      "punpckhwd   %%xmm1,%%xmm6                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm6,0x10(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(dst_argb),   // %1
+        "+r"(width)       // %2
+      : "r"(matrix_argb)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
 
 #ifdef HAS_ARGBQUANTIZEROW_SSE2
 // Quantize 4 ARGB pixels (16 bytes).
-void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
-                          int interval_offset, int width) {
-  asm volatile (
-    "movd      %2,%%xmm2                       \n"
-    "movd      %3,%%xmm3                       \n"
-    "movd      %4,%%xmm4                       \n"
-    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
-    "pshufd    $0x44,%%xmm2,%%xmm2             \n"
-    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
-    "pshufd    $0x44,%%xmm3,%%xmm3             \n"
-    "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
-    "pshufd    $0x44,%%xmm4,%%xmm4             \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-    "pcmpeqb   %%xmm6,%%xmm6                   \n"
-    "pslld     $0x18,%%xmm6                    \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
-    "punpckhbw %%xmm5,%%xmm1                   \n"
-    "pmulhuw   %%xmm2,%%xmm1                   \n"
-    "pmullw    %%xmm3,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm7         \n"
-    "pmullw    %%xmm3,%%xmm1                   \n"
-    "pand      %%xmm6,%%xmm7                   \n"
-    "paddw     %%xmm4,%%xmm0                   \n"
-    "paddw     %%xmm4,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "por       %%xmm7,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "sub       $0x4,%1                         \n"
-    "jg        1b                              \n"
-  : "+r"(dst_argb),       // %0
-    "+r"(width)           // %1
-  : "r"(scale),           // %2
-    "r"(interval_size),   // %3
-    "r"(interval_offset)  // %4
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void ARGBQuantizeRow_SSE2(uint8_t* dst_argb,
+                          int scale,
+                          int interval_size,
+                          int interval_offset,
+                          int width) {
+  asm volatile(
+      "movd        %2,%%xmm2                     \n"
+      "movd        %3,%%xmm3                     \n"
+      "movd        %4,%%xmm4                     \n"
+      "pshuflw     $0x40,%%xmm2,%%xmm2           \n"
+      "pshufd      $0x44,%%xmm2,%%xmm2           \n"
+      "pshuflw     $0x40,%%xmm3,%%xmm3           \n"
+      "pshufd      $0x44,%%xmm3,%%xmm3           \n"
+      "pshuflw     $0x40,%%xmm4,%%xmm4           \n"
+      "pshufd      $0x44,%%xmm4,%%xmm4           \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "pslld       $0x18,%%xmm6                  \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "punpckhbw   %%xmm5,%%xmm1                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "pmullw      %%xmm3,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm7                   \n"
+      "pmullw      %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm6,%%xmm7                 \n"
+      "paddw       %%xmm4,%%xmm0                 \n"
+      "paddw       %%xmm4,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "por         %%xmm7,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x4,%1                       \n"
+      "jg          1b                            \n"
+      : "+r"(dst_argb),       // %0
+        "+r"(width)           // %1
+      : "r"(scale),           // %2
+        "r"(interval_size),   // %3
+        "r"(interval_offset)  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBQUANTIZEROW_SSE2
 
 #ifdef HAS_ARGBSHADEROW_SSE2
 // Shade 4 pixels at a time by specified value.
-void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value) {
-  asm volatile (
-    "movd      %3,%%xmm2                       \n"
-    "punpcklbw %%xmm2,%%xmm2                   \n"
-    "punpcklqdq %%xmm2,%%xmm2                  \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "punpckhbw %%xmm1,%%xmm1                   \n"
-    "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "pmulhuw   %%xmm2,%%xmm1                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  : "r"(value)       // %3
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2"
-  );
+void ARGBShadeRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value) {
+  asm volatile(
+      "movd        %3,%%xmm2                     \n"
+      "punpcklbw   %%xmm2,%%xmm2                 \n"
+      "punpcklqdq  %%xmm2,%%xmm2                 \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(value)       // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_ARGBSHADEROW_SSE2
 
 #ifdef HAS_ARGBMULTIPLYROW_SSE2
 // Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
-void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    "pxor      %%xmm5,%%xmm5                  \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "movdqu    %%xmm0,%%xmm1                   \n"
-    "movdqu    %%xmm2,%%xmm3                   \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "punpckhbw %%xmm1,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "punpckhbw %%xmm5,%%xmm3                   \n"
-    "pmulhuw   %%xmm2,%%xmm0                   \n"
-    "pmulhuw   %%xmm3,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void ARGBMultiplyRow_SSE2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm2                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqu      %%xmm0,%%xmm1                 \n"
+      "movdqu      %%xmm2,%%xmm3                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpckhbw   %%xmm5,%%xmm3                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm3,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_ARGBMULTIPLYROW_SSE2
 
 #ifdef HAS_ARGBMULTIPLYROW_AVX2
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm1        \n"
-    "lea        " MEMLEA(0x20,0) ",%0          \n"
-    "vmovdqu    " MEMACCESS(1) ",%%ymm3        \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "vpunpcklbw %%ymm1,%%ymm1,%%ymm0           \n"
-    "vpunpckhbw %%ymm1,%%ymm1,%%ymm1           \n"
-    "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
-    "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
-    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
-    "lea       " MEMLEA(0x20,2) ",%2           \n"
-    "sub        $0x8,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "memory", "cc"
-#if defined(__AVX2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
-  );
+void ARGBMultiplyRow_AVX2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm1                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "vmovdqu     (%1),%%ymm3                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "vpunpcklbw  %%ymm1,%%ymm1,%%ymm0          \n"
+      "vpunpckhbw  %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpcklbw  %%ymm5,%%ymm3,%%ymm2          \n"
+      "vpunpckhbw  %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_ARGBMULTIPLYROW_AVX2
 
 #ifdef HAS_ARGBADDROW_SSE2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
-void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1"
-  );
+void ARGBAddRow_SSE2(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 #endif  // HAS_ARGBADDROW_SSE2
 
 #ifdef HAS_ARGBADDROW_AVX2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
-void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "lea        " MEMLEA(0x20,0) ",%0          \n"
-    "vpaddusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
-    "lea        " MEMLEA(0x20,2) ",%2          \n"
-    "sub        $0x8,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "memory", "cc"
-    , "xmm0"
-  );
+void ARGBAddRow_AVX2(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpaddusb    (%1),%%ymm0,%%ymm0            \n"
+      "lea         0x20(%1),%1                   \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "memory", "cc", "xmm0");
 }
 #endif  // HAS_ARGBADDROW_AVX2
 
 #ifdef HAS_ARGBSUBTRACTROW_SSE2
 // Subtract 2 rows of ARGB pixels, 4 pixels at a time.
-void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "psubusb   %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1"
-  );
+void ARGBSubtractRow_SSE2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "psubusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 #endif  // HAS_ARGBSUBTRACTROW_SSE2
 
 #ifdef HAS_ARGBSUBTRACTROW_AVX2
 // Subtract 2 rows of ARGB pixels, 8 pixels at a time.
-void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "lea        " MEMLEA(0x20,0) ",%0          \n"
-    "vpsubusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
-    "lea        " MEMLEA(0x20,2) ",%2          \n"
-    "sub        $0x8,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "memory", "cc"
-    , "xmm0"
-  );
+void ARGBSubtractRow_AVX2(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpsubusb    (%1),%%ymm0,%%ymm0            \n"
+      "lea         0x20(%1),%1                   \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "memory", "cc", "xmm0");
 }
 #endif  // HAS_ARGBSUBTRACTROW_AVX2
 
@@ -4336,52 +8110,53 @@ void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
 // -1  0  1
 // -2  0  2
 // -1  0  1
-void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobelx, int width) {
-  asm volatile (
-    "sub       %0,%1                           \n"
-    "sub       %0,%2                           \n"
-    "sub       %0,%3                           \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    "movq      " MEMACCESS2(0x2,0) ",%%xmm1    \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "psubw     %%xmm1,%%xmm0                   \n"
-    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
-    MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "psubw     %%xmm2,%%xmm1                   \n"
-    MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
-    MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm3                   \n"
-    "psubw     %%xmm3,%%xmm2                   \n"
-    "paddw     %%xmm2,%%xmm0                   \n"
-    "paddw     %%xmm1,%%xmm0                   \n"
-    "paddw     %%xmm1,%%xmm0                   \n"
-    "pxor      %%xmm1,%%xmm1                   \n"
-    "psubw     %%xmm0,%%xmm1                   \n"
-    "pmaxsw    %%xmm1,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "sub       $0x8,%4                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_y0),      // %0
-    "+r"(src_y1),      // %1
-    "+r"(src_y2),      // %2
-    "+r"(dst_sobelx),  // %3
-    "+r"(width)        // %4
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void SobelXRow_SSE2(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    const uint8_t* src_y2,
+                    uint8_t* dst_sobelx,
+                    int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movq        0x2(%0),%%xmm1                \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "psubw       %%xmm1,%%xmm0                 \n"
+      "movq        0x00(%0,%1,1),%%xmm1          \n"
+      "movq        0x02(%0,%1,1),%%xmm2          \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "psubw       %%xmm2,%%xmm1                 \n"
+      "movq        0x00(%0,%2,1),%%xmm2          \n"
+      "movq        0x02(%0,%2,1),%%xmm3          \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm3                 \n"
+      "psubw       %%xmm3,%%xmm2                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "psubw       %%xmm0,%%xmm1                 \n"
+      "pmaxsw      %%xmm1,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,0x00(%0,%3,1)          \n"
+      "lea         0x8(%0),%0                    \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_y0),      // %0
+        "+r"(src_y1),      // %1
+        "+r"(src_y2),      // %2
+        "+r"(dst_sobelx),  // %3
+        "+r"(width)        // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SOBELXROW_SSE2
 
@@ -4390,50 +8165,50 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
 // -1 -2 -1
 //  0  0  0
 //  1  2  1
-void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width) {
-  asm volatile (
-    "sub       %0,%1                           \n"
-    "sub       %0,%2                           \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "psubw     %%xmm1,%%xmm0                   \n"
-    "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
-    MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "psubw     %%xmm2,%%xmm1                   \n"
-    "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
-    MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm3                   \n"
-    "psubw     %%xmm3,%%xmm2                   \n"
-    "paddw     %%xmm2,%%xmm0                   \n"
-    "paddw     %%xmm1,%%xmm0                   \n"
-    "paddw     %%xmm1,%%xmm0                   \n"
-    "pxor      %%xmm1,%%xmm1                   \n"
-    "psubw     %%xmm0,%%xmm1                   \n"
-    "pmaxsw    %%xmm1,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "sub       $0x8,%3                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_y0),      // %0
-    "+r"(src_y1),      // %1
-    "+r"(dst_sobely),  // %2
-    "+r"(width)        // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void SobelYRow_SSE2(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    uint8_t* dst_sobely,
+                    int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movq        0x00(%0,%1,1),%%xmm1          \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "psubw       %%xmm1,%%xmm0                 \n"
+      "movq        0x1(%0),%%xmm1                \n"
+      "movq        0x01(%0,%1,1),%%xmm2          \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "psubw       %%xmm2,%%xmm1                 \n"
+      "movq        0x2(%0),%%xmm2                \n"
+      "movq        0x02(%0,%1,1),%%xmm3          \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm3                 \n"
+      "psubw       %%xmm3,%%xmm2                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "psubw       %%xmm0,%%xmm1                 \n"
+      "pmaxsw      %%xmm1,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,0x00(%0,%2,1)          \n"
+      "lea         0x8(%0),%0                    \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_y0),      // %0
+        "+r"(src_y1),      // %1
+        "+r"(dst_sobely),  // %2
+        "+r"(width)        // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SOBELYROW_SSE2
 
@@ -4443,79 +8218,79 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
 // R = Sobel
 // G = Sobel
 // B = Sobel
-void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                   uint8* dst_argb, int width) {
-  asm volatile (
-    "sub       %0,%1                           \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pslld     $0x18,%%xmm5                    \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "punpcklbw %%xmm0,%%xmm2                   \n"
-    "punpckhbw %%xmm0,%%xmm0                   \n"
-    "movdqa    %%xmm2,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm1                   \n"
-    "punpckhwd %%xmm2,%%xmm2                   \n"
-    "por       %%xmm5,%%xmm1                   \n"
-    "por       %%xmm5,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm3                   \n"
-    "punpcklwd %%xmm0,%%xmm3                   \n"
-    "punpckhwd %%xmm0,%%xmm0                   \n"
-    "por       %%xmm5,%%xmm3                   \n"
-    "por       %%xmm5,%%xmm0                   \n"
-    "movdqu    %%xmm1," MEMACCESS(2) "         \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "   \n"
-    "movdqu    %%xmm3," MEMACCESS2(0x20,2) "   \n"
-    "movdqu    %%xmm0," MEMACCESS2(0x30,2) "   \n"
-    "lea       " MEMLEA(0x40,2) ",%2           \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(width)        // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void SobelRow_SSE2(const uint8_t* src_sobelx,
+                   const uint8_t* src_sobely,
+                   uint8_t* dst_argb,
+                   int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0x18,%%xmm5                  \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklbw   %%xmm0,%%xmm2                 \n"
+      "punpckhbw   %%xmm0,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "punpcklwd   %%xmm2,%%xmm1                 \n"
+      "punpckhwd   %%xmm2,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm3                 \n"
+      "punpcklwd   %%xmm0,%%xmm3                 \n"
+      "punpckhwd   %%xmm0,%%xmm0                 \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm1,(%2)                   \n"
+      "movdqu      %%xmm2,0x10(%2)               \n"
+      "movdqu      %%xmm3,0x20(%2)               \n"
+      "movdqu      %%xmm0,0x30(%2)               \n"
+      "lea         0x40(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(width)        // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SOBELROW_SSE2
 
 #ifdef HAS_SOBELTOPLANEROW_SSE2
 // Adds Sobel X and Sobel Y and stores Sobel into a plane.
-void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_y, int width) {
-  asm volatile (
-    "sub       %0,%1                           \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "pslld     $0x18,%%xmm5                    \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_y),       // %2
-    "+r"(width)        // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1"
-  );
+void SobelToPlaneRow_SSE2(const uint8_t* src_sobelx,
+                          const uint8_t* src_sobely,
+                          uint8_t* dst_y,
+                          int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0x18,%%xmm5                  \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_y),       // %2
+        "+r"(width)        // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1");
 }
 #endif  // HAS_SOBELTOPLANEROW_SSE2
 
@@ -4525,1007 +8300,1429 @@ void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 // R = Sobel X
 // G = Sobel
 // B = Sobel Y
-void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    "sub       %0,%1                           \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-
-    // 8 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "paddusb   %%xmm1,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm3                   \n"
-    "punpcklbw %%xmm5,%%xmm3                   \n"
-    "punpckhbw %%xmm5,%%xmm0                   \n"
-    "movdqa    %%xmm1,%%xmm4                   \n"
-    "punpcklbw %%xmm2,%%xmm4                   \n"
-    "punpckhbw %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm4,%%xmm6                   \n"
-    "punpcklwd %%xmm3,%%xmm6                   \n"
-    "punpckhwd %%xmm3,%%xmm4                   \n"
-    "movdqa    %%xmm1,%%xmm7                   \n"
-    "punpcklwd %%xmm0,%%xmm7                   \n"
-    "punpckhwd %%xmm0,%%xmm1                   \n"
-    "movdqu    %%xmm6," MEMACCESS(2) "         \n"
-    "movdqu    %%xmm4," MEMACCESS2(0x10,2) "   \n"
-    "movdqu    %%xmm7," MEMACCESS2(0x20,2) "   \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x30,2) "   \n"
-    "lea       " MEMLEA(0x40,2) ",%2           \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(width)        // %3
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+void SobelXYRow_SSE2(const uint8_t* src_sobelx,
+                     const uint8_t* src_sobely,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+      // 8 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "paddusb     %%xmm1,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm3                 \n"
+      "punpcklbw   %%xmm5,%%xmm3                 \n"
+      "punpckhbw   %%xmm5,%%xmm0                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "punpcklbw   %%xmm2,%%xmm4                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "punpcklwd   %%xmm3,%%xmm6                 \n"
+      "punpckhwd   %%xmm3,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm7                 \n"
+      "punpcklwd   %%xmm0,%%xmm7                 \n"
+      "punpckhwd   %%xmm0,%%xmm1                 \n"
+      "movdqu      %%xmm6,(%2)                   \n"
+      "movdqu      %%xmm4,0x10(%2)               \n"
+      "movdqu      %%xmm7,0x20(%2)               \n"
+      "movdqu      %%xmm1,0x30(%2)               \n"
+      "lea         0x40(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(width)        // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_SOBELXYROW_SSE2
 
 #ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
 // Creates a table of cumulative sums where each value is a sum of all values
 // above and to the left of the value, inclusive of the value.
-void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
-                                  const int32* previous_cumsum, int width) {
-  asm volatile (
-    "pxor      %%xmm0,%%xmm0                   \n"
-    "pxor      %%xmm1,%%xmm1                   \n"
-    "sub       $0x4,%3                         \n"
-    "jl        49f                             \n"
-    "test      $0xf,%1                         \n"
-    "jne       49f                             \n"
-
-  // 4 pixel loop                              \n"
-    LABELALIGN
-  "40:                                         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm2,%%xmm4                   \n"
-    "punpcklbw %%xmm1,%%xmm2                   \n"
-    "movdqa    %%xmm2,%%xmm3                   \n"
-    "punpcklwd %%xmm1,%%xmm2                   \n"
-    "punpckhwd %%xmm1,%%xmm3                   \n"
-    "punpckhbw %%xmm1,%%xmm4                   \n"
-    "movdqa    %%xmm4,%%xmm5                   \n"
-    "punpcklwd %%xmm1,%%xmm4                   \n"
-    "punpckhwd %%xmm1,%%xmm5                   \n"
-    "paddd     %%xmm2,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
-    "paddd     %%xmm0,%%xmm2                   \n"
-    "paddd     %%xmm3,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
-    "paddd     %%xmm0,%%xmm3                   \n"
-    "paddd     %%xmm4,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
-    "paddd     %%xmm0,%%xmm4                   \n"
-    "paddd     %%xmm5,%%xmm0                   \n"
-    "movdqu    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
-    "lea       " MEMLEA(0x40,2) ",%2           \n"
-    "paddd     %%xmm0,%%xmm5                   \n"
-    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
-    "movdqu    %%xmm4," MEMACCESS2(0x20,1) "   \n"
-    "movdqu    %%xmm5," MEMACCESS2(0x30,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x4,%3                         \n"
-    "jge       40b                             \n"
-
-  "49:                                         \n"
-    "add       $0x3,%3                         \n"
-    "jl        19f                             \n"
-
-  // 1 pixel loop                              \n"
-    LABELALIGN
-  "10:                                         \n"
-    "movd      " MEMACCESS(0) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x4,0) ",%0            \n"
-    "punpcklbw %%xmm1,%%xmm2                   \n"
-    "punpcklwd %%xmm1,%%xmm2                   \n"
-    "paddd     %%xmm2,%%xmm0                   \n"
-    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "paddd     %%xmm0,%%xmm2                   \n"
-    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x1,%3                         \n"
-    "jge       10b                             \n"
-
-  "19:                                         \n"
-  : "+r"(row),  // %0
-    "+r"(cumsum),  // %1
-    "+r"(previous_cumsum),  // %2
-    "+r"(width)  // %3
-  :
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ComputeCumulativeSumRow_SSE2(const uint8_t* row,
+                                  int32_t* cumsum,
+                                  const int32_t* previous_cumsum,
+                                  int width) {
+  asm volatile(
+      "pxor        %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "sub         $0x4,%3                       \n"
+      "jl          49f                           \n"
+      "test        $0xf,%1                       \n"
+      "jne         49f                           \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "40:                                       \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm1,%%xmm2                 \n"
+      "punpckhwd   %%xmm1,%%xmm3                 \n"
+      "punpckhbw   %%xmm1,%%xmm4                 \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "punpcklwd   %%xmm1,%%xmm4                 \n"
+      "punpckhwd   %%xmm1,%%xmm5                 \n"
+      "paddd       %%xmm2,%%xmm0                 \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm0                 \n"
+      "movdqu      0x10(%2),%%xmm3               \n"
+      "paddd       %%xmm0,%%xmm3                 \n"
+      "paddd       %%xmm4,%%xmm0                 \n"
+      "movdqu      0x20(%2),%%xmm4               \n"
+      "paddd       %%xmm0,%%xmm4                 \n"
+      "paddd       %%xmm5,%%xmm0                 \n"
+      "movdqu      0x30(%2),%%xmm5               \n"
+      "lea         0x40(%2),%2                   \n"
+      "paddd       %%xmm0,%%xmm5                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "movdqu      %%xmm3,0x10(%1)               \n"
+      "movdqu      %%xmm4,0x20(%1)               \n"
+      "movdqu      %%xmm5,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         40b                           \n"
+
+      "49:                                       \n"
+      "add         $0x3,%3                       \n"
+      "jl          19f                           \n"
+
+      // 1 pixel loop.
+      LABELALIGN
+      "10:                                       \n"
+      "movd        (%0),%%xmm2                   \n"
+      "lea         0x4(%0),%0                    \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "punpcklwd   %%xmm1,%%xmm2                 \n"
+      "paddd       %%xmm2,%%xmm0                 \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x1,%3                       \n"
+      "jge         10b                           \n"
+
+      "19:                                       \n"
+      : "+r"(row),              // %0
+        "+r"(cumsum),           // %1
+        "+r"(previous_cumsum),  // %2
+        "+r"(width)             // %3
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
 
 #ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
-void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
-                                    int width, int area, uint8* dst,
+void CumulativeSumToAverageRow_SSE2(const int32_t* topleft,
+                                    const int32_t* botleft,
+                                    int width,
+                                    int area,
+                                    uint8_t* dst,
                                     int count) {
-  asm volatile (
-    "movd      %5,%%xmm5                       \n"
-    "cvtdq2ps  %%xmm5,%%xmm5                   \n"
-    "rcpss     %%xmm5,%%xmm4                   \n"
-    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
-    "sub       $0x4,%3                         \n"
-    "jl        49f                             \n"
-    "cmpl      $0x80,%5                        \n"
-    "ja        40f                             \n"
-
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    "pcmpeqb   %%xmm6,%%xmm6                   \n"
-    "psrld     $0x10,%%xmm6                    \n"
-    "cvtdq2ps  %%xmm6,%%xmm6                   \n"
-    "addps     %%xmm6,%%xmm5                   \n"
-    "mulps     %%xmm4,%%xmm5                   \n"
-    "cvtps2dq  %%xmm5,%%xmm5                   \n"
-    "packssdw  %%xmm5,%%xmm5                   \n"
-
-  // 4 pixel small loop                        \n"
-    LABELALIGN
-  "4:                                         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
-    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
-    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
-    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
-    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
-    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
-    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
-    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
-    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
-    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
-    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "packssdw  %%xmm1,%%xmm0                   \n"
-    "packssdw  %%xmm3,%%xmm2                   \n"
-    "pmulhuw   %%xmm5,%%xmm0                   \n"
-    "pmulhuw   %%xmm5,%%xmm2                   \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jge       4b                              \n"
-    "jmp       49f                             \n"
-
-  // 4 pixel loop                              \n"
-    LABELALIGN
-  "40:                                         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
-    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
-    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
-    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
-    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
-    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
-    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
-    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
-    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
-    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
-    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
-    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
-    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
-    "cvtdq2ps  %%xmm1,%%xmm1                   \n"
-    "mulps     %%xmm4,%%xmm0                   \n"
-    "mulps     %%xmm4,%%xmm1                   \n"
-    "cvtdq2ps  %%xmm2,%%xmm2                   \n"
-    "cvtdq2ps  %%xmm3,%%xmm3                   \n"
-    "mulps     %%xmm4,%%xmm2                   \n"
-    "mulps     %%xmm4,%%xmm3                   \n"
-    "cvtps2dq  %%xmm0,%%xmm0                   \n"
-    "cvtps2dq  %%xmm1,%%xmm1                   \n"
-    "cvtps2dq  %%xmm2,%%xmm2                   \n"
-    "cvtps2dq  %%xmm3,%%xmm3                   \n"
-    "packssdw  %%xmm1,%%xmm0                   \n"
-    "packssdw  %%xmm3,%%xmm2                   \n"
-    "packuswb  %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jge       40b                             \n"
-
-  "49:                                         \n"
-    "add       $0x3,%3                         \n"
-    "jl        19f                             \n"
-
-  // 1 pixel loop                              \n"
-    LABELALIGN
-  "10:                                         \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
-    "mulps     %%xmm4,%%xmm0                   \n"
-    "cvtps2dq  %%xmm0,%%xmm0                   \n"
-    "packssdw  %%xmm0,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movd      %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x4,2) ",%2            \n"
-    "sub       $0x1,%3                         \n"
-    "jge       10b                             \n"
-  "19:                                         \n"
-  : "+r"(topleft),  // %0
-    "+r"(botleft),  // %1
-    "+r"(dst),      // %2
-    "+rm"(count)    // %3
-  : "r"((intptr_t)(width)),  // %4
-    "rm"(area)     // %5
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+  asm volatile(
+      "movd        %5,%%xmm5                     \n"
+      "cvtdq2ps    %%xmm5,%%xmm5                 \n"
+      "rcpss       %%xmm5,%%xmm4                 \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "sub         $0x4,%3                       \n"
+      "jl          49f                           \n"
+      "cmpl        $0x80,%5                      \n"
+      "ja          40f                           \n"
+
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrld       $0x10,%%xmm6                  \n"
+      "cvtdq2ps    %%xmm6,%%xmm6                 \n"
+      "addps       %%xmm6,%%xmm5                 \n"
+      "mulps       %%xmm4,%%xmm5                 \n"
+      "cvtps2dq    %%xmm5,%%xmm5                 \n"
+      "packssdw    %%xmm5,%%xmm5                 \n"
+
+      // 4 pixel small loop.
+      LABELALIGN
+      "4:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "psubd       0x00(%0,%4,4),%%xmm0          \n"
+      "psubd       0x10(%0,%4,4),%%xmm1          \n"
+      "psubd       0x20(%0,%4,4),%%xmm2          \n"
+      "psubd       0x30(%0,%4,4),%%xmm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "psubd       (%1),%%xmm0                   \n"
+      "psubd       0x10(%1),%%xmm1               \n"
+      "psubd       0x20(%1),%%xmm2               \n"
+      "psubd       0x30(%1),%%xmm3               \n"
+      "paddd       0x00(%1,%4,4),%%xmm0          \n"
+      "paddd       0x10(%1,%4,4),%%xmm1          \n"
+      "paddd       0x20(%1,%4,4),%%xmm2          \n"
+      "paddd       0x30(%1,%4,4),%%xmm3          \n"
+      "lea         0x40(%1),%1                   \n"
+      "packssdw    %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm3,%%xmm2                 \n"
+      "pmulhuw     %%xmm5,%%xmm0                 \n"
+      "pmulhuw     %%xmm5,%%xmm2                 \n"
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         4b                            \n"
+      "jmp         49f                           \n"
+
+      // 4 pixel loop
+      LABELALIGN
+      "40:                                       \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "psubd       0x00(%0,%4,4),%%xmm0          \n"
+      "psubd       0x10(%0,%4,4),%%xmm1          \n"
+      "psubd       0x20(%0,%4,4),%%xmm2          \n"
+      "psubd       0x30(%0,%4,4),%%xmm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "psubd       (%1),%%xmm0                   \n"
+      "psubd       0x10(%1),%%xmm1               \n"
+      "psubd       0x20(%1),%%xmm2               \n"
+      "psubd       0x30(%1),%%xmm3               \n"
+      "paddd       0x00(%1,%4,4),%%xmm0          \n"
+      "paddd       0x10(%1,%4,4),%%xmm1          \n"
+      "paddd       0x20(%1,%4,4),%%xmm2          \n"
+      "paddd       0x30(%1,%4,4),%%xmm3          \n"
+      "lea         0x40(%1),%1                   \n"
+      "cvtdq2ps    %%xmm0,%%xmm0                 \n"
+      "cvtdq2ps    %%xmm1,%%xmm1                 \n"
+      "mulps       %%xmm4,%%xmm0                 \n"
+      "mulps       %%xmm4,%%xmm1                 \n"
+      "cvtdq2ps    %%xmm2,%%xmm2                 \n"
+      "cvtdq2ps    %%xmm3,%%xmm3                 \n"
+      "mulps       %%xmm4,%%xmm2                 \n"
+      "mulps       %%xmm4,%%xmm3                 \n"
+      "cvtps2dq    %%xmm0,%%xmm0                 \n"
+      "cvtps2dq    %%xmm1,%%xmm1                 \n"
+      "cvtps2dq    %%xmm2,%%xmm2                 \n"
+      "cvtps2dq    %%xmm3,%%xmm3                 \n"
+      "packssdw    %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm3,%%xmm2                 \n"
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         40b                           \n"
+
+      "49:                                       \n"
+      "add         $0x3,%3                       \n"
+      "jl          19f                           \n"
+
+      // 1 pixel loop
+      LABELALIGN
+      "10:                                       \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "psubd       0x00(%0,%4,4),%%xmm0          \n"
+      "lea         0x10(%0),%0                   \n"
+      "psubd       (%1),%%xmm0                   \n"
+      "paddd       0x00(%1,%4,4),%%xmm0          \n"
+      "lea         0x10(%1),%1                   \n"
+      "cvtdq2ps    %%xmm0,%%xmm0                 \n"
+      "mulps       %%xmm4,%%xmm0                 \n"
+      "cvtps2dq    %%xmm0,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "lea         0x4(%2),%2                    \n"
+      "sub         $0x1,%3                       \n"
+      "jge         10b                           \n"
+      "19:                                       \n"
+      : "+r"(topleft),           // %0
+        "+r"(botleft),           // %1
+        "+r"(dst),               // %2
+        "+rm"(count)             // %3
+      : "r"((intptr_t)(width)),  // %4
+        "rm"(area)               // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
 
 #ifdef HAS_ARGBAFFINEROW_SSE2
 // Copy ARGB pixels from source image with slope to a row of destination.
 LIBYUV_API
-void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
-                        uint8* dst_argb, const float* src_dudv, int width) {
+void ARGBAffineRow_SSE2(const uint8_t* src_argb,
+                        int src_argb_stride,
+                        uint8_t* dst_argb,
+                        const float* src_dudv,
+                        int width) {
   intptr_t src_argb_stride_temp = src_argb_stride;
   intptr_t temp;
-  asm volatile (
-    "movq      " MEMACCESS(3) ",%%xmm2         \n"
-    "movq      " MEMACCESS2(0x08,3) ",%%xmm7   \n"
-    "shl       $0x10,%1                        \n"
-    "add       $0x4,%1                         \n"
-    "movd      %1,%%xmm5                       \n"
-    "sub       $0x4,%4                         \n"
-    "jl        49f                             \n"
-
-    "pshufd    $0x44,%%xmm7,%%xmm7             \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    "movdqa    %%xmm2,%%xmm0                   \n"
-    "addps     %%xmm7,%%xmm0                   \n"
-    "movlhps   %%xmm0,%%xmm2                   \n"
-    "movdqa    %%xmm7,%%xmm4                   \n"
-    "addps     %%xmm4,%%xmm4                   \n"
-    "movdqa    %%xmm2,%%xmm3                   \n"
-    "addps     %%xmm4,%%xmm3                   \n"
-    "addps     %%xmm4,%%xmm4                   \n"
-
-  // 4 pixel loop                              \n"
-    LABELALIGN
-  "40:                                         \n"
-    "cvttps2dq %%xmm2,%%xmm0                   \n"  // x, y float to int first 2
-    "cvttps2dq %%xmm3,%%xmm1                   \n"  // x, y float to int next 2
-    "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
-    "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x * 4 + y * stride
-    "movd      %%xmm0,%k1                      \n"
-    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-    "movd      %%xmm0,%k5                      \n"
-    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
-    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
-    "punpckldq %%xmm6,%%xmm1                   \n"
-    "addps     %%xmm4,%%xmm2                   \n"
-    "movq      %%xmm1," MEMACCESS(2) "         \n"
-    "movd      %%xmm0,%k1                      \n"
-    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-    "movd      %%xmm0,%k5                      \n"
-    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
-    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
-    "punpckldq %%xmm6,%%xmm0                   \n"
-    "addps     %%xmm4,%%xmm3                   \n"
-    "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%4                         \n"
-    "jge       40b                             \n"
-
-  "49:                                         \n"
-    "add       $0x3,%4                         \n"
-    "jl        19f                             \n"
-
-  // 1 pixel loop                              \n"
-    LABELALIGN
-  "10:                                         \n"
-    "cvttps2dq %%xmm2,%%xmm0                   \n"
-    "packssdw  %%xmm0,%%xmm0                   \n"
-    "pmaddwd   %%xmm5,%%xmm0                   \n"
-    "addps     %%xmm7,%%xmm2                   \n"
-    "movd      %%xmm0,%k1                      \n"
-    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
-    "movd      %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x04,2) ",%2           \n"
-    "sub       $0x1,%4                         \n"
-    "jge       10b                             \n"
-  "19:                                         \n"
-  : "+r"(src_argb),  // %0
-    "+r"(src_argb_stride_temp),  // %1
-    "+r"(dst_argb),  // %2
-    "+r"(src_dudv),  // %3
-    "+rm"(width),    // %4
-    "=&r"(temp)      // %5
-  :
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+  asm volatile(
+      "movq        (%3),%%xmm2                   \n"
+      "movq        0x08(%3),%%xmm7               \n"
+      "shl         $0x10,%1                      \n"
+      "add         $0x4,%1                       \n"
+      "movd        %1,%%xmm5                     \n"
+      "sub         $0x4,%4                       \n"
+      "jl          49f                           \n"
+
+      "pshufd      $0x44,%%xmm7,%%xmm7           \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "addps       %%xmm7,%%xmm0                 \n"
+      "movlhps     %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm7,%%xmm4                 \n"
+      "addps       %%xmm4,%%xmm4                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "addps       %%xmm4,%%xmm3                 \n"
+      "addps       %%xmm4,%%xmm4                 \n"
+
+      // 4 pixel loop
+      LABELALIGN
+      "40:                                       \n"
+      "cvttps2dq   %%xmm2,%%xmm0                 \n"  // x,y float->int first 2
+      "cvttps2dq   %%xmm3,%%xmm1                 \n"  // x,y float->int next 2
+      "packssdw    %%xmm1,%%xmm0                 \n"  // x, y as 8 shorts
+      "pmaddwd     %%xmm5,%%xmm0                 \n"  // off = x*4 + y*stride
+      "movd        %%xmm0,%k1                    \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+      "movd        %%xmm0,%k5                    \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+      "movd        0x00(%0,%1,1),%%xmm1          \n"
+      "movd        0x00(%0,%5,1),%%xmm6          \n"
+      "punpckldq   %%xmm6,%%xmm1                 \n"
+      "addps       %%xmm4,%%xmm2                 \n"
+      "movq        %%xmm1,(%2)                   \n"
+      "movd        %%xmm0,%k1                    \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+      "movd        %%xmm0,%k5                    \n"
+      "movd        0x00(%0,%1,1),%%xmm0          \n"
+      "movd        0x00(%0,%5,1),%%xmm6          \n"
+      "punpckldq   %%xmm6,%%xmm0                 \n"
+      "addps       %%xmm4,%%xmm3                 \n"
+      "movq        %%xmm0,0x08(%2)               \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%4                       \n"
+      "jge         40b                           \n"
+
+      "49:                                       \n"
+      "add         $0x3,%4                       \n"
+      "jl          19f                           \n"
+
+      // 1 pixel loop
+      LABELALIGN
+      "10:                                       \n"
+      "cvttps2dq   %%xmm2,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "pmaddwd     %%xmm5,%%xmm0                 \n"
+      "addps       %%xmm7,%%xmm2                 \n"
+      "movd        %%xmm0,%k1                    \n"
+      "movd        0x00(%0,%1,1),%%xmm0          \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "lea         0x04(%2),%2                   \n"
+      "sub         $0x1,%4                       \n"
+      "jge         10b                           \n"
+      "19:                                       \n"
+      : "+r"(src_argb),              // %0
+        "+r"(src_argb_stride_temp),  // %1
+        "+r"(dst_argb),              // %2
+        "+r"(src_dudv),              // %3
+        "+rm"(width),                // %4
+        "=&r"(temp)                  // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBAFFINEROW_SSE2
 
 #ifdef HAS_INTERPOLATEROW_SSSE3
 // Bilinear filter 16x2 -> 16x1
-void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                          ptrdiff_t src_stride, int dst_width,
+void InterpolateRow_SSSE3(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          int width,
                           int source_y_fraction) {
-  asm volatile (
-    "sub       %1,%0                           \n"
-    "cmp       $0x0,%3                         \n"
-    "je        100f                            \n"
-    "cmp       $0x80,%3                        \n"
-    "je        50f                             \n"
-
-    "movd      %3,%%xmm0                       \n"
-    "neg       %3                              \n"
-    "add       $0x100,%3                       \n"
-    "movd      %3,%%xmm5                       \n"
-    "punpcklbw %%xmm0,%%xmm5                   \n"
-    "punpcklwd %%xmm5,%%xmm5                   \n"
-    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-    "mov       $0x80808080,%%eax               \n"
-    "movd      %%eax,%%xmm4                    \n"
-    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
-
-    // General purpose row blend.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
-    "movdqa     %%xmm0,%%xmm1                  \n"
-    "punpcklbw  %%xmm2,%%xmm0                  \n"
-    "punpckhbw  %%xmm2,%%xmm1                  \n"
-    "psubb      %%xmm4,%%xmm0                  \n"
-    "psubb      %%xmm4,%%xmm1                  \n"
-    "movdqa     %%xmm5,%%xmm2                  \n"
-    "movdqa     %%xmm5,%%xmm3                  \n"
-    "pmaddubsw  %%xmm0,%%xmm2                  \n"
-    "pmaddubsw  %%xmm1,%%xmm3                  \n"
-    "paddw      %%xmm4,%%xmm2                  \n"
-    "paddw      %%xmm4,%%xmm3                  \n"
-    "psrlw      $0x8,%%xmm2                    \n"
-    "psrlw      $0x8,%%xmm3                    \n"
-    "packuswb   %%xmm3,%%xmm2                  \n"
-    MEMOPMEM(movdqu,xmm2,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-    "jmp       99f                             \n"
-
-    // Blend 50 / 50.
-    LABELALIGN
-  "50:                                         \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        50b                             \n"
-    "jmp       99f                             \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-    LABELALIGN
-  "100:                                        \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        100b                            \n"
-
-  "99:                                         \n"
-  : "+r"(dst_ptr),     // %0
-    "+r"(src_ptr),     // %1
-    "+rm"(dst_width),  // %2
-    "+r"(source_y_fraction)  // %3
-  : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc", "eax", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+  asm volatile(
+      "sub         %1,%0                         \n"
+      "cmp         $0x0,%3                       \n"
+      "je          100f                          \n"
+      "cmp         $0x80,%3                      \n"
+      "je          50f                           \n"
+
+      "movd        %3,%%xmm0                     \n"
+      "neg         %3                            \n"
+      "add         $0x100,%3                     \n"
+      "movd        %3,%%xmm5                     \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"
+      "punpcklwd   %%xmm5,%%xmm5                 \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "mov         $0x80808080,%%eax             \n"
+      "movd        %%eax,%%xmm4                  \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+
+      // General purpose row blend.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      0x00(%1,%4,1),%%xmm2          \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "psubb       %%xmm4,%%xmm0                 \n"
+      "psubb       %%xmm4,%%xmm1                 \n"
+      "movdqa      %%xmm5,%%xmm2                 \n"
+      "movdqa      %%xmm5,%%xmm3                 \n"
+      "pmaddubsw   %%xmm0,%%xmm2                 \n"
+      "pmaddubsw   %%xmm1,%%xmm3                 \n"
+      "paddw       %%xmm4,%%xmm2                 \n"
+      "paddw       %%xmm4,%%xmm3                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "psrlw       $0x8,%%xmm3                   \n"
+      "packuswb    %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm2,0x00(%1,%0,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "jmp         99f                           \n"
+
+      // Blend 50 / 50.
+      LABELALIGN
+      "50:                                       \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      0x00(%1,%4,1),%%xmm1          \n"
+      "pavgb       %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,0x00(%1,%0,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          50b                           \n"
+      "jmp         99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      LABELALIGN
+      "100:                                      \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      %%xmm0,0x00(%1,%0,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),               // %0
+        "+r"(src_ptr),               // %1
+        "+rm"(width),                // %2
+        "+r"(source_y_fraction)      // %3
+      : "r"((intptr_t)(src_stride))  // %4
+      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_INTERPOLATEROW_SSSE3
 
 #ifdef HAS_INTERPOLATEROW_AVX2
 // Bilinear filter 32x2 -> 32x1
-void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
-                         ptrdiff_t src_stride, int dst_width,
+void InterpolateRow_AVX2(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         int width,
                          int source_y_fraction) {
-  asm volatile (
-    "cmp       $0x0,%3                         \n"
-    "je        100f                            \n"
-    "sub       %1,%0                           \n"
-    "cmp       $0x80,%3                        \n"
-    "je        50f                             \n"
-
-    "vmovd      %3,%%xmm0                      \n"
-    "neg        %3                             \n"
-    "add        $0x100,%3                      \n"
-    "vmovd      %3,%%xmm5                      \n"
-    "vpunpcklbw %%xmm0,%%xmm5,%%xmm5           \n"
-    "vpunpcklwd %%xmm5,%%xmm5,%%xmm5           \n"
-    "vbroadcastss %%xmm5,%%ymm5                \n"
-    "mov        $0x80808080,%%eax              \n"
-    "vmovd      %%eax,%%xmm4                   \n"
-    "vbroadcastss %%xmm4,%%ymm4                \n"
-
-    // General purpose row blend.
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
-    MEMOPREG(vmovdqu,0x00,1,4,1,ymm2)
-    "vpunpckhbw %%ymm2,%%ymm0,%%ymm1           \n"
-    "vpunpcklbw %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpsubb     %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpsubb     %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm1,%%ymm5,%%ymm1           \n"
-    "vpmaddubsw %%ymm0,%%ymm5,%%ymm0           \n"
-    "vpaddw     %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpaddw     %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
-    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "jmp       99f                             \n"
-
-    // Blend 50 / 50.
-    LABELALIGN
-  "50:                                         \n"
-    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
-    VMEMOPREG(vpavgb,0x00,1,4,1,ymm0,ymm0)     // vpavgb (%1,%4,1),%%ymm0,%%ymm0
-    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        50b                             \n"
-    "jmp       99f                             \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-    LABELALIGN
-  "100:                                        \n"
-    "rep movsb " MEMMOVESTRING(1,0) "          \n"
-    "jmp       999f                            \n"
-
-  "99:                                         \n"
-    "vzeroupper                                \n"
-  "999:                                        \n"
-  : "+D"(dst_ptr),    // %0
-    "+S"(src_ptr),    // %1
-    "+cm"(dst_width),  // %2
-    "+r"(source_y_fraction)  // %3
-  : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc", "eax", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm4", "xmm5"
-  );
+  asm volatile(
+      "sub         %1,%0                         \n"
+      "cmp         $0x0,%3                       \n"
+      "je          100f                          \n"
+      "cmp         $0x80,%3                      \n"
+      "je          50f                           \n"
+
+      "vmovd       %3,%%xmm0                     \n"
+      "neg         %3                            \n"
+      "add         $0x100,%3                     \n"
+      "vmovd       %3,%%xmm5                     \n"
+      "vpunpcklbw  %%xmm0,%%xmm5,%%xmm5          \n"
+      "vpunpcklwd  %%xmm5,%%xmm5,%%xmm5          \n"
+      "vbroadcastss %%xmm5,%%ymm5                \n"
+      "mov         $0x80808080,%%eax             \n"
+      "vmovd       %%eax,%%xmm4                  \n"
+      "vbroadcastss %%xmm4,%%ymm4                \n"
+
+      // General purpose row blend.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%1),%%ymm0                   \n"
+      "vmovdqu     0x00(%1,%4,1),%%ymm2          \n"
+      "vpunpckhbw  %%ymm2,%%ymm0,%%ymm1          \n"
+      "vpunpcklbw  %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpsubb      %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpsubb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm1,%%ymm5,%%ymm1          \n"
+      "vpmaddubsw  %%ymm0,%%ymm5,%%ymm0          \n"
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpaddw      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,0x00(%1,%0,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "jmp         99f                           \n"
+
+      // Blend 50 / 50.
+      LABELALIGN
+      "50:                                       \n"
+      "vmovdqu     (%1),%%ymm0                   \n"
+      "vpavgb      0x00(%1,%4,1),%%ymm0,%%ymm0   \n"
+      "vmovdqu     %%ymm0,0x00(%1,%0,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          50b                           \n"
+      "jmp         99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      LABELALIGN
+      "100:                                      \n"
+      "vmovdqu     (%1),%%ymm0                   \n"
+      "vmovdqu     %%ymm0,0x00(%1,%0,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          100b                          \n"
+
+      "99:                                       \n"
+      "vzeroupper                                \n"
+      : "+r"(dst_ptr),               // %0
+        "+r"(src_ptr),               // %1
+        "+r"(width),                 // %2
+        "+r"(source_y_fraction)      // %3
+      : "r"((intptr_t)(src_stride))  // %4
+      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm4", "xmm5");
 }
 #endif  // HAS_INTERPOLATEROW_AVX2
 
 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
 // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
-void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                          const uint8* shuffler, int width) {
-  asm volatile (
-    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pshufb    %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  : "r"(shuffler)    // %3
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm5"
-  );
+void ARGBShuffleRow_SSSE3(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          const uint8_t* shuffler,
+                          int width) {
+  asm volatile(
+
+      "movdqu      (%3),%%xmm5                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(shuffler)    // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 #endif  // HAS_ARGBSHUFFLEROW_SSSE3
 
 #ifdef HAS_ARGBSHUFFLEROW_AVX2
 // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
-void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width) {
-  asm volatile (
-    "vbroadcastf128 " MEMACCESS(3) ",%%ymm5    \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
-    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
-    "lea       " MEMLEA(0x40,0) ",%0           \n"
-    "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
-    "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
-    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
-    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  : "r"(shuffler)    // %3
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm5"
-  );
+void ARGBShuffleRow_AVX2(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width) {
+  asm volatile(
+
+      "vbroadcastf128 (%3),%%ymm5                \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(shuffler)    // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
 }
 #endif  // HAS_ARGBSHUFFLEROW_AVX2
 
-#ifdef HAS_ARGBSHUFFLEROW_SSE2
-// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
-void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width) {
-  uintptr_t pixel_temp;
-  asm volatile (
-    "pxor      %%xmm5,%%xmm5                   \n"
-    "mov       " MEMACCESS(4) ",%k2            \n"
-    "cmp       $0x3000102,%k2                  \n"
-    "je        3012f                           \n"
-    "cmp       $0x10203,%k2                    \n"
-    "je        123f                            \n"
-    "cmp       $0x30201,%k2                    \n"
-    "je        321f                            \n"
-    "cmp       $0x2010003,%k2                  \n"
-    "je        2103f                           \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movzb     " MEMACCESS(4) ",%2             \n"
-    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
-    "mov       %b2," MEMACCESS(1) "            \n"
-    "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
-    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
-    "mov       %b2," MEMACCESS2(0x1,1) "       \n"
-    "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
-    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
-    "mov       %b2," MEMACCESS2(0x2,1) "       \n"
-    "movzb     " MEMACCESS2(0x3,4) ",%2        \n"
-    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
-    "mov       %b2," MEMACCESS2(0x3,1) "       \n"
-    "lea       " MEMLEA(0x4,0) ",%0            \n"
-    "lea       " MEMLEA(0x4,1) ",%1            \n"
-    "sub       $0x1,%3                         \n"
-    "jg        1b                              \n"
-    "jmp       99f                             \n"
-
-    LABELALIGN
-  "123:                                        \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpckhbw %%xmm5,%%xmm1                   \n"
-    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
-    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
-    "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
-    "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        123b                            \n"
-    "jmp       99f                             \n"
-
-    LABELALIGN
-  "321:                                        \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpckhbw %%xmm5,%%xmm1                   \n"
-    "pshufhw   $0x39,%%xmm0,%%xmm0             \n"
-    "pshuflw   $0x39,%%xmm0,%%xmm0             \n"
-    "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
-    "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        321b                            \n"
-    "jmp       99f                             \n"
-
-    LABELALIGN
-  "2103:                                       \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpckhbw %%xmm5,%%xmm1                   \n"
-    "pshufhw   $0x93,%%xmm0,%%xmm0             \n"
-    "pshuflw   $0x93,%%xmm0,%%xmm0             \n"
-    "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
-    "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        2103b                           \n"
-    "jmp       99f                             \n"
-
-    LABELALIGN
-  "3012:                                       \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpckhbw %%xmm5,%%xmm1                   \n"
-    "pshufhw   $0xc6,%%xmm0,%%xmm0             \n"
-    "pshuflw   $0xc6,%%xmm0,%%xmm0             \n"
-    "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
-    "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        3012b                           \n"
-
-  "99:                                         \n"
-  : "+r"(src_argb),     // %0
-    "+r"(dst_argb),     // %1
-    "=&d"(pixel_temp),  // %2
-    "+r"(width)         // %3
-  : "r"(shuffler)       // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm5"
-  );
-}
-#endif  // HAS_ARGBSHUFFLEROW_SSE2
-
 #ifdef HAS_I422TOYUY2ROW_SSE2
-void I422ToYUY2Row_SSE2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_frame, int width) {
- asm volatile (
-    "sub       %1,%2                             \n"
-    LABELALIGN
-  "1:                                            \n"
-    "movq      " MEMACCESS(1) ",%%xmm2           \n"
-    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
-    "lea       " MEMLEA(0x8,1) ",%1              \n"
-    "punpcklbw %%xmm3,%%xmm2                     \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
-    "lea       " MEMLEA(0x10,0) ",%0             \n"
-    "movdqa    %%xmm0,%%xmm1                     \n"
-    "punpcklbw %%xmm2,%%xmm0                     \n"
-    "punpckhbw %%xmm2,%%xmm1                     \n"
-    "movdqu    %%xmm0," MEMACCESS(3) "           \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,3) "     \n"
-    "lea       " MEMLEA(0x20,3) ",%3             \n"
-    "sub       $0x10,%4                          \n"
-    "jg         1b                               \n"
-    : "+r"(src_y),  // %0
-      "+r"(src_u),  // %1
-      "+r"(src_v),  // %2
-      "+r"(dst_frame),  // %3
-      "+rm"(width)  // %4
-    :
-    : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3"
-  );
+void I422ToYUY2Row_SSE2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%1),%%xmm2                   \n"
+      "movq        0x00(%1,%2,1),%%xmm1          \n"
+      "add         $0x8,%1                       \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "add         $0x10,%0                      \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%3)                   \n"
+      "movdqu      %%xmm1,0x10(%3)               \n"
+      "lea         0x20(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_yuy2),  // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_I422TOYUY2ROW_SSE2
 
 #ifdef HAS_I422TOUYVYROW_SSE2
-void I422ToUYVYRow_SSE2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_frame, int width) {
- asm volatile (
-    "sub        %1,%2                            \n"
-    LABELALIGN
-  "1:                                            \n"
-    "movq      " MEMACCESS(1) ",%%xmm2           \n"
-    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
-    "lea       " MEMLEA(0x8,1) ",%1              \n"
-    "punpcklbw %%xmm3,%%xmm2                     \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
-    "movdqa    %%xmm2,%%xmm1                     \n"
-    "lea       " MEMLEA(0x10,0) ",%0             \n"
-    "punpcklbw %%xmm0,%%xmm1                     \n"
-    "punpckhbw %%xmm0,%%xmm2                     \n"
-    "movdqu    %%xmm1," MEMACCESS(3) "           \n"
-    "movdqu    %%xmm2," MEMACCESS2(0x10,3) "     \n"
-    "lea       " MEMLEA(0x20,3) ",%3             \n"
-    "sub       $0x10,%4                          \n"
-    "jg         1b                               \n"
-    : "+r"(src_y),  // %0
-      "+r"(src_u),  // %1
-      "+r"(src_v),  // %2
-      "+r"(dst_frame),  // %3
-      "+rm"(width)  // %4
-    :
-    : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3"
-  );
+void I422ToUYVYRow_SSE2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%1),%%xmm2                   \n"
+      "movq        0x00(%1,%2,1),%%xmm1          \n"
+      "add         $0x8,%1                       \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "add         $0x10,%0                      \n"
+      "punpcklbw   %%xmm0,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm1,(%3)                   \n"
+      "movdqu      %%xmm2,0x10(%3)               \n"
+      "lea         0x20(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_uyvy),  // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_I422TOUYVYROW_SSE2
 
+#ifdef HAS_I422TOYUY2ROW_AVX2
+void I422ToYUY2Row_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbw   (%1),%%ymm1                   \n"
+      "vpmovzxbw   0x00(%1,%2,1),%%ymm2          \n"
+      "add         $0x10,%1                      \n"
+      "vpsllw      $0x8,%%ymm2,%%ymm2            \n"
+      "vpor        %%ymm1,%%ymm2,%%ymm2          \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "add         $0x20,%0                      \n"
+      "vpunpcklbw  %%ymm2,%%ymm0,%%ymm1          \n"
+      "vpunpckhbw  %%ymm2,%%ymm0,%%ymm2          \n"
+      "vextractf128 $0x0,%%ymm1,(%3)             \n"
+      "vextractf128 $0x0,%%ymm2,0x10(%3)         \n"
+      "vextractf128 $0x1,%%ymm1,0x20(%3)         \n"
+      "vextractf128 $0x1,%%ymm2,0x30(%3)         \n"
+      "lea         0x40(%3),%3                   \n"
+      "sub         $0x20,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_yuy2),  // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_I422TOYUY2ROW_AVX2
+
+#ifdef HAS_I422TOUYVYROW_AVX2
+void I422ToUYVYRow_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbw   (%1),%%ymm1                   \n"
+      "vpmovzxbw   0x00(%1,%2,1),%%ymm2          \n"
+      "add         $0x10,%1                      \n"
+      "vpsllw      $0x8,%%ymm2,%%ymm2            \n"
+      "vpor        %%ymm1,%%ymm2,%%ymm2          \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "add         $0x20,%0                      \n"
+      "vpunpcklbw  %%ymm0,%%ymm2,%%ymm1          \n"
+      "vpunpckhbw  %%ymm0,%%ymm2,%%ymm2          \n"
+      "vextractf128 $0x0,%%ymm1,(%3)             \n"
+      "vextractf128 $0x0,%%ymm2,0x10(%3)         \n"
+      "vextractf128 $0x1,%%ymm1,0x20(%3)         \n"
+      "vextractf128 $0x1,%%ymm2,0x30(%3)         \n"
+      "lea         0x40(%3),%3                   \n"
+      "sub         $0x20,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_uyvy),  // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_I422TOUYVYROW_AVX2
+
 #ifdef HAS_ARGBPOLYNOMIALROW_SSE2
-void ARGBPolynomialRow_SSE2(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
+void ARGBPolynomialRow_SSE2(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const float* poly,
                             int width) {
-  asm volatile (
-    "pxor      %%xmm3,%%xmm3                   \n"
-
-    // 2 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "punpcklbw %%xmm3,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm4                   \n"
-    "punpcklwd %%xmm3,%%xmm0                   \n"
-    "punpckhwd %%xmm3,%%xmm4                   \n"
-    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
-    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "movdqa    %%xmm4,%%xmm5                   \n"
-    "mulps     " MEMACCESS2(0x10,3) ",%%xmm0   \n"
-    "mulps     " MEMACCESS2(0x10,3) ",%%xmm4   \n"
-    "addps     " MEMACCESS(3) ",%%xmm0         \n"
-    "addps     " MEMACCESS(3) ",%%xmm4         \n"
-    "movdqa    %%xmm1,%%xmm2                   \n"
-    "movdqa    %%xmm5,%%xmm6                   \n"
-    "mulps     %%xmm1,%%xmm2                   \n"
-    "mulps     %%xmm5,%%xmm6                   \n"
-    "mulps     %%xmm2,%%xmm1                   \n"
-    "mulps     %%xmm6,%%xmm5                   \n"
-    "mulps     " MEMACCESS2(0x20,3) ",%%xmm2   \n"
-    "mulps     " MEMACCESS2(0x20,3) ",%%xmm6   \n"
-    "mulps     " MEMACCESS2(0x30,3) ",%%xmm1   \n"
-    "mulps     " MEMACCESS2(0x30,3) ",%%xmm5   \n"
-    "addps     %%xmm2,%%xmm0                   \n"
-    "addps     %%xmm6,%%xmm4                   \n"
-    "addps     %%xmm1,%%xmm0                   \n"
-    "addps     %%xmm5,%%xmm4                   \n"
-    "cvttps2dq %%xmm0,%%xmm0                   \n"
-    "cvttps2dq %%xmm4,%%xmm4                   \n"
-    "packuswb  %%xmm4,%%xmm0                   \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x2,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  : "r"(poly)        // %3
-  : "memory", "cc"
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+  asm volatile(
+
+      "pxor        %%xmm3,%%xmm3                 \n"
+
+      // 2 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "punpcklbw   %%xmm3,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "punpcklwd   %%xmm3,%%xmm0                 \n"
+      "punpckhwd   %%xmm3,%%xmm4                 \n"
+      "cvtdq2ps    %%xmm0,%%xmm0                 \n"
+      "cvtdq2ps    %%xmm4,%%xmm4                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "mulps       0x10(%3),%%xmm0               \n"
+      "mulps       0x10(%3),%%xmm4               \n"
+      "addps       (%3),%%xmm0                   \n"
+      "addps       (%3),%%xmm4                   \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "movdqa      %%xmm5,%%xmm6                 \n"
+      "mulps       %%xmm1,%%xmm2                 \n"
+      "mulps       %%xmm5,%%xmm6                 \n"
+      "mulps       %%xmm2,%%xmm1                 \n"
+      "mulps       %%xmm6,%%xmm5                 \n"
+      "mulps       0x20(%3),%%xmm2               \n"
+      "mulps       0x20(%3),%%xmm6               \n"
+      "mulps       0x30(%3),%%xmm1               \n"
+      "mulps       0x30(%3),%%xmm5               \n"
+      "addps       %%xmm2,%%xmm0                 \n"
+      "addps       %%xmm6,%%xmm4                 \n"
+      "addps       %%xmm1,%%xmm0                 \n"
+      "addps       %%xmm5,%%xmm4                 \n"
+      "cvttps2dq   %%xmm0,%%xmm0                 \n"
+      "cvttps2dq   %%xmm4,%%xmm4                 \n"
+      "packuswb    %%xmm4,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x2,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(poly)        // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_SSE2
 
 #ifdef HAS_ARGBPOLYNOMIALROW_AVX2
-void ARGBPolynomialRow_AVX2(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
+void ARGBPolynomialRow_AVX2(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const float* poly,
                             int width) {
-  asm volatile (
-    "vbroadcastf128 " MEMACCESS(3) ",%%ymm4     \n"
-    "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n"
-    "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n"
-    "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n"
-
-    // 2 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "vpmovzxbd   " MEMACCESS(0) ",%%ymm0       \n"  // 2 ARGB pixels
-    "lea         " MEMLEA(0x8,0) ",%0          \n"
-    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
-    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
-    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
-    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
-    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
-    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
-    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
-    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
-    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
-    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
-    "vmovq       %%xmm0," MEMACCESS(1) "       \n"
-    "lea         " MEMLEA(0x8,1) ",%1          \n"
-    "sub         $0x2,%2                       \n"
-    "jg          1b                            \n"
-    "vzeroupper                                \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  : "r"(poly)        // %3
-  : "memory", "cc",
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+  asm volatile(
+      "vbroadcastf128 (%3),%%ymm4                \n"
+      "vbroadcastf128 0x10(%3),%%ymm5            \n"
+      "vbroadcastf128 0x20(%3),%%ymm6            \n"
+      "vbroadcastf128 0x30(%3),%%ymm7            \n"
+
+      // 2 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbd   (%0),%%ymm0                   \n"  // 2 ARGB pixels
+      "lea         0x8(%0),%0                    \n"
+      "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+      "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+      "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+      "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+      "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+      "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X *
+                                                      // X
+      "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+      "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+      "vmovq       %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x2,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(poly)        // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_AVX2
 
+#ifdef HAS_HALFFLOATROW_SSE2
+static float kScaleBias = 1.9259299444e-34f;
+void HalfFloatRow_SSE2(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width) {
+  scale *= kScaleBias;
+  asm volatile(
+      "movd        %3,%%xmm4                     \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "sub         %0,%1                         \n"
+
+      // 16 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm2                   \n"  // 8 shorts
+      "add         $0x10,%0                      \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm5,%%xmm2                 \n"  // 8 ints in xmm2/1
+      "cvtdq2ps    %%xmm2,%%xmm2                 \n"  // 8 floats
+      "punpckhwd   %%xmm5,%%xmm3                 \n"
+      "cvtdq2ps    %%xmm3,%%xmm3                 \n"
+      "mulps       %%xmm4,%%xmm2                 \n"
+      "mulps       %%xmm4,%%xmm3                 \n"
+      "psrld       $0xd,%%xmm2                   \n"
+      "psrld       $0xd,%%xmm3                   \n"
+      "packssdw    %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm2,-0x10(%0,%1,1)         \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      : "m"(scale)   // %3
+      : "memory", "cc", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_HALFFLOATROW_SSE2
+
+#ifdef HAS_HALFFLOATROW_AVX2
+void HalfFloatRow_AVX2(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width) {
+  scale *= kScaleBias;
+  asm volatile(
+      "vbroadcastss %3, %%ymm4                   \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+      "sub         %0,%1                         \n"
+
+      // 16 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm2                   \n"  // 16 shorts
+      "add         $0x20,%0                      \n"
+      "vpunpckhwd  %%ymm5,%%ymm2,%%ymm3          \n"  // mutates
+      "vpunpcklwd  %%ymm5,%%ymm2,%%ymm2          \n"
+      "vcvtdq2ps   %%ymm3,%%ymm3                 \n"
+      "vcvtdq2ps   %%ymm2,%%ymm2                 \n"
+      "vmulps      %%ymm3,%%ymm4,%%ymm3          \n"
+      "vmulps      %%ymm2,%%ymm4,%%ymm2          \n"
+      "vpsrld      $0xd,%%ymm3,%%ymm3            \n"
+      "vpsrld      $0xd,%%ymm2,%%ymm2            \n"
+      "vpackssdw   %%ymm3, %%ymm2, %%ymm2        \n"  // unmutates
+      "vmovdqu     %%ymm2,-0x20(%0,%1,1)         \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+#if defined(__x86_64__)
+      : "x"(scale)  // %3
+#else
+      : "m"(scale)            // %3
+#endif
+      : "memory", "cc", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_HALFFLOATROW_AVX2
+
+#ifdef HAS_HALFFLOATROW_F16C
+void HalfFloatRow_F16C(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width) {
+  asm volatile(
+      "vbroadcastss %3, %%ymm4                   \n"
+      "sub         %0,%1                         \n"
+
+      // 16 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxwd   (%0),%%ymm2                   \n"  // 16 shorts -> 16 ints
+      "vpmovzxwd   0x10(%0),%%ymm3               \n"
+      "vcvtdq2ps   %%ymm2,%%ymm2                 \n"
+      "vcvtdq2ps   %%ymm3,%%ymm3                 \n"
+      "vmulps      %%ymm2,%%ymm4,%%ymm2          \n"
+      "vmulps      %%ymm3,%%ymm4,%%ymm3          \n"
+      "vcvtps2ph   $3, %%ymm2, %%xmm2            \n"
+      "vcvtps2ph   $3, %%ymm3, %%xmm3            \n"
+      "vmovdqu     %%xmm2,0x00(%0,%1,1)          \n"
+      "vmovdqu     %%xmm3,0x10(%0,%1,1)          \n"
+      "add         $0x20,%0                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+#if defined(__x86_64__)
+      : "x"(scale)  // %3
+#else
+      : "m"(scale)            // %3
+#endif
+      : "memory", "cc", "xmm2", "xmm3", "xmm4");
+}
+#endif  // HAS_HALFFLOATROW_F16C
+
+#ifdef HAS_HALFFLOATROW_F16C
+void HalfFloat1Row_F16C(const uint16_t* src, uint16_t* dst, float, int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      // 16 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxwd   (%0),%%ymm2                   \n"  // 16 shorts -> 16 ints
+      "vpmovzxwd   0x10(%0),%%ymm3               \n"
+      "vcvtdq2ps   %%ymm2,%%ymm2                 \n"
+      "vcvtdq2ps   %%ymm3,%%ymm3                 \n"
+      "vcvtps2ph   $3, %%ymm2, %%xmm2            \n"
+      "vcvtps2ph   $3, %%ymm3, %%xmm3            \n"
+      "vmovdqu     %%xmm2,0x00(%0,%1,1)          \n"
+      "vmovdqu     %%xmm3,0x10(%0,%1,1)          \n"
+      "add         $0x20,%0                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "memory", "cc", "xmm2", "xmm3");
+}
+#endif  // HAS_HALFFLOATROW_F16C
+
 #ifdef HAS_ARGBCOLORTABLEROW_X86
 // Tranform ARGB pixels with color table.
-void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+void ARGBColorTableRow_X86(uint8_t* dst_argb,
+                           const uint8_t* table_argb,
                            int width) {
   uintptr_t pixel_temp;
-  asm volatile (
-    // 1 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movzb     " MEMACCESS(0) ",%1             \n"
-    "lea       " MEMLEA(0x4,0) ",%0            \n"
-    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
-    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
-    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
-    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
-    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
-    "movzb     " MEMACCESS2(-0x1,0) ",%1       \n"
-    MEMOPARG(movzb,0x03,3,1,4,1) "             \n"  // movzb 0x3(%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x1,0) "      \n"
-    "dec       %2                              \n"
-    "jg        1b                              \n"
-  : "+r"(dst_argb),     // %0
-    "=&d"(pixel_temp),  // %1
-    "+r"(width)         // %2
-  : "r"(table_argb)     // %3
-  : "memory", "cc");
+  asm volatile(
+      // 1 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movzb       (%0),%1                       \n"
+      "lea         0x4(%0),%0                    \n"
+      "movzb       0x00(%3,%1,4),%1              \n"
+      "mov         %b1,-0x4(%0)                  \n"
+      "movzb       -0x3(%0),%1                   \n"
+      "movzb       0x01(%3,%1,4),%1              \n"
+      "mov         %b1,-0x3(%0)                  \n"
+      "movzb       -0x2(%0),%1                   \n"
+      "movzb       0x02(%3,%1,4),%1              \n"
+      "mov         %b1,-0x2(%0)                  \n"
+      "movzb       -0x1(%0),%1                   \n"
+      "movzb       0x03(%3,%1,4),%1              \n"
+      "mov         %b1,-0x1(%0)                  \n"
+      "dec         %2                            \n"
+      "jg          1b                            \n"
+      : "+r"(dst_argb),     // %0
+        "=&d"(pixel_temp),  // %1
+        "+r"(width)         // %2
+      : "r"(table_argb)     // %3
+      : "memory", "cc");
 }
 #endif  // HAS_ARGBCOLORTABLEROW_X86
 
 #ifdef HAS_RGBCOLORTABLEROW_X86
 // Tranform RGB pixels with color table.
-void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+void RGBColorTableRow_X86(uint8_t* dst_argb,
+                          const uint8_t* table_argb,
+                          int width) {
   uintptr_t pixel_temp;
-  asm volatile (
-    // 1 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movzb     " MEMACCESS(0) ",%1             \n"
-    "lea       " MEMLEA(0x4,0) ",%0            \n"
-    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
-    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
-    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
-    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
-    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
-    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
-    "dec       %2                              \n"
-    "jg        1b                              \n"
-  : "+r"(dst_argb),     // %0
-    "=&d"(pixel_temp),  // %1
-    "+r"(width)         // %2
-  : "r"(table_argb)     // %3
-  : "memory", "cc");
+  asm volatile(
+      // 1 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movzb       (%0),%1                       \n"
+      "lea         0x4(%0),%0                    \n"
+      "movzb       0x00(%3,%1,4),%1              \n"
+      "mov         %b1,-0x4(%0)                  \n"
+      "movzb       -0x3(%0),%1                   \n"
+      "movzb       0x01(%3,%1,4),%1              \n"
+      "mov         %b1,-0x3(%0)                  \n"
+      "movzb       -0x2(%0),%1                   \n"
+      "movzb       0x02(%3,%1,4),%1              \n"
+      "mov         %b1,-0x2(%0)                  \n"
+      "dec         %2                            \n"
+      "jg          1b                            \n"
+      : "+r"(dst_argb),     // %0
+        "=&d"(pixel_temp),  // %1
+        "+r"(width)         // %2
+      : "r"(table_argb)     // %3
+      : "memory", "cc");
 }
 #endif  // HAS_RGBCOLORTABLEROW_X86
 
 #ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
 // Tranform RGB pixels with luma table.
-void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb,
+                                 uint8_t* dst_argb,
                                  int width,
-                                 const uint8* luma, uint32 lumacoeff) {
+                                 const uint8_t* luma,
+                                 uint32_t lumacoeff) {
   uintptr_t pixel_temp;
   uintptr_t table_temp;
-  asm volatile (
-    "movd      %6,%%xmm3                       \n"
-    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-    "pcmpeqb   %%xmm4,%%xmm4                   \n"
-    "psllw     $0x8,%%xmm4                     \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-
-    // 4 pixel loop.
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(2) ",%%xmm0         \n"
-    "pmaddubsw %%xmm3,%%xmm0                   \n"
-    "phaddw    %%xmm0,%%xmm0                   \n"
-    "pand      %%xmm4,%%xmm0                   \n"
-    "punpcklwd %%xmm5,%%xmm0                   \n"
-    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-    "add       %5,%1                           \n"
-    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-
-    "movzb     " MEMACCESS(2) ",%0             \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS(3) "            \n"
-    "movzb     " MEMACCESS2(0x1,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x1,3) "       \n"
-    "movzb     " MEMACCESS2(0x2,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x2,3) "       \n"
-    "movzb     " MEMACCESS2(0x3,2) ",%0        \n"
-    "mov       %b0," MEMACCESS2(0x3,3) "       \n"
-
-    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-    "add       %5,%1                           \n"
-    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-
-    "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x4,3) "       \n"
-    "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x5,3) "       \n"
-    "movzb     " MEMACCESS2(0x6,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x6,3) "       \n"
-    "movzb     " MEMACCESS2(0x7,2) ",%0        \n"
-    "mov       %b0," MEMACCESS2(0x7,3) "       \n"
-
-    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-    "add       %5,%1                           \n"
-    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-
-    "movzb     " MEMACCESS2(0x8,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x8,3) "       \n"
-    "movzb     " MEMACCESS2(0x9,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0x9,3) "       \n"
-    "movzb     " MEMACCESS2(0xa,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0xa,3) "       \n"
-    "movzb     " MEMACCESS2(0xb,2) ",%0        \n"
-    "mov       %b0," MEMACCESS2(0xb,3) "       \n"
-
-    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-    "add       %5,%1                           \n"
-
-    "movzb     " MEMACCESS2(0xc,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0xc,3) "       \n"
-    "movzb     " MEMACCESS2(0xd,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0xd,3) "       \n"
-    "movzb     " MEMACCESS2(0xe,2) ",%0        \n"
-    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
-    "mov       %b0," MEMACCESS2(0xe,3) "       \n"
-    "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
-    "mov       %b0," MEMACCESS2(0xf,3) "       \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "lea       " MEMLEA(0x10,3) ",%3           \n"
-    "sub       $0x4,%4                         \n"
-    "jg        1b                              \n"
-  : "=&d"(pixel_temp),  // %0
-    "=&a"(table_temp),  // %1
-    "+r"(src_argb),     // %2
-    "+r"(dst_argb),     // %3
-    "+rm"(width)        // %4
-  : "r"(luma),          // %5
-    "rm"(lumacoeff)     // %6
-  : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5"
-  );
+  asm volatile(
+      "movd        %6,%%xmm3                     \n"
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psllw       $0x8,%%xmm4                   \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      // 4 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%2),%%xmm0                   \n"
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"
+      "phaddw      %%xmm0,%%xmm0                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "punpcklwd   %%xmm5,%%xmm0                 \n"
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+
+      "movzb       (%2),%0                       \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,(%3)                      \n"
+      "movzb       0x1(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x1(%3)                   \n"
+      "movzb       0x2(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x2(%3)                   \n"
+      "movzb       0x3(%2),%0                    \n"
+      "mov         %b0,0x3(%3)                   \n"
+
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+
+      "movzb       0x4(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x4(%3)                   \n"
+      "movzb       0x5(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x5(%3)                   \n"
+      "movzb       0x6(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x6(%3)                   \n"
+      "movzb       0x7(%2),%0                    \n"
+      "mov         %b0,0x7(%3)                   \n"
+
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+
+      "movzb       0x8(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x8(%3)                   \n"
+      "movzb       0x9(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x9(%3)                   \n"
+      "movzb       0xa(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xa(%3)                   \n"
+      "movzb       0xb(%2),%0                    \n"
+      "mov         %b0,0xb(%3)                   \n"
+
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+
+      "movzb       0xc(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xc(%3)                   \n"
+      "movzb       0xd(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xd(%3)                   \n"
+      "movzb       0xe(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xe(%3)                   \n"
+      "movzb       0xf(%2),%0                    \n"
+      "mov         %b0,0xf(%3)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "lea         0x10(%3),%3                   \n"
+      "sub         $0x4,%4                       \n"
+      "jg          1b                            \n"
+      : "=&d"(pixel_temp),  // %0
+        "=&a"(table_temp),  // %1
+        "+r"(src_argb),     // %2
+        "+r"(dst_argb),     // %3
+        "+rm"(width)        // %4
+      : "r"(luma),          // %5
+        "rm"(lumacoeff)     // %6
+      : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
 
+static const uvec8 kYUV24Shuffle[3] = {
+    {8, 9, 0, 8, 9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12},
+    {9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15},
+    {2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15, 6, 14, 15, 7}};
+
+// Convert biplanar NV21 to packed YUV24
+// NV21 has VU in memory for chroma.
+// YUV24 is VUY in memory
+void NV21ToYUV24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_vu,
+                          uint8_t* dst_yuv24,
+                          int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "movdqa      (%4),%%xmm4                   \n"  // 3 shuffler constants
+      "movdqa      16(%4),%%xmm5                 \n"
+      "movdqa      32(%4),%%xmm6                 \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm2                   \n"  // load 16 Y values
+      "movdqu      (%0,%1),%%xmm3                \n"  // load 8 VU values
+      "lea         16(%0),%0                     \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "shufps      $0x44,%%xmm3,%%xmm0           \n"  // Y 0..7,  UV 0..3
+      "shufps      $0x99,%%xmm3,%%xmm1           \n"  // Y 4..11, UV 2..5
+      "shufps      $0xee,%%xmm3,%%xmm2           \n"  // Y 8..15, UV 4..7
+      "pshufb      %%xmm4, %%xmm0                \n"  // weave into YUV24
+      "pshufb      %%xmm5, %%xmm1                \n"
+      "pshufb      %%xmm6, %%xmm2                \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "movdqu      %%xmm1,16(%2)                 \n"
+      "movdqu      %%xmm2,32(%2)                 \n"
+      "lea         48(%2),%2                     \n"
+      "sub         $16,%3                        \n"  // 16 pixels per loop
+      "jg          1b                            \n"
+      : "+r"(src_y),            // %0
+        "+r"(src_vu),           // %1
+        "+r"(dst_yuv24),        // %2
+        "+r"(width)             // %3
+      : "r"(&kYUV24Shuffle[0])  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+// Convert biplanar NV21 to packed YUV24
+// NV21 has VU in memory for chroma.
+// YUV24 is VUY in memory
+void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "vbroadcastf128 (%4),%%ymm4                \n"  // 3 shuffler constants
+      "vbroadcastf128 16(%4),%%ymm5              \n"
+      "vbroadcastf128 32(%4),%%ymm6              \n"
+
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm2                   \n"  // load 32 Y values
+      "vmovdqu     (%0,%1),%%ymm3                \n"  // load 16 VU values
+      "lea         32(%0),%0                     \n"
+      "vshufps     $0x44,%%ymm3,%%ymm2,%%ymm0    \n"  // Y 0..7,  UV 0..3
+      "vshufps     $0x99,%%ymm3,%%ymm2,%%ymm1    \n"  // Y 4..11, UV 2..5
+      "vshufps     $0xee,%%ymm3,%%ymm2,%%ymm2    \n"  // Y 8..15, UV 4..7
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"  // weave into YUV24
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm2,%%ymm2          \n"
+      "vperm2i128  $0x20,%%ymm1,%%ymm0,%%ymm3    \n"
+      "vperm2i128  $0x30,%%ymm0,%%ymm2,%%ymm0    \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm1,%%ymm1    \n"
+      "vmovdqu     %%ymm3,(%2)                   \n"
+      "vmovdqu     %%ymm0,32(%2)                 \n"
+      "vmovdqu     %%ymm1,64(%2)                 \n"
+      "lea         96(%2),%2                     \n"
+      "sub         $32,%3                        \n"  // 32 pixels per loop
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_y),            // %0
+        "+r"(src_vu),           // %1
+        "+r"(dst_yuv24),        // %2
+        "+r"(width)             // %3
+      : "r"(&kYUV24Shuffle[0])  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+#ifdef HAS_NV21ToYUV24ROW_AVX512
+// The following VMBI VEX256 code tests okay with the intelsde emulator.
+static const lvec8 kYUV24Perm[3] = {
+    {32, 33, 0,  32, 33, 1,  34, 35, 2,  34, 35, 3,  36, 37, 4,  36,
+     37, 5,  38, 39, 6,  38, 39, 7,  40, 41, 8,  40, 41, 9,  42, 43},
+    {10, 42, 43, 11, 44, 45, 12, 44, 45, 13, 46, 47, 14, 46, 47, 15,
+     48, 49, 16, 48, 49, 17, 50, 51, 18, 50, 51, 19, 52, 53, 20, 52},
+    {53, 21, 54, 55, 22, 54, 55, 23, 56, 57, 24, 56, 57, 25, 58, 59,
+     26, 58, 59, 27, 60, 61, 28, 60, 61, 29, 62, 63, 30, 62, 63, 31}};
+
+void NV21ToYUV24Row_AVX512(const uint8_t* src_y,
+                           const uint8_t* src_vu,
+                           uint8_t* dst_yuv24,
+                           int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "vmovdqa     (%4),%%ymm4                   \n"  // 3 shuffler constants
+      "vmovdqa     32(%4),%%ymm5                 \n"
+      "vmovdqa     64(%4),%%ymm6                 \n" LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm2                   \n"  // load 32 Y values
+      "vmovdqu     (%0,%1),%%ymm3                \n"  // load 16 VU values
+      "lea         32(%0),%0                     \n"
+      "vmovdqa     %%ymm2, %%ymm0                \n"
+      "vmovdqa     %%ymm2, %%ymm1                \n"
+      "vpermt2b    %%ymm3,%%ymm4,%%ymm0          \n"
+      "vpermt2b    %%ymm3,%%ymm5,%%ymm1          \n"
+      "vpermt2b    %%ymm3,%%ymm6,%%ymm2          \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "vmovdqu     %%ymm1,32(%2)                 \n"
+      "vmovdqu     %%ymm2,64(%2)                 \n"
+      "lea         96(%2),%2                     \n"
+      "sub         $32,%3                        \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_y),         // %0
+        "+r"(src_vu),        // %1
+        "+r"(dst_yuv24),     // %2
+        "+r"(width)          // %3
+      : "r"(&kYUV24Perm[0])  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+#endif  // HAS_NV21ToYUV24ROW_AVX512
+
+#ifdef HAS_SWAPUVROW_SSSE3
+
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleUVToVU = {1u, 0u, 3u,  2u,  5u,  4u,  7u,  6u,
+                                     9u, 8u, 11u, 10u, 13u, 12u, 15u, 14u};
+
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+
+      "movdqu      %3,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),        // %0
+        "+r"(dst_vu),        // %1
+        "+r"(width)          // %2
+      : "m"(kShuffleUVToVU)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
+}
+#endif  // HAS_SWAPUVROW_SSSE3
+
+#ifdef HAS_SWAPUVROW_AVX2
+void SwapUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uv),        // %0
+        "+r"(dst_vu),        // %1
+        "+r"(width)          // %2
+      : "m"(kShuffleUVToVU)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
+}
+#endif  // HAS_SWAPUVROW_AVX2
+
+void HalfMergeUVRow_SSSE3(const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          uint8_t* dst_uv,
+                          int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // load 16 U values
+      "movdqu      (%1),%%xmm1                   \n"  // load 16 V values
+      "movdqu      0(%0,%4,1),%%xmm2             \n"  // 16 from next row
+      "movdqu      0(%1,%5,1),%%xmm3             \n"
+      "lea         0x10(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"  // half size
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "lea         0x10(%1),%1                   \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x1,%%xmm0                   \n"
+      "psrlw       $0x1,%%xmm1                   \n"
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"  // store 8 UV pixels
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"  // 16 src pixels per loop
+      "jg          1b                            \n"
+      : "+r"(src_u),                    // %0
+        "+r"(src_v),                    // %1
+        "+r"(dst_uv),                   // %2
+        "+r"(width)                     // %3
+      : "r"((intptr_t)(src_stride_u)),  // %4
+        "r"((intptr_t)(src_stride_v))   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+
+void HalfMergeUVRow_AVX2(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // load 32 U values
+      "vmovdqu     (%1),%%ymm1                   \n"  // load 32 V values
+      "vmovdqu     0(%0,%4,1),%%ymm2             \n"  // 32 from next row
+      "vmovdqu     0(%1,%5,1),%%ymm3             \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"  // half size
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "lea         0x20(%1),%1                   \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpsrlw      $0x1,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x1,%%ymm1,%%ymm1            \n"
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpavgw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"  // store 16 UV pixels
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x20,%3                      \n"  // 32 src pixels per loop
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_u),                    // %0
+        "+r"(src_v),                    // %1
+        "+r"(dst_uv),                   // %2
+        "+r"(width)                     // %3
+      : "r"((intptr_t)(src_stride_u)),  // %4
+        "r"((intptr_t)(src_stride_v))   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+
+void ClampFloatToZero_SSE2(const float* src_x, float* dst_y, int width) {
+  asm volatile(
+      "pxor        %%xmm1,%%xmm1                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movd        (%0),%%xmm0                   \n"  // load float
+      "maxss       %%xmm1, %%xmm0                \n"  // clamp to zero
+      "add         4, %0                         \n"
+      "movd        %%xmm0, (%1)                  \n"  // store float
+      "add         4, %1                         \n"
+      "sub         $0x4,%2                       \n"  // 1 float per loop
+      "jg          1b                            \n"
+      : "+r"(src_x),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+
 #endif  // defined(__x86_64__) || defined(__i386__)
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/row_lasx.cc b/media/libyuv/source/row_lasx.cc
new file mode 100644
index 0000000000..29ac9254d9
--- /dev/null
+++ b/media/libyuv/source/row_lasx.cc
@@ -0,0 +1,2304 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_LASX) && defined(__loongarch_asx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ALPHA_VAL (-1)
+
+// Fill YUV -> RGB conversion constants into vectors
+#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, yg, yb) \
+  {                                                      \
+    ub = __lasx_xvreplgr2vr_h(yuvconst->kUVToB[0]);      \
+    vr = __lasx_xvreplgr2vr_h(yuvconst->kUVToR[1]);      \
+    ug = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[0]);      \
+    vg = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[1]);      \
+    yg = __lasx_xvreplgr2vr_h(yuvconst->kYToRgb[0]);     \
+    yb = __lasx_xvreplgr2vr_w(yuvconst->kYBiasToRgb[0]); \
+  }
+
+// Load 32 YUV422 pixel data
+#define READYUV422_D(psrc_y, psrc_u, psrc_v, out_y, uv_l, uv_h) \
+  {                                                             \
+    __m256i temp0, temp1;                                       \
+                                                                \
+    DUP2_ARG2(__lasx_xvld, psrc_y, 0, psrc_u, 0, out_y, temp0); \
+    temp1 = __lasx_xvld(psrc_v, 0);                             \
+    temp0 = __lasx_xvsub_b(temp0, const_0x80);                  \
+    temp1 = __lasx_xvsub_b(temp1, const_0x80);                  \
+    temp0 = __lasx_vext2xv_h_b(temp0);                          \
+    temp1 = __lasx_vext2xv_h_b(temp1);                          \
+    uv_l = __lasx_xvilvl_h(temp0, temp1);                       \
+    uv_h = __lasx_xvilvh_h(temp0, temp1);                       \
+  }
+
+// Load 16 YUV422 pixel data
+#define READYUV422(psrc_y, psrc_u, psrc_v, out_y, uv) \
+  {                                                   \
+    __m256i temp0, temp1;                             \
+                                                      \
+    out_y = __lasx_xvld(psrc_y, 0);                   \
+    temp0 = __lasx_xvldrepl_d(psrc_u, 0);             \
+    temp1 = __lasx_xvldrepl_d(psrc_v, 0);             \
+    uv = __lasx_xvilvl_b(temp0, temp1);               \
+    uv = __lasx_xvsub_b(uv, const_0x80);              \
+    uv = __lasx_vext2xv_h_b(uv);                      \
+  }
+
+// Convert 16 pixels of YUV420 to RGB.
+#define YUVTORGB_D(in_y, in_uvl, in_uvh, ubvr, ugvg, yg, yb, b_l, b_h, g_l,   \
+                   g_h, r_l, r_h)                                             \
+  {                                                                           \
+    __m256i u_l, u_h, v_l, v_h;                                               \
+    __m256i yl_ev, yl_od, yh_ev, yh_od;                                       \
+    __m256i temp0, temp1, temp2, temp3;                                       \
+                                                                              \
+    temp0 = __lasx_xvilvl_b(in_y, in_y);                                      \
+    temp1 = __lasx_xvilvh_b(in_y, in_y);                                      \
+    yl_ev = __lasx_xvmulwev_w_hu_h(temp0, yg);                                \
+    yl_od = __lasx_xvmulwod_w_hu_h(temp0, yg);                                \
+    yh_ev = __lasx_xvmulwev_w_hu_h(temp1, yg);                                \
+    yh_od = __lasx_xvmulwod_w_hu_h(temp1, yg);                                \
+    DUP4_ARG2(__lasx_xvsrai_w, yl_ev, 16, yl_od, 16, yh_ev, 16, yh_od, 16,    \
+              yl_ev, yl_od, yh_ev, yh_od);                                    \
+    yl_ev = __lasx_xvadd_w(yl_ev, yb);                                        \
+    yl_od = __lasx_xvadd_w(yl_od, yb);                                        \
+    yh_ev = __lasx_xvadd_w(yh_ev, yb);                                        \
+    yh_od = __lasx_xvadd_w(yh_od, yb);                                        \
+    v_l = __lasx_xvmulwev_w_h(in_uvl, ubvr);                                  \
+    u_l = __lasx_xvmulwod_w_h(in_uvl, ubvr);                                  \
+    v_h = __lasx_xvmulwev_w_h(in_uvh, ubvr);                                  \
+    u_h = __lasx_xvmulwod_w_h(in_uvh, ubvr);                                  \
+    temp0 = __lasx_xvadd_w(yl_ev, u_l);                                       \
+    temp1 = __lasx_xvadd_w(yl_od, u_l);                                       \
+    temp2 = __lasx_xvadd_w(yh_ev, u_h);                                       \
+    temp3 = __lasx_xvadd_w(yh_od, u_h);                                       \
+    DUP4_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp2, 6, temp3, 6, temp0, \
+              temp1, temp2, temp3);                                           \
+    DUP4_ARG1(__lasx_xvclip255_w, temp0, temp1, temp2, temp3, temp0, temp1,   \
+              temp2, temp3);                                                  \
+    b_l = __lasx_xvpackev_h(temp1, temp0);                                    \
+    b_h = __lasx_xvpackev_h(temp3, temp2);                                    \
+    temp0 = __lasx_xvadd_w(yl_ev, v_l);                                       \
+    temp1 = __lasx_xvadd_w(yl_od, v_l);                                       \
+    temp2 = __lasx_xvadd_w(yh_ev, v_h);                                       \
+    temp3 = __lasx_xvadd_w(yh_od, v_h);                                       \
+    DUP4_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp2, 6, temp3, 6, temp0, \
+              temp1, temp2, temp3);                                           \
+    DUP4_ARG1(__lasx_xvclip255_w, temp0, temp1, temp2, temp3, temp0, temp1,   \
+              temp2, temp3);                                                  \
+    r_l = __lasx_xvpackev_h(temp1, temp0);                                    \
+    r_h = __lasx_xvpackev_h(temp3, temp2);                                    \
+    DUP2_ARG2(__lasx_xvdp2_w_h, in_uvl, ugvg, in_uvh, ugvg, u_l, u_h);        \
+    temp0 = __lasx_xvsub_w(yl_ev, u_l);                                       \
+    temp1 = __lasx_xvsub_w(yl_od, u_l);                                       \
+    temp2 = __lasx_xvsub_w(yh_ev, u_h);                                       \
+    temp3 = __lasx_xvsub_w(yh_od, u_h);                                       \
+    DUP4_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp2, 6, temp3, 6, temp0, \
+              temp1, temp2, temp3);                                           \
+    DUP4_ARG1(__lasx_xvclip255_w, temp0, temp1, temp2, temp3, temp0, temp1,   \
+              temp2, temp3);                                                  \
+    g_l = __lasx_xvpackev_h(temp1, temp0);                                    \
+    g_h = __lasx_xvpackev_h(temp3, temp2);                                    \
+  }
+
+// Convert 8 pixels of YUV420 to RGB.
+#define YUVTORGB(in_y, in_uv, ubvr, ugvg, yg, yb, out_b, out_g, out_r) \
+  {                                                                    \
+    __m256i u_l, v_l, yl_ev, yl_od;                                    \
+    __m256i temp0, temp1;                                              \
+                                                                       \
+    in_y = __lasx_xvpermi_d(in_y, 0xD8);                               \
+    temp0 = __lasx_xvilvl_b(in_y, in_y);                               \
+    yl_ev = __lasx_xvmulwev_w_hu_h(temp0, yg);                         \
+    yl_od = __lasx_xvmulwod_w_hu_h(temp0, yg);                         \
+    DUP2_ARG2(__lasx_xvsrai_w, yl_ev, 16, yl_od, 16, yl_ev, yl_od);    \
+    yl_ev = __lasx_xvadd_w(yl_ev, yb);                                 \
+    yl_od = __lasx_xvadd_w(yl_od, yb);                                 \
+    v_l = __lasx_xvmulwev_w_h(in_uv, ubvr);                            \
+    u_l = __lasx_xvmulwod_w_h(in_uv, ubvr);                            \
+    temp0 = __lasx_xvadd_w(yl_ev, u_l);                                \
+    temp1 = __lasx_xvadd_w(yl_od, u_l);                                \
+    DUP2_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp0, temp1);      \
+    DUP2_ARG1(__lasx_xvclip255_w, temp0, temp1, temp0, temp1);         \
+    out_b = __lasx_xvpackev_h(temp1, temp0);                           \
+    temp0 = __lasx_xvadd_w(yl_ev, v_l);                                \
+    temp1 = __lasx_xvadd_w(yl_od, v_l);                                \
+    DUP2_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp0, temp1);      \
+    DUP2_ARG1(__lasx_xvclip255_w, temp0, temp1, temp0, temp1);         \
+    out_r = __lasx_xvpackev_h(temp1, temp0);                           \
+    u_l = __lasx_xvdp2_w_h(in_uv, ugvg);                               \
+    temp0 = __lasx_xvsub_w(yl_ev, u_l);                                \
+    temp1 = __lasx_xvsub_w(yl_od, u_l);                                \
+    DUP2_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp0, temp1);      \
+    DUP2_ARG1(__lasx_xvclip255_w, temp0, temp1, temp0, temp1);         \
+    out_g = __lasx_xvpackev_h(temp1, temp0);                           \
+  }
+
+// Pack and Store 16 ARGB values.
+#define STOREARGB_D(a_l, a_h, r_l, r_h, g_l, g_h, b_l, b_h, pdst_argb) \
+  {                                                                    \
+    __m256i temp0, temp1, temp2, temp3;                                \
+                                                                       \
+    temp0 = __lasx_xvpackev_b(g_l, b_l);                               \
+    temp1 = __lasx_xvpackev_b(a_l, r_l);                               \
+    temp2 = __lasx_xvpackev_b(g_h, b_h);                               \
+    temp3 = __lasx_xvpackev_b(a_h, r_h);                               \
+    r_l = __lasx_xvilvl_h(temp1, temp0);                               \
+    r_h = __lasx_xvilvh_h(temp1, temp0);                               \
+    g_l = __lasx_xvilvl_h(temp3, temp2);                               \
+    g_h = __lasx_xvilvh_h(temp3, temp2);                               \
+    temp0 = __lasx_xvpermi_q(r_h, r_l, 0x20);                          \
+    temp1 = __lasx_xvpermi_q(g_h, g_l, 0x20);                          \
+    temp2 = __lasx_xvpermi_q(r_h, r_l, 0x31);                          \
+    temp3 = __lasx_xvpermi_q(g_h, g_l, 0x31);                          \
+    __lasx_xvst(temp0, pdst_argb, 0);                                  \
+    __lasx_xvst(temp1, pdst_argb, 32);                                 \
+    __lasx_xvst(temp2, pdst_argb, 64);                                 \
+    __lasx_xvst(temp3, pdst_argb, 96);                                 \
+    pdst_argb += 128;                                                  \
+  }
+
+// Pack and Store 8 ARGB values.
+#define STOREARGB(in_a, in_r, in_g, in_b, pdst_argb) \
+  {                                                  \
+    __m256i temp0, temp1, temp2, temp3;              \
+                                                     \
+    temp0 = __lasx_xvpackev_b(in_g, in_b);           \
+    temp1 = __lasx_xvpackev_b(in_a, in_r);           \
+    temp2 = __lasx_xvilvl_h(temp1, temp0);           \
+    temp3 = __lasx_xvilvh_h(temp1, temp0);           \
+    temp0 = __lasx_xvpermi_q(temp3, temp2, 0x20);    \
+    temp1 = __lasx_xvpermi_q(temp3, temp2, 0x31);    \
+    __lasx_xvst(temp0, pdst_argb, 0);                \
+    __lasx_xvst(temp1, pdst_argb, 32);               \
+    pdst_argb += 64;                                 \
+  }
+
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _reg0, _reg1) \
+  {                                                                     \
+    __m256i _tmp0, _tmp1, _tmp2, _tmp3;                                 \
+    _tmp0 = __lasx_xvaddwev_h_bu(_tmpb, _nexb);                         \
+    _tmp1 = __lasx_xvaddwod_h_bu(_tmpb, _nexb);                         \
+    _tmp2 = __lasx_xvaddwev_h_bu(_tmpg, _nexg);                         \
+    _tmp3 = __lasx_xvaddwod_h_bu(_tmpg, _nexg);                         \
+    _reg0 = __lasx_xvaddwev_h_bu(_tmpr, _nexr);                         \
+    _reg1 = __lasx_xvaddwod_h_bu(_tmpr, _nexr);                         \
+    _tmpb = __lasx_xvavgr_hu(_tmp0, _tmp1);                             \
+    _tmpg = __lasx_xvavgr_hu(_tmp2, _tmp3);                             \
+    _tmpr = __lasx_xvavgr_hu(_reg0, _reg1);                             \
+    _reg0 = __lasx_xvmadd_h(const_8080, const_112, _tmpb);              \
+    _reg1 = __lasx_xvmadd_h(const_8080, const_112, _tmpr);              \
+    _reg0 = __lasx_xvmsub_h(_reg0, const_74, _tmpg);                    \
+    _reg1 = __lasx_xvmsub_h(_reg1, const_94, _tmpg);                    \
+    _reg0 = __lasx_xvmsub_h(_reg0, const_38, _tmpr);                    \
+    _reg1 = __lasx_xvmsub_h(_reg1, const_18, _tmpb);                    \
+  }
+
+void MirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width) {
+  int x;
+  int len = width / 64;
+  __m256i src0, src1;
+  __m256i shuffler = {0x08090A0B0C0D0E0F, 0x0001020304050607,
+                      0x08090A0B0C0D0E0F, 0x0001020304050607};
+  src += width - 64;
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src, 0, src, 32, src0, src1);
+    DUP2_ARG3(__lasx_xvshuf_b, src0, src0, shuffler, src1, src1, shuffler, src0,
+              src1);
+    src0 = __lasx_xvpermi_q(src0, src0, 0x01);
+    src1 = __lasx_xvpermi_q(src1, src1, 0x01);
+    __lasx_xvst(src1, dst, 0);
+    __lasx_xvst(src0, dst, 32);
+    dst += 64;
+    src -= 64;
+  }
+}
+
+void MirrorUVRow_LASX(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src, dst;
+  __m256i shuffler = {0x0004000500060007, 0x0000000100020003,
+                      0x0004000500060007, 0x0000000100020003};
+
+  src_uv += (width - 16) << 1;
+  for (x = 0; x < len; x++) {
+    src = __lasx_xvld(src_uv, 0);
+    dst = __lasx_xvshuf_h(shuffler, src, src);
+    dst = __lasx_xvpermi_q(dst, dst, 0x01);
+    __lasx_xvst(dst, dst_uv, 0);
+    src_uv -= 32;
+    dst_uv += 32;
+  }
+}
+
+void ARGBMirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1;
+  __m256i dst0, dst1;
+  __m256i shuffler = {0x0B0A09080F0E0D0C, 0x0302010007060504,
+                      0x0B0A09080F0E0D0C, 0x0302010007060504};
+  src += (width * 4) - 64;
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src, 0, src, 32, src0, src1);
+    DUP2_ARG3(__lasx_xvshuf_b, src0, src0, shuffler, src1, src1, shuffler, src0,
+              src1);
+    dst1 = __lasx_xvpermi_q(src0, src0, 0x01);
+    dst0 = __lasx_xvpermi_q(src1, src1, 0x01);
+    __lasx_xvst(dst0, dst, 0);
+    __lasx_xvst(dst1, dst, 32);
+    dst += 64;
+    src -= 64;
+  }
+}
+
+void I422ToYUY2Row_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i src_u0, src_v0, src_y0, vec_uv0;
+  __m256i vec_yuy2_0, vec_yuy2_1;
+  __m256i dst_yuy2_0, dst_yuy2_1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_u, 0, src_v, 0, src_u0, src_v0);
+    src_y0 = __lasx_xvld(src_y, 0);
+    src_u0 = __lasx_xvpermi_d(src_u0, 0xD8);
+    src_v0 = __lasx_xvpermi_d(src_v0, 0xD8);
+    vec_uv0 = __lasx_xvilvl_b(src_v0, src_u0);
+    vec_yuy2_0 = __lasx_xvilvl_b(vec_uv0, src_y0);
+    vec_yuy2_1 = __lasx_xvilvh_b(vec_uv0, src_y0);
+    dst_yuy2_0 = __lasx_xvpermi_q(vec_yuy2_1, vec_yuy2_0, 0x20);
+    dst_yuy2_1 = __lasx_xvpermi_q(vec_yuy2_1, vec_yuy2_0, 0x31);
+    __lasx_xvst(dst_yuy2_0, dst_yuy2, 0);
+    __lasx_xvst(dst_yuy2_1, dst_yuy2, 32);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_yuy2 += 64;
+  }
+}
+
+void I422ToUYVYRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i src_u0, src_v0, src_y0, vec_uv0;
+  __m256i vec_uyvy0, vec_uyvy1;
+  __m256i dst_uyvy0, dst_uyvy1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_u, 0, src_v, 0, src_u0, src_v0);
+    src_y0 = __lasx_xvld(src_y, 0);
+    src_u0 = __lasx_xvpermi_d(src_u0, 0xD8);
+    src_v0 = __lasx_xvpermi_d(src_v0, 0xD8);
+    vec_uv0 = __lasx_xvilvl_b(src_v0, src_u0);
+    vec_uyvy0 = __lasx_xvilvl_b(src_y0, vec_uv0);
+    vec_uyvy1 = __lasx_xvilvh_b(src_y0, vec_uv0);
+    dst_uyvy0 = __lasx_xvpermi_q(vec_uyvy1, vec_uyvy0, 0x20);
+    dst_uyvy1 = __lasx_xvpermi_q(vec_uyvy1, vec_uyvy0, 0x31);
+    __lasx_xvst(dst_uyvy0, dst_uyvy, 0);
+    __lasx_xvst(dst_uyvy1, dst_uyvy, 32);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_uyvy += 64;
+  }
+}
+
+void I422ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_ug, vec_vr, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    STOREARGB_D(alpha, alpha, r_l, r_h, g_l, g_h, b_l, b_h, dst_argb);
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I422ToRGBARow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    STOREARGB_D(r_l, r_h, g_l, g_h, b_l, b_h, alpha, alpha, dst_argb);
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I422AlphaToARGBRow_LASX(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) {
+  int x;
+  int len = width / 32;
+  int res = width & 31;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h, a_l, a_h;
+
+    y = __lasx_xvld(src_a, 0);
+    a_l = __lasx_xvilvl_b(zero, y);
+    a_h = __lasx_xvilvh_b(zero, y);
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    STOREARGB_D(a_l, a_h, r_l, r_h, g_l, g_h, b_l, b_h, dst_argb);
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+    src_a += 32;
+  }
+  if (res) {
+    __m256i y, uv, r, g, b, a;
+    a = __lasx_xvld(src_a, 0);
+    a = __lasx_vext2xv_hu_bu(a);
+    READYUV422(src_y, src_u, src_v, y, uv);
+    YUVTORGB(y, uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b, g, r);
+    STOREARGB(a, r, g, b, dst_argb);
+  }
+}
+
+void I422ToRGB24Row_LASX(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int32_t width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i shuffler0 = {0x0504120302100100, 0x0A18090816070614,
+                       0x0504120302100100, 0x0A18090816070614};
+  __m256i shuffler1 = {0x1E0F0E1C0D0C1A0B, 0x1E0F0E1C0D0C1A0B,
+                       0x1E0F0E1C0D0C1A0B, 0x1E0F0E1C0D0C1A0B};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i temp0, temp1, temp2, temp3;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    temp0 = __lasx_xvpackev_b(g_l, b_l);
+    temp1 = __lasx_xvpackev_b(g_h, b_h);
+    DUP4_ARG3(__lasx_xvshuf_b, r_l, temp0, shuffler1, r_h, temp1, shuffler1,
+              r_l, temp0, shuffler0, r_h, temp1, shuffler0, temp2, temp3, temp0,
+              temp1);
+
+    b_l = __lasx_xvilvl_d(temp1, temp2);
+    b_h = __lasx_xvilvh_d(temp3, temp1);
+    temp1 = __lasx_xvpermi_q(b_l, temp0, 0x20);
+    temp2 = __lasx_xvpermi_q(temp0, b_h, 0x30);
+    temp3 = __lasx_xvpermi_q(b_h, b_l, 0x31);
+    __lasx_xvst(temp1, dst_argb, 0);
+    __lasx_xvst(temp2, dst_argb, 32);
+    __lasx_xvst(temp3, dst_argb, 64);
+    dst_argb += 96;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+// TODO(fbarchard): Consider AND instead of shift to isolate 5 upper bits of R.
+void I422ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i dst_l, dst_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    b_l = __lasx_xvsrli_h(b_l, 3);
+    b_h = __lasx_xvsrli_h(b_h, 3);
+    g_l = __lasx_xvsrli_h(g_l, 2);
+    g_h = __lasx_xvsrli_h(g_h, 2);
+    r_l = __lasx_xvsrli_h(r_l, 3);
+    r_h = __lasx_xvsrli_h(r_h, 3);
+    r_l = __lasx_xvslli_h(r_l, 11);
+    r_h = __lasx_xvslli_h(r_h, 11);
+    g_l = __lasx_xvslli_h(g_l, 5);
+    g_h = __lasx_xvslli_h(g_h, 5);
+    r_l = __lasx_xvor_v(r_l, g_l);
+    r_l = __lasx_xvor_v(r_l, b_l);
+    r_h = __lasx_xvor_v(r_h, g_h);
+    r_h = __lasx_xvor_v(r_h, b_h);
+    dst_l = __lasx_xvpermi_q(r_h, r_l, 0x20);
+    dst_h = __lasx_xvpermi_q(r_h, r_l, 0x31);
+    __lasx_xvst(dst_l, dst_rgb565, 0);
+    __lasx_xvst(dst_h, dst_rgb565, 32);
+    dst_rgb565 += 64;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+// TODO(fbarchard): Consider AND instead of shift to isolate 4 upper bits of G.
+void I422ToARGB4444Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = {0xF000F000F000F000, 0xF000F000F000F000, 0xF000F000F000F000,
+                   0xF000F000F000F000};
+  __m256i mask = {0x00F000F000F000F0, 0x00F000F000F000F0, 0x00F000F000F000F0,
+                  0x00F000F000F000F0};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i dst_l, dst_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    b_l = __lasx_xvsrli_h(b_l, 4);
+    b_h = __lasx_xvsrli_h(b_h, 4);
+    r_l = __lasx_xvsrli_h(r_l, 4);
+    r_h = __lasx_xvsrli_h(r_h, 4);
+    g_l = __lasx_xvand_v(g_l, mask);
+    g_h = __lasx_xvand_v(g_h, mask);
+    r_l = __lasx_xvslli_h(r_l, 8);
+    r_h = __lasx_xvslli_h(r_h, 8);
+    r_l = __lasx_xvor_v(r_l, alpha);
+    r_h = __lasx_xvor_v(r_h, alpha);
+    r_l = __lasx_xvor_v(r_l, g_l);
+    r_h = __lasx_xvor_v(r_h, g_h);
+    r_l = __lasx_xvor_v(r_l, b_l);
+    r_h = __lasx_xvor_v(r_h, b_h);
+    dst_l = __lasx_xvpermi_q(r_h, r_l, 0x20);
+    dst_h = __lasx_xvpermi_q(r_h, r_l, 0x31);
+    __lasx_xvst(dst_l, dst_argb4444, 0);
+    __lasx_xvst(dst_h, dst_argb4444, 32);
+    dst_argb4444 += 64;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I422ToARGB1555Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = {0x8000800080008000, 0x8000800080008000, 0x8000800080008000,
+                   0x8000800080008000};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i dst_l, dst_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    b_l = __lasx_xvsrli_h(b_l, 3);
+    b_h = __lasx_xvsrli_h(b_h, 3);
+    g_l = __lasx_xvsrli_h(g_l, 3);
+    g_h = __lasx_xvsrli_h(g_h, 3);
+    g_l = __lasx_xvslli_h(g_l, 5);
+    g_h = __lasx_xvslli_h(g_h, 5);
+    r_l = __lasx_xvsrli_h(r_l, 3);
+    r_h = __lasx_xvsrli_h(r_h, 3);
+    r_l = __lasx_xvslli_h(r_l, 10);
+    r_h = __lasx_xvslli_h(r_h, 10);
+    r_l = __lasx_xvor_v(r_l, alpha);
+    r_h = __lasx_xvor_v(r_h, alpha);
+    r_l = __lasx_xvor_v(r_l, g_l);
+    r_h = __lasx_xvor_v(r_h, g_h);
+    r_l = __lasx_xvor_v(r_l, b_l);
+    r_h = __lasx_xvor_v(r_h, b_h);
+    dst_l = __lasx_xvpermi_q(r_h, r_l, 0x20);
+    dst_h = __lasx_xvpermi_q(r_h, r_l, 0x31);
+    __lasx_xvst(dst_l, dst_argb1555, 0);
+    __lasx_xvst(dst_h, dst_argb1555, 32);
+    dst_argb1555 += 64;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void YUY2ToYRow_LASX(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_yuy2, 0, src_yuy2, 32, src0, src1);
+    dst0 = __lasx_xvpickev_b(src1, src0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_yuy2 += 64;
+    dst_y += 32;
+  }
+}
+
+void YUY2ToUVRow_LASX(const uint8_t* src_yuy2,
+                      int src_stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_yuy2_next = src_yuy2 + src_stride_yuy2;
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_yuy2, 0, src_yuy2, 32, src_yuy2_next, 0,
+              src_yuy2_next, 32, src0, src1, src2, src3);
+    src0 = __lasx_xvpickod_b(src1, src0);
+    src1 = __lasx_xvpickod_b(src3, src2);
+    tmp0 = __lasx_xvavgr_bu(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_yuy2 += 64;
+    src_yuy2_next += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void YUY2ToUV422Row_LASX(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_yuy2, 0, src_yuy2, 32, src0, src1);
+    tmp0 = __lasx_xvpickod_b(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_yuy2 += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void UYVYToYRow_LASX(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_uyvy, 0, src_uyvy, 32, src0, src1);
+    dst0 = __lasx_xvpickod_b(src1, src0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_uyvy += 64;
+    dst_y += 32;
+  }
+}
+
+void UYVYToUVRow_LASX(const uint8_t* src_uyvy,
+                      int src_stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_uyvy_next = src_uyvy + src_stride_uyvy;
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3, tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_uyvy, 0, src_uyvy, 32, src_uyvy_next, 0,
+              src_uyvy_next, 32, src0, src1, src2, src3);
+    src0 = __lasx_xvpickev_b(src1, src0);
+    src1 = __lasx_xvpickev_b(src3, src2);
+    tmp0 = __lasx_xvavgr_bu(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_uyvy += 64;
+    src_uyvy_next += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void UYVYToUV422Row_LASX(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_uyvy, 0, src_uyvy, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_uyvy += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBToUVRow_LASX(const uint8_t* src_argb0,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  int len = width / 32;
+  const uint8_t* src_argb1 = src_argb0 + src_stride_argb;
+
+  __m256i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m256i vec0, vec1, vec2, vec3;
+  __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, dst0, dst1;
+  __m256i const_0x70 = {0x0038003800380038, 0x0038003800380038,
+                        0x0038003800380038, 0x0038003800380038};
+  __m256i const_0x4A = {0x0025002500250025, 0x0025002500250025,
+                        0x0025002500250025, 0x0025002500250025};
+  __m256i const_0x26 = {0x0013001300130013, 0x0013001300130013,
+                        0x0013001300130013, 0x0013001300130013};
+  __m256i const_0x5E = {0x002f002f002f002f, 0x002f002f002f002f,
+                        0x002f002f002f002f, 0x002f002f002f002f};
+  __m256i const_0x12 = {0x0009000900090009, 0x0009000900090009,
+                        0x0009000900090009, 0x0009000900090009};
+  __m256i control = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+                     0x0000000700000003};
+  __m256i const_0x8080 = {0x8080808080808080, 0x8080808080808080,
+                          0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb0, 0, src_argb0, 32, src_argb0, 64,
+              src_argb0, 96, src0, src1, src2, src3);
+    DUP4_ARG2(__lasx_xvld, src_argb1, 0, src_argb1, 32, src_argb1, 64,
+              src_argb1, 96, src4, src5, src6, src7);
+    vec0 = __lasx_xvaddwev_h_bu(src0, src4);
+    vec1 = __lasx_xvaddwev_h_bu(src1, src5);
+    vec2 = __lasx_xvaddwev_h_bu(src2, src6);
+    vec3 = __lasx_xvaddwev_h_bu(src3, src7);
+    tmp0 = __lasx_xvpickev_h(vec1, vec0);
+    tmp1 = __lasx_xvpickev_h(vec3, vec2);
+    tmp2 = __lasx_xvpickod_h(vec1, vec0);
+    tmp3 = __lasx_xvpickod_h(vec3, vec2);
+    vec0 = __lasx_xvaddwod_h_bu(src0, src4);
+    vec1 = __lasx_xvaddwod_h_bu(src1, src5);
+    vec2 = __lasx_xvaddwod_h_bu(src2, src6);
+    vec3 = __lasx_xvaddwod_h_bu(src3, src7);
+    tmp4 = __lasx_xvpickev_h(vec1, vec0);
+    tmp5 = __lasx_xvpickev_h(vec3, vec2);
+    vec0 = __lasx_xvpickev_h(tmp1, tmp0);
+    vec1 = __lasx_xvpickod_h(tmp1, tmp0);
+    src0 = __lasx_xvavgr_h(vec0, vec1);
+    vec0 = __lasx_xvpickev_h(tmp3, tmp2);
+    vec1 = __lasx_xvpickod_h(tmp3, tmp2);
+    src1 = __lasx_xvavgr_h(vec0, vec1);
+    vec0 = __lasx_xvpickev_h(tmp5, tmp4);
+    vec1 = __lasx_xvpickod_h(tmp5, tmp4);
+    src2 = __lasx_xvavgr_h(vec0, vec1);
+    dst0 = __lasx_xvmadd_h(const_0x8080, src0, const_0x70);
+    dst0 = __lasx_xvmsub_h(dst0, src2, const_0x4A);
+    dst0 = __lasx_xvmsub_h(dst0, src1, const_0x26);
+    dst1 = __lasx_xvmadd_h(const_0x8080, src1, const_0x70);
+    dst1 = __lasx_xvmsub_h(dst1, src2, const_0x5E);
+    dst1 = __lasx_xvmsub_h(dst1, src0, const_0x12);
+    dst0 = __lasx_xvperm_w(dst0, control);
+    dst1 = __lasx_xvperm_w(dst1, control);
+    dst0 = __lasx_xvssrani_b_h(dst0, dst0, 8);
+    dst1 = __lasx_xvssrani_b_h(dst1, dst1, 8);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_argb0 += 128;
+    src_argb1 += 128;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBToRGB24Row_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  int len = (width / 32) - 1;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i shuf = {0x0908060504020100, 0x000000000E0D0C0A, 0x0908060504020100,
+                  0x000000000E0D0C0A};
+  __m256i control = {0x0000000100000000, 0x0000000400000002, 0x0000000600000005,
+                     0x0000000700000003};
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+    tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+    tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+    tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+    tmp0 = __lasx_xvperm_w(tmp0, control);
+    tmp1 = __lasx_xvperm_w(tmp1, control);
+    tmp2 = __lasx_xvperm_w(tmp2, control);
+    tmp3 = __lasx_xvperm_w(tmp3, control);
+    __lasx_xvst(tmp0, dst_rgb, 0);
+    __lasx_xvst(tmp1, dst_rgb, 24);
+    __lasx_xvst(tmp2, dst_rgb, 48);
+    __lasx_xvst(tmp3, dst_rgb, 72);
+    dst_rgb += 96;
+    src_argb += 128;
+  }
+  DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb, 96,
+            src0, src1, src2, src3);
+  tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+  tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+  tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+  tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+  tmp0 = __lasx_xvperm_w(tmp0, control);
+  tmp1 = __lasx_xvperm_w(tmp1, control);
+  tmp2 = __lasx_xvperm_w(tmp2, control);
+  tmp3 = __lasx_xvperm_w(tmp3, control);
+  __lasx_xvst(tmp0, dst_rgb, 0);
+  __lasx_xvst(tmp1, dst_rgb, 24);
+  __lasx_xvst(tmp2, dst_rgb, 48);
+  dst_rgb += 72;
+  __lasx_xvstelm_d(tmp3, dst_rgb, 0, 0);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 8, 1);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 16, 2);
+}
+
+void ARGBToRAWRow_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  int len = (width / 32) - 1;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i shuf = {0x090A040506000102, 0x000000000C0D0E08, 0x090A040506000102,
+                  0x000000000C0D0E08};
+  __m256i control = {0x0000000100000000, 0x0000000400000002, 0x0000000600000005,
+                     0x0000000700000003};
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+    tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+    tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+    tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+    tmp0 = __lasx_xvperm_w(tmp0, control);
+    tmp1 = __lasx_xvperm_w(tmp1, control);
+    tmp2 = __lasx_xvperm_w(tmp2, control);
+    tmp3 = __lasx_xvperm_w(tmp3, control);
+    __lasx_xvst(tmp0, dst_rgb, 0);
+    __lasx_xvst(tmp1, dst_rgb, 24);
+    __lasx_xvst(tmp2, dst_rgb, 48);
+    __lasx_xvst(tmp3, dst_rgb, 72);
+    dst_rgb += 96;
+    src_argb += 128;
+  }
+  DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb, 96,
+            src0, src1, src2, src3);
+  tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+  tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+  tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+  tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+  tmp0 = __lasx_xvperm_w(tmp0, control);
+  tmp1 = __lasx_xvperm_w(tmp1, control);
+  tmp2 = __lasx_xvperm_w(tmp2, control);
+  tmp3 = __lasx_xvperm_w(tmp3, control);
+  __lasx_xvst(tmp0, dst_rgb, 0);
+  __lasx_xvst(tmp1, dst_rgb, 24);
+  __lasx_xvst(tmp2, dst_rgb, 48);
+  dst_rgb += 72;
+  __lasx_xvstelm_d(tmp3, dst_rgb, 0, 0);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 8, 1);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 16, 2);
+}
+
+void ARGBToRGB565Row_LASX(const uint8_t* src_argb,
+                          uint8_t* dst_rgb,
+                          int width) {
+  int x;
+  int len = width / 16;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i src0, src1, tmp0, tmp1, dst0;
+  __m256i shift = {0x0300030003000300, 0x0300030003000300, 0x0300030003000300,
+                   0x0300030003000300};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp0 = __lasx_xvsrli_b(tmp0, 3);
+    tmp1 = __lasx_xvpackev_b(zero, tmp1);
+    tmp1 = __lasx_xvsrli_h(tmp1, 2);
+    tmp0 = __lasx_xvsll_b(tmp0, shift);
+    tmp1 = __lasx_xvslli_h(tmp1, 5);
+    dst0 = __lasx_xvor_v(tmp0, tmp1);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    dst_rgb += 32;
+    src_argb += 64;
+  }
+}
+
+void ARGBToARGB1555Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width) {
+  int x;
+  int len = width / 16;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i src0, src1, tmp0, tmp1, tmp2, tmp3, dst0;
+  __m256i shift1 = {0x0703070307030703, 0x0703070307030703, 0x0703070307030703,
+                    0x0703070307030703};
+  __m256i shift2 = {0x0200020002000200, 0x0200020002000200, 0x0200020002000200,
+                    0x0200020002000200};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp0 = __lasx_xvsrli_b(tmp0, 3);
+    tmp1 = __lasx_xvsrl_b(tmp1, shift1);
+    tmp0 = __lasx_xvsll_b(tmp0, shift2);
+    tmp2 = __lasx_xvpackev_b(zero, tmp1);
+    tmp3 = __lasx_xvpackod_b(zero, tmp1);
+    tmp2 = __lasx_xvslli_h(tmp2, 5);
+    tmp3 = __lasx_xvslli_h(tmp3, 15);
+    dst0 = __lasx_xvor_v(tmp0, tmp2);
+    dst0 = __lasx_xvor_v(dst0, tmp3);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    dst_rgb += 32;
+    src_argb += 64;
+  }
+}
+
+void ARGBToARGB4444Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp1 = __lasx_xvandi_b(tmp1, 0xF0);
+    tmp0 = __lasx_xvsrli_b(tmp0, 4);
+    dst0 = __lasx_xvor_v(tmp1, tmp0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    dst_rgb += 32;
+    src_argb += 64;
+  }
+}
+
+void ARGBToUV444Row_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int32_t width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, reg2, reg3, dst0, dst1;
+  __m256i const_112 = __lasx_xvldi(112);
+  __m256i const_74 = __lasx_xvldi(74);
+  __m256i const_38 = __lasx_xvldi(38);
+  __m256i const_94 = __lasx_xvldi(94);
+  __m256i const_18 = __lasx_xvldi(18);
+  __m256i const_0x8080 = {0x8080808080808080, 0x8080808080808080,
+                          0x8080808080808080, 0x8080808080808080};
+  __m256i control = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+                     0x0000000700000003};
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvpickev_h(src1, src0);
+    tmp1 = __lasx_xvpickod_h(src1, src0);
+    tmp2 = __lasx_xvpickev_h(src3, src2);
+    tmp3 = __lasx_xvpickod_h(src3, src2);
+    reg0 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp0, const_112);
+    reg1 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp2, const_112);
+    reg2 = __lasx_xvmulwod_h_bu(tmp0, const_74);
+    reg3 = __lasx_xvmulwod_h_bu(tmp2, const_74);
+    reg2 = __lasx_xvmaddwev_h_bu(reg2, tmp1, const_38);
+    reg3 = __lasx_xvmaddwev_h_bu(reg3, tmp3, const_38);
+    reg0 = __lasx_xvsub_h(reg0, reg2);
+    reg1 = __lasx_xvsub_h(reg1, reg3);
+    dst0 = __lasx_xvssrani_b_h(reg1, reg0, 8);
+    dst0 = __lasx_xvperm_w(dst0, control);
+    reg0 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp1, const_112);
+    reg1 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp3, const_112);
+    reg2 = __lasx_xvmulwev_h_bu(tmp0, const_18);
+    reg3 = __lasx_xvmulwev_h_bu(tmp2, const_18);
+    reg2 = __lasx_xvmaddwod_h_bu(reg2, tmp0, const_94);
+    reg3 = __lasx_xvmaddwod_h_bu(reg3, tmp2, const_94);
+    reg0 = __lasx_xvsub_h(reg0, reg2);
+    reg1 = __lasx_xvsub_h(reg1, reg3);
+    dst1 = __lasx_xvssrani_b_h(reg1, reg0, 8);
+    dst1 = __lasx_xvperm_w(dst1, control);
+    __lasx_xvst(dst0, dst_u, 0);
+    __lasx_xvst(dst1, dst_v, 0);
+    dst_u += 32;
+    dst_v += 32;
+    src_argb += 128;
+  }
+}
+
+void ARGBMultiplyRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  int len = width / 8;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i src0, src1, dst0, dst1;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb0, 0, src_argb1, 0, src0, src1);
+    tmp0 = __lasx_xvilvl_b(src0, src0);
+    tmp1 = __lasx_xvilvh_b(src0, src0);
+    tmp2 = __lasx_xvilvl_b(zero, src1);
+    tmp3 = __lasx_xvilvh_b(zero, src1);
+    dst0 = __lasx_xvmuh_hu(tmp0, tmp2);
+    dst1 = __lasx_xvmuh_hu(tmp1, tmp3);
+    dst0 = __lasx_xvpickev_b(dst1, dst0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb0 += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBAddRow_LASX(const uint8_t* src_argb0,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width) {
+  int x;
+  int len = width / 8;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb0, 0, src_argb1, 0, src0, src1);
+    dst0 = __lasx_xvsadd_bu(src0, src1);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb0 += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBSubtractRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  int len = width / 8;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb0, 0, src_argb1, 0, src0, src1);
+    dst0 = __lasx_xvssub_bu(src0, src1);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb0 += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBAttenuateRow_LASX(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1;
+  __m256i reg0, reg1, reg2, reg3, reg4, reg5;
+  __m256i b, g, r, a, dst0, dst1;
+  __m256i control = {0x0005000100040000, 0x0007000300060002, 0x0005000100040000,
+                     0x0007000300060002};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    b = __lasx_xvpackev_b(tmp0, tmp0);
+    r = __lasx_xvpackod_b(tmp0, tmp0);
+    g = __lasx_xvpackev_b(tmp1, tmp1);
+    a = __lasx_xvpackod_b(tmp1, tmp1);
+    reg0 = __lasx_xvmulwev_w_hu(b, a);
+    reg1 = __lasx_xvmulwod_w_hu(b, a);
+    reg2 = __lasx_xvmulwev_w_hu(r, a);
+    reg3 = __lasx_xvmulwod_w_hu(r, a);
+    reg4 = __lasx_xvmulwev_w_hu(g, a);
+    reg5 = __lasx_xvmulwod_w_hu(g, a);
+    reg0 = __lasx_xvssrani_h_w(reg1, reg0, 24);
+    reg2 = __lasx_xvssrani_h_w(reg3, reg2, 24);
+    reg4 = __lasx_xvssrani_h_w(reg5, reg4, 24);
+    reg0 = __lasx_xvshuf_h(control, reg0, reg0);
+    reg2 = __lasx_xvshuf_h(control, reg2, reg2);
+    reg4 = __lasx_xvshuf_h(control, reg4, reg4);
+    tmp0 = __lasx_xvpackev_b(reg4, reg0);
+    tmp1 = __lasx_xvpackev_b(a, reg2);
+    dst0 = __lasx_xvilvl_h(tmp1, tmp0);
+    dst1 = __lasx_xvilvh_h(tmp1, tmp0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    dst_argb += 64;
+    src_argb += 64;
+  }
+}
+
+void ARGBToRGB565DitherRow_LASX(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1, dst0;
+  __m256i b, g, r;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i vec_dither = __lasx_xvldrepl_w(&dither4, 0);
+
+  vec_dither = __lasx_xvilvl_b(zero, vec_dither);
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    b = __lasx_xvpackev_b(zero, tmp0);
+    r = __lasx_xvpackod_b(zero, tmp0);
+    g = __lasx_xvpackev_b(zero, tmp1);
+    b = __lasx_xvadd_h(b, vec_dither);
+    g = __lasx_xvadd_h(g, vec_dither);
+    r = __lasx_xvadd_h(r, vec_dither);
+    DUP2_ARG1(__lasx_xvclip255_h, b, g, b, g);
+    r = __lasx_xvclip255_h(r);
+    b = __lasx_xvsrai_h(b, 3);
+    g = __lasx_xvsrai_h(g, 2);
+    r = __lasx_xvsrai_h(r, 3);
+    g = __lasx_xvslli_h(g, 5);
+    r = __lasx_xvslli_h(r, 11);
+    dst0 = __lasx_xvor_v(b, g);
+    dst0 = __lasx_xvor_v(dst0, r);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    src_argb += 64;
+    dst_rgb += 32;
+  }
+}
+
+void ARGBShuffleRow_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, dst0, dst1;
+  __m256i shuf = {0x0404040400000000, 0x0C0C0C0C08080808, 0x0404040400000000,
+                  0x0C0C0C0C08080808};
+  __m256i temp = __lasx_xvldrepl_w(shuffler, 0);
+
+  shuf = __lasx_xvadd_b(shuf, temp);
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    dst0 = __lasx_xvshuf_b(src0, src0, shuf);
+    dst1 = __lasx_xvshuf_b(src1, src1, shuf);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    src_argb += 64;
+    dst_argb += 64;
+  }
+}
+
+void ARGBShadeRow_LASX(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value) {
+  int x;
+  int len = width / 8;
+  __m256i src0, dst0, tmp0, tmp1;
+  __m256i vec_value = __lasx_xvreplgr2vr_w(value);
+
+  vec_value = __lasx_xvilvl_b(vec_value, vec_value);
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb, 0);
+    tmp0 = __lasx_xvilvl_b(src0, src0);
+    tmp1 = __lasx_xvilvh_b(src0, src0);
+    tmp0 = __lasx_xvmuh_hu(tmp0, vec_value);
+    tmp1 = __lasx_xvmuh_hu(tmp1, vec_value);
+    dst0 = __lasx_xvpickod_b(tmp1, tmp0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBGrayRow_LASX(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1;
+  __m256i reg0, reg1, reg2, dst0, dst1;
+  __m256i const_128 = __lasx_xvldi(0x480);
+  __m256i const_150 = __lasx_xvldi(0x96);
+  __m256i const_br = {0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D,
+                      0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    reg0 = __lasx_xvdp2_h_bu(tmp0, const_br);
+    reg1 = __lasx_xvmaddwev_h_bu(const_128, tmp1, const_150);
+    reg2 = __lasx_xvadd_h(reg0, reg1);
+    tmp0 = __lasx_xvpackod_b(reg2, reg2);
+    tmp1 = __lasx_xvpackod_b(tmp1, reg2);
+    dst0 = __lasx_xvilvl_h(tmp1, tmp0);
+    dst1 = __lasx_xvilvh_h(tmp1, tmp0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    src_argb += 64;
+    dst_argb += 64;
+  }
+}
+
+void ARGBSepiaRow_LASX(uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1;
+  __m256i reg0, reg1, spb, spg, spr;
+  __m256i dst0, dst1;
+  __m256i spb_g = __lasx_xvldi(68);
+  __m256i spg_g = __lasx_xvldi(88);
+  __m256i spr_g = __lasx_xvldi(98);
+  __m256i spb_br = {0x2311231123112311, 0x2311231123112311, 0x2311231123112311,
+                    0x2311231123112311};
+  __m256i spg_br = {0x2D162D162D162D16, 0x2D162D162D162D16, 0x2D162D162D162D16,
+                    0x2D162D162D162D16};
+  __m256i spr_br = {0x3218321832183218, 0x3218321832183218, 0x3218321832183218,
+                    0x3218321832183218};
+  __m256i shuff = {0x1706150413021100, 0x1F0E1D0C1B0A1908, 0x1706150413021100,
+                   0x1F0E1D0C1B0A1908};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, dst_argb, 0, dst_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    DUP2_ARG2(__lasx_xvdp2_h_bu, tmp0, spb_br, tmp0, spg_br, spb, spg);
+    spr = __lasx_xvdp2_h_bu(tmp0, spr_br);
+    spb = __lasx_xvmaddwev_h_bu(spb, tmp1, spb_g);
+    spg = __lasx_xvmaddwev_h_bu(spg, tmp1, spg_g);
+    spr = __lasx_xvmaddwev_h_bu(spr, tmp1, spr_g);
+    spb = __lasx_xvsrli_h(spb, 7);
+    spg = __lasx_xvsrli_h(spg, 7);
+    spr = __lasx_xvsrli_h(spr, 7);
+    spg = __lasx_xvsat_hu(spg, 7);
+    spr = __lasx_xvsat_hu(spr, 7);
+    reg0 = __lasx_xvpackev_b(spg, spb);
+    reg1 = __lasx_xvshuf_b(tmp1, spr, shuff);
+    dst0 = __lasx_xvilvl_h(reg1, reg0);
+    dst1 = __lasx_xvilvh_h(reg1, reg0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    dst_argb += 64;
+  }
+}
+
+void ARGB4444ToARGBRow_LASX(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, reg2, reg3;
+  __m256i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb4444, 0);
+    src1 = __lasx_xvld(src_argb4444, 32);
+    DUP4_ARG2(__lasx_xvandi_b, src0, 0x0F, src0, 0xF0, src1, 0x0F, src1, 0xF0,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG2(__lasx_xvslli_b, tmp0, 4, tmp2, 4, reg0, reg2);
+    DUP2_ARG2(__lasx_xvsrli_b, tmp1, 4, tmp3, 4, reg1, reg3);
+    DUP4_ARG2(__lasx_xvor_v, tmp0, reg0, tmp1, reg1, tmp2, reg2, tmp3, reg3,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG2(__lasx_xvilvl_b, tmp1, tmp0, tmp3, tmp2, reg0, reg2);
+    DUP2_ARG2(__lasx_xvilvh_b, tmp1, tmp0, tmp3, tmp2, reg1, reg3);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x20, reg1, reg0, 0x31, reg3, reg2,
+              0x20, reg3, reg2, 0x31, dst0, dst1, dst2, dst3);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    __lasx_xvst(dst2, dst_argb, 64);
+    __lasx_xvst(dst3, dst_argb, 96);
+    src_argb4444 += 64;
+    dst_argb += 128;
+  }
+}
+
+void ARGB1555ToARGBRow_LASX(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr, tmpa;
+  __m256i reg0, reg1, reg2, reg3;
+  __m256i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb1555, 0);
+    src1 = __lasx_xvld(src_argb1555, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpg = __lasx_xvsrli_b(tmp0, 5);
+    reg0 = __lasx_xvandi_b(tmp1, 0x03);
+    reg0 = __lasx_xvslli_b(reg0, 3);
+    tmpg = __lasx_xvor_v(tmpg, reg0);
+    reg1 = __lasx_xvandi_b(tmp1, 0x7C);
+    tmpr = __lasx_xvsrli_b(reg1, 2);
+    tmpa = __lasx_xvsrli_b(tmp1, 7);
+    tmpa = __lasx_xvneg_b(tmpa);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvslli_b(tmpg, 3);
+    reg2 = __lasx_xvslli_b(tmpr, 3);
+    tmpb = __lasx_xvsrli_b(tmpb, 2);
+    tmpg = __lasx_xvsrli_b(tmpg, 2);
+    tmpr = __lasx_xvsrli_b(tmpr, 2);
+    tmpb = __lasx_xvor_v(reg0, tmpb);
+    tmpg = __lasx_xvor_v(reg1, tmpg);
+    tmpr = __lasx_xvor_v(reg2, tmpr);
+    DUP2_ARG2(__lasx_xvilvl_b, tmpg, tmpb, tmpa, tmpr, reg0, reg1);
+    DUP2_ARG2(__lasx_xvilvh_b, tmpg, tmpb, tmpa, tmpr, reg2, reg3);
+    dst0 = __lasx_xvilvl_h(reg1, reg0);
+    dst1 = __lasx_xvilvh_h(reg1, reg0);
+    dst2 = __lasx_xvilvl_h(reg3, reg2);
+    dst3 = __lasx_xvilvh_h(reg3, reg2);
+    DUP4_ARG3(__lasx_xvpermi_q, dst1, dst0, 0x20, dst1, dst0, 0x31, dst3, dst2,
+              0x20, dst3, dst2, 0x31, reg0, reg1, reg2, reg3);
+    __lasx_xvst(reg0, dst_argb, 0);
+    __lasx_xvst(reg1, dst_argb, 32);
+    __lasx_xvst(reg2, dst_argb, 64);
+    __lasx_xvst(reg3, dst_argb, 96);
+    src_argb1555 += 64;
+    dst_argb += 128;
+  }
+}
+
+void RGB565ToARGBRow_LASX(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m256i reg0, reg1, reg2, reg3, dst0, dst1, dst2, dst3;
+  __m256i alpha = __lasx_xvldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_rgb565, 0);
+    src1 = __lasx_xvld(src_rgb565, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpr = __lasx_xvandi_b(tmp1, 0xF8);
+    reg1 = __lasx_xvandi_b(tmp1, 0x07);
+    reg0 = __lasx_xvsrli_b(tmp0, 5);
+    reg1 = __lasx_xvslli_b(reg1, 3);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvsrli_b(tmpb, 2);
+    tmpb = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpg, 2);
+    reg1 = __lasx_xvsrli_b(tmpg, 4);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvsrli_b(tmpr, 5);
+    tmpr = __lasx_xvor_v(tmpr, reg0);
+    DUP2_ARG2(__lasx_xvilvl_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst0 = __lasx_xvilvl_h(reg1, reg0);
+    dst1 = __lasx_xvilvh_h(reg1, reg0);
+    DUP2_ARG2(__lasx_xvilvh_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst2 = __lasx_xvilvl_h(reg1, reg0);
+    dst3 = __lasx_xvilvh_h(reg1, reg0);
+    DUP4_ARG3(__lasx_xvpermi_q, dst1, dst0, 0x20, dst1, dst0, 0x31, dst3, dst2,
+              0x20, dst3, dst2, 0x31, reg0, reg1, reg2, reg3);
+    __lasx_xvst(reg0, dst_argb, 0);
+    __lasx_xvst(reg1, dst_argb, 32);
+    __lasx_xvst(reg2, dst_argb, 64);
+    __lasx_xvst(reg3, dst_argb, 96);
+    src_rgb565 += 64;
+    dst_argb += 128;
+  }
+}
+
+void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2;
+  __m256i tmp0, tmp1, tmp2;
+  __m256i dst0, dst1, dst2, dst3;
+  __m256i reg0, reg1, reg2, reg3;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514, 0x131211100F0E0D0C,
+                   0x1B1A191817161514};
+  __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100, 0x1F1E1D1C1B1A1918,
+                   0x0706050403020100};
+  __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C, 0x0B0A090807060504,
+                   0x131211100F0E0D0C};
+  __m256i shuf3 = {0x1005040310020100, 0x100B0A0910080706, 0x1005040310020100,
+                   0x100B0A0910080706};
+
+  for (x = 0; x < len; x++) {
+    reg0 = __lasx_xvld(src_rgb24, 0);
+    reg1 = __lasx_xvld(src_rgb24, 32);
+    reg2 = __lasx_xvld(src_rgb24, 64);
+    src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
+    src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
+    src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0,
+              tmp1);
+    tmp2 = __lasx_xvshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lasx_xvshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, reg0, reg1, reg2, reg3);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x20, reg3, reg2, 0x20, reg1, reg0,
+              0x31, reg3, reg2, 0x31, dst0, dst1, dst2, dst3);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    __lasx_xvst(dst2, dst_argb, 64);
+    __lasx_xvst(dst3, dst_argb, 96);
+    src_rgb24 += 96;
+    dst_argb += 128;
+  }
+}
+
+void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2;
+  __m256i tmp0, tmp1, tmp2, reg0, reg1, reg2, reg3;
+  __m256i dst0, dst1, dst2, dst3;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514, 0x131211100F0E0D0C,
+                   0x1B1A191817161514};
+  __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100, 0x1F1E1D1C1B1A1918,
+                   0x0706050403020100};
+  __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C, 0x0B0A090807060504,
+                   0x131211100F0E0D0C};
+  __m256i shuf3 = {0x1003040510000102, 0x10090A0B10060708, 0x1003040510000102,
+                   0x10090A0B10060708};
+
+  for (x = 0; x < len; x++) {
+    reg0 = __lasx_xvld(src_raw, 0);
+    reg1 = __lasx_xvld(src_raw, 32);
+    reg2 = __lasx_xvld(src_raw, 64);
+    src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
+    src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
+    src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0,
+              tmp1);
+    tmp2 = __lasx_xvshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lasx_xvshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, reg0, reg1, reg2, reg3);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x20, reg3, reg2, 0x20, reg1, reg0,
+              0x31, reg3, reg2, 0x31, dst0, dst1, dst2, dst3);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    __lasx_xvst(dst2, dst_argb, 64);
+    __lasx_xvst(dst3, dst_argb, 96);
+    src_raw += 96;
+    dst_argb += 128;
+  }
+}
+
+void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m256i reg0, reg1, reg2, dst0;
+  __m256i const_66 = __lasx_xvldi(66);
+  __m256i const_129 = __lasx_xvldi(129);
+  __m256i const_25 = __lasx_xvldi(25);
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb1555, 0);
+    src1 = __lasx_xvld(src_argb1555, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpg = __lasx_xvsrli_b(tmp0, 5);
+    reg0 = __lasx_xvandi_b(tmp1, 0x03);
+    reg0 = __lasx_xvslli_b(reg0, 3);
+    tmpg = __lasx_xvor_v(tmpg, reg0);
+    reg1 = __lasx_xvandi_b(tmp1, 0x7C);
+    tmpr = __lasx_xvsrli_b(reg1, 2);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvslli_b(tmpg, 3);
+    reg2 = __lasx_xvslli_b(tmpr, 3);
+    tmpb = __lasx_xvsrli_b(tmpb, 2);
+    tmpg = __lasx_xvsrli_b(tmpg, 2);
+    tmpr = __lasx_xvsrli_b(tmpr, 2);
+    tmpb = __lasx_xvor_v(reg0, tmpb);
+    tmpg = __lasx_xvor_v(reg1, tmpg);
+    tmpr = __lasx_xvor_v(reg2, tmpr);
+    reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lasx_xvmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lasx_xvpackod_b(reg1, reg0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_argb1555 += 64;
+    dst_y += 32;
+  }
+}
+
+void ARGB1555ToUVRow_LASX(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  int x;
+  int len = width / 32;
+  const uint8_t* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i reg0, reg1, reg2, reg3, dst0;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb1555, 0, src_argb1555, 32, next_argb1555, 0,
+              next_argb1555, 32, src0, src1, src2, src3);
+    DUP2_ARG2(__lasx_xvpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lasx_xvpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    nexb = __lasx_xvandi_b(tmp2, 0x1F);
+    tmpg = __lasx_xvsrli_b(tmp0, 5);
+    nexg = __lasx_xvsrli_b(tmp2, 5);
+    reg0 = __lasx_xvandi_b(tmp1, 0x03);
+    reg2 = __lasx_xvandi_b(tmp3, 0x03);
+    reg0 = __lasx_xvslli_b(reg0, 3);
+    reg2 = __lasx_xvslli_b(reg2, 3);
+    tmpg = __lasx_xvor_v(tmpg, reg0);
+    nexg = __lasx_xvor_v(nexg, reg2);
+    reg1 = __lasx_xvandi_b(tmp1, 0x7C);
+    reg3 = __lasx_xvandi_b(tmp3, 0x7C);
+    tmpr = __lasx_xvsrli_b(reg1, 2);
+    nexr = __lasx_xvsrli_b(reg3, 2);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvslli_b(tmpg, 3);
+    reg2 = __lasx_xvslli_b(tmpr, 3);
+    tmpb = __lasx_xvsrli_b(tmpb, 2);
+    tmpg = __lasx_xvsrli_b(tmpg, 2);
+    tmpr = __lasx_xvsrli_b(tmpr, 2);
+    tmpb = __lasx_xvor_v(reg0, tmpb);
+    tmpg = __lasx_xvor_v(reg1, tmpg);
+    tmpr = __lasx_xvor_v(reg2, tmpr);
+    reg0 = __lasx_xvslli_b(nexb, 3);
+    reg1 = __lasx_xvslli_b(nexg, 3);
+    reg2 = __lasx_xvslli_b(nexr, 3);
+    nexb = __lasx_xvsrli_b(nexb, 2);
+    nexg = __lasx_xvsrli_b(nexg, 2);
+    nexr = __lasx_xvsrli_b(nexr, 2);
+    nexb = __lasx_xvor_v(reg0, nexb);
+    nexg = __lasx_xvor_v(reg1, nexg);
+    nexr = __lasx_xvor_v(reg2, nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    reg0 = __lasx_xvpermi_d(reg0, 0xD8);
+    reg1 = __lasx_xvpermi_d(reg1, 0xD8);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    src_argb1555 += 64;
+    next_argb1555 += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void RGB565ToYRow_LASX(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m256i reg0, reg1, dst0;
+  __m256i const_66 = __lasx_xvldi(66);
+  __m256i const_129 = __lasx_xvldi(129);
+  __m256i const_25 = __lasx_xvldi(25);
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_rgb565, 0);
+    src1 = __lasx_xvld(src_rgb565, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpr = __lasx_xvandi_b(tmp1, 0xF8);
+    reg1 = __lasx_xvandi_b(tmp1, 0x07);
+    reg0 = __lasx_xvsrli_b(tmp0, 5);
+    reg1 = __lasx_xvslli_b(reg1, 3);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvsrli_b(tmpb, 2);
+    tmpb = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpg, 2);
+    reg1 = __lasx_xvsrli_b(tmpg, 4);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvsrli_b(tmpr, 5);
+    tmpr = __lasx_xvor_v(tmpr, reg0);
+    reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lasx_xvmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lasx_xvpackod_b(reg1, reg0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    dst_y += 32;
+    src_rgb565 += 64;
+  }
+}
+
+void RGB565ToUVRow_LASX(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  int x;
+  int len = width / 32;
+  const uint8_t* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i reg0, reg1, reg2, reg3, dst0;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_rgb565, 0, src_rgb565, 32, next_rgb565, 0,
+              next_rgb565, 32, src0, src1, src2, src3);
+    DUP2_ARG2(__lasx_xvpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lasx_xvpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpr = __lasx_xvandi_b(tmp1, 0xF8);
+    nexb = __lasx_xvandi_b(tmp2, 0x1F);
+    nexr = __lasx_xvandi_b(tmp3, 0xF8);
+    reg1 = __lasx_xvandi_b(tmp1, 0x07);
+    reg3 = __lasx_xvandi_b(tmp3, 0x07);
+    reg0 = __lasx_xvsrli_b(tmp0, 5);
+    reg1 = __lasx_xvslli_b(reg1, 3);
+    reg2 = __lasx_xvsrli_b(tmp2, 5);
+    reg3 = __lasx_xvslli_b(reg3, 3);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    nexg = __lasx_xvor_v(reg2, reg3);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvsrli_b(tmpb, 2);
+    reg2 = __lasx_xvslli_b(nexb, 3);
+    reg3 = __lasx_xvsrli_b(nexb, 2);
+    tmpb = __lasx_xvor_v(reg1, reg0);
+    nexb = __lasx_xvor_v(reg2, reg3);
+    reg0 = __lasx_xvslli_b(tmpg, 2);
+    reg1 = __lasx_xvsrli_b(tmpg, 4);
+    reg2 = __lasx_xvslli_b(nexg, 2);
+    reg3 = __lasx_xvsrli_b(nexg, 4);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    nexg = __lasx_xvor_v(reg2, reg3);
+    reg0 = __lasx_xvsrli_b(tmpr, 5);
+    reg2 = __lasx_xvsrli_b(nexr, 5);
+    tmpr = __lasx_xvor_v(tmpr, reg0);
+    nexr = __lasx_xvor_v(nexr, reg2);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    reg0 = __lasx_xvpermi_d(reg0, 0xD8);
+    reg1 = __lasx_xvpermi_d(reg1, 0xD8);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    dst_u += 16;
+    dst_v += 16;
+    src_rgb565 += 64;
+    next_rgb565 += 64;
+  }
+}
+
+void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  const uint8_t* next_rgb24 = src_rgb24 + src_stride_rgb24;
+  int len = width / 32;
+  __m256i src0, src1, src2, reg0, reg1, reg2;
+  __m256i nex0, nex1, nex2, dst0, tmp0, tmp1, tmp2;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+  __m256i shuff0_b = {0x15120F0C09060300, 0x00000000001E1B18,
+                      0x15120F0C09060300, 0x00000000001E1B18};
+  __m256i shuff1_b = {0x0706050403020100, 0x1D1A1714110A0908,
+                      0x0706050403020100, 0x1D1A1714110A0908};
+  __m256i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19,
+                      0x1613100D0A070401, 0x00000000001F1C19};
+  __m256i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908,
+                      0x0706050403020100, 0x1E1B1815120A0908};
+  __m256i shuff0_r = {0x1714110E0B080502, 0x0000000000001D1A,
+                      0x1714110E0B080502, 0x0000000000001D1A};
+  __m256i shuff1_r = {0x0706050403020100, 0x1F1C191613100908,
+                      0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_rgb24, 0, src_rgb24, 32, src_rgb24, 64,
+              next_rgb24, 0, reg0, reg1, reg2, tmp0);
+    DUP2_ARG2(__lasx_xvld, next_rgb24, 32, next_rgb24, 64, tmp1, tmp2);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2, reg1,
+              0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
+    DUP2_ARG3(__lasx_xvpermi_q, tmp2, tmp0, 0x21, tmp2, tmp1, 0x30, nex1, nex2);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    src_rgb24 += 96;
+    next_rgb24 += 96;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void RAWToUVRow_LASX(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_raw = src_raw + src_stride_raw;
+  int len = width / 32;
+  __m256i src0, src1, src2, reg0, reg1, reg2;
+  __m256i nex0, nex1, nex2, dst0, tmp0, tmp1, tmp2;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+  __m256i shuff0_r = {0x15120F0C09060300, 0x00000000001E1B18,
+                      0x15120F0C09060300, 0x00000000001E1B18};
+  __m256i shuff1_r = {0x0706050403020100, 0x1D1A1714110A0908,
+                      0x0706050403020100, 0x1D1A1714110A0908};
+  __m256i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19,
+                      0x1613100D0A070401, 0x00000000001F1C19};
+  __m256i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908,
+                      0x0706050403020100, 0x1E1B1815120A0908};
+  __m256i shuff0_b = {0x1714110E0B080502, 0x0000000000001D1A,
+                      0x1714110E0B080502, 0x0000000000001D1A};
+  __m256i shuff1_b = {0x0706050403020100, 0x1F1C191613100908,
+                      0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_raw, 0, src_raw, 32, src_raw, 64, next_raw, 0,
+              reg0, reg1, reg2, tmp0);
+    DUP2_ARG2(__lasx_xvld, next_raw, 32, next_raw, 64, tmp1, tmp2);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2, reg1,
+              0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
+    DUP2_ARG3(__lasx_xvpermi_q, tmp2, tmp0, 0x21, tmp2, tmp1, 0x30, nex1, nex2);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    src_raw += 96;
+    next_raw += 96;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void NV12ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m256i vec_yg, vec_yb, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_vrub, vec_vgug, vec_y, vec_vu;
+  __m256i out_b, out_g, out_r;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = __lasx_xvldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_vrub = __lasx_xvilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lasx_xvilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lasx_xvld(src_y, 0);
+    vec_vu = __lasx_xvld(src_uv, 0);
+    vec_vu = __lasx_xvsub_b(vec_vu, const_0x80);
+    vec_vu = __lasx_vext2xv_h_b(vec_vu);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_r, out_g,
+             out_b);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 16;
+    src_uv += 16;
+  }
+}
+
+void NV12ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  int len = width / 16;
+  __m256i vec_yg, vec_yb, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_vrub, vec_vgug, vec_y, vec_vu;
+  __m256i out_b, out_g, out_r;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_vrub = __lasx_xvilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lasx_xvilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lasx_xvld(src_y, 0);
+    vec_vu = __lasx_xvld(src_uv, 0);
+    vec_vu = __lasx_xvsub_b(vec_vu, const_0x80);
+    vec_vu = __lasx_vext2xv_h_b(vec_vu);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_r, out_g,
+             out_b);
+    out_b = __lasx_xvsrli_h(out_b, 3);
+    out_g = __lasx_xvsrli_h(out_g, 2);
+    out_r = __lasx_xvsrli_h(out_r, 3);
+    out_g = __lasx_xvslli_h(out_g, 5);
+    out_r = __lasx_xvslli_h(out_r, 11);
+    out_r = __lasx_xvor_v(out_r, out_g);
+    out_r = __lasx_xvor_v(out_r, out_b);
+    __lasx_xvst(out_r, dst_rgb565, 0);
+    src_y += 16;
+    src_uv += 16;
+    dst_rgb565 += 32;
+  }
+}
+
+void NV21ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m256i vec_yg, vec_yb, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg, vec_y, vec_uv;
+  __m256i out_b, out_g, out_r;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = __lasx_xvldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lasx_xvld(src_y, 0);
+    vec_uv = __lasx_xvld(src_uv, 0);
+    vec_uv = __lasx_xvsub_b(vec_uv, const_0x80);
+    vec_uv = __lasx_vext2xv_h_b(vec_uv);
+    YUVTORGB(vec_y, vec_uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 16;
+    src_uv += 16;
+  }
+}
+
+struct RgbConstants {
+  uint8_t kRGBToY[4];
+  uint16_t kAddY;
+  uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+                                                        128,
+                                                        0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+                                                        0x1080,
+                                                        0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+                                                      0x1080,
+                                                      0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+static void ARGBToYMatrixRow_LASX(const uint8_t* src_argb,
+                                  uint8_t* dst_y,
+                                  int width,
+                                  const struct RgbConstants* rgbconstants) {
+  int32_t shuff[8] = {0, 4, 1, 5, 2, 6, 3, 7};
+  asm volatile(
+      "xvldrepl.b      $xr0,  %3,    0             \n\t"  // load rgbconstants
+      "xvldrepl.b      $xr1,  %3,    1             \n\t"  // load rgbconstants
+      "xvldrepl.b      $xr2,  %3,    2             \n\t"  // load rgbconstants
+      "xvldrepl.h      $xr3,  %3,    4             \n\t"  // load rgbconstants
+      "xvld            $xr20, %4,    0             \n\t"  // load shuff
+      "1:                                          \n\t"
+      "xvld            $xr4,  %0,    0             \n\t"
+      "xvld            $xr5,  %0,    32            \n\t"
+      "xvld            $xr6,  %0,    64            \n\t"
+      "xvld            $xr7,  %0,    96            \n\t"  // load 32 pixels of
+                                                          // ARGB
+      "xvor.v          $xr12, $xr3,  $xr3          \n\t"
+      "xvor.v          $xr13, $xr3,  $xr3          \n\t"
+      "addi.d          %2,    %2,    -32           \n\t"  // 32 processed per
+                                                          // loop.
+      "xvpickev.b      $xr8,  $xr5,  $xr4          \n\t"  // BR
+      "xvpickev.b      $xr10, $xr7,  $xr6          \n\t"
+      "xvpickod.b      $xr9,  $xr5,  $xr4          \n\t"  // GA
+      "xvpickod.b      $xr11, $xr7,  $xr6          \n\t"
+      "xvmaddwev.h.bu  $xr12, $xr8,  $xr0          \n\t"  // B
+      "xvmaddwev.h.bu  $xr13, $xr10, $xr0          \n\t"
+      "xvmaddwev.h.bu  $xr12, $xr9,  $xr1          \n\t"  // G
+      "xvmaddwev.h.bu  $xr13, $xr11, $xr1          \n\t"
+      "xvmaddwod.h.bu  $xr12, $xr8,  $xr2          \n\t"  // R
+      "xvmaddwod.h.bu  $xr13, $xr10, $xr2          \n\t"
+      "addi.d          %0,    %0,    128           \n\t"
+      "xvpickod.b      $xr10, $xr13, $xr12         \n\t"
+      "xvperm.w        $xr11, $xr10, $xr20         \n\t"
+      "xvst            $xr11, %1,    0             \n\t"
+      "addi.d          %1,    %1,    32            \n\t"
+      "bnez            %2,    1b                   \n\t"
+      : "+&r"(src_argb),  // %0
+        "+&r"(dst_y),     // %1
+        "+&r"(width)      // %2
+      : "r"(rgbconstants), "r"(shuff)
+      : "memory");
+}
+
+void ARGBToYRow_LASX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_LASX(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_LASX(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_LASX(src_argb, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void ABGRToYRow_LASX(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_LASX(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+void ABGRToYJRow_LASX(const uint8_t* src_abgr, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_LASX(src_abgr, dst_yj, width, &kRawJPEGConstants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+static void RGBAToYMatrixRow_LASX(const uint8_t* src_rgba,
+                                  uint8_t* dst_y,
+                                  int width,
+                                  const struct RgbConstants* rgbconstants) {
+  int32_t shuff[8] = {0, 4, 1, 5, 2, 6, 3, 7};
+  asm volatile(
+      "xvldrepl.b      $xr0,  %3,    0             \n\t"  // load rgbconstants
+      "xvldrepl.b      $xr1,  %3,    1             \n\t"  // load rgbconstants
+      "xvldrepl.b      $xr2,  %3,    2             \n\t"  // load rgbconstants
+      "xvldrepl.h      $xr3,  %3,    4             \n\t"  // load rgbconstants
+      "xvld            $xr20, %4,    0             \n\t"  // load shuff
+      "1:                                          \n\t"
+      "xvld            $xr4,  %0,    0             \n\t"
+      "xvld            $xr5,  %0,    32            \n\t"
+      "xvld            $xr6,  %0,    64            \n\t"
+      "xvld            $xr7,  %0,    96            \n\t"  // load 32 pixels of
+                                                          // RGBA
+      "xvor.v          $xr12, $xr3,  $xr3          \n\t"
+      "xvor.v          $xr13, $xr3,  $xr3          \n\t"
+      "addi.d          %2,    %2,    -32           \n\t"  // 32 processed per
+                                                          // loop.
+      "xvpickev.b      $xr8,  $xr5,  $xr4          \n\t"  // AG
+      "xvpickev.b      $xr10, $xr7,  $xr6          \n\t"
+      "xvpickod.b      $xr9,  $xr5,  $xr4          \n\t"  // BR
+      "xvpickod.b      $xr11, $xr7,  $xr6          \n\t"
+      "xvmaddwev.h.bu  $xr12, $xr9,  $xr0          \n\t"  // B
+      "xvmaddwev.h.bu  $xr13, $xr11, $xr0          \n\t"
+      "xvmaddwod.h.bu  $xr12, $xr8,  $xr1          \n\t"  // G
+      "xvmaddwod.h.bu  $xr13, $xr10, $xr1          \n\t"
+      "xvmaddwod.h.bu  $xr12, $xr9,  $xr2          \n\t"  // R
+      "xvmaddwod.h.bu  $xr13, $xr11, $xr2          \n\t"
+      "addi.d          %0,    %0,    128           \n\t"
+      "xvpickod.b      $xr10, $xr13, $xr12         \n\t"
+      "xvperm.w        $xr11, $xr10, $xr20         \n\t"
+      "xvst            $xr11, %1,    0             \n\t"
+      "addi.d          %1,    %1,    32            \n\t"
+      "bnez            %2,    1b                   \n\t"
+      : "+&r"(src_rgba),  // %0
+        "+&r"(dst_y),     // %1
+        "+&r"(width)      // %2
+      : "r"(rgbconstants), "r"(shuff)
+      : "memory");
+}
+
+void RGBAToYRow_LASX(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_LASX(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_LASX(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+  RGBAToYMatrixRow_LASX(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_LASX(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_LASX(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+static void RGBToYMatrixRow_LASX(const uint8_t* src_rgba,
+                                 uint8_t* dst_y,
+                                 int width,
+                                 const struct RgbConstants* rgbconstants) {
+  int8_t shuff[128] = {
+      0,  2,  3,  5,  6,  8, 9,  11, 12, 14, 15, 17, 18, 20, 21, 23,
+      0,  2,  3,  5,  6,  8, 9,  11, 12, 14, 15, 17, 18, 20, 21, 23,
+      24, 26, 27, 29, 30, 0, 1,  3,  4,  6,  7,  9,  10, 12, 13, 15,
+      24, 26, 27, 29, 30, 0, 1,  3,  4,  6,  7,  9,  10, 12, 13, 15,
+      1,  0,  4,  0,  7,  0, 10, 0,  13, 0,  16, 0,  19, 0,  22, 0,
+      1,  0,  4,  0,  7,  0, 10, 0,  13, 0,  16, 0,  19, 0,  22, 0,
+      25, 0,  28, 0,  31, 0, 2,  0,  5,  0,  8,  0,  11, 0,  14, 0,
+      25, 0,  28, 0,  31, 0, 2,  0,  5,  0,  8,  0,  11, 0,  14, 0};
+  asm volatile(
+      "xvldrepl.b      $xr0,  %3,    0             \n\t"  // load rgbconstants
+      "xvldrepl.b      $xr1,  %3,    1             \n\t"  // load rgbconstants
+      "xvldrepl.b      $xr2,  %3,    2             \n\t"  // load rgbconstants
+      "xvldrepl.h      $xr3,  %3,    4             \n\t"  // load rgbconstants
+      "xvld            $xr4,  %4,    0             \n\t"  // load shuff
+      "xvld            $xr5,  %4,    32            \n\t"
+      "xvld            $xr6,  %4,    64            \n\t"
+      "xvld            $xr7,  %4,    96            \n\t"
+      "1:                                          \n\t"
+      "xvld            $xr8,  %0,    0             \n\t"
+      "xvld            $xr9,  %0,    32            \n\t"
+      "xvld            $xr10, %0,    64            \n\t"  // load 32 pixels of
+                                                          // RGB
+      "xvor.v          $xr12, $xr3,  $xr3          \n\t"
+      "xvor.v          $xr13, $xr3,  $xr3          \n\t"
+      "xvor.v          $xr11, $xr9,  $xr9          \n\t"
+      "addi.d          %2,    %2,    -32           \n\t"  // 32 processed per
+                                                          // loop.
+      "xvpermi.q       $xr9,  $xr8,  0x30          \n\t"  // src0
+      "xvpermi.q       $xr8,  $xr10, 0x03          \n\t"  // src1
+      "xvpermi.q       $xr10, $xr11, 0x30          \n\t"  // src2
+      "xvshuf.b        $xr14, $xr8,  $xr9,  $xr4   \n\t"
+      "xvshuf.b        $xr15, $xr8,  $xr10, $xr5   \n\t"
+      "xvshuf.b        $xr16, $xr8,  $xr9,  $xr6   \n\t"
+      "xvshuf.b        $xr17, $xr8,  $xr10, $xr7   \n\t"
+      "xvmaddwev.h.bu  $xr12, $xr16, $xr1          \n\t"  // G
+      "xvmaddwev.h.bu  $xr13, $xr17, $xr1          \n\t"
+      "xvmaddwev.h.bu  $xr12, $xr14, $xr0          \n\t"  // B
+      "xvmaddwev.h.bu  $xr13, $xr15, $xr0          \n\t"
+      "xvmaddwod.h.bu  $xr12, $xr14, $xr2          \n\t"  // R
+      "xvmaddwod.h.bu  $xr13, $xr15, $xr2          \n\t"
+      "addi.d          %0,    %0,    96            \n\t"
+      "xvpickod.b      $xr10, $xr13, $xr12         \n\t"
+      "xvst            $xr10, %1,    0             \n\t"
+      "addi.d          %1,    %1,    32            \n\t"
+      "bnez            %2,    1b                   \n\t"
+      : "+&r"(src_rgba),    // %0
+        "+&r"(dst_y),       // %1
+        "+&r"(width)        // %2
+      : "r"(rgbconstants),  // %3
+        "r"(shuff)          // %4
+      : "memory");
+}
+
+void RGB24ToYJRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_LASX(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void RAWToYJRow_LASX(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_LASX(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_LASX(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_LASX(src_raw, dst_y, width, &kRawI601Constants);
+}
+
+void ARGBToUVJRow_LASX(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  const uint8_t* next_argb = src_argb + src_stride_argb;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3;
+  __m256i nex0, nex1, nex2, nex3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, dst0;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i const_63 = __lasx_xvldi(0x43F);
+  __m256i const_42 = __lasx_xvldi(0x42A);
+  __m256i const_21 = __lasx_xvldi(0x415);
+  __m256i const_53 = __lasx_xvldi(0x435);
+  __m256i const_10 = __lasx_xvldi(0x40A);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+  __m256i shuff = {0x1614060412100200, 0x1E1C0E0C1A180A08, 0x1715070513110301,
+                   0x1F1D0F0D1B190B09};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    DUP4_ARG2(__lasx_xvld, next_argb, 0, next_argb, 32, next_argb, 64,
+              next_argb, 96, nex0, nex1, nex2, nex3);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp2 = __lasx_xvpickev_b(src3, src2);
+    tmp3 = __lasx_xvpickod_b(src3, src2);
+    tmpr = __lasx_xvpickod_b(tmp2, tmp0);
+    tmpb = __lasx_xvpickev_b(tmp2, tmp0);
+    tmpg = __lasx_xvpickev_b(tmp3, tmp1);
+    tmp0 = __lasx_xvpickev_b(nex1, nex0);
+    tmp1 = __lasx_xvpickod_b(nex1, nex0);
+    tmp2 = __lasx_xvpickev_b(nex3, nex2);
+    tmp3 = __lasx_xvpickod_b(nex3, nex2);
+    nexr = __lasx_xvpickod_b(tmp2, tmp0);
+    nexb = __lasx_xvpickev_b(tmp2, tmp0);
+    nexg = __lasx_xvpickev_b(tmp3, tmp1);
+    tmp0 = __lasx_xvaddwev_h_bu(tmpb, nexb);
+    tmp1 = __lasx_xvaddwod_h_bu(tmpb, nexb);
+    tmp2 = __lasx_xvaddwev_h_bu(tmpg, nexg);
+    tmp3 = __lasx_xvaddwod_h_bu(tmpg, nexg);
+    reg0 = __lasx_xvaddwev_h_bu(tmpr, nexr);
+    reg1 = __lasx_xvaddwod_h_bu(tmpr, nexr);
+    tmpb = __lasx_xvavgr_hu(tmp0, tmp1);
+    tmpg = __lasx_xvavgr_hu(tmp2, tmp3);
+    tmpr = __lasx_xvavgr_hu(reg0, reg1);
+    reg0 = __lasx_xvmadd_h(const_8080, const_63, tmpb);
+    reg1 = __lasx_xvmadd_h(const_8080, const_63, tmpr);
+    reg0 = __lasx_xvmsub_h(reg0, const_42, tmpg);
+    reg1 = __lasx_xvmsub_h(reg1, const_53, tmpg);
+    reg0 = __lasx_xvmsub_h(reg0, const_21, tmpr);
+    reg1 = __lasx_xvmsub_h(reg1, const_10, tmpb);
+    dst0 = __lasx_xvpackod_b(reg1, reg0);
+    tmp0 = __lasx_xvpermi_d(dst0, 0x44);
+    tmp1 = __lasx_xvpermi_d(dst0, 0xEE);
+    dst0 = __lasx_xvshuf_b(tmp1, tmp0, shuff);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 2);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 1);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    dst_u += 16;
+    dst_v += 16;
+    src_argb += 128;
+    next_argb += 128;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LASX) && defined(__loongarch_asx)
diff --git a/media/libyuv/source/row_lsx.cc b/media/libyuv/source/row_lsx.cc
new file mode 100644
index 0000000000..9c1e16f22e
--- /dev/null
+++ b/media/libyuv/source/row_lsx.cc
@@ -0,0 +1,1861 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Fill YUV -> RGB conversion constants into vectors
+#define YUVTORGB_SETUP(yuvconst, vr, ub, vg, ug, yg, yb) \
+  {                                                      \
+    ub = __lsx_vreplgr2vr_h(yuvconst->kUVToB[0]);        \
+    vr = __lsx_vreplgr2vr_h(yuvconst->kUVToR[1]);        \
+    ug = __lsx_vreplgr2vr_h(yuvconst->kUVToG[0]);        \
+    vg = __lsx_vreplgr2vr_h(yuvconst->kUVToG[1]);        \
+    yg = __lsx_vreplgr2vr_h(yuvconst->kYToRgb[0]);       \
+    yb = __lsx_vreplgr2vr_w(yuvconst->kYBiasToRgb[0]);   \
+  }
+
+// Convert 8 pixels of YUV420 to RGB.
+#define YUVTORGB(in_y, in_vu, vrub, vgug, yg, yb, out_b, out_g, out_r) \
+  {                                                                    \
+    __m128i y_ev, y_od, u_l, v_l;                                      \
+    __m128i tmp0, tmp1, tmp2, tmp3;                                    \
+                                                                       \
+    tmp0 = __lsx_vilvl_b(in_y, in_y);                                  \
+    y_ev = __lsx_vmulwev_w_hu_h(tmp0, yg);                             \
+    y_od = __lsx_vmulwod_w_hu_h(tmp0, yg);                             \
+    y_ev = __lsx_vsrai_w(y_ev, 16);                                    \
+    y_od = __lsx_vsrai_w(y_od, 16);                                    \
+    y_ev = __lsx_vadd_w(y_ev, yb);                                     \
+    y_od = __lsx_vadd_w(y_od, yb);                                     \
+    in_vu = __lsx_vilvl_b(zero, in_vu);                                \
+    in_vu = __lsx_vsub_h(in_vu, const_80);                             \
+    u_l = __lsx_vmulwev_w_h(in_vu, vrub);                              \
+    v_l = __lsx_vmulwod_w_h(in_vu, vrub);                              \
+    tmp0 = __lsx_vadd_w(y_ev, u_l);                                    \
+    tmp1 = __lsx_vadd_w(y_od, u_l);                                    \
+    tmp2 = __lsx_vadd_w(y_ev, v_l);                                    \
+    tmp3 = __lsx_vadd_w(y_od, v_l);                                    \
+    tmp0 = __lsx_vsrai_w(tmp0, 6);                                     \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                     \
+    tmp2 = __lsx_vsrai_w(tmp2, 6);                                     \
+    tmp3 = __lsx_vsrai_w(tmp3, 6);                                     \
+    tmp0 = __lsx_vclip255_w(tmp0);                                     \
+    tmp1 = __lsx_vclip255_w(tmp1);                                     \
+    tmp2 = __lsx_vclip255_w(tmp2);                                     \
+    tmp3 = __lsx_vclip255_w(tmp3);                                     \
+    out_b = __lsx_vpackev_h(tmp1, tmp0);                               \
+    out_r = __lsx_vpackev_h(tmp3, tmp2);                               \
+    tmp0 = __lsx_vdp2_w_h(in_vu, vgug);                                \
+    tmp1 = __lsx_vsub_w(y_ev, tmp0);                                   \
+    tmp2 = __lsx_vsub_w(y_od, tmp0);                                   \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                     \
+    tmp2 = __lsx_vsrai_w(tmp2, 6);                                     \
+    tmp1 = __lsx_vclip255_w(tmp1);                                     \
+    tmp2 = __lsx_vclip255_w(tmp2);                                     \
+    out_g = __lsx_vpackev_h(tmp2, tmp1);                               \
+  }
+
+// Convert I444 pixels of YUV420 to RGB.
+#define I444TORGB(in_yy, in_u, in_v, ub, vr, ugvg, yg, yb, out_b, out_g, \
+                  out_r)                                                 \
+  {                                                                      \
+    __m128i y_ev, y_od, u_ev, v_ev, u_od, v_od;                          \
+    __m128i tmp0, tmp1, tmp2, tmp3;                                      \
+                                                                         \
+    y_ev = __lsx_vmulwev_w_hu_h(in_yy, yg);                              \
+    y_od = __lsx_vmulwod_w_hu_h(in_yy, yg);                              \
+    y_ev = __lsx_vsrai_w(y_ev, 16);                                      \
+    y_od = __lsx_vsrai_w(y_od, 16);                                      \
+    y_ev = __lsx_vadd_w(y_ev, yb);                                       \
+    y_od = __lsx_vadd_w(y_od, yb);                                       \
+    in_u = __lsx_vsub_h(in_u, const_80);                                 \
+    in_v = __lsx_vsub_h(in_v, const_80);                                 \
+    u_ev = __lsx_vmulwev_w_h(in_u, ub);                                  \
+    u_od = __lsx_vmulwod_w_h(in_u, ub);                                  \
+    v_ev = __lsx_vmulwev_w_h(in_v, vr);                                  \
+    v_od = __lsx_vmulwod_w_h(in_v, vr);                                  \
+    tmp0 = __lsx_vadd_w(y_ev, u_ev);                                     \
+    tmp1 = __lsx_vadd_w(y_od, u_od);                                     \
+    tmp2 = __lsx_vadd_w(y_ev, v_ev);                                     \
+    tmp3 = __lsx_vadd_w(y_od, v_od);                                     \
+    tmp0 = __lsx_vsrai_w(tmp0, 6);                                       \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                       \
+    tmp2 = __lsx_vsrai_w(tmp2, 6);                                       \
+    tmp3 = __lsx_vsrai_w(tmp3, 6);                                       \
+    tmp0 = __lsx_vclip255_w(tmp0);                                       \
+    tmp1 = __lsx_vclip255_w(tmp1);                                       \
+    tmp2 = __lsx_vclip255_w(tmp2);                                       \
+    tmp3 = __lsx_vclip255_w(tmp3);                                       \
+    out_b = __lsx_vpackev_h(tmp1, tmp0);                                 \
+    out_r = __lsx_vpackev_h(tmp3, tmp2);                                 \
+    u_ev = __lsx_vpackev_h(in_u, in_v);                                  \
+    u_od = __lsx_vpackod_h(in_u, in_v);                                  \
+    v_ev = __lsx_vdp2_w_h(u_ev, ugvg);                                   \
+    v_od = __lsx_vdp2_w_h(u_od, ugvg);                                   \
+    tmp0 = __lsx_vsub_w(y_ev, v_ev);                                     \
+    tmp1 = __lsx_vsub_w(y_od, v_od);                                     \
+    tmp0 = __lsx_vsrai_w(tmp0, 6);                                       \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                       \
+    tmp0 = __lsx_vclip255_w(tmp0);                                       \
+    tmp1 = __lsx_vclip255_w(tmp1);                                       \
+    out_g = __lsx_vpackev_h(tmp1, tmp0);                                 \
+  }
+
+// Pack and Store 8 ARGB values.
+#define STOREARGB(in_a, in_r, in_g, in_b, pdst_argb) \
+  {                                                  \
+    __m128i temp0, temp1;                            \
+    __m128i dst0, dst1;                              \
+                                                     \
+    temp0 = __lsx_vpackev_b(in_g, in_b);             \
+    temp1 = __lsx_vpackev_b(in_a, in_r);             \
+    dst0 = __lsx_vilvl_h(temp1, temp0);              \
+    dst1 = __lsx_vilvh_h(temp1, temp0);              \
+    __lsx_vst(dst0, pdst_argb, 0);                   \
+    __lsx_vst(dst1, pdst_argb, 16);                  \
+    pdst_argb += 32;                                 \
+  }
+
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _dst0) \
+  {                                                              \
+    __m128i _tmp0, _tmp1, _tmp2, _tmp3;                          \
+    __m128i _reg0, _reg1;                                        \
+    _tmp0 = __lsx_vaddwev_h_bu(_tmpb, _nexb);                    \
+    _tmp1 = __lsx_vaddwod_h_bu(_tmpb, _nexb);                    \
+    _tmp2 = __lsx_vaddwev_h_bu(_tmpg, _nexg);                    \
+    _tmp3 = __lsx_vaddwod_h_bu(_tmpg, _nexg);                    \
+    _reg0 = __lsx_vaddwev_h_bu(_tmpr, _nexr);                    \
+    _reg1 = __lsx_vaddwod_h_bu(_tmpr, _nexr);                    \
+    _tmpb = __lsx_vavgr_hu(_tmp0, _tmp1);                        \
+    _tmpg = __lsx_vavgr_hu(_tmp2, _tmp3);                        \
+    _tmpr = __lsx_vavgr_hu(_reg0, _reg1);                        \
+    _reg0 = __lsx_vmadd_h(const_8080, const_112, _tmpb);         \
+    _reg1 = __lsx_vmadd_h(const_8080, const_112, _tmpr);         \
+    _reg0 = __lsx_vmsub_h(_reg0, const_74, _tmpg);               \
+    _reg1 = __lsx_vmsub_h(_reg1, const_94, _tmpg);               \
+    _reg0 = __lsx_vmsub_h(_reg0, const_38, _tmpr);               \
+    _reg1 = __lsx_vmsub_h(_reg1, const_18, _tmpb);               \
+    _dst0 = __lsx_vpickod_b(_reg1, _reg0);                       \
+  }
+
+void ARGB4444ToARGBRow_LSX(const uint8_t* src_argb4444,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_argb4444, 0);
+    src1 = __lsx_vld(src_argb4444, 16);
+    tmp0 = __lsx_vandi_b(src0, 0x0F);
+    tmp1 = __lsx_vandi_b(src0, 0xF0);
+    tmp2 = __lsx_vandi_b(src1, 0x0F);
+    tmp3 = __lsx_vandi_b(src1, 0xF0);
+    reg0 = __lsx_vslli_b(tmp0, 4);
+    reg2 = __lsx_vslli_b(tmp2, 4);
+    reg1 = __lsx_vsrli_b(tmp1, 4);
+    reg3 = __lsx_vsrli_b(tmp3, 4);
+    DUP4_ARG2(__lsx_vor_v, tmp0, reg0, tmp1, reg1, tmp2, reg2, tmp3, reg3, tmp0,
+              tmp1, tmp2, tmp3);
+    dst0 = __lsx_vilvl_b(tmp1, tmp0);
+    dst2 = __lsx_vilvl_b(tmp3, tmp2);
+    dst1 = __lsx_vilvh_b(tmp1, tmp0);
+    dst3 = __lsx_vilvh_b(tmp3, tmp2);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_argb4444 += 32;
+  }
+}
+
+void ARGB1555ToARGBRow_LSX(const uint8_t* src_argb1555,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr, tmpa;
+  __m128i reg0, reg1, reg2;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_argb1555, 0);
+    src1 = __lsx_vld(src_argb1555, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpg = __lsx_vsrli_b(tmp0, 5);
+    reg0 = __lsx_vandi_b(tmp1, 0x03);
+    reg0 = __lsx_vslli_b(reg0, 3);
+    tmpg = __lsx_vor_v(tmpg, reg0);
+    reg1 = __lsx_vandi_b(tmp1, 0x7C);
+    tmpr = __lsx_vsrli_b(reg1, 2);
+    tmpa = __lsx_vsrli_b(tmp1, 7);
+    tmpa = __lsx_vneg_b(tmpa);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vslli_b(tmpg, 3);
+    reg2 = __lsx_vslli_b(tmpr, 3);
+    tmpb = __lsx_vsrli_b(tmpb, 2);
+    tmpg = __lsx_vsrli_b(tmpg, 2);
+    tmpr = __lsx_vsrli_b(tmpr, 2);
+    tmpb = __lsx_vor_v(reg0, tmpb);
+    tmpg = __lsx_vor_v(reg1, tmpg);
+    tmpr = __lsx_vor_v(reg2, tmpr);
+    DUP2_ARG2(__lsx_vilvl_b, tmpg, tmpb, tmpa, tmpr, reg0, reg1);
+    dst0 = __lsx_vilvl_h(reg1, reg0);
+    dst1 = __lsx_vilvh_h(reg1, reg0);
+    DUP2_ARG2(__lsx_vilvh_b, tmpg, tmpb, tmpa, tmpr, reg0, reg1);
+    dst2 = __lsx_vilvl_h(reg1, reg0);
+    dst3 = __lsx_vilvh_h(reg1, reg0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_argb1555 += 32;
+  }
+}
+
+void RGB565ToARGBRow_LSX(const uint8_t* src_rgb565,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m128i reg0, reg1, dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb565, 0);
+    src1 = __lsx_vld(src_rgb565, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpr = __lsx_vandi_b(tmp1, 0xF8);
+    reg1 = __lsx_vandi_b(tmp1, 0x07);
+    reg0 = __lsx_vsrli_b(tmp0, 5);
+    reg1 = __lsx_vslli_b(reg1, 3);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vsrli_b(tmpb, 2);
+    tmpb = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpg, 2);
+    reg1 = __lsx_vsrli_b(tmpg, 4);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vsrli_b(tmpr, 5);
+    tmpr = __lsx_vor_v(tmpr, reg0);
+    DUP2_ARG2(__lsx_vilvl_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst0 = __lsx_vilvl_h(reg1, reg0);
+    dst1 = __lsx_vilvh_h(reg1, reg0);
+    DUP2_ARG2(__lsx_vilvh_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst2 = __lsx_vilvl_h(reg1, reg0);
+    dst3 = __lsx_vilvh_h(reg1, reg0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_rgb565 += 32;
+  }
+}
+
+void RGB24ToARGBRow_LSX(const uint8_t* src_rgb24,
+                        uint8_t* dst_argb,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i tmp0, tmp1, tmp2;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514};
+  __m128i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100};
+  __m128i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C};
+  __m128i shuf3 = {0x1005040310020100, 0x100B0A0910080706};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb24, 0);
+    src1 = __lsx_vld(src_rgb24, 16);
+    src2 = __lsx_vld(src_rgb24, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0, tmp1);
+    tmp2 = __lsx_vshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lsx_vshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_rgb24 += 48;
+  }
+}
+
+void RAWToARGBRow_LSX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i tmp0, tmp1, tmp2;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514};
+  __m128i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100};
+  __m128i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C};
+  __m128i shuf3 = {0x1003040510000102, 0x10090A0B10060708};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_raw, 0);
+    src1 = __lsx_vld(src_raw, 16);
+    src2 = __lsx_vld(src_raw, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0, tmp1);
+    tmp2 = __lsx_vshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lsx_vshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_raw += 48;
+  }
+}
+
+void ARGB1555ToYRow_LSX(const uint8_t* src_argb1555,
+                        uint8_t* dst_y,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m128i reg0, reg1, reg2, dst0;
+  __m128i const_66 = __lsx_vldi(66);
+  __m128i const_129 = __lsx_vldi(129);
+  __m128i const_25 = __lsx_vldi(25);
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_argb1555, 0);
+    src1 = __lsx_vld(src_argb1555, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpg = __lsx_vsrli_b(tmp0, 5);
+    reg0 = __lsx_vandi_b(tmp1, 0x03);
+    reg0 = __lsx_vslli_b(reg0, 3);
+    tmpg = __lsx_vor_v(tmpg, reg0);
+    reg1 = __lsx_vandi_b(tmp1, 0x7C);
+    tmpr = __lsx_vsrli_b(reg1, 2);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vslli_b(tmpg, 3);
+    reg2 = __lsx_vslli_b(tmpr, 3);
+    tmpb = __lsx_vsrli_b(tmpb, 2);
+    tmpg = __lsx_vsrli_b(tmpg, 2);
+    tmpr = __lsx_vsrli_b(tmpr, 2);
+    tmpb = __lsx_vor_v(reg0, tmpb);
+    tmpg = __lsx_vor_v(reg1, tmpg);
+    tmpr = __lsx_vor_v(reg2, tmpr);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lsx_vmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lsx_vpackod_b(reg1, reg0);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_argb1555 += 32;
+  }
+}
+
+void ARGB1555ToUVRow_LSX(const uint8_t* src_argb1555,
+                         int src_stride_argb1555,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  int len = width / 16;
+  const uint8_t* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i reg0, reg1, reg2, reg3, dst0;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb1555, 0, src_argb1555, 16, next_argb1555, 0,
+              next_argb1555, 16, src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    nexb = __lsx_vandi_b(tmp2, 0x1F);
+    tmpg = __lsx_vsrli_b(tmp0, 5);
+    nexg = __lsx_vsrli_b(tmp2, 5);
+    reg0 = __lsx_vandi_b(tmp1, 0x03);
+    reg2 = __lsx_vandi_b(tmp3, 0x03);
+    reg0 = __lsx_vslli_b(reg0, 3);
+    reg2 = __lsx_vslli_b(reg2, 3);
+    tmpg = __lsx_vor_v(tmpg, reg0);
+    nexg = __lsx_vor_v(nexg, reg2);
+    reg1 = __lsx_vandi_b(tmp1, 0x7C);
+    reg3 = __lsx_vandi_b(tmp3, 0x7C);
+    tmpr = __lsx_vsrli_b(reg1, 2);
+    nexr = __lsx_vsrli_b(reg3, 2);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vslli_b(tmpg, 3);
+    reg2 = __lsx_vslli_b(tmpr, 3);
+    tmpb = __lsx_vsrli_b(tmpb, 2);
+    tmpg = __lsx_vsrli_b(tmpg, 2);
+    tmpr = __lsx_vsrli_b(tmpr, 2);
+    tmpb = __lsx_vor_v(reg0, tmpb);
+    tmpg = __lsx_vor_v(reg1, tmpg);
+    tmpr = __lsx_vor_v(reg2, tmpr);
+    reg0 = __lsx_vslli_b(nexb, 3);
+    reg1 = __lsx_vslli_b(nexg, 3);
+    reg2 = __lsx_vslli_b(nexr, 3);
+    nexb = __lsx_vsrli_b(nexb, 2);
+    nexg = __lsx_vsrli_b(nexg, 2);
+    nexr = __lsx_vsrli_b(nexr, 2);
+    nexb = __lsx_vor_v(reg0, nexb);
+    nexg = __lsx_vor_v(reg1, nexg);
+    nexr = __lsx_vor_v(reg2, nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_argb1555 += 32;
+    next_argb1555 += 32;
+  }
+}
+
+void RGB565ToYRow_LSX(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m128i reg0, reg1, dst0;
+  __m128i const_66 = __lsx_vldi(66);
+  __m128i const_129 = __lsx_vldi(129);
+  __m128i const_25 = __lsx_vldi(25);
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb565, 0);
+    src1 = __lsx_vld(src_rgb565, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpr = __lsx_vandi_b(tmp1, 0xF8);
+    reg1 = __lsx_vandi_b(tmp1, 0x07);
+    reg0 = __lsx_vsrli_b(tmp0, 5);
+    reg1 = __lsx_vslli_b(reg1, 3);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vsrli_b(tmpb, 2);
+    tmpb = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpg, 2);
+    reg1 = __lsx_vsrli_b(tmpg, 4);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vsrli_b(tmpr, 5);
+    tmpr = __lsx_vor_v(tmpr, reg0);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lsx_vmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lsx_vpackod_b(reg1, reg0);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_rgb565 += 32;
+  }
+}
+
+void RGB565ToUVRow_LSX(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  int len = width / 16;
+  const uint8_t* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i reg0, reg1, reg2, reg3, dst0;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_rgb565, 0, src_rgb565, 16, next_rgb565, 0,
+              next_rgb565, 16, src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpr = __lsx_vandi_b(tmp1, 0xF8);
+    nexb = __lsx_vandi_b(tmp2, 0x1F);
+    nexr = __lsx_vandi_b(tmp3, 0xF8);
+    reg1 = __lsx_vandi_b(tmp1, 0x07);
+    reg3 = __lsx_vandi_b(tmp3, 0x07);
+    reg0 = __lsx_vsrli_b(tmp0, 5);
+    reg1 = __lsx_vslli_b(reg1, 3);
+    reg2 = __lsx_vsrli_b(tmp2, 5);
+    reg3 = __lsx_vslli_b(reg3, 3);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    nexg = __lsx_vor_v(reg2, reg3);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vsrli_b(tmpb, 2);
+    reg2 = __lsx_vslli_b(nexb, 3);
+    reg3 = __lsx_vsrli_b(nexb, 2);
+    tmpb = __lsx_vor_v(reg1, reg0);
+    nexb = __lsx_vor_v(reg2, reg3);
+    reg0 = __lsx_vslli_b(tmpg, 2);
+    reg1 = __lsx_vsrli_b(tmpg, 4);
+    reg2 = __lsx_vslli_b(nexg, 2);
+    reg3 = __lsx_vsrli_b(nexg, 4);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    nexg = __lsx_vor_v(reg2, reg3);
+    reg0 = __lsx_vsrli_b(tmpr, 5);
+    reg2 = __lsx_vsrli_b(nexr, 5);
+    tmpr = __lsx_vor_v(tmpr, reg0);
+    nexr = __lsx_vor_v(nexr, reg2);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_rgb565 += 32;
+    next_rgb565 += 32;
+  }
+}
+
+void RGB24ToUVRow_LSX(const uint8_t* src_rgb24,
+                      int src_stride_rgb24,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  const uint8_t* next_rgb24 = src_rgb24 + src_stride_rgb24;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i nex0, nex1, nex2, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+  __m128i shuff0_b = {0x15120F0C09060300, 0x00000000001E1B18};
+  __m128i shuff1_b = {0x0706050403020100, 0x1D1A1714110A0908};
+  __m128i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19};
+  __m128i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908};
+  __m128i shuff0_r = {0x1714110E0B080502, 0x0000000000001D1A};
+  __m128i shuff1_r = {0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb24, 0);
+    src1 = __lsx_vld(src_rgb24, 16);
+    src2 = __lsx_vld(src_rgb24, 32);
+    nex0 = __lsx_vld(next_rgb24, 0);
+    nex1 = __lsx_vld(next_rgb24, 16);
+    nex2 = __lsx_vld(next_rgb24, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_rgb24 += 48;
+    next_rgb24 += 48;
+  }
+}
+
+void RAWToUVRow_LSX(const uint8_t* src_raw,
+                    int src_stride_raw,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  int x;
+  const uint8_t* next_raw = src_raw + src_stride_raw;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i nex0, nex1, nex2, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+  __m128i shuff0_r = {0x15120F0C09060300, 0x00000000001E1B18};
+  __m128i shuff1_r = {0x0706050403020100, 0x1D1A1714110A0908};
+  __m128i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19};
+  __m128i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908};
+  __m128i shuff0_b = {0x1714110E0B080502, 0x0000000000001D1A};
+  __m128i shuff1_b = {0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_raw, 0);
+    src1 = __lsx_vld(src_raw, 16);
+    src2 = __lsx_vld(src_raw, 32);
+    nex0 = __lsx_vld(next_raw, 0);
+    nex1 = __lsx_vld(next_raw, 16);
+    nex2 = __lsx_vld(next_raw, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_raw += 48;
+    next_raw += 48;
+  }
+}
+
+void NV12ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_vu = __lsx_vld(src_uv, 0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 8;
+    src_uv += 8;
+  }
+}
+
+void NV12ToRGB565Row_LSX(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  int x;
+  int len = width / 8;
+  __m128i vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_vu = __lsx_vld(src_uv, 0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    out_b = __lsx_vsrli_h(out_b, 3);
+    out_g = __lsx_vsrli_h(out_g, 2);
+    out_r = __lsx_vsrli_h(out_r, 3);
+    out_g = __lsx_vslli_h(out_g, 5);
+    out_r = __lsx_vslli_h(out_r, 11);
+    out_r = __lsx_vor_v(out_r, out_g);
+    out_r = __lsx_vor_v(out_r, out_b);
+    __lsx_vst(out_r, dst_rgb565, 0);
+    src_y += 8;
+    src_uv += 8;
+    dst_rgb565 += 16;
+  }
+}
+
+void NV21ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i vec_y, vec_uv;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_ubvr, vec_ugvg;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_ubvr = __lsx_vilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lsx_vilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_uv = __lsx_vld(src_vu, 0);
+    YUVTORGB(vec_y, vec_uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, out_r, out_g,
+             out_b);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 8;
+    src_vu += 8;
+  }
+}
+
+void SobelRow_LSX(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, tmp0;
+  __m128i out0, out1, out2, out3;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i shuff0 = {0x1001010110000000, 0x1003030310020202};
+  __m128i shuff1 = __lsx_vaddi_bu(shuff0, 0x04);
+  __m128i shuff2 = __lsx_vaddi_bu(shuff1, 0x04);
+  __m128i shuff3 = __lsx_vaddi_bu(shuff2, 0x04);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_sobelx, 0);
+    src1 = __lsx_vld(src_sobely, 0);
+    tmp0 = __lsx_vsadd_bu(src0, src1);
+    DUP4_ARG3(__lsx_vshuf_b, alpha, tmp0, shuff0, alpha, tmp0, shuff1, alpha,
+              tmp0, shuff2, alpha, tmp0, shuff3, out0, out1, out2, out3);
+    __lsx_vst(out0, dst_argb, 0);
+    __lsx_vst(out1, dst_argb, 16);
+    __lsx_vst(out2, dst_argb, 32);
+    __lsx_vst(out3, dst_argb, 48);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void SobelToPlaneRow_LSX(const uint8_t* src_sobelx,
+                         const uint8_t* src_sobely,
+                         uint8_t* dst_y,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m128i src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_sobelx, 0, src_sobelx, 16, src0, src1);
+    DUP2_ARG2(__lsx_vld, src_sobely, 0, src_sobely, 16, src2, src3);
+    dst0 = __lsx_vsadd_bu(src0, src2);
+    dst1 = __lsx_vsadd_bu(src1, src3);
+    __lsx_vst(dst0, dst_y, 0);
+    __lsx_vst(dst1, dst_y, 16);
+    src_sobelx += 32;
+    src_sobely += 32;
+    dst_y += 32;
+  }
+}
+
+void SobelXYRow_LSX(const uint8_t* src_sobelx,
+                    const uint8_t* src_sobely,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  int len = width / 16;
+  __m128i src_r, src_b, src_g;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    src_r = __lsx_vld(src_sobelx, 0);
+    src_b = __lsx_vld(src_sobely, 0);
+    src_g = __lsx_vsadd_bu(src_r, src_b);
+    tmp0 = __lsx_vilvl_b(src_g, src_b);
+    tmp1 = __lsx_vilvh_b(src_g, src_b);
+    tmp2 = __lsx_vilvl_b(alpha, src_r);
+    tmp3 = __lsx_vilvh_b(alpha, src_r);
+    dst0 = __lsx_vilvl_h(tmp2, tmp0);
+    dst1 = __lsx_vilvh_h(tmp2, tmp0);
+    dst2 = __lsx_vilvl_h(tmp3, tmp1);
+    dst3 = __lsx_vilvh_h(tmp3, tmp1);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void BGRAToUVRow_LSX(const uint8_t* src_bgra,
+                     int src_stride_bgra,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_bgra = src_bgra + src_stride_bgra;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_bgra, 0, src_bgra, 16, src_bgra, 32, src_bgra, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_bgra, 0, next_bgra, 16, next_bgra, 32, next_bgra,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickev_b(src1, src0);
+    tmp2 = __lsx_vpickod_b(src3, src2);
+    tmp3 = __lsx_vpickev_b(src3, src2);
+    tmpb = __lsx_vpickod_b(tmp2, tmp0);
+    tmpr = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickod_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickod_b(nex1, nex0);
+    tmp1 = __lsx_vpickev_b(nex1, nex0);
+    tmp2 = __lsx_vpickod_b(nex3, nex2);
+    tmp3 = __lsx_vpickev_b(nex3, nex2);
+    nexb = __lsx_vpickod_b(tmp2, tmp0);
+    nexr = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickod_b(tmp3, tmp1);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_bgra += 64;
+    next_bgra += 64;
+  }
+}
+
+void ABGRToUVRow_LSX(const uint8_t* src_abgr,
+                     int src_stride_abgr,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_abgr = src_abgr + src_stride_abgr;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_abgr, 0, src_abgr, 16, src_abgr, 32, src_abgr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_abgr, 0, next_abgr, 16, next_abgr, 32, next_abgr,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmp2 = __lsx_vpickev_b(src3, src2);
+    tmp3 = __lsx_vpickod_b(src3, src2);
+    tmpb = __lsx_vpickod_b(tmp2, tmp0);
+    tmpr = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickev_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickev_b(nex1, nex0);
+    tmp1 = __lsx_vpickod_b(nex1, nex0);
+    tmp2 = __lsx_vpickev_b(nex3, nex2);
+    tmp3 = __lsx_vpickod_b(nex3, nex2);
+    nexb = __lsx_vpickod_b(tmp2, tmp0);
+    nexr = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickev_b(tmp3, tmp1);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_abgr += 64;
+    next_abgr += 64;
+  }
+}
+
+void RGBAToUVRow_LSX(const uint8_t* src_rgba,
+                     int src_stride_rgba,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_rgba = src_rgba + src_stride_rgba;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_rgba, 0, src_rgba, 16, src_rgba, 32, src_rgba, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_rgba, 0, next_rgba, 16, next_rgba, 32, next_rgba,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickev_b(src1, src0);
+    tmp2 = __lsx_vpickod_b(src3, src2);
+    tmp3 = __lsx_vpickev_b(src3, src2);
+    tmpr = __lsx_vpickod_b(tmp2, tmp0);
+    tmpb = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickod_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickod_b(nex1, nex0);
+    tmp1 = __lsx_vpickev_b(nex1, nex0);
+    tmp2 = __lsx_vpickod_b(nex3, nex2);
+    tmp3 = __lsx_vpickev_b(nex3, nex2);
+    nexr = __lsx_vpickod_b(tmp2, tmp0);
+    nexb = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickod_b(tmp3, tmp1);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_rgba += 64;
+    next_rgba += 64;
+  }
+}
+
+void ARGBToUVJRow_LSX(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  const uint8_t* next_argb = src_argb + src_stride_argb;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_63 = __lsx_vldi(0x43F);
+  __m128i const_42 = __lsx_vldi(0x42A);
+  __m128i const_21 = __lsx_vldi(0x415);
+  __m128i const_53 = __lsx_vldi(0x435);
+  __m128i const_10 = __lsx_vldi(0x40A);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb, 32, src_argb, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_argb, 0, next_argb, 16, next_argb, 32, next_argb,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmp2 = __lsx_vpickev_b(src3, src2);
+    tmp3 = __lsx_vpickod_b(src3, src2);
+    tmpr = __lsx_vpickod_b(tmp2, tmp0);
+    tmpb = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickev_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickev_b(nex1, nex0);
+    tmp1 = __lsx_vpickod_b(nex1, nex0);
+    tmp2 = __lsx_vpickev_b(nex3, nex2);
+    tmp3 = __lsx_vpickod_b(nex3, nex2);
+    nexr = __lsx_vpickod_b(tmp2, tmp0);
+    nexb = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickev_b(tmp3, tmp1);
+    tmp0 = __lsx_vaddwev_h_bu(tmpb, nexb);
+    tmp1 = __lsx_vaddwod_h_bu(tmpb, nexb);
+    tmp2 = __lsx_vaddwev_h_bu(tmpg, nexg);
+    tmp3 = __lsx_vaddwod_h_bu(tmpg, nexg);
+    reg0 = __lsx_vaddwev_h_bu(tmpr, nexr);
+    reg1 = __lsx_vaddwod_h_bu(tmpr, nexr);
+    tmpb = __lsx_vavgr_hu(tmp0, tmp1);
+    tmpg = __lsx_vavgr_hu(tmp2, tmp3);
+    tmpr = __lsx_vavgr_hu(reg0, reg1);
+    reg0 = __lsx_vmadd_h(const_8080, const_63, tmpb);
+    reg1 = __lsx_vmadd_h(const_8080, const_63, tmpr);
+    reg0 = __lsx_vmsub_h(reg0, const_42, tmpg);
+    reg1 = __lsx_vmsub_h(reg1, const_53, tmpg);
+    reg0 = __lsx_vmsub_h(reg0, const_21, tmpr);
+    reg1 = __lsx_vmsub_h(reg1, const_10, tmpb);
+    dst0 = __lsx_vpickod_b(reg1, reg0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_argb += 64;
+    next_argb += 64;
+  }
+}
+
+void I444ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 16;
+  __m128i vec_y, vec_u, vec_v, out_b, out_g, out_r;
+  __m128i vec_yl, vec_yh, vec_ul, vec_vl, vec_uh, vec_vh;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb, vec_ugvg;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_ugvg = __lsx_vilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_u = __lsx_vld(src_u, 0);
+    vec_v = __lsx_vld(src_v, 0);
+    vec_yl = __lsx_vilvl_b(vec_y, vec_y);
+    vec_ul = __lsx_vilvl_b(zero, vec_u);
+    vec_vl = __lsx_vilvl_b(zero, vec_v);
+    I444TORGB(vec_yl, vec_ul, vec_vl, vec_ub, vec_vr, vec_ugvg, vec_yg, vec_yb,
+              out_b, out_g, out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    vec_yh = __lsx_vilvh_b(vec_y, vec_y);
+    vec_uh = __lsx_vilvh_b(zero, vec_u);
+    vec_vh = __lsx_vilvh_b(zero, vec_v);
+    I444TORGB(vec_yh, vec_uh, vec_vh, vec_ub, vec_vr, vec_ugvg, vec_yg, vec_yb,
+              out_b, out_g, out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 16;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I400ToARGBRow_LSX(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 16;
+  __m128i vec_y, vec_yl, vec_yh, out0;
+  __m128i y_ev, y_od, dst0, dst1, dst2, dst3;
+  __m128i temp0, temp1;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i vec_yg = __lsx_vreplgr2vr_h(yuvconstants->kYToRgb[0]);
+  __m128i vec_yb = __lsx_vreplgr2vr_w(yuvconstants->kYBiasToRgb[0]);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_yl = __lsx_vilvl_b(vec_y, vec_y);
+    y_ev = __lsx_vmulwev_w_hu_h(vec_yl, vec_yg);
+    y_od = __lsx_vmulwod_w_hu_h(vec_yl, vec_yg);
+    y_ev = __lsx_vsrai_w(y_ev, 16);
+    y_od = __lsx_vsrai_w(y_od, 16);
+    y_ev = __lsx_vadd_w(y_ev, vec_yb);
+    y_od = __lsx_vadd_w(y_od, vec_yb);
+    y_ev = __lsx_vsrai_w(y_ev, 6);
+    y_od = __lsx_vsrai_w(y_od, 6);
+    y_ev = __lsx_vclip255_w(y_ev);
+    y_od = __lsx_vclip255_w(y_od);
+    out0 = __lsx_vpackev_h(y_od, y_ev);
+    temp0 = __lsx_vpackev_b(out0, out0);
+    temp1 = __lsx_vpackev_b(alpha, out0);
+    dst0 = __lsx_vilvl_h(temp1, temp0);
+    dst1 = __lsx_vilvh_h(temp1, temp0);
+    vec_yh = __lsx_vilvh_b(vec_y, vec_y);
+    y_ev = __lsx_vmulwev_w_hu_h(vec_yh, vec_yg);
+    y_od = __lsx_vmulwod_w_hu_h(vec_yh, vec_yg);
+    y_ev = __lsx_vsrai_w(y_ev, 16);
+    y_od = __lsx_vsrai_w(y_od, 16);
+    y_ev = __lsx_vadd_w(y_ev, vec_yb);
+    y_od = __lsx_vadd_w(y_od, vec_yb);
+    y_ev = __lsx_vsrai_w(y_ev, 6);
+    y_od = __lsx_vsrai_w(y_od, 6);
+    y_ev = __lsx_vclip255_w(y_ev);
+    y_od = __lsx_vclip255_w(y_od);
+    out0 = __lsx_vpackev_h(y_od, y_ev);
+    temp0 = __lsx_vpackev_b(out0, out0);
+    temp1 = __lsx_vpackev_b(alpha, out0);
+    dst2 = __lsx_vilvl_h(temp1, temp0);
+    dst3 = __lsx_vilvh_h(temp1, temp0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_y += 16;
+  }
+}
+
+void J400ToARGBRow_LSX(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m128i vec_y, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    tmp0 = __lsx_vilvl_b(vec_y, vec_y);
+    tmp1 = __lsx_vilvh_b(vec_y, vec_y);
+    tmp2 = __lsx_vilvl_b(alpha, vec_y);
+    tmp3 = __lsx_vilvh_b(alpha, vec_y);
+    dst0 = __lsx_vilvl_h(tmp2, tmp0);
+    dst1 = __lsx_vilvh_h(tmp2, tmp0);
+    dst2 = __lsx_vilvl_h(tmp3, tmp1);
+    dst3 = __lsx_vilvh_h(tmp3, tmp1);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_y += 16;
+  }
+}
+
+void YUY2ToARGBRow_LSX(const uint8_t* src_yuy2,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i zero = __lsx_vldi(0);
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_yuy2, 0);
+    vec_y = __lsx_vpickev_b(src0, src0);
+    vec_vu = __lsx_vpickod_b(src0, src0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_yuy2 += 16;
+  }
+}
+
+void UYVYToARGBRow_LSX(const uint8_t* src_uyvy,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i zero = __lsx_vldi(0);
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_uyvy, 0);
+    vec_y = __lsx_vpickod_b(src0, src0);
+    vec_vu = __lsx_vpickev_b(src0, src0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_uyvy += 16;
+  }
+}
+
+void InterpolateRow_LSX(uint8_t* dst_ptr,
+                        const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int32_t source_y_fraction) {
+  int x;
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8_t* nex_ptr = src_ptr + src_stride;
+  uint16_t y_fractions;
+  int len = width / 32;
+  __m128i src0, src1, nex0, nex1;
+  __m128i dst0, dst1, y_frac;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i const_128 = __lsx_vldi(0x480);
+
+  if (y1_fraction == 0) {
+    for (x = 0; x < len; x++) {
+      DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+      __lsx_vst(src0, dst_ptr, 0);
+      __lsx_vst(src1, dst_ptr, 16);
+      src_ptr += 32;
+      dst_ptr += 32;
+    }
+    return;
+  }
+
+  if (y1_fraction == 128) {
+    for (x = 0; x < len; x++) {
+      DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+      DUP2_ARG2(__lsx_vld, nex_ptr, 0, nex_ptr, 16, nex0, nex1);
+      dst0 = __lsx_vavgr_bu(src0, nex0);
+      dst1 = __lsx_vavgr_bu(src1, nex1);
+      __lsx_vst(dst0, dst_ptr, 0);
+      __lsx_vst(dst1, dst_ptr, 16);
+      src_ptr += 32;
+      nex_ptr += 32;
+      dst_ptr += 32;
+    }
+    return;
+  }
+
+  y_fractions = (uint16_t)(y0_fraction + (y1_fraction << 8));
+  y_frac = __lsx_vreplgr2vr_h(y_fractions);
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+    DUP2_ARG2(__lsx_vld, nex_ptr, 0, nex_ptr, 16, nex0, nex1);
+    tmp0 = __lsx_vilvl_b(nex0, src0);
+    tmp1 = __lsx_vilvh_b(nex0, src0);
+    tmp2 = __lsx_vilvl_b(nex1, src1);
+    tmp3 = __lsx_vilvh_b(nex1, src1);
+    tmp0 = __lsx_vdp2add_h_bu(const_128, tmp0, y_frac);
+    tmp1 = __lsx_vdp2add_h_bu(const_128, tmp1, y_frac);
+    tmp2 = __lsx_vdp2add_h_bu(const_128, tmp2, y_frac);
+    tmp3 = __lsx_vdp2add_h_bu(const_128, tmp3, y_frac);
+    dst0 = __lsx_vsrlni_b_h(tmp1, tmp0, 8);
+    dst1 = __lsx_vsrlni_b_h(tmp3, tmp2, 8);
+    __lsx_vst(dst0, dst_ptr, 0);
+    __lsx_vst(dst1, dst_ptr, 16);
+    src_ptr += 32;
+    nex_ptr += 32;
+    dst_ptr += 32;
+  }
+}
+
+void ARGBSetRow_LSX(uint8_t* dst_argb, uint32_t v32, int width) {
+  int x;
+  int len = width / 4;
+  __m128i dst0 = __lsx_vreplgr2vr_w(v32);
+
+  for (x = 0; x < len; x++) {
+    __lsx_vst(dst0, dst_argb, 0);
+    dst_argb += 16;
+  }
+}
+
+void RAWToRGB24Row_LSX(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i dst0, dst1, dst2;
+  __m128i shuf0 = {0x0708030405000102, 0x110C0D0E090A0B06};
+  __m128i shuf1 = {0x1516171213140F10, 0x1F1E1B1C1D18191A};
+  __m128i shuf2 = {0x090405060102031E, 0x0D0E0F0A0B0C0708};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_raw, 0, src_raw, 16, src0, src1);
+    src2 = __lsx_vld(src_raw, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuf0, src1, src0, shuf1, dst0, dst1);
+    dst2 = __lsx_vshuf_b(src1, src2, shuf2);
+    dst1 = __lsx_vinsgr2vr_b(dst1, src_raw[32], 0x0E);
+    __lsx_vst(dst0, dst_rgb24, 0);
+    __lsx_vst(dst1, dst_rgb24, 16);
+    __lsx_vst(dst2, dst_rgb24, 32);
+    dst_rgb24 += 48;
+    src_raw += 48;
+  }
+}
+
+void MergeUVRow_LSX(const uint8_t* src_u,
+                    const uint8_t* src_v,
+                    uint8_t* dst_uv,
+                    int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_u, 0, src_v, 0, src0, src1);
+    dst0 = __lsx_vilvl_b(src1, src0);
+    dst1 = __lsx_vilvh_b(src1, src0);
+    __lsx_vst(dst0, dst_uv, 0);
+    __lsx_vst(dst1, dst_uv, 16);
+    src_u += 16;
+    src_v += 16;
+    dst_uv += 32;
+  }
+}
+
+void ARGBExtractAlphaRow_LSX(const uint8_t* src_argb,
+                             uint8_t* dst_a,
+                             int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, tmp0, tmp1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb, 32, src_argb, 48,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src3, src2);
+    dst0 = __lsx_vpickod_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst_a, 0);
+    src_argb += 64;
+    dst_a += 16;
+  }
+}
+
+void ARGBBlendRow_LSX(const uint8_t* src_argb,
+                      const uint8_t* src_argb1,
+                      uint8_t* dst_argb,
+                      int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, dst0, dst1;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i a0, a1, a2, a3;
+  __m128i const_256 = __lsx_vldi(0x500);
+  __m128i zero = __lsx_vldi(0);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i control = {0xFF000000FF000000, 0xFF000000FF000000};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb1, 0, src_argb1, 16,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vshuf4i_b(src0, 0xFF);
+    tmp1 = __lsx_vshuf4i_b(src1, 0xFF);
+    a0 = __lsx_vilvl_b(zero, tmp0);
+    a1 = __lsx_vilvh_b(zero, tmp0);
+    a2 = __lsx_vilvl_b(zero, tmp1);
+    a3 = __lsx_vilvh_b(zero, tmp1);
+    reg0 = __lsx_vilvl_b(zero, src2);
+    reg1 = __lsx_vilvh_b(zero, src2);
+    reg2 = __lsx_vilvl_b(zero, src3);
+    reg3 = __lsx_vilvh_b(zero, src3);
+    DUP4_ARG2(__lsx_vsub_h, const_256, a0, const_256, a1, const_256, a2,
+              const_256, a3, a0, a1, a2, a3);
+    DUP4_ARG2(__lsx_vmul_h, a0, reg0, a1, reg1, a2, reg2, a3, reg3, reg0, reg1,
+              reg2, reg3);
+    DUP2_ARG3(__lsx_vsrani_b_h, reg1, reg0, 8, reg3, reg2, 8, dst0, dst1);
+    dst0 = __lsx_vsadd_bu(dst0, src0);
+    dst1 = __lsx_vsadd_bu(dst1, src1);
+    dst0 = __lsx_vbitsel_v(dst0, alpha, control);
+    dst1 = __lsx_vbitsel_v(dst1, alpha, control);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    src_argb += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBQuantizeRow_LSX(uint8_t* dst_argb,
+                         int scale,
+                         int interval_size,
+                         int interval_offset,
+                         int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i vec_size = __lsx_vreplgr2vr_b(interval_size);
+  __m128i vec_offset = __lsx_vreplgr2vr_b(interval_offset);
+  __m128i vec_scale = __lsx_vreplgr2vr_w(scale);
+  __m128i zero = __lsx_vldi(0);
+  __m128i control = {0xFF000000FF000000, 0xFF000000FF000000};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, dst_argb, 0, dst_argb, 16, dst_argb, 32, dst_argb, 48,
+              src0, src1, src2, src3);
+    reg0 = __lsx_vilvl_b(zero, src0);
+    reg1 = __lsx_vilvh_b(zero, src0);
+    reg2 = __lsx_vilvl_b(zero, src1);
+    reg3 = __lsx_vilvh_b(zero, src1);
+    reg4 = __lsx_vilvl_b(zero, src2);
+    reg5 = __lsx_vilvh_b(zero, src2);
+    reg6 = __lsx_vilvl_b(zero, src3);
+    reg7 = __lsx_vilvh_b(zero, src3);
+    tmp0 = __lsx_vilvl_h(zero, reg0);
+    tmp1 = __lsx_vilvh_h(zero, reg0);
+    tmp2 = __lsx_vilvl_h(zero, reg1);
+    tmp3 = __lsx_vilvh_h(zero, reg1);
+    tmp4 = __lsx_vilvl_h(zero, reg2);
+    tmp5 = __lsx_vilvh_h(zero, reg2);
+    tmp6 = __lsx_vilvl_h(zero, reg3);
+    tmp7 = __lsx_vilvh_h(zero, reg3);
+    DUP4_ARG2(__lsx_vmul_w, tmp0, vec_scale, tmp1, vec_scale, tmp2, vec_scale,
+              tmp3, vec_scale, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vmul_w, tmp4, vec_scale, tmp5, vec_scale, tmp6, vec_scale,
+              tmp7, vec_scale, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vsrani_h_w, tmp1, tmp0, 16, tmp3, tmp2, 16, tmp5, tmp4, 16,
+              tmp7, tmp6, 16, reg0, reg1, reg2, reg3);
+    dst0 = __lsx_vpickev_b(reg1, reg0);
+    dst1 = __lsx_vpickev_b(reg3, reg2);
+    tmp0 = __lsx_vilvl_h(zero, reg4);
+    tmp1 = __lsx_vilvh_h(zero, reg4);
+    tmp2 = __lsx_vilvl_h(zero, reg5);
+    tmp3 = __lsx_vilvh_h(zero, reg5);
+    tmp4 = __lsx_vilvl_h(zero, reg6);
+    tmp5 = __lsx_vilvh_h(zero, reg6);
+    tmp6 = __lsx_vilvl_h(zero, reg7);
+    tmp7 = __lsx_vilvh_h(zero, reg7);
+    DUP4_ARG2(__lsx_vmul_w, tmp0, vec_scale, tmp1, vec_scale, tmp2, vec_scale,
+              tmp3, vec_scale, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vmul_w, tmp4, vec_scale, tmp5, vec_scale, tmp6, vec_scale,
+              tmp7, vec_scale, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vsrani_h_w, tmp1, tmp0, 16, tmp3, tmp2, 16, tmp5, tmp4, 16,
+              tmp7, tmp6, 16, reg0, reg1, reg2, reg3);
+    dst2 = __lsx_vpickev_b(reg1, reg0);
+    dst3 = __lsx_vpickev_b(reg3, reg2);
+    DUP4_ARG2(__lsx_vmul_b, dst0, vec_size, dst1, vec_size, dst2, vec_size,
+              dst3, vec_size, dst0, dst1, dst2, dst3);
+    DUP4_ARG2(__lsx_vadd_b, dst0, vec_offset, dst1, vec_offset, dst2,
+              vec_offset, dst3, vec_offset, dst0, dst1, dst2, dst3);
+    DUP4_ARG3(__lsx_vbitsel_v, dst0, src0, control, dst1, src1, control, dst2,
+              src2, control, dst3, src3, control, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+  }
+}
+
+void ARGBColorMatrixRow_LSX(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const int8_t* matrix_argb,
+                            int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, src1, tmp0, tmp1, dst0, dst1;
+  __m128i tmp_b, tmp_g, tmp_r, tmp_a;
+  __m128i reg_b, reg_g, reg_r, reg_a;
+  __m128i matrix_b = __lsx_vldrepl_w(matrix_argb, 0);
+  __m128i matrix_g = __lsx_vldrepl_w(matrix_argb, 4);
+  __m128i matrix_r = __lsx_vldrepl_w(matrix_argb, 8);
+  __m128i matrix_a = __lsx_vldrepl_w(matrix_argb, 12);
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src0, src1);
+    DUP4_ARG2(__lsx_vdp2_h_bu_b, src0, matrix_b, src0, matrix_g, src0, matrix_r,
+              src0, matrix_a, tmp_b, tmp_g, tmp_r, tmp_a);
+    DUP4_ARG2(__lsx_vdp2_h_bu_b, src1, matrix_b, src1, matrix_g, src1, matrix_r,
+              src1, matrix_a, reg_b, reg_g, reg_r, reg_a);
+    DUP4_ARG2(__lsx_vhaddw_w_h, tmp_b, tmp_b, tmp_g, tmp_g, tmp_r, tmp_r, tmp_a,
+              tmp_a, tmp_b, tmp_g, tmp_r, tmp_a);
+    DUP4_ARG2(__lsx_vhaddw_w_h, reg_b, reg_b, reg_g, reg_g, reg_r, reg_r, reg_a,
+              reg_a, reg_b, reg_g, reg_r, reg_a);
+    DUP4_ARG2(__lsx_vsrai_w, tmp_b, 6, tmp_g, 6, tmp_r, 6, tmp_a, 6, tmp_b,
+              tmp_g, tmp_r, tmp_a);
+    DUP4_ARG2(__lsx_vsrai_w, reg_b, 6, reg_g, 6, reg_r, 6, reg_a, 6, reg_b,
+              reg_g, reg_r, reg_a);
+    DUP4_ARG1(__lsx_vclip255_w, tmp_b, tmp_g, tmp_r, tmp_a, tmp_b, tmp_g, tmp_r,
+              tmp_a)
+    DUP4_ARG1(__lsx_vclip255_w, reg_b, reg_g, reg_r, reg_a, reg_b, reg_g, reg_r,
+              reg_a)
+    DUP4_ARG2(__lsx_vpickev_h, reg_b, tmp_b, reg_g, tmp_g, reg_r, tmp_r, reg_a,
+              tmp_a, tmp_b, tmp_g, tmp_r, tmp_a);
+    tmp0 = __lsx_vpackev_b(tmp_g, tmp_b);
+    tmp1 = __lsx_vpackev_b(tmp_a, tmp_r);
+    dst0 = __lsx_vilvl_h(tmp1, tmp0);
+    dst1 = __lsx_vilvh_h(tmp1, tmp0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void SplitUVRow_LSX(const uint8_t* src_uv,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  int x;
+  int len = width / 32;
+  __m128i src0, src1, src2, src3;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_uv, 0, src_uv, 16, src_uv, 32, src_uv, 48, src0,
+              src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, dst0, dst1);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, dst2, dst3);
+    __lsx_vst(dst0, dst_u, 0);
+    __lsx_vst(dst1, dst_u, 16);
+    __lsx_vst(dst2, dst_v, 0);
+    __lsx_vst(dst3, dst_v, 16);
+    src_uv += 64;
+    dst_u += 32;
+    dst_v += 32;
+  }
+}
+
+void SetRow_LSX(uint8_t* dst, uint8_t v8, int width) {
+  int x;
+  int len = width / 16;
+  __m128i dst0 = __lsx_vreplgr2vr_b(v8);
+
+  for (x = 0; x < len; x++) {
+    __lsx_vst(dst0, dst, 0);
+    dst += 16;
+  }
+}
+
+void MirrorSplitUVRow_LSX(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  int x;
+  int len = width / 32;
+  __m128i src0, src1, src2, src3;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i shuff0 = {0x10121416181A1C1E, 0x00020406080A0C0E};
+  __m128i shuff1 = {0x11131517191B1D1F, 0x01030507090B0D0F};
+
+  src_uv += (width << 1);
+  for (x = 0; x < len; x++) {
+    src_uv -= 64;
+    DUP4_ARG2(__lsx_vld, src_uv, 0, src_uv, 16, src_uv, 32, src_uv, 48, src2,
+              src3, src0, src1);
+    DUP4_ARG3(__lsx_vshuf_b, src1, src0, shuff1, src3, src2, shuff1, src1, src0,
+              shuff0, src3, src2, shuff0, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_v, 0);
+    __lsx_vst(dst1, dst_v, 16);
+    __lsx_vst(dst2, dst_u, 0);
+    __lsx_vst(dst3, dst_u, 16);
+    dst_u += 32;
+    dst_v += 32;
+  }
+}
+
+void HalfFloatRow_LSX(const uint16_t* src,
+                      uint16_t* dst,
+                      float scale,
+                      int width) {
+  int x;
+  int len = width / 32;
+  float mult = 1.9259299444e-34f * scale;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128 vec_mult = (__m128)__lsx_vldrepl_w(&mult, 0);
+  __m128i zero = __lsx_vldi(0);
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src, 0, src, 16, src, 32, src, 48, src0, src1, src2,
+              src3);
+    DUP4_ARG2(__lsx_vilvl_h, zero, src0, zero, src1, zero, src2, zero, src3,
+              tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vilvh_h, zero, src0, zero, src1, zero, src2, zero, src3,
+              tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG1(__lsx_vffint_s_wu, tmp0, tmp2, tmp4, tmp6, reg0, reg2, reg4,
+              reg6);
+    DUP4_ARG1(__lsx_vffint_s_wu, tmp1, tmp3, tmp5, tmp7, reg1, reg3, reg5,
+              reg7);
+    DUP4_ARG2(__lsx_vfmul_s, reg0, vec_mult, reg1, vec_mult, reg2, vec_mult,
+              reg3, vec_mult, reg0, reg1, reg2, reg3);
+    DUP4_ARG2(__lsx_vfmul_s, reg4, vec_mult, reg5, vec_mult, reg6, vec_mult,
+              reg7, vec_mult, reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vsrli_w, (v4u32)reg0, 13, (v4u32)reg1, 13, (v4u32)reg2, 13,
+              (v4u32)reg3, 13, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vsrli_w, (v4u32)reg4, 13, (v4u32)reg5, 13, (v4u32)reg6, 13,
+              (v4u32)reg7, 13, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG2(__lsx_vpickev_h, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
+              dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    __lsx_vst(dst2, dst, 32);
+    __lsx_vst(dst3, dst, 48);
+    src += 32;
+    dst += 32;
+  }
+}
+
+struct RgbConstants {
+  uint8_t kRGBToY[4];
+  uint16_t kAddY;
+  uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+                                                        128,
+                                                        0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+                                                        0x1080,
+                                                        0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+                                                      0x1080,
+                                                      0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+static void ARGBToYMatrixRow_LSX(const uint8_t* src_argb,
+                                 uint8_t* dst_y,
+                                 int width,
+                                 const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vldrepl.b      $vr0,  %3,    0             \n\t"  // load rgbconstants
+      "vldrepl.b      $vr1,  %3,    1             \n\t"  // load rgbconstants
+      "vldrepl.b      $vr2,  %3,    2             \n\t"  // load rgbconstants
+      "vldrepl.h      $vr3,  %3,    4             \n\t"  // load rgbconstants
+      "1:                                         \n\t"
+      "vld            $vr4,  %0,    0             \n\t"
+      "vld            $vr5,  %0,    16            \n\t"
+      "vld            $vr6,  %0,    32            \n\t"
+      "vld            $vr7,  %0,    48            \n\t"  // load 16 pixels of
+                                                         // ARGB
+      "vor.v          $vr12, $vr3,  $vr3          \n\t"
+      "vor.v          $vr13, $vr3,  $vr3          \n\t"
+      "addi.d         %2,    %2,    -16           \n\t"  // 16 processed per
+                                                         // loop.
+      "vpickev.b      $vr8,  $vr5,  $vr4          \n\t"  // BR
+      "vpickev.b      $vr10, $vr7,  $vr6          \n\t"
+      "vpickod.b      $vr9,  $vr5,  $vr4          \n\t"  // GA
+      "vpickod.b      $vr11, $vr7,  $vr6          \n\t"
+      "vmaddwev.h.bu  $vr12, $vr8,  $vr0          \n\t"  // B
+      "vmaddwev.h.bu  $vr13, $vr10, $vr0          \n\t"
+      "vmaddwev.h.bu  $vr12, $vr9,  $vr1          \n\t"  // G
+      "vmaddwev.h.bu  $vr13, $vr11, $vr1          \n\t"
+      "vmaddwod.h.bu  $vr12, $vr8,  $vr2          \n\t"  // R
+      "vmaddwod.h.bu  $vr13, $vr10, $vr2          \n\t"
+      "addi.d         %0,    %0,    64            \n\t"
+      "vpickod.b      $vr10, $vr13, $vr12         \n\t"
+      "vst            $vr10, %1,    0             \n\t"
+      "addi.d         %1,    %1,    16            \n\t"
+      "bnez           %2,    1b                   \n\t"
+      : "+&r"(src_argb),  // %0
+        "+&r"(dst_y),     // %1
+        "+&r"(width)      // %2
+      : "r"(rgbconstants)
+      : "memory");
+}
+
+void ARGBToYRow_LSX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_LSX(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_LSX(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_LSX(src_argb, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void ABGRToYRow_LSX(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_LSX(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+void ABGRToYJRow_LSX(const uint8_t* src_abgr, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_LSX(src_abgr, dst_yj, width, &kRawJPEGConstants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+static void RGBAToYMatrixRow_LSX(const uint8_t* src_rgba,
+                                 uint8_t* dst_y,
+                                 int width,
+                                 const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vldrepl.b      $vr0,  %3,    0             \n\t"  // load rgbconstants
+      "vldrepl.b      $vr1,  %3,    1             \n\t"  // load rgbconstants
+      "vldrepl.b      $vr2,  %3,    2             \n\t"  // load rgbconstants
+      "vldrepl.h      $vr3,  %3,    4             \n\t"  // load rgbconstants
+      "1:                                         \n\t"
+      "vld            $vr4,  %0,    0             \n\t"
+      "vld            $vr5,  %0,    16            \n\t"
+      "vld            $vr6,  %0,    32            \n\t"
+      "vld            $vr7,  %0,    48            \n\t"  // load 16 pixels of
+                                                         // RGBA
+      "vor.v          $vr12, $vr3,  $vr3          \n\t"
+      "vor.v          $vr13, $vr3,  $vr3          \n\t"
+      "addi.d         %2,    %2,    -16           \n\t"  // 16 processed per
+                                                         // loop.
+      "vpickev.b      $vr8,  $vr5,  $vr4          \n\t"  // AG
+      "vpickev.b      $vr10, $vr7,  $vr6          \n\t"
+      "vpickod.b      $vr9,  $vr5,  $vr4          \n\t"  // BR
+      "vpickod.b      $vr11, $vr7,  $vr6          \n\t"
+      "vmaddwev.h.bu  $vr12, $vr9,  $vr0          \n\t"  // B
+      "vmaddwev.h.bu  $vr13, $vr11, $vr0          \n\t"
+      "vmaddwod.h.bu  $vr12, $vr8,  $vr1          \n\t"  // G
+      "vmaddwod.h.bu  $vr13, $vr10, $vr1          \n\t"
+      "vmaddwod.h.bu  $vr12, $vr9,  $vr2          \n\t"  // R
+      "vmaddwod.h.bu  $vr13, $vr11, $vr2          \n\t"
+      "addi.d         %0,    %0,    64            \n\t"
+      "vpickod.b      $vr10, $vr13, $vr12         \n\t"
+      "vst            $vr10, %1,    0             \n\t"
+      "addi.d         %1,    %1,    16            \n\t"
+      "bnez           %2,    1b                   \n\t"
+      : "+&r"(src_rgba),  // %0
+        "+&r"(dst_y),     // %1
+        "+&r"(width)      // %2
+      : "r"(rgbconstants)
+      : "memory");
+}
+
+void RGBAToYRow_LSX(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_LSX(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_LSX(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+  RGBAToYMatrixRow_LSX(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_LSX(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_LSX(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+static void RGBToYMatrixRow_LSX(const uint8_t* src_rgba,
+                                uint8_t* dst_y,
+                                int width,
+                                const struct RgbConstants* rgbconstants) {
+  int8_t shuff[64] = {0,  2,  3,  5,  6,  8,  9,  11, 12, 14, 15, 17, 18,
+                      20, 21, 23, 24, 26, 27, 29, 30, 0,  1,  3,  4,  6,
+                      7,  9,  10, 12, 13, 15, 1,  0,  4,  0,  7,  0,  10,
+                      0,  13, 0,  16, 0,  19, 0,  22, 0,  25, 0,  28, 0,
+                      31, 0,  2,  0,  5,  0,  8,  0,  11, 0,  14, 0};
+  asm volatile(
+      "vldrepl.b      $vr0,  %3,    0             \n\t"  // load rgbconstants
+      "vldrepl.b      $vr1,  %3,    1             \n\t"  // load rgbconstants
+      "vldrepl.b      $vr2,  %3,    2             \n\t"  // load rgbconstants
+      "vldrepl.h      $vr3,  %3,    4             \n\t"  // load rgbconstants
+      "vld            $vr4,  %4,    0             \n\t"  // load shuff
+      "vld            $vr5,  %4,    16            \n\t"
+      "vld            $vr6,  %4,    32            \n\t"
+      "vld            $vr7,  %4,    48            \n\t"
+      "1:                                         \n\t"
+      "vld            $vr8,  %0,    0             \n\t"
+      "vld            $vr9,  %0,    16            \n\t"
+      "vld            $vr10, %0,    32            \n\t"  // load 16 pixels of
+                                                         // RGB
+      "vor.v          $vr12, $vr3,  $vr3          \n\t"
+      "vor.v          $vr13, $vr3,  $vr3          \n\t"
+      "addi.d         %2,    %2,    -16           \n\t"  // 16 processed per
+                                                         // loop.
+      "vshuf.b        $vr14, $vr9,  $vr8,  $vr4   \n\t"
+      "vshuf.b        $vr15, $vr9,  $vr10, $vr5   \n\t"
+      "vshuf.b        $vr16, $vr9,  $vr8,  $vr6   \n\t"
+      "vshuf.b        $vr17, $vr9,  $vr10, $vr7   \n\t"
+      "vmaddwev.h.bu  $vr12, $vr16, $vr1          \n\t"  // G
+      "vmaddwev.h.bu  $vr13, $vr17, $vr1          \n\t"
+      "vmaddwev.h.bu  $vr12, $vr14, $vr0          \n\t"  // B
+      "vmaddwev.h.bu  $vr13, $vr15, $vr0          \n\t"
+      "vmaddwod.h.bu  $vr12, $vr14, $vr2          \n\t"  // R
+      "vmaddwod.h.bu  $vr13, $vr15, $vr2          \n\t"
+      "addi.d         %0,    %0,    48            \n\t"
+      "vpickod.b      $vr10, $vr13, $vr12         \n\t"
+      "vst            $vr10, %1,    0             \n\t"
+      "addi.d         %1,    %1,    16            \n\t"
+      "bnez           %2,    1b                   \n\t"
+      : "+&r"(src_rgba),    // %0
+        "+&r"(dst_y),       // %1
+        "+&r"(width)        // %2
+      : "r"(rgbconstants),  // %3
+        "r"(shuff)          // %4
+      : "memory");
+}
+
+void RGB24ToYJRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_LSX(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void RAWToYJRow_LSX(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_LSX(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_LSX(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_LSX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_LSX(src_raw, dst_y, width, &kRawI601Constants);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
diff --git a/media/libyuv/source/row_mips.cc b/media/libyuv/source/row_mips.cc
deleted file mode 100644
index 285f0b5adc..0000000000
--- a/media/libyuv/source/row_mips.cc
+++ /dev/null
@@ -1,782 +0,0 @@
-/*
- *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// The following are available on Mips platforms:
-#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32)
-
-#ifdef HAS_COPYROW_MIPS
-void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
-  __asm__ __volatile__ (
-    ".set      noreorder                         \n"
-    ".set      noat                              \n"
-    "slti      $at, %[count], 8                  \n"
-    "bne       $at ,$zero, $last8                \n"
-    "xor       $t8, %[src], %[dst]               \n"
-    "andi      $t8, $t8, 0x3                     \n"
-
-    "bne       $t8, $zero, unaligned             \n"
-    "negu      $a3, %[dst]                       \n"
-    // make dst/src aligned
-    "andi      $a3, $a3, 0x3                     \n"
-    "beq       $a3, $zero, $chk16w               \n"
-    // word-aligned now count is the remining bytes count
-    "subu     %[count], %[count], $a3            \n"
-
-    "lwr       $t8, 0(%[src])                    \n"
-    "addu      %[src], %[src], $a3               \n"
-    "swr       $t8, 0(%[dst])                    \n"
-    "addu      %[dst], %[dst], $a3               \n"
-
-    // Now the dst/src are mutually word-aligned with word-aligned addresses
-    "$chk16w:                                    \n"
-    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
-    // t8 is the byte count after 64-byte chunks
-    "beq       %[count], $t8, chk8w              \n"
-    // There will be at most 1 32-byte chunk after it
-    "subu      $a3, %[count], $t8                \n"  // the reminder
-    // Here a3 counts bytes in 16w chunks
-    "addu      $a3, %[dst], $a3                  \n"
-    // Now a3 is the final dst after 64-byte chunks
-    "addu      $t0, %[dst], %[count]             \n"
-    // t0 is the "past the end" address
-
-    // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past
-    // the "t0-32" address
-    // This means: for x=128 the last "safe" a1 address is "t0-160"
-    // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
-    // we will use "pref 30,128(a1)", so "t0-160" is the limit
-    "subu      $t9, $t0, 160                     \n"
-    // t9 is the "last safe pref 30,128(a1)" address
-    "pref      0, 0(%[src])                      \n"  // first line of src
-    "pref      0, 32(%[src])                     \n"  // second line of src
-    "pref      0, 64(%[src])                     \n"
-    "pref      30, 32(%[dst])                    \n"
-    // In case the a1 > t9 don't use "pref 30" at all
-    "sgtu      $v1, %[dst], $t9                  \n"
-    "bgtz      $v1, $loop16w                     \n"
-    "nop                                         \n"
-    // otherwise, start with using pref30
-    "pref      30, 64(%[dst])                    \n"
-    "$loop16w:                                    \n"
-    "pref      0, 96(%[src])                     \n"
-    "lw        $t0, 0(%[src])                    \n"
-    "bgtz      $v1, $skip_pref30_96              \n"  // skip
-    "lw        $t1, 4(%[src])                    \n"
-    "pref      30, 96(%[dst])                    \n"  // continue
-    "$skip_pref30_96:                            \n"
-    "lw        $t2, 8(%[src])                    \n"
-    "lw        $t3, 12(%[src])                   \n"
-    "lw        $t4, 16(%[src])                   \n"
-    "lw        $t5, 20(%[src])                   \n"
-    "lw        $t6, 24(%[src])                   \n"
-    "lw        $t7, 28(%[src])                   \n"
-    "pref      0, 128(%[src])                    \n"
-    //  bring the next lines of src, addr 128
-    "sw        $t0, 0(%[dst])                    \n"
-    "sw        $t1, 4(%[dst])                    \n"
-    "sw        $t2, 8(%[dst])                    \n"
-    "sw        $t3, 12(%[dst])                   \n"
-    "sw        $t4, 16(%[dst])                   \n"
-    "sw        $t5, 20(%[dst])                   \n"
-    "sw        $t6, 24(%[dst])                   \n"
-    "sw        $t7, 28(%[dst])                   \n"
-    "lw        $t0, 32(%[src])                   \n"
-    "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
-    "lw        $t1, 36(%[src])                   \n"
-    "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
-    "$skip_pref30_128:                           \n"
-    "lw        $t2, 40(%[src])                   \n"
-    "lw        $t3, 44(%[src])                   \n"
-    "lw        $t4, 48(%[src])                   \n"
-    "lw        $t5, 52(%[src])                   \n"
-    "lw        $t6, 56(%[src])                   \n"
-    "lw        $t7, 60(%[src])                   \n"
-    "pref      0, 160(%[src])                    \n"
-    // bring the next lines of src, addr 160
-    "sw        $t0, 32(%[dst])                   \n"
-    "sw        $t1, 36(%[dst])                   \n"
-    "sw        $t2, 40(%[dst])                   \n"
-    "sw        $t3, 44(%[dst])                   \n"
-    "sw        $t4, 48(%[dst])                   \n"
-    "sw        $t5, 52(%[dst])                   \n"
-    "sw        $t6, 56(%[dst])                   \n"
-    "sw        $t7, 60(%[dst])                   \n"
-
-    "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
-    "sgtu      $v1, %[dst], $t9                  \n"
-    "bne       %[dst], $a3, $loop16w             \n"
-    " addiu    %[src], %[src], 64                \n"  // adding 64 to src
-    "move      %[count], $t8                     \n"
-
-    // Here we have src and dest word-aligned but less than 64-bytes to go
-
-    "chk8w:                                      \n"
-    "pref      0, 0x0(%[src])                    \n"
-    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
-    // the t8 is the reminder count past 32-bytes
-    "beq       %[count], $t8, chk1w              \n"
-    // count=t8,no 32-byte chunk
-    " nop                                        \n"
-
-    "lw        $t0, 0(%[src])                    \n"
-    "lw        $t1, 4(%[src])                    \n"
-    "lw        $t2, 8(%[src])                    \n"
-    "lw        $t3, 12(%[src])                   \n"
-    "lw        $t4, 16(%[src])                   \n"
-    "lw        $t5, 20(%[src])                   \n"
-    "lw        $t6, 24(%[src])                   \n"
-    "lw        $t7, 28(%[src])                   \n"
-    "addiu     %[src], %[src], 32                \n"
-
-    "sw        $t0, 0(%[dst])                    \n"
-    "sw        $t1, 4(%[dst])                    \n"
-    "sw        $t2, 8(%[dst])                    \n"
-    "sw        $t3, 12(%[dst])                   \n"
-    "sw        $t4, 16(%[dst])                   \n"
-    "sw        $t5, 20(%[dst])                   \n"
-    "sw        $t6, 24(%[dst])                   \n"
-    "sw        $t7, 28(%[dst])                   \n"
-    "addiu     %[dst], %[dst], 32                \n"
-
-    "chk1w:                                      \n"
-    "andi      %[count], $t8, 0x3                \n"
-    // now count is the reminder past 1w chunks
-    "beq       %[count], $t8, $last8             \n"
-    " subu     $a3, $t8, %[count]                \n"
-    // a3 is count of bytes in 1w chunks
-    "addu      $a3, %[dst], $a3                  \n"
-    // now a3 is the dst address past the 1w chunks
-    // copying in words (4-byte chunks)
-    "$wordCopy_loop:                             \n"
-    "lw        $t3, 0(%[src])                    \n"
-    // the first t3 may be equal t0 ... optimize?
-    "addiu     %[src], %[src],4                  \n"
-    "addiu     %[dst], %[dst],4                  \n"
-    "bne       %[dst], $a3,$wordCopy_loop        \n"
-    " sw       $t3, -4(%[dst])                   \n"
-
-    // For the last (<8) bytes
-    "$last8:                                     \n"
-    "blez      %[count], leave                   \n"
-    " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
-    "$last8loop:                                 \n"
-    "lb        $v1, 0(%[src])                    \n"
-    "addiu     %[src], %[src], 1                 \n"
-    "addiu     %[dst], %[dst], 1                 \n"
-    "bne       %[dst], $a3, $last8loop           \n"
-    " sb       $v1, -1(%[dst])                   \n"
-
-    "leave:                                      \n"
-    "  j       $ra                               \n"
-    "  nop                                       \n"
-
-    //
-    // UNALIGNED case
-    //
-
-    "unaligned:                                  \n"
-    // got here with a3="negu a1"
-    "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
-    "beqz      $a3, $ua_chk16w                   \n"
-    " subu     %[count], %[count], $a3           \n"
-    // bytes left after initial a3 bytes
-    "lwr       $v1, 0(%[src])                    \n"
-    "lwl       $v1, 3(%[src])                    \n"
-    "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
-    "swr       $v1, 0(%[dst])                    \n"
-    "addu      %[dst], %[dst], $a3               \n"
-    // below the dst will be word aligned (NOTE1)
-    "$ua_chk16w:                                 \n"
-    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
-    // t8 is the byte count after 64-byte chunks
-    "beq       %[count], $t8, ua_chk8w           \n"
-    // if a2==t8, no 64-byte chunks
-    // There will be at most 1 32-byte chunk after it
-    "subu      $a3, %[count], $t8                \n"  // the reminder
-    // Here a3 counts bytes in 16w chunks
-    "addu      $a3, %[dst], $a3                  \n"
-    // Now a3 is the final dst after 64-byte chunks
-    "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
-    "subu      $t9, $t0, 160                     \n"
-    // t9 is the "last safe pref 30,128(a1)" address
-    "pref      0, 0(%[src])                      \n"  // first line of src
-    "pref      0, 32(%[src])                     \n"  // second line  addr 32
-    "pref      0, 64(%[src])                     \n"
-    "pref      30, 32(%[dst])                    \n"
-    // safe, as we have at least 64 bytes ahead
-    // In case the a1 > t9 don't use "pref 30" at all
-    "sgtu      $v1, %[dst], $t9                  \n"
-    "bgtz      $v1, $ua_loop16w                  \n"
-    // skip "pref 30,64(a1)" for too short arrays
-    " nop                                        \n"
-    // otherwise, start with using pref30
-    "pref      30, 64(%[dst])                    \n"
-    "$ua_loop16w:                                \n"
-    "pref      0, 96(%[src])                     \n"
-    "lwr       $t0, 0(%[src])                    \n"
-    "lwl       $t0, 3(%[src])                    \n"
-    "lwr       $t1, 4(%[src])                    \n"
-    "bgtz      $v1, $ua_skip_pref30_96           \n"
-    " lwl      $t1, 7(%[src])                    \n"
-    "pref      30, 96(%[dst])                    \n"
-    // continue setting up the dest, addr 96
-    "$ua_skip_pref30_96:                         \n"
-    "lwr       $t2, 8(%[src])                    \n"
-    "lwl       $t2, 11(%[src])                   \n"
-    "lwr       $t3, 12(%[src])                   \n"
-    "lwl       $t3, 15(%[src])                   \n"
-    "lwr       $t4, 16(%[src])                   \n"
-    "lwl       $t4, 19(%[src])                   \n"
-    "lwr       $t5, 20(%[src])                   \n"
-    "lwl       $t5, 23(%[src])                   \n"
-    "lwr       $t6, 24(%[src])                   \n"
-    "lwl       $t6, 27(%[src])                   \n"
-    "lwr       $t7, 28(%[src])                   \n"
-    "lwl       $t7, 31(%[src])                   \n"
-    "pref      0, 128(%[src])                    \n"
-    // bring the next lines of src, addr 128
-    "sw        $t0, 0(%[dst])                    \n"
-    "sw        $t1, 4(%[dst])                    \n"
-    "sw        $t2, 8(%[dst])                    \n"
-    "sw        $t3, 12(%[dst])                   \n"
-    "sw        $t4, 16(%[dst])                   \n"
-    "sw        $t5, 20(%[dst])                   \n"
-    "sw        $t6, 24(%[dst])                   \n"
-    "sw        $t7, 28(%[dst])                   \n"
-    "lwr       $t0, 32(%[src])                   \n"
-    "lwl       $t0, 35(%[src])                   \n"
-    "lwr       $t1, 36(%[src])                   \n"
-    "bgtz      $v1, ua_skip_pref30_128           \n"
-    " lwl      $t1, 39(%[src])                   \n"
-    "pref      30, 128(%[dst])                   \n"
-    // continue setting up the dest, addr 128
-    "ua_skip_pref30_128:                         \n"
-
-    "lwr       $t2, 40(%[src])                   \n"
-    "lwl       $t2, 43(%[src])                   \n"
-    "lwr       $t3, 44(%[src])                   \n"
-    "lwl       $t3, 47(%[src])                   \n"
-    "lwr       $t4, 48(%[src])                   \n"
-    "lwl       $t4, 51(%[src])                   \n"
-    "lwr       $t5, 52(%[src])                   \n"
-    "lwl       $t5, 55(%[src])                   \n"
-    "lwr       $t6, 56(%[src])                   \n"
-    "lwl       $t6, 59(%[src])                   \n"
-    "lwr       $t7, 60(%[src])                   \n"
-    "lwl       $t7, 63(%[src])                   \n"
-    "pref      0, 160(%[src])                    \n"
-    // bring the next lines of src, addr 160
-    "sw        $t0, 32(%[dst])                   \n"
-    "sw        $t1, 36(%[dst])                   \n"
-    "sw        $t2, 40(%[dst])                   \n"
-    "sw        $t3, 44(%[dst])                   \n"
-    "sw        $t4, 48(%[dst])                   \n"
-    "sw        $t5, 52(%[dst])                   \n"
-    "sw        $t6, 56(%[dst])                   \n"
-    "sw        $t7, 60(%[dst])                   \n"
-
-    "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
-    "sgtu      $v1,%[dst],$t9                    \n"
-    "bne       %[dst],$a3,$ua_loop16w            \n"
-    " addiu    %[src],%[src],64                  \n"  // adding 64 to src
-    "move      %[count],$t8                      \n"
-
-    // Here we have src and dest word-aligned but less than 64-bytes to go
-
-    "ua_chk8w:                                   \n"
-    "pref      0, 0x0(%[src])                    \n"
-    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
-    // the t8 is the reminder count
-    "beq       %[count], $t8, $ua_chk1w          \n"
-    // when count==t8, no 32-byte chunk
-
-    "lwr       $t0, 0(%[src])                    \n"
-    "lwl       $t0, 3(%[src])                    \n"
-    "lwr       $t1, 4(%[src])                    \n"
-    "lwl       $t1, 7(%[src])                    \n"
-    "lwr       $t2, 8(%[src])                    \n"
-    "lwl       $t2, 11(%[src])                   \n"
-    "lwr       $t3, 12(%[src])                   \n"
-    "lwl       $t3, 15(%[src])                   \n"
-    "lwr       $t4, 16(%[src])                   \n"
-    "lwl       $t4, 19(%[src])                   \n"
-    "lwr       $t5, 20(%[src])                   \n"
-    "lwl       $t5, 23(%[src])                   \n"
-    "lwr       $t6, 24(%[src])                   \n"
-    "lwl       $t6, 27(%[src])                   \n"
-    "lwr       $t7, 28(%[src])                   \n"
-    "lwl       $t7, 31(%[src])                   \n"
-    "addiu     %[src], %[src], 32                \n"
-
-    "sw        $t0, 0(%[dst])                    \n"
-    "sw        $t1, 4(%[dst])                    \n"
-    "sw        $t2, 8(%[dst])                    \n"
-    "sw        $t3, 12(%[dst])                   \n"
-    "sw        $t4, 16(%[dst])                   \n"
-    "sw        $t5, 20(%[dst])                   \n"
-    "sw        $t6, 24(%[dst])                   \n"
-    "sw        $t7, 28(%[dst])                   \n"
-    "addiu     %[dst], %[dst], 32                \n"
-
-    "$ua_chk1w:                                  \n"
-    "andi      %[count], $t8, 0x3                \n"
-    // now count is the reminder past 1w chunks
-    "beq       %[count], $t8, ua_smallCopy       \n"
-    "subu      $a3, $t8, %[count]                \n"
-    // a3 is count of bytes in 1w chunks
-    "addu      $a3, %[dst], $a3                  \n"
-    // now a3 is the dst address past the 1w chunks
-
-    // copying in words (4-byte chunks)
-    "$ua_wordCopy_loop:                          \n"
-    "lwr       $v1, 0(%[src])                    \n"
-    "lwl       $v1, 3(%[src])                    \n"
-    "addiu     %[src], %[src], 4                 \n"
-    "addiu     %[dst], %[dst], 4                 \n"
-    // note: dst=a1 is word aligned here, see NOTE1
-    "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
-    " sw       $v1,-4(%[dst])                    \n"
-
-    // Now less than 4 bytes (value in count) left to copy
-    "ua_smallCopy:                               \n"
-    "beqz      %[count], leave                   \n"
-    " addu     $a3, %[dst], %[count]             \n" // a3 = last dst address
-    "$ua_smallCopy_loop:                         \n"
-    "lb        $v1, 0(%[src])                    \n"
-    "addiu     %[src], %[src], 1                 \n"
-    "addiu     %[dst], %[dst], 1                 \n"
-    "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
-    " sb       $v1, -1(%[dst])                   \n"
-
-    "j         $ra                               \n"
-    " nop                                        \n"
-    ".set      at                                \n"
-    ".set      reorder                           \n"
-       : [dst] "+r" (dst), [src] "+r" (src)
-       : [count] "r" (count)
-       : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7",
-       "t8", "t9", "a3", "v1", "at"
-  );
-}
-#endif  // HAS_COPYROW_MIPS
-
-// DSPR2 functions
-#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
-    (__mips_dsp_rev >= 2) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
-
-void SplitUVRow_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                      int width) {
-  __asm__ __volatile__ (
-    ".set push                                     \n"
-    ".set noreorder                                \n"
-    "srl             $t4, %[width], 4              \n"  // multiplies of 16
-    "blez            $t4, 2f                       \n"
-    " andi           %[width], %[width], 0xf       \n"  // residual
-
-  "1:                                              \n"
-    "addiu           $t4, $t4, -1                  \n"
-    "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
-    "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
-    "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
-    "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
-    "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
-    "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 | U10
-    "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 | U12
-    "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 | U14
-    "addiu           %[src_uv], %[src_uv], 32      \n"
-    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
-    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
-    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
-    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
-    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
-    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
-    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
-    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
-    "sw              $t9, 0(%[dst_v])              \n"
-    "sw              $t0, 0(%[dst_u])              \n"
-    "sw              $t1, 4(%[dst_v])              \n"
-    "sw              $t2, 4(%[dst_u])              \n"
-    "sw              $t3, 8(%[dst_v])              \n"
-    "sw              $t5, 8(%[dst_u])              \n"
-    "sw              $t6, 12(%[dst_v])             \n"
-    "sw              $t7, 12(%[dst_u])             \n"
-    "addiu           %[dst_v], %[dst_v], 16        \n"
-    "bgtz            $t4, 1b                       \n"
-    " addiu          %[dst_u], %[dst_u], 16        \n"
-
-    "beqz            %[width], 3f                  \n"
-    " nop                                          \n"
-
-  "2:                                              \n"
-    "lbu             $t0, 0(%[src_uv])             \n"
-    "lbu             $t1, 1(%[src_uv])             \n"
-    "addiu           %[src_uv], %[src_uv], 2       \n"
-    "addiu           %[width], %[width], -1        \n"
-    "sb              $t0, 0(%[dst_u])              \n"
-    "sb              $t1, 0(%[dst_v])              \n"
-    "addiu           %[dst_u], %[dst_u], 1         \n"
-    "bgtz            %[width], 2b                  \n"
-    " addiu          %[dst_v], %[dst_v], 1         \n"
-
-  "3:                                              \n"
-    ".set pop                                      \n"
-     : [src_uv] "+r" (src_uv),
-       [width] "+r" (width),
-       [dst_u] "+r" (dst_u),
-       [dst_v] "+r" (dst_v)
-     :
-     : "t0", "t1", "t2", "t3",
-     "t4", "t5", "t6", "t7", "t8", "t9"
-  );
-}
-
-void MirrorRow_DSPR2(const uint8* src, uint8* dst, int width) {
-  __asm__ __volatile__ (
-    ".set push                             \n"
-    ".set noreorder                        \n"
-
-    "srl       $t4, %[width], 4            \n"  // multiplies of 16
-    "andi      $t5, %[width], 0xf          \n"
-    "blez      $t4, 2f                     \n"
-    " addu     %[src], %[src], %[width]    \n"  // src += width
-
-   "1:                                     \n"
-    "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
-    "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
-    "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
-    "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
-    "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
-    "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
-    "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
-    "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
-    "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
-    "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
-    "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
-    "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
-    "addiu     %[src], %[src], -16         \n"
-    "addiu     $t4, $t4, -1                \n"
-    "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
-    "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
-    "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
-    "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
-    "bgtz      $t4, 1b                     \n"
-    " addiu    %[dst], %[dst], 16          \n"
-    "beqz      $t5, 3f                     \n"
-    " nop                                  \n"
-
-   "2:                                     \n"
-    "lbu       $t0, -1(%[src])             \n"
-    "addiu     $t5, $t5, -1                \n"
-    "addiu     %[src], %[src], -1          \n"
-    "sb        $t0, 0(%[dst])              \n"
-    "bgez      $t5, 2b                     \n"
-    " addiu    %[dst], %[dst], 1           \n"
-
-   "3:                                     \n"
-    ".set pop                              \n"
-      : [src] "+r" (src), [dst] "+r" (dst)
-      : [width] "r" (width)
-      : "t0", "t1", "t2", "t3", "t4", "t5"
-  );
-}
-
-void MirrorUVRow_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                       int width) {
-  int x;
-  int y;
-  __asm__ __volatile__ (
-    ".set push                                    \n"
-    ".set noreorder                               \n"
-
-    "addu            $t4, %[width], %[width]      \n"
-    "srl             %[x], %[width], 4            \n"
-    "andi            %[y], %[width], 0xf          \n"
-    "blez            %[x], 2f                     \n"
-    " addu           %[src_uv], %[src_uv], $t4    \n"
-
-   "1:                                            \n"
-    "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
-    "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
-    "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
-    "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
-    "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
-    "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
-    "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
-    "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
-
-    "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
-    "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
-    "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
-    "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
-    "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
-    "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
-    "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
-    "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
-    "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
-    "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
-    "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
-    "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
-    "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
-    "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
-    "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
-    "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
-    "addiu           %[src_uv], %[src_uv], -32    \n"
-    "addiu           %[x], %[x], -1               \n"
-    "swr             $t4, 0(%[dst_u])             \n"
-    "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
-    "swr             $t6, 0(%[dst_v])             \n"
-    "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
-    "swr             $t2, 4(%[dst_u])             \n"
-    "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
-    "swr             $t3, 4(%[dst_v])             \n"
-    "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
-    "swr             $t0, 8(%[dst_u])             \n"
-    "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
-    "swr             $t1, 8(%[dst_v])             \n"
-    "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
-    "swr             $t9, 12(%[dst_u])            \n"
-    "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
-    "swr             $t5, 12(%[dst_v])            \n"
-    "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
-    "addiu           %[dst_v], %[dst_v], 16       \n"
-    "bgtz            %[x], 1b                     \n"
-    " addiu          %[dst_u], %[dst_u], 16       \n"
-    "beqz            %[y], 3f                     \n"
-    " nop                                         \n"
-    "b               2f                           \n"
-    " nop                                         \n"
-
-   "2:                                            \n"
-    "lbu             $t0, -2(%[src_uv])           \n"
-    "lbu             $t1, -1(%[src_uv])           \n"
-    "addiu           %[src_uv], %[src_uv], -2     \n"
-    "addiu           %[y], %[y], -1               \n"
-    "sb              $t0, 0(%[dst_u])             \n"
-    "sb              $t1, 0(%[dst_v])             \n"
-    "addiu           %[dst_u], %[dst_u], 1        \n"
-    "bgtz            %[y], 2b                     \n"
-    " addiu          %[dst_v], %[dst_v], 1        \n"
-
-   "3:                                            \n"
-    ".set pop                                     \n"
-      : [src_uv] "+r" (src_uv),
-        [dst_u] "+r" (dst_u),
-        [dst_v] "+r" (dst_v),
-        [x] "=&r" (x),
-        [y] "=&r" (y)
-      : [width] "r" (width)
-      : "t0", "t1", "t2", "t3", "t4",
-      "t5", "t7", "t8", "t9"
-  );
-}
-
-// Convert (4 Y and 2 VU) I422 and arrange RGB values into
-// t5 = | 0 | B0 | 0 | b0 |
-// t4 = | 0 | B1 | 0 | b1 |
-// t9 = | 0 | G0 | 0 | g0 |
-// t8 = | 0 | G1 | 0 | g1 |
-// t2 = | 0 | R0 | 0 | r0 |
-// t1 = | 0 | R1 | 0 | r1 |
-#define YUVTORGB                                                               \
-      "lw                $t0, 0(%[y_buf])       \n"                            \
-      "lhu               $t1, 0(%[u_buf])       \n"                            \
-      "lhu               $t2, 0(%[v_buf])       \n"                            \
-      "preceu.ph.qbr     $t1, $t1               \n"                            \
-      "preceu.ph.qbr     $t2, $t2               \n"                            \
-      "preceu.ph.qbra    $t3, $t0               \n"                            \
-      "preceu.ph.qbla    $t0, $t0               \n"                            \
-      "subu.ph           $t1, $t1, $s5          \n"                            \
-      "subu.ph           $t2, $t2, $s5          \n"                            \
-      "subu.ph           $t3, $t3, $s4          \n"                            \
-      "subu.ph           $t0, $t0, $s4          \n"                            \
-      "mul.ph            $t3, $t3, $s0          \n"                            \
-      "mul.ph            $t0, $t0, $s0          \n"                            \
-      "shll.ph           $t4, $t1, 0x7          \n"                            \
-      "subu.ph           $t4, $t4, $t1          \n"                            \
-      "mul.ph            $t6, $t1, $s1          \n"                            \
-      "mul.ph            $t1, $t2, $s2          \n"                            \
-      "addq_s.ph         $t5, $t4, $t3          \n"                            \
-      "addq_s.ph         $t4, $t4, $t0          \n"                            \
-      "shra.ph           $t5, $t5, 6            \n"                            \
-      "shra.ph           $t4, $t4, 6            \n"                            \
-      "addiu             %[u_buf], 2            \n"                            \
-      "addiu             %[v_buf], 2            \n"                            \
-      "addu.ph           $t6, $t6, $t1          \n"                            \
-      "mul.ph            $t1, $t2, $s3          \n"                            \
-      "addu.ph           $t9, $t6, $t3          \n"                            \
-      "addu.ph           $t8, $t6, $t0          \n"                            \
-      "shra.ph           $t9, $t9, 6            \n"                            \
-      "shra.ph           $t8, $t8, 6            \n"                            \
-      "addu.ph           $t2, $t1, $t3          \n"                            \
-      "addu.ph           $t1, $t1, $t0          \n"                            \
-      "shra.ph           $t2, $t2, 6            \n"                            \
-      "shra.ph           $t1, $t1, 6            \n"                            \
-      "subu.ph           $t5, $t5, $s5          \n"                            \
-      "subu.ph           $t4, $t4, $s5          \n"                            \
-      "subu.ph           $t9, $t9, $s5          \n"                            \
-      "subu.ph           $t8, $t8, $s5          \n"                            \
-      "subu.ph           $t2, $t2, $s5          \n"                            \
-      "subu.ph           $t1, $t1, $s5          \n"                            \
-      "shll_s.ph         $t5, $t5, 8            \n"                            \
-      "shll_s.ph         $t4, $t4, 8            \n"                            \
-      "shll_s.ph         $t9, $t9, 8            \n"                            \
-      "shll_s.ph         $t8, $t8, 8            \n"                            \
-      "shll_s.ph         $t2, $t2, 8            \n"                            \
-      "shll_s.ph         $t1, $t1, 8            \n"                            \
-      "shra.ph           $t5, $t5, 8            \n"                            \
-      "shra.ph           $t4, $t4, 8            \n"                            \
-      "shra.ph           $t9, $t9, 8            \n"                            \
-      "shra.ph           $t8, $t8, 8            \n"                            \
-      "shra.ph           $t2, $t2, 8            \n"                            \
-      "shra.ph           $t1, $t1, 8            \n"                            \
-      "addu.ph           $t5, $t5, $s5          \n"                            \
-      "addu.ph           $t4, $t4, $s5          \n"                            \
-      "addu.ph           $t9, $t9, $s5          \n"                            \
-      "addu.ph           $t8, $t8, $s5          \n"                            \
-      "addu.ph           $t2, $t2, $s5          \n"                            \
-      "addu.ph           $t1, $t1, $s5          \n"
-
-// TODO(fbarchard): accept yuv conversion constants.
-void I422ToARGBRow_DSPR2(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* rgb_buf,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
-  __asm__ __volatile__ (
-    ".set push                                \n"
-    ".set noreorder                           \n"
-    "beqz              %[width], 2f           \n"
-    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
-    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
-    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
-    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
-    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
-    "repl.ph           $s5, 128               \n"  // |128|128| // clipping
-    "lui               $s6, 0xff00            \n"
-    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|ff|
-
-   "1:                                        \n"
-      YUVTORGB
-// Arranging into argb format
-    "precr.qb.ph       $t4, $t8, $t4          \n"  // |G1|g1|B1|b1|
-    "precr.qb.ph       $t5, $t9, $t5          \n"  // |G0|g0|B0|b0|
-    "addiu             %[width], -4           \n"
-    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |G1|B1|G0|B0|
-    "precr.qb.ph       $t9, $t4, $t5          \n"  // |g1|b1|g0|b0|
-    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
-
-    "addiu             %[y_buf], 4            \n"
-    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
-    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
-    "or                $t1, $t1, $s6          \n"  // |ff|R1|ff|R0|
-    "or                $t2, $t2, $s6          \n"  // |ff|r1|ff|r0|
-    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|r1|g1|b1|
-    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|R1|G1|B1|
-    "sll               $t9, $t9, 16           \n"
-    "sll               $t8, $t8, 16           \n"
-    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|r0|g0|b0|
-    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|R0|G0|B0|
-// Store results.
-    "sw                $t2, 0(%[rgb_buf])     \n"
-    "sw                $t0, 4(%[rgb_buf])     \n"
-    "sw                $t1, 8(%[rgb_buf])     \n"
-    "sw                $t3, 12(%[rgb_buf])    \n"
-    "bnez              %[width], 1b           \n"
-    " addiu            %[rgb_buf], 16         \n"
-   "2:                                        \n"
-    ".set pop                                 \n"
-      :[y_buf] "+r" (y_buf),
-       [u_buf] "+r" (u_buf),
-       [v_buf] "+r" (v_buf),
-       [width] "+r" (width),
-       [rgb_buf] "+r" (rgb_buf)
-      :
-      : "t0", "t1",  "t2", "t3",  "t4", "t5",
-      "t6", "t7", "t8", "t9",
-      "s0", "s1", "s2", "s3",
-      "s4", "s5", "s6"
-  );
-}
-
-// Bilinear filter 8x2 -> 8x1
-void InterpolateRow_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
-                          ptrdiff_t src_stride, int dst_width,
-                          int source_y_fraction) {
-    int y0_fraction = 256 - source_y_fraction;
-    const uint8* src_ptr1 = src_ptr + src_stride;
-
-  __asm__ __volatile__ (
-     ".set push                                           \n"
-     ".set noreorder                                      \n"
-
-     "replv.ph          $t0, %[y0_fraction]               \n"
-     "replv.ph          $t1, %[source_y_fraction]         \n"
-
-   "1:                                                    \n"
-     "lw                $t2, 0(%[src_ptr])                \n"
-     "lw                $t3, 0(%[src_ptr1])               \n"
-     "lw                $t4, 4(%[src_ptr])                \n"
-     "lw                $t5, 4(%[src_ptr1])               \n"
-     "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
-     "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
-     "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
-     "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
-     "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
-     "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
-     "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
-     "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
-     "addq.ph           $t6, $t6, $t8                     \n"
-     "addq.ph           $t7, $t7, $t9                     \n"
-     "addq.ph           $t2, $t2, $t4                     \n"
-     "addq.ph           $t3, $t3, $t5                     \n"
-     "shra.ph           $t6, $t6, 8                       \n"
-     "shra.ph           $t7, $t7, 8                       \n"
-     "shra.ph           $t2, $t2, 8                       \n"
-     "shra.ph           $t3, $t3, 8                       \n"
-     "precr.qb.ph       $t6, $t6, $t7                     \n"
-     "precr.qb.ph       $t2, $t2, $t3                     \n"
-     "addiu             %[src_ptr], %[src_ptr], 8         \n"
-     "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
-     "addiu             %[dst_width], %[dst_width], -8    \n"
-     "sw                $t6, 0(%[dst_ptr])                \n"
-     "sw                $t2, 4(%[dst_ptr])                \n"
-     "bgtz              %[dst_width], 1b                  \n"
-     " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
-
-     ".set pop                                            \n"
-  : [dst_ptr] "+r" (dst_ptr),
-    [src_ptr1] "+r" (src_ptr1),
-    [src_ptr] "+r" (src_ptr),
-    [dst_width] "+r" (dst_width)
-  : [source_y_fraction] "r" (source_y_fraction),
-    [y0_fraction] "r" (y0_fraction),
-    [src_stride] "r" (src_stride)
-  : "t0", "t1", "t2", "t3", "t4", "t5",
-    "t6", "t7", "t8", "t9"
-  );
-}
-#endif  // __mips_dsp_rev >= 2
-
-#endif  // defined(__mips__)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/media/libyuv/source/row_mmi.cc b/media/libyuv/source/row_mmi.cc
new file mode 100644
index 0000000000..362fd1cfcc
--- /dev/null
+++ b/media/libyuv/source/row_mmi.cc
@@ -0,0 +1,7842 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "libyuv/row.h"
+
+#include <string.h>  // For memcpy and memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Mips MMI.
+#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+// clang-format off
+
+void RGB24ToARGBRow_MMI(const uint8_t* src_rgb24,
+                        uint8_t* dst_argb,
+                        int width) {
+  uint64_t src0, src1, dest;
+  const uint64_t mask = 0xff000000ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gslwlc1    %[src0],         0x03(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gslwlc1    %[src1],         0x06(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src1],         0x03(%[src_ptr])                 \n\t"
+
+      "or         %[src0],         %[src0],           %[mask]       \n\t"
+      "or         %[src1],         %[src1],           %[mask]       \n\t"
+      "punpcklwd  %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "gslwlc1    %[src0],         0x09(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src0],         0x06(%[src_ptr])                 \n\t"
+      "gslwlc1    %[src1],         0x0c(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src1],         0x09(%[src_ptr])                 \n\t"
+
+      "or         %[src0],         %[src0],           %[mask]       \n\t"
+      "or         %[src1],         %[src1],           %[mask]       \n\t"
+      "punpcklwd  %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x0c          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_rgb24), [dst_ptr] "r"(dst_argb), [width] "r"(width),
+        [mask] "f"(mask)
+      : "memory");
+}
+
+void RAWToARGBRow_MMI(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  uint64_t src0, src1, dest;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0xff000000ULL;
+  const uint64_t mask2 = 0xc6;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gslwlc1    %[src0],         0x03(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gslwlc1    %[src1],         0x06(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src1],         0x03(%[src_ptr])                 \n\t"
+
+      "or         %[src0],         %[src0],           %[mask1]      \n\t"
+      "punpcklbh  %[src0],         %[src0],           %[mask0]      \n\t"
+      "pshufh     %[src0],         %[src0],           %[mask2]      \n\t"
+      "or         %[src1],         %[src1],           %[mask1]      \n\t"
+      "punpcklbh  %[src1],         %[src1],           %[mask0]      \n\t"
+      "pshufh     %[src1],         %[src1],           %[mask2]      \n\t"
+      "packushb   %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "gslwlc1    %[src0],         0x09(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src0],         0x06(%[src_ptr])                 \n\t"
+      "gslwlc1    %[src1],         0x0c(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src1],         0x09(%[src_ptr])                 \n\t"
+
+      "or         %[src0],         %[src0],           %[mask1]      \n\t"
+      "punpcklbh  %[src0],         %[src0],           %[mask0]      \n\t"
+      "pshufh     %[src0],         %[src0],           %[mask2]      \n\t"
+      "or         %[src1],         %[src1],           %[mask1]      \n\t"
+      "punpcklbh  %[src1],         %[src1],           %[mask0]      \n\t"
+      "pshufh     %[src1],         %[src1],           %[mask2]      \n\t"
+      "packushb   %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x0c          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_raw), [dst_ptr] "r"(dst_argb), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [mask2] "f"(mask2), [width] "r"(width)
+      : "memory");
+}
+
+void RAWToRGB24Row_MMI(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
+  uint64_t src0, src1;
+  uint64_t ftmp[4];
+  uint64_t mask0 = 0xc6;
+  uint64_t mask1 = 0x6c;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_raw])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_raw])                 \n\t"
+      "gslwrc1    %[src1],         0x08(%[src_raw])                 \n\t"
+      "gslwlc1    %[src1],         0x0b(%[src_raw])                 \n\t"
+
+      "punpcklbh  %[ftmp0],        %[src0],           %[zero]       \n\t"
+      "pshufh     %[ftmp0],        %[ftmp0],          %[mask0]      \n\t"
+      "punpckhbh  %[ftmp1],        %[src0],           %[zero]       \n\t"
+      "punpcklbh  %[src1],         %[src1],           %[zero]       \n\t"
+      "pextrh     %[ftmp2],        %[ftmp0],          %[three]      \n\t"
+      "pextrh     %[ftmp3],        %[ftmp1],          %[one]        \n\t"
+      "pinsrh_3   %[ftmp0],        %[ftmp0],          %[ftmp3]      \n\t"
+      "pextrh     %[ftmp3],        %[ftmp1],          %[two]        \n\t"
+      "pinsrh_1   %[ftmp1],        %[ftmp1],          %[ftmp2]      \n\t"
+      "pshufh     %[src1],         %[src1],           %[mask1]      \n\t"
+      "pextrh     %[ftmp2],        %[src1],           %[zero]       \n\t"
+      "pinsrh_2   %[ftmp1],        %[ftmp1],          %[ftmp2]      \n\t"
+      "pinsrh_0   %[src1],         %[src1],           %[ftmp3]      \n\t"
+      "packushb   %[ftmp0],        %[ftmp0],          %[ftmp1]      \n\t"
+      "packushb   %[src1],         %[src1],           %[zero]       \n\t"
+
+      "gssdrc1    %[ftmp0],        0x00(%[dst_rgb24])               \n\t"
+      "gssdlc1    %[ftmp0],        0x07(%[dst_rgb24])               \n\t"
+      "gsswrc1    %[src1],         0x08(%[dst_rgb24])               \n\t"
+      "gsswlc1    %[src1],         0x0b(%[dst_rgb24])               \n\t"
+
+      "daddiu     %[src_raw],      %[src_raw],        0x0c          \n\t"
+      "daddiu     %[dst_rgb24],    %[dst_rgb24],      0x0c          \n\t"
+      "daddiu     %[width],        %[width],         -0x04          \n\t"
+      "bgtz       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [ftmp0] "=&f"(ftmp[0]),
+        [ftmp1] "=&f"(ftmp[1]), [ftmp2] "=&f"(ftmp[2]), [ftmp3] "=&f"(ftmp[3])
+      : [src_raw] "r"(src_raw), [dst_rgb24] "r"(dst_rgb24), [width] "r"(width),
+        [mask0] "f"(mask0), [mask1] "f"(mask1), [zero] "f"(0x00),
+        [one] "f"(0x01), [two] "f"(0x02), [three] "f"(0x03)
+      : "memory");
+}
+
+void RGB565ToARGBRow_MMI(const uint8_t* src_rgb565,
+                         uint8_t* dst_argb,
+                         int width) {
+  uint64_t ftmp[5];
+  uint64_t c0 = 0x001f001f001f001f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t c2 = 0x0007000700070007;
+  __asm__ volatile(
+      "1:                                                      \n\t"
+      "gsldrc1   %[src0],       0x00(%[src_rgb565])            \n\t"
+      "gsldlc1   %[src0],       0x07(%[src_rgb565])            \n\t"
+      "psrlh     %[src1],       %[src0],             %[eight]  \n\t"
+      "and       %[b],          %[src0],             %[c0]     \n\t"
+      "and       %[src0],       %[src0],             %[c1]     \n\t"
+      "psrlh     %[src0],       %[src0],             %[five]   \n\t"
+      "and       %[g],          %[src1],             %[c2]     \n\t"
+      "psllh     %[g],          %[g],                %[three]  \n\t"
+      "or        %[g],          %[src0],             %[g]      \n\t"
+      "psrlh     %[r],          %[src1],             %[three]  \n\t"
+      "psllh     %[src0],       %[b],                %[three]  \n\t"
+      "psrlh     %[src1],       %[b],                %[two]    \n\t"
+      "or        %[b],          %[src0],             %[src1]   \n\t"
+      "psllh     %[src0],       %[g],                %[two]    \n\t"
+      "psrlh     %[src1],       %[g],                %[four]   \n\t"
+      "or        %[g],          %[src0],             %[src1]   \n\t"
+      "psllh     %[src0],       %[r],                %[three]  \n\t"
+      "psrlh     %[src1],       %[r],                %[two]    \n\t"
+      "or        %[r],          %[src0],             %[src1]   \n\t"
+      "packushb  %[b],          %[b],                %[r]      \n\t"
+      "packushb  %[g],          %[g],                %[c1]     \n\t"
+      "punpcklbh %[src0],       %[b],                %[g]      \n\t"
+      "punpckhbh %[src1],       %[b],                %[g]      \n\t"
+      "punpcklhw %[r],          %[src0],             %[src1]   \n\t"
+      "gssdrc1   %[r],          0x00(%[dst_argb])              \n\t"
+      "gssdlc1   %[r],          0x07(%[dst_argb])              \n\t"
+      "punpckhhw %[r],          %[src0],             %[src1]   \n\t"
+      "gssdrc1   %[r],          0x08(%[dst_argb])              \n\t"
+      "gssdlc1   %[r],          0x0f(%[dst_argb])              \n\t"
+      "daddiu    %[src_rgb565], %[src_rgb565],       0x08      \n\t"
+      "daddiu    %[dst_argb],   %[dst_argb],         0x10      \n\t"
+      "daddiu    %[width],      %[width],           -0x04      \n\t"
+      "bgtz      %[width],     1b                              \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [b] "=&f"(ftmp[2]),
+        [g] "=&f"(ftmp[3]), [r] "=&f"(ftmp[4])
+      : [src_rgb565] "r"(src_rgb565), [dst_argb] "r"(dst_argb),
+        [width] "r"(width), [c0] "f"(c0), [c1] "f"(c1), [c2] "f"(c2),
+        [eight] "f"(0x08), [five] "f"(0x05), [three] "f"(0x03), [two] "f"(0x02),
+        [four] "f"(0x04)
+      : "memory");
+}
+
+void ARGB1555ToARGBRow_MMI(const uint8_t* src_argb1555,
+                           uint8_t* dst_argb,
+                           int width) {
+  uint64_t ftmp[6];
+  uint64_t c0 = 0x001f001f001f001f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t c2 = 0x0003000300030003;
+  uint64_t c3 = 0x007c007c007c007c;
+  uint64_t c4 = 0x0001000100010001;
+  __asm__ volatile(
+      "1:                                                         \n\t"
+      "gsldrc1   %[src0],         0x00(%[src_argb1555])           \n\t"
+      "gsldlc1   %[src0],         0x07(%[src_argb1555])           \n\t"
+      "psrlh     %[src1],         %[src0],              %[eight]  \n\t"
+      "and       %[b],            %[src0],              %[c0]     \n\t"
+      "and       %[src0],         %[src0],              %[c1]     \n\t"
+      "psrlh     %[src0],         %[src0],              %[five]   \n\t"
+      "and       %[g],            %[src1],              %[c2]     \n\t"
+      "psllh     %[g],            %[g],                 %[three]  \n\t"
+      "or        %[g],            %[src0],              %[g]      \n\t"
+      "and       %[r],            %[src1],              %[c3]     \n\t"
+      "psrlh     %[r],            %[r],                 %[two]    \n\t"
+      "psrlh     %[a],            %[src1],              %[seven]  \n\t"
+      "psllh     %[src0],         %[b],                 %[three]  \n\t"
+      "psrlh     %[src1],         %[b],                 %[two]    \n\t"
+      "or        %[b],            %[src0],              %[src1]   \n\t"
+      "psllh     %[src0],         %[g],                 %[three]  \n\t"
+      "psrlh     %[src1],         %[g],                 %[two]    \n\t"
+      "or        %[g],            %[src0],              %[src1]   \n\t"
+      "psllh     %[src0],         %[r],                 %[three]  \n\t"
+      "psrlh     %[src1],         %[r],                 %[two]    \n\t"
+      "or        %[r],            %[src0],              %[src1]   \n\t"
+      "xor       %[a],            %[a],                 %[c1]     \n\t"
+      "paddb     %[a],            %[a],                 %[c4]     \n\t"
+      "packushb  %[b],            %[b],                 %[r]      \n\t"
+      "packushb  %[g],            %[g],                 %[a]      \n\t"
+      "punpcklbh %[src0],         %[b],                 %[g]      \n\t"
+      "punpckhbh %[src1],         %[b],                 %[g]      \n\t"
+      "punpcklhw %[r],            %[src0],              %[src1]   \n\t"
+      "gssdrc1   %[r],            0x00(%[dst_argb])               \n\t"
+      "gssdlc1   %[r],            0x07(%[dst_argb])               \n\t"
+      "punpckhhw %[r],            %[src0],              %[src1]   \n\t"
+      "gssdrc1   %[r],            0x08(%[dst_argb])               \n\t"
+      "gssdlc1   %[r],            0x0f(%[dst_argb])               \n\t"
+      "daddiu    %[src_argb1555], %[src_argb1555],      0x08      \n\t"
+      "daddiu    %[dst_argb],     %[dst_argb],          0x10      \n\t"
+      "daddiu    %[width],        %[width],            -0x04      \n\t"
+      "bgtz      %[width],        1b                              \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [b] "=&f"(ftmp[2]),
+        [g] "=&f"(ftmp[3]), [r] "=&f"(ftmp[4]), [a] "=&f"(ftmp[5])
+      : [src_argb1555] "r"(src_argb1555), [dst_argb] "r"(dst_argb),
+        [width] "r"(width), [c0] "f"(c0), [c1] "f"(c1), [c2] "f"(c2),
+        [c3] "f"(c3), [c4] "f"(c4), [eight] "f"(0x08), [five] "f"(0x05),
+        [three] "f"(0x03), [two] "f"(0x02), [seven] "f"(0x07)
+      : "memory");
+}
+
+void ARGB4444ToARGBRow_MMI(const uint8_t* src_argb4444,
+                           uint8_t* dst_argb,
+                           int width) {
+  uint64_t ftmp[6];
+  uint64_t c0 = 0x000f000f000f000f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  __asm__ volatile(
+      "1:                                                          \n\t"
+      "gsldrc1   %[src0],         0x00(%[src_argb4444])            \n\t"
+      "gsldlc1   %[src0],         0x07(%[src_argb4444])            \n\t"
+      "psrlh     %[src1],         %[src0],              %[eight]   \n\t"
+      "and       %[b],            %[src0],              %[c0]      \n\t"
+      "and       %[src0],         %[src0],              %[c1]      \n\t"
+      "psrlh     %[g],            %[src0],              %[four]    \n\t"
+      "and       %[r],            %[src1],              %[c0]      \n\t"
+      "psrlh     %[a],            %[src1],              %[four]    \n\t"
+      "psllh     %[src0],         %[b],                 %[four]    \n\t"
+      "or        %[b],            %[src0],              %[b]       \n\t"
+      "psllh     %[src0],         %[g],                 %[four]    \n\t"
+      "or        %[g],            %[src0],              %[g]       \n\t"
+      "psllh     %[src0],         %[r],                 %[four]    \n\t"
+      "or        %[r],            %[src0],              %[r]       \n\t"
+      "psllh     %[src0],         %[a],                 %[four]    \n\t"
+      "or        %[a],            %[src0],              %[a]       \n\t"
+      "packushb  %[b],            %[b],                 %[r]       \n\t"
+      "packushb  %[g],            %[g],                 %[a]       \n\t"
+      "punpcklbh %[src0],         %[b],                 %[g]       \n\t"
+      "punpckhbh %[src1],         %[b],                 %[g]       \n\t"
+      "punpcklhw %[r],            %[src0],              %[src1]    \n\t"
+      "gssdrc1   %[r],            0x00(%[dst_argb])                \n\t"
+      "gssdlc1   %[r],            0x07(%[dst_argb])                \n\t"
+      "punpckhhw %[r],            %[src0],              %[src1]    \n\t"
+      "gssdrc1   %[r],            0x08(%[dst_argb])                \n\t"
+      "gssdlc1   %[r],            0x0f(%[dst_argb])                \n\t"
+      "daddiu    %[src_argb4444], %[src_argb4444],      0x08       \n\t"
+      "daddiu    %[dst_argb],     %[dst_argb],          0x10       \n\t"
+      "daddiu    %[width],        %[width],            -0x04       \n\t"
+      "bgtz      %[width],        1b                               \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [b] "=&f"(ftmp[2]),
+        [g] "=&f"(ftmp[3]), [r] "=&f"(ftmp[4]), [a] "=&f"(ftmp[5])
+      : [src_argb4444] "r"(src_argb4444), [dst_argb] "r"(dst_argb),
+        [width] "r"(width), [c0] "f"(c0), [c1] "f"(c1), [eight] "f"(0x08),
+        [four] "f"(0x04)
+      : "memory");
+}
+
+void ARGBToRGB24Row_MMI(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  uint64_t src;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gslwlc1    %[src],          0x03(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "gsswlc1    %[src],          0x03(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[src],          0x00(%[dst_ptr])                 \n\t"
+
+      "gslwlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src],          0x04(%[src_ptr])                 \n\t"
+      "gsswlc1    %[src],          0x06(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[src],          0x03(%[dst_ptr])                 \n\t"
+
+      "gslwlc1    %[src],          0x0b(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src],          0x08(%[src_ptr])                 \n\t"
+      "gsswlc1    %[src],          0x09(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[src],          0x06(%[dst_ptr])                 \n\t"
+
+      "gslwlc1    %[src],          0x0f(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src],          0x0c(%[src_ptr])                 \n\t"
+      "gsswlc1    %[src],          0x0c(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[src],          0x09(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x0c          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_rgb), [width] "r"(width)
+      : "memory");
+}
+
+void ARGBToRAWRow_MMI(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  uint64_t src0, src1;
+  uint64_t ftmp[3];
+  uint64_t mask0 = 0xc6;
+  uint64_t mask1 = 0x18;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb])                \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb])                \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_argb])                \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_argb])                \n\t"
+
+      "punpcklbh  %[ftmp0],        %[src0],           %[zero]       \n\t"
+      "pshufh     %[ftmp0],        %[ftmp0],          %[mask0]      \n\t"
+      "punpckhbh  %[ftmp1],        %[src0],           %[zero]       \n\t"
+      "punpcklbh  %[ftmp2],        %[src1],           %[zero]       \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]       \n\t"
+
+      "pextrh     %[src0],         %[ftmp1],          %[two]        \n\t"
+      "pinsrh_3   %[ftmp0],        %[ftmp0],          %[src0]       \n\t"
+      "pshufh     %[ftmp1],        %[ftmp1],          %[one]        \n\t"
+
+      "pextrh     %[src0],         %[ftmp2],          %[two]        \n\t"
+      "pinsrh_2   %[ftmp1],        %[ftmp1],          %[src0]       \n\t"
+      "pextrh     %[src0],         %[ftmp2],          %[one]        \n\t"
+      "pinsrh_3   %[ftmp1],        %[ftmp1],          %[src0]       \n\t"
+      "pextrh     %[src0],         %[ftmp2],          %[zero]       \n\t"
+      "pshufh     %[src1],         %[src1],           %[mask1]      \n\t"
+      "pinsrh_0   %[src1],         %[src1],           %[src0]       \n\t"
+      "packushb   %[ftmp0],        %[ftmp0],          %[ftmp1]      \n\t"
+      "packushb   %[src1],         %[src1],           %[zero]       \n\t"
+
+      "gssdrc1    %[ftmp0],        0x00(%[dst_rgb])                 \n\t"
+      "gssdlc1    %[ftmp0],        0x07(%[dst_rgb])                 \n\t"
+      "gsswrc1    %[src1],         0x08(%[dst_rgb])                 \n\t"
+      "gsswlc1    %[src1],         0x0b(%[dst_rgb])                 \n\t"
+
+      "daddiu     %[src_argb],     %[src_argb],       0x10          \n\t"
+      "daddiu     %[dst_rgb],      %[dst_rgb],        0x0c          \n\t"
+      "daddiu     %[width],        %[width],         -0x04          \n\t"
+      "bgtz       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [ftmp0] "=&f"(ftmp[0]),
+        [ftmp1] "=&f"(ftmp[1]), [ftmp2] "=&f"(ftmp[2])
+      : [src_argb] "r"(src_argb), [dst_rgb] "r"(dst_rgb), [width] "r"(width),
+        [mask0] "f"(mask0), [mask1] "f"(mask1), [zero] "f"(0x00),
+        [one] "f"(0x01), [two] "f"(0x02)
+      : "memory");
+}
+
+void ARGBToRGB565Row_MMI(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  uint64_t src0, src1;
+  uint64_t ftmp[3];
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb])                \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb])                \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_argb])                \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_argb])                \n\t"
+
+      "punpcklbh  %[b],            %[src0],           %[src1]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[src1]       \n\t"
+      "punpcklbh  %[src0],         %[b],              %[g]          \n\t"
+      "punpckhbh  %[src1],         %[b],              %[g]          \n\t"
+      "punpcklbh  %[b],            %[src0],           %[zero]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[zero]       \n\t"
+      "punpcklbh  %[r],            %[src1],           %[zero]       \n\t"
+
+      "psrlh      %[b],            %[b],              %[three]      \n\t"
+      "psrlh      %[g],            %[g],              %[two]        \n\t"
+      "psrlh      %[r],            %[r],              %[three]      \n\t"
+
+      "psllh      %[g],            %[g],              %[five]       \n\t"
+      "psllh      %[r],            %[r],              %[eleven]     \n\t"
+      "or         %[b],            %[b],              %[g]          \n\t"
+      "or         %[b],            %[b],              %[r]          \n\t"
+
+      "gssdrc1    %[b],            0x00(%[dst_rgb])                 \n\t"
+      "gssdlc1    %[b],            0x07(%[dst_rgb])                 \n\t"
+
+      "daddiu     %[src_argb],     %[src_argb],       0x10          \n\t"
+      "daddiu     %[dst_rgb],      %[dst_rgb],        0x08          \n\t"
+      "daddiu     %[width],        %[width],         -0x04          \n\t"
+      "bgtz       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [b] "=&f"(ftmp[0]),
+        [g] "=&f"(ftmp[1]), [r] "=&f"(ftmp[2])
+      : [src_argb] "r"(src_argb), [dst_rgb] "r"(dst_rgb), [width] "r"(width),
+        [zero] "f"(0x00), [two] "f"(0x02), [three] "f"(0x03), [five] "f"(0x05),
+        [eleven] "f"(0x0b)
+      : "memory");
+}
+
+// dither4 is a row of 4 values from 4x4 dither matrix.
+// The 4x4 matrix contains values to increase RGB.  When converting to
+// fewer bits (565) this provides an ordered dither.
+// The order in the 4x4 matrix in first byte is upper left.
+// The 4 values are passed as an int, then referenced as an array, so
+// endian will not affect order of the original matrix.  But the dither4
+// will containing the first pixel in the lower byte for little endian
+// or the upper byte for big endian.
+void ARGBToRGB565DitherRow_MMI(const uint8_t* src_argb,
+                               uint8_t* dst_rgb,
+                               const uint32_t dither4,
+                               int width) {
+  uint64_t src0, src1;
+  uint64_t ftmp[3];
+  uint64_t c0 = 0x00ff00ff00ff00ff;
+
+  __asm__ volatile(
+      "punpcklbh  %[dither],       %[dither],         %[zero]       \n\t"
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb])                \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb])                \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_argb])                \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_argb])                \n\t"
+
+      "punpcklbh  %[b],            %[src0],           %[src1]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[src1]       \n\t"
+      "punpcklbh  %[src0],         %[b],              %[g]          \n\t"
+      "punpckhbh  %[src1],         %[b],              %[g]          \n\t"
+      "punpcklbh  %[b],            %[src0],           %[zero]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[zero]       \n\t"
+      "punpcklbh  %[r],            %[src1],           %[zero]       \n\t"
+
+      "paddh      %[b],            %[b],              %[dither]     \n\t"
+      "paddh      %[g],            %[g],              %[dither]     \n\t"
+      "paddh      %[r],            %[r],              %[dither]     \n\t"
+      "pcmpgth    %[src0],         %[b],              %[c0]         \n\t"
+      "or         %[src0],         %[src0],           %[b]          \n\t"
+      "and        %[b],            %[src0],           %[c0]         \n\t"
+      "pcmpgth    %[src0],         %[g],              %[c0]         \n\t"
+      "or         %[src0],         %[src0],           %[g]          \n\t"
+      "and        %[g],            %[src0],           %[c0]         \n\t"
+      "pcmpgth    %[src0],         %[r],              %[c0]         \n\t"
+      "or         %[src0],         %[src0],           %[r]          \n\t"
+      "and        %[r],            %[src0],           %[c0]         \n\t"
+
+      "psrlh      %[b],            %[b],              %[three]      \n\t"
+      "psrlh      %[g],            %[g],              %[two]        \n\t"
+      "psrlh      %[r],            %[r],              %[three]      \n\t"
+
+      "psllh      %[g],            %[g],              %[five]       \n\t"
+      "psllh      %[r],            %[r],              %[eleven]     \n\t"
+      "or         %[b],            %[b],              %[g]          \n\t"
+      "or         %[b],            %[b],              %[r]          \n\t"
+
+      "gssdrc1    %[b],            0x00(%[dst_rgb])                 \n\t"
+      "gssdlc1    %[b],            0x07(%[dst_rgb])                 \n\t"
+
+      "daddiu     %[src_argb],     %[src_argb],       0x10          \n\t"
+      "daddiu     %[dst_rgb],      %[dst_rgb],        0x08          \n\t"
+      "daddiu     %[width],        %[width],         -0x04          \n\t"
+      "bgtz       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [b] "=&f"(ftmp[0]),
+        [g] "=&f"(ftmp[1]), [r] "=&f"(ftmp[2])
+      : [src_argb] "r"(src_argb), [dst_rgb] "r"(dst_rgb), [width] "r"(width),
+        [dither] "f"(dither4), [c0] "f"(c0), [zero] "f"(0x00), [two] "f"(0x02),
+        [three] "f"(0x03), [five] "f"(0x05), [eleven] "f"(0x0b)
+      : "memory");
+}
+
+void ARGBToARGB1555Row_MMI(const uint8_t* src_argb,
+                           uint8_t* dst_rgb,
+                           int width) {
+  uint64_t src0, src1;
+  uint64_t ftmp[4];
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb])                \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb])                \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_argb])                \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_argb])                \n\t"
+
+      "punpcklbh  %[b],            %[src0],           %[src1]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[src1]       \n\t"
+      "punpcklbh  %[src0],         %[b],              %[g]          \n\t"
+      "punpckhbh  %[src1],         %[b],              %[g]          \n\t"
+      "punpcklbh  %[b],            %[src0],           %[zero]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[zero]       \n\t"
+      "punpcklbh  %[r],            %[src1],           %[zero]       \n\t"
+      "punpckhbh  %[a],            %[src1],           %[zero]       \n\t"
+
+      "psrlh      %[b],            %[b],              %[three]      \n\t"
+      "psrlh      %[g],            %[g],              %[three]      \n\t"
+      "psrlh      %[r],            %[r],              %[three]      \n\t"
+      "psrlh      %[a],            %[a],              %[seven]      \n\t"
+
+      "psllh      %[g],            %[g],              %[five]       \n\t"
+      "psllh      %[r],            %[r],              %[ten]        \n\t"
+      "psllh      %[a],            %[a],              %[fifteen]    \n\t"
+      "or         %[b],            %[b],              %[g]          \n\t"
+      "or         %[b],            %[b],              %[r]          \n\t"
+      "or         %[b],            %[b],              %[a]          \n\t"
+
+      "gssdrc1    %[b],            0x00(%[dst_rgb])                 \n\t"
+      "gssdlc1    %[b],            0x07(%[dst_rgb])                 \n\t"
+
+      "daddiu     %[src_argb],     %[src_argb],       0x10          \n\t"
+      "daddiu     %[dst_rgb],      %[dst_rgb],        0x08          \n\t"
+      "daddiu     %[width],        %[width],         -0x04          \n\t"
+      "bgtz       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [b] "=&f"(ftmp[0]),
+        [g] "=&f"(ftmp[1]), [r] "=&f"(ftmp[2]), [a] "=&f"(ftmp[3])
+      : [src_argb] "r"(src_argb), [dst_rgb] "r"(dst_rgb), [width] "r"(width),
+        [zero] "f"(0x00), [three] "f"(0x03), [five] "f"(0x05),
+        [seven] "f"(0x07), [ten] "f"(0x0a), [fifteen] "f"(0x0f)
+      : "memory");
+}
+
+void ARGBToARGB4444Row_MMI(const uint8_t* src_argb,
+                           uint8_t* dst_rgb,
+                           int width) {
+  uint64_t src0, src1;
+  uint64_t ftmp[4];
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb])                \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb])                \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_argb])                \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_argb])                \n\t"
+
+      "punpcklbh  %[b],            %[src0],           %[src1]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[src1]       \n\t"
+      "punpcklbh  %[src0],         %[b],              %[g]          \n\t"
+      "punpckhbh  %[src1],         %[b],              %[g]          \n\t"
+      "punpcklbh  %[b],            %[src0],           %[zero]       \n\t"
+      "punpckhbh  %[g],            %[src0],           %[zero]       \n\t"
+      "punpcklbh  %[r],            %[src1],           %[zero]       \n\t"
+      "punpckhbh  %[a],            %[src1],           %[zero]       \n\t"
+
+      "psrlh      %[b],            %[b],              %[four]       \n\t"
+      "psrlh      %[g],            %[g],              %[four]       \n\t"
+      "psrlh      %[r],            %[r],              %[four]       \n\t"
+      "psrlh      %[a],            %[a],              %[four]       \n\t"
+
+      "psllh      %[g],            %[g],              %[four]       \n\t"
+      "psllh      %[r],            %[r],              %[eight]      \n\t"
+      "psllh      %[a],            %[a],              %[twelve]     \n\t"
+      "or         %[b],            %[b],              %[g]          \n\t"
+      "or         %[b],            %[b],              %[r]          \n\t"
+      "or         %[b],            %[b],              %[a]          \n\t"
+
+      "gssdrc1    %[b],            0x00(%[dst_rgb])                 \n\t"
+      "gssdlc1    %[b],            0x07(%[dst_rgb])                 \n\t"
+
+      "daddiu     %[src_argb],     %[src_argb],       0x10          \n\t"
+      "daddiu     %[dst_rgb],      %[dst_rgb],        0x08          \n\t"
+      "daddiu     %[width],        %[width],         -0x04          \n\t"
+      "bgtz       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [b] "=&f"(ftmp[0]),
+        [g] "=&f"(ftmp[1]), [r] "=&f"(ftmp[2]), [a] "=&f"(ftmp[3])
+      : [src_argb] "r"(src_argb), [dst_rgb] "r"(dst_rgb), [width] "r"(width),
+        [zero] "f"(0x00), [four] "f"(0x04), [eight] "f"(0x08),
+        [twelve] "f"(0x0c)
+      : "memory");
+}
+
+void ARGBToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest0, dest1, dest2, dest3;
+  const uint64_t value = 0x1080;
+  const uint64_t mask = 0x0001004200810019;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest1],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest2],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest3],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[dest3],          %[src]        \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[eight]      \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[src_hi],       %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
+
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
+      "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3)
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
+        [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
+        [zero] "f"(0x00)
+      : "memory");
+}
+
+void ARGBToUVRow_MMI(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0013002500380002;
+  const uint64_t mask_v = 0x00020038002f0009;
+
+  __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest0_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest1_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest2_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest3_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void BGRAToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest0, dest1, dest2, dest3;
+  const uint64_t value = 0x1080;
+  const uint64_t mask = 0x0019008100420001;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest1],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest2],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest3],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[dest3],          %[src]        \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[eight]      \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[src_hi],       %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
+
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
+      "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3)
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
+        [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
+        [zero] "f"(0x00)
+      : "memory");
+}
+
+void BGRAToUVRow_MMI(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0002003800250013;
+  const uint64_t mask_v = 0x0009002f00380002;
+
+  __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
+      "pinsrh_0   %[dest0_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
+      "pinsrh_0   %[dest1_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
+      "pinsrh_0   %[dest2_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]        \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
+      "pinsrh_0   %[dest3_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void ABGRToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest0, dest1, dest2, dest3;
+  const uint64_t value = 0x1080;
+  const uint64_t mask = 0x0001001900810042;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest1],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest2],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest3],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[dest3],          %[src]        \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[eight]      \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[src_hi],       %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
+
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
+      "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3)
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
+        [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
+        [zero] "f"(0x00)
+      : "memory");
+}
+
+void ABGRToUVRow_MMI(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0002003800250013;
+  const uint64_t mask_v = 0x0009002F00380002;
+
+  __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest0_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest0_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest0_v],      %[dest0_v],        %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest1_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest1_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest1_v],      %[dest1_v],        %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest2_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest2_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest2_v],      %[dest2_v],        %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest3_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest3_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest3_v],      %[dest3_v],        %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void RGBAToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest0, dest1, dest2, dest3;
+  const uint64_t value = 0x1080;
+  const uint64_t mask = 0x0042008100190001;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest1],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest2],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest3],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[dest3],          %[src]        \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[eight]      \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[src_hi],       %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
+
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
+      "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3)
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
+        [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
+        [zero] "f"(0x00)
+      : "memory");
+}
+
+void RGBAToUVRow_MMI(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0013002500380002;
+  const uint64_t mask_v = 0x00020038002f0009;
+
+  __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[dest0_u],      %[src0],           %[value]          \n\t"
+      "dsrl       %[dest0_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest0_v],      %[dest0_v],        %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[dest1_u],      %[src0],           %[value]          \n\t"
+      "dsrl       %[dest1_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest1_v],      %[dest1_v],        %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[dest2_u],      %[src0],           %[value]          \n\t"
+      "dsrl       %[dest2_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest2_v],      %[dest2_v],        %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[dest3_u],      %[src0],           %[value]          \n\t"
+      "dsrl       %[dest3_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[dest3_v],      %[dest3_v],        %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void RGB24ToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest0, dest1, dest2, dest3;
+  const uint64_t value = 0x1080;
+  const uint64_t mask = 0x0001004200810019;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x0d(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x06(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest1],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x13(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x0c(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest2],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x19(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x12(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest3],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[dest3],          %[src]        \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[eight]      \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[src_hi],       %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
+
+      "daddiu     %[src_argb],    %[src_argb],      0x18          \n\t"
+      "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3)
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
+        [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
+        [zero] "f"(0x00)
+      : "memory");
+}
+
+void RGB24ToUVRow_MMI(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0013002500380002;
+  const uint64_t mask_v = 0x00020038002f0009;
+
+  __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest0_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x06(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0d(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x06(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0d(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x0c(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x13(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x0c(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x13(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest1_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x12(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x19(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x12(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x19(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest2_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x1e(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x25(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x1e(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x25(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x24(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2b(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x24(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2b(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest3_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x2a(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x31(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x2a(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x31(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x30              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void RAWToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest0, dest1, dest2, dest3;
+  const uint64_t value = 0x1080;
+  const uint64_t mask = 0x0001001900810042;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x0d(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x06(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest1],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x13(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x0c(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest2],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
+
+      "gsldlc1    %[src],          0x19(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x12(%[src_argb])               \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
+      "dsll       %[src],          %[src],            %[eight]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[zero]       \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask]       \n\t"
+      "punpcklwd  %[src],          %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[dest3],        %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[dest3],          %[src]        \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[eight]      \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[src_hi],       %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest0],        %[src_lo],         %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
+
+      "daddiu     %[src_argb],    %[src_argb],      0x18          \n\t"
+      "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3)
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
+        [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
+        [zero] "f"(0x00)
+      : "memory");
+}
+
+void RAWToUVRow_MMI(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0002003800250013;
+  const uint64_t mask_v = 0x0009002f00380002;
+
+  __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest0_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest0_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest0_v],      %[dest0_v],        %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x06(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0d(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x06(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0d(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x0c(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x13(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x0c(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x13(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest1_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest1_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest1_v],      %[dest1_v],        %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x12(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x19(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x12(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x19(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest2_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest2_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest2_v],      %[dest2_v],        %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x1e(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x25(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x1e(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x25(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x24(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2b(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x24(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2b(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[dest3_u],      %[src0],           %[value]          \n\t"
+      "dsll       %[dest3_v],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest3_v],      %[dest3_v],        %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x2a(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x31(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x2a(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x31(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "dsll       %[src0],         %[src0],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src_hi]         \n\t"
+      "punpcklbh  %[src_lo],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
+      "dsll       %[src1],         %[src1],           %[eight]          \n\t"
+      "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
+      "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
+      "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x30              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void ARGBToYJRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest, dest0, dest1, dest2, dest3;
+  uint64_t tmp0, tmp1;
+  const uint64_t shift = 0x08;
+  const uint64_t value = 0x80;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x0001004D0096001DULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask1]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask1]      \n\t"
+      "punpcklwd  %[tmp0],         %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[tmp1],         %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest0],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[shift]      \n\t"
+
+      "gsldlc1    %[src],          0x0f(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x08(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask1]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask1]      \n\t"
+      "punpcklwd  %[tmp0],         %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[tmp1],         %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest1],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[shift]      \n\t"
+
+      "gsldlc1    %[src],          0x17(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x10(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask1]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask1]      \n\t"
+      "punpcklwd  %[tmp0],         %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[tmp1],         %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest2],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest2],        %[dest2],          %[shift]      \n\t"
+
+      "gsldlc1    %[src],          0x1f(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x18(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask1]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]      \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask1]      \n\t"
+      "punpcklwd  %[tmp0],         %[src_lo],         %[src_hi]     \n\t"
+      "punpckhwd  %[tmp1],         %[src_lo],         %[src_hi]     \n\t"
+      "paddw      %[dest3],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest3],        %[dest3],          %[shift]      \n\t"
+
+      "packsswh   %[tmp0],         %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[tmp1],         %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest],         %[tmp0],           %[tmp1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x20          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest), [src_hi] "=&f"(src_hi),
+        [src_lo] "=&f"(src_lo), [dest0] "=&f"(dest0), [dest1] "=&f"(dest1),
+        [dest2] "=&f"(dest2), [dest3] "=&f"(dest3), [tmp0] "=&f"(tmp0),
+        [tmp1] "=&f"(tmp1)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_y), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [shift] "f"(shift), [value] "f"(value),
+        [width] "r"(width)
+      : "memory");
+}
+
+void ARGBToUVJRow_MMI(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  uint64_t src_rgb1;
+  uint64_t ftmp[12];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0015002a003f0002;
+  const uint64_t mask_v = 0x0002003f0035000a;
+
+  __asm__ volatile(
+      "1:                                                               \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest0_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest1_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest2_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest3_v],      %[src0],           %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
+      "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
+      "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
+      "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
+      "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x10              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src_rgb1] "=&r"(src_rgb1), [src0] "=&f"(ftmp[0]),
+        [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]), [src_hi] "=&f"(ftmp[3]),
+        [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
+        [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
+        [zero] "f"(0x00), [eight] "f"(0x08),
+        [sixteen] "f"(0x10)
+      : "memory");
+}
+
+void RGB565ToYRow_MMI(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  uint64_t ftmp[11];
+  const uint64_t value = 0x1080108010801080;
+  const uint64_t mask = 0x0001004200810019;
+  uint64_t c0 = 0x001f001f001f001f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t c2 = 0x0007000700070007;
+  __asm__ volatile(
+      "1:                                                            \n\t"
+      "gsldrc1    %[src0],        0x00(%[src_rgb565])                \n\t"
+      "gsldlc1    %[src0],        0x07(%[src_rgb565])                \n\t"
+      "psrlh      %[src1],        %[src0],             %[eight]      \n\t"
+      "and        %[b],           %[src0],             %[c0]         \n\t"
+      "and        %[src0],        %[src0],             %[c1]         \n\t"
+      "psrlh      %[src0],        %[src0],             %[five]       \n\t"
+      "and        %[g],           %[src1],             %[c2]         \n\t"
+      "psllh      %[g],           %[g],                %[three]      \n\t"
+      "or         %[g],           %[src0],             %[g]          \n\t"
+      "psrlh      %[r],           %[src1],             %[three]      \n\t"
+      "psllh      %[src0],        %[b],                %[three]      \n\t"
+      "psrlh      %[src1],        %[b],                %[two]        \n\t"
+      "or         %[b],           %[src0],             %[src1]       \n\t"
+      "psllh      %[src0],        %[g],                %[two]        \n\t"
+      "psrlh      %[src1],        %[g],                %[four]       \n\t"
+      "or         %[g],           %[src0],             %[src1]       \n\t"
+      "psllh      %[src0],        %[r],                %[three]      \n\t"
+      "psrlh      %[src1],        %[r],                %[two]        \n\t"
+      "or         %[r],           %[src0],             %[src1]       \n\t"
+      "punpcklhw  %[src0],        %[b],                %[r]          \n\t"
+      "punpcklhw  %[src1],        %[g],                %[value]      \n\t"
+      "punpcklhw  %[src_lo],      %[src0],             %[src1]       \n\t"
+      "punpckhhw  %[src_hi],      %[src0],             %[src1]       \n\t"
+      "pmaddhw    %[src_lo],      %[src_lo],           %[mask]       \n\t"
+      "pmaddhw    %[src_hi],      %[src_hi],           %[mask]       \n\t"
+      "punpcklwd  %[src0],        %[src_lo],           %[src_hi]     \n\t"
+      "punpckhwd  %[src1],        %[src_lo],           %[src_hi]     \n\t"
+      "paddw      %[dest0],       %[src0],             %[src1]       \n\t"
+      "psrlw      %[dest0],       %[dest0],            %[eight]      \n\t"
+
+      "punpckhhw  %[src0],        %[b],                %[r]          \n\t"
+      "punpckhhw  %[src1],        %[g],                %[value]      \n\t"
+      "punpcklhw  %[src_lo],      %[src0],             %[src1]       \n\t"
+      "punpckhhw  %[src_hi],      %[src0],             %[src1]       \n\t"
+      "pmaddhw    %[src_lo],      %[src_lo],           %[mask]       \n\t"
+      "pmaddhw    %[src_hi],      %[src_hi],           %[mask]       \n\t"
+      "punpcklwd  %[src0],        %[src_lo],           %[src_hi]     \n\t"
+      "punpckhwd  %[src1],        %[src_lo],           %[src_hi]     \n\t"
+      "paddw      %[dest1],       %[src0],             %[src1]       \n\t"
+      "psrlw      %[dest1],       %[dest1],            %[eight]      \n\t"
+
+      "gsldrc1    %[src0],        0x08(%[src_rgb565])                \n\t"
+      "gsldlc1    %[src0],        0x0f(%[src_rgb565])                \n\t"
+      "psrlh      %[src1],        %[src0],             %[eight]      \n\t"
+      "and        %[b],           %[src0],             %[c0]         \n\t"
+      "and        %[src0],        %[src0],             %[c1]         \n\t"
+      "psrlh      %[src0],        %[src0],             %[five]       \n\t"
+      "and        %[g],           %[src1],             %[c2]         \n\t"
+      "psllh      %[g],           %[g],                %[three]      \n\t"
+      "or         %[g],           %[src0],             %[g]          \n\t"
+      "psrlh      %[r],           %[src1],             %[three]      \n\t"
+      "psllh      %[src0],        %[b],                %[three]      \n\t"
+      "psrlh      %[src1],        %[b],                %[two]        \n\t"
+      "or         %[b],           %[src0],             %[src1]       \n\t"
+      "psllh      %[src0],        %[g],                %[two]        \n\t"
+      "psrlh      %[src1],        %[g],                %[four]       \n\t"
+      "or         %[g],           %[src0],             %[src1]       \n\t"
+      "psllh      %[src0],        %[r],                %[three]      \n\t"
+      "psrlh      %[src1],        %[r],                %[two]        \n\t"
+      "or         %[r],           %[src0],             %[src1]       \n\t"
+      "punpcklhw  %[src0],        %[b],                %[r]          \n\t"
+      "punpcklhw  %[src1],        %[g],                %[value]      \n\t"
+      "punpcklhw  %[src_lo],      %[src0],             %[src1]       \n\t"
+      "punpckhhw  %[src_hi],      %[src0],             %[src1]       \n\t"
+      "pmaddhw    %[src_lo],      %[src_lo],           %[mask]       \n\t"
+      "pmaddhw    %[src_hi],      %[src_hi],           %[mask]       \n\t"
+      "punpcklwd  %[src0],        %[src_lo],           %[src_hi]     \n\t"
+      "punpckhwd  %[src1],        %[src_lo],           %[src_hi]     \n\t"
+      "paddw      %[dest2],       %[src0],             %[src1]       \n\t"
+      "psrlw      %[dest2],       %[dest2],            %[eight]      \n\t"
+
+      "punpckhhw  %[src0],        %[b],                %[r]          \n\t"
+      "punpckhhw  %[src1],        %[g],                %[value]      \n\t"
+      "punpcklhw  %[src_lo],      %[src0],             %[src1]       \n\t"
+      "punpckhhw  %[src_hi],      %[src0],             %[src1]       \n\t"
+      "pmaddhw    %[src_lo],      %[src_lo],           %[mask]       \n\t"
+      "pmaddhw    %[src_hi],      %[src_hi],           %[mask]       \n\t"
+      "punpcklwd  %[src0],        %[src_lo],           %[src_hi]     \n\t"
+      "punpckhwd  %[src1],        %[src_lo],           %[src_hi]     \n\t"
+      "paddw      %[dest3],       %[src0],             %[src1]       \n\t"
+      "psrlw      %[dest3],       %[dest3],            %[eight]      \n\t"
+
+      "packsswh   %[src_lo],      %[dest0],            %[dest1]      \n\t"
+      "packsswh   %[src_hi],      %[dest2],            %[dest3]      \n\t"
+      "packushb   %[dest0],       %[src_lo],           %[src_hi]     \n\t"
+      "gssdlc1    %[dest0],       0x07(%[dst_y])                     \n\t"
+      "gssdrc1    %[dest0],       0x00(%[dst_y])                     \n\t"
+
+      "daddiu    %[src_rgb565],   %[src_rgb565],       0x10          \n\t"
+      "daddiu    %[dst_y],        %[dst_y],            0x08          \n\t"
+      "daddiu    %[width],        %[width],           -0x08          \n\t"
+      "bgtz      %[width],        1b                                 \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]),
+        [src_hi] "=&f"(ftmp[3]), [b] "=&f"(ftmp[4]), [g] "=&f"(ftmp[5]),
+        [r] "=&f"(ftmp[6]), [dest0] "=&f"(ftmp[7]), [dest1] "=&f"(ftmp[8]),
+        [dest2] "=&f"(ftmp[9]), [dest3] "=&f"(ftmp[10])
+      : [src_rgb565] "r"(src_rgb565), [dst_y] "r"(dst_y), [value] "f"(value),
+        [width] "r"(width), [c0] "f"(c0), [c1] "f"(c1), [c2] "f"(c2),
+        [mask] "f"(mask), [eight] "f"(0x08), [five] "f"(0x05),
+        [three] "f"(0x03), [two] "f"(0x02), [four] "f"(0x04)
+      : "memory");
+}
+
+void ARGB1555ToYRow_MMI(const uint8_t* src_argb1555,
+                        uint8_t* dst_y,
+                        int width) {
+  uint64_t ftmp[11];
+  const uint64_t value = 0x1080108010801080;
+  const uint64_t mask = 0x0001004200810019;
+  uint64_t c0 = 0x001f001f001f001f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t c2 = 0x0003000300030003;
+  uint64_t c3 = 0x007c007c007c007c;
+  __asm__ volatile(
+      "1:                                                            \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb1555])             \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb1555])             \n\t"
+      "psrlh      %[src1],         %[src0],              %[eight]    \n\t"
+      "and        %[b],            %[src0],              %[c0]       \n\t"
+      "and        %[src0],         %[src0],              %[c1]       \n\t"
+      "psrlh      %[src0],         %[src0],              %[five]     \n\t"
+      "and        %[g],            %[src1],              %[c2]       \n\t"
+      "psllh      %[g],            %[g],                 %[three]    \n\t"
+      "or         %[g],            %[src0],              %[g]        \n\t"
+      "and        %[r],            %[src1],              %[c3]       \n\t"
+      "psrlh      %[r],            %[r],                 %[two]      \n\t"
+      "psllh      %[src0],         %[b],                 %[three]    \n\t"
+      "psrlh      %[src1],         %[b],                 %[two]      \n\t"
+      "or         %[b],            %[src0],              %[src1]     \n\t"
+      "psllh      %[src0],         %[g],                 %[three]    \n\t"
+      "psrlh      %[src1],         %[g],                 %[two]      \n\t"
+      "or         %[g],            %[src0],              %[src1]     \n\t"
+      "psllh      %[src0],         %[r],                 %[three]    \n\t"
+      "psrlh      %[src1],         %[r],                 %[two]      \n\t"
+      "or         %[r],            %[src0],              %[src1]     \n\t"
+      "punpcklhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpcklhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest0],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest0],        %[dest0],             %[eight]    \n\t"
+
+      "punpckhhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpckhhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest1],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest1],        %[dest1],             %[eight]    \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_argb1555])             \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_argb1555])             \n\t"
+      "psrlh      %[src1],         %[src0],              %[eight]    \n\t"
+      "and        %[b],            %[src0],              %[c0]       \n\t"
+      "and        %[src0],         %[src0],              %[c1]       \n\t"
+      "psrlh      %[src0],         %[src0],              %[five]     \n\t"
+      "and        %[g],            %[src1],              %[c2]       \n\t"
+      "psllh      %[g],            %[g],                 %[three]    \n\t"
+      "or         %[g],            %[src0],              %[g]        \n\t"
+      "and        %[r],            %[src1],              %[c3]       \n\t"
+      "psrlh      %[r],            %[r],                 %[two]      \n\t"
+      "psllh      %[src0],         %[b],                 %[three]    \n\t"
+      "psrlh      %[src1],         %[b],                 %[two]      \n\t"
+      "or         %[b],            %[src0],              %[src1]     \n\t"
+      "psllh      %[src0],         %[g],                 %[three]    \n\t"
+      "psrlh      %[src1],         %[g],                 %[two]      \n\t"
+      "or         %[g],            %[src0],              %[src1]     \n\t"
+      "psllh      %[src0],         %[r],                 %[three]    \n\t"
+      "psrlh      %[src1],         %[r],                 %[two]      \n\t"
+      "or         %[r],            %[src0],              %[src1]     \n\t"
+      "punpcklhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpcklhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest2],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest2],        %[dest2],             %[eight]    \n\t"
+
+      "punpckhhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpckhhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest3],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest3],        %[dest3],             %[eight]    \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],             %[dest1]    \n\t"
+      "packsswh   %[src_hi],       %[dest2],             %[dest3]    \n\t"
+      "packushb   %[dest0],        %[src_lo],            %[src_hi]   \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                    \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                    \n\t"
+
+      "daddiu     %[src_argb1555], %[src_argb1555],      0x10        \n\t"
+      "daddiu     %[dst_y],        %[dst_y],             0x08        \n\t"
+      "daddiu     %[width],        %[width],            -0x08        \n\t"
+      "bgtz       %[width],        1b                                \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]),
+        [src_hi] "=&f"(ftmp[3]), [b] "=&f"(ftmp[4]), [g] "=&f"(ftmp[5]),
+        [r] "=&f"(ftmp[6]), [dest0] "=&f"(ftmp[7]), [dest1] "=&f"(ftmp[8]),
+        [dest2] "=&f"(ftmp[9]), [dest3] "=&f"(ftmp[10])
+      : [src_argb1555] "r"(src_argb1555), [dst_y] "r"(dst_y),
+        [width] "r"(width), [value] "f"(value), [mask] "f"(mask), [c0] "f"(c0),
+        [c1] "f"(c1), [c2] "f"(c2), [c3] "f"(c3), [eight] "f"(0x08),
+        [five] "f"(0x05), [three] "f"(0x03), [two] "f"(0x02), [seven] "f"(0x07)
+      : "memory");
+}
+
+void ARGB4444ToYRow_MMI(const uint8_t* src_argb4444,
+                        uint8_t* dst_y,
+                        int width) {
+  uint64_t ftmp[11];
+  uint64_t value = 0x1080108010801080;
+  uint64_t mask = 0x0001004200810019;
+  uint64_t c0 = 0x000f000f000f000f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  __asm__ volatile(
+      "1:                                                            \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb4444])             \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb4444])             \n\t"
+      "psrlh      %[src1],         %[src0],              %[eight]    \n\t"
+      "and        %[b],            %[src0],              %[c0]       \n\t"
+      "and        %[src0],         %[src0],              %[c1]       \n\t"
+      "psrlh      %[g],            %[src0],              %[four]     \n\t"
+      "and        %[r],            %[src1],              %[c0]       \n\t"
+      "psllh      %[src0],         %[b],                 %[four]     \n\t"
+      "or         %[b],            %[src0],              %[b]        \n\t"
+      "psllh      %[src0],         %[g],                 %[four]     \n\t"
+      "or         %[g],            %[src0],              %[g]        \n\t"
+      "psllh      %[src0],         %[r],                 %[four]     \n\t"
+      "or         %[r],            %[src0],              %[r]        \n\t"
+      "punpcklhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpcklhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest0],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest0],        %[dest0],             %[eight]    \n\t"
+
+      "punpckhhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpckhhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest1],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest1],        %[dest1],             %[eight]    \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_argb4444])             \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_argb4444])             \n\t"
+      "psrlh      %[src1],         %[src0],              %[eight]    \n\t"
+      "and        %[b],            %[src0],              %[c0]       \n\t"
+      "and        %[src0],         %[src0],              %[c1]       \n\t"
+      "psrlh      %[g],            %[src0],              %[four]     \n\t"
+      "and        %[r],            %[src1],              %[c0]       \n\t"
+      "psllh      %[src0],         %[b],                 %[four]     \n\t"
+      "or         %[b],            %[src0],              %[b]        \n\t"
+      "psllh      %[src0],         %[g],                 %[four]     \n\t"
+      "or         %[g],            %[src0],              %[g]        \n\t"
+      "psllh      %[src0],         %[r],                 %[four]     \n\t"
+      "or         %[r],            %[src0],              %[r]        \n\t"
+      "punpcklhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpcklhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest2],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest2],        %[dest2],             %[eight]    \n\t"
+
+      "punpckhhw  %[src0],         %[b],                 %[r]        \n\t"
+      "punpckhhw  %[src1],         %[g],                 %[value]    \n\t"
+      "punpcklhw  %[src_lo],       %[src0],              %[src1]     \n\t"
+      "punpckhhw  %[src_hi],       %[src0],              %[src1]     \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],            %[mask]     \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],            %[mask]     \n\t"
+      "punpcklwd  %[src0],         %[src_lo],            %[src_hi]   \n\t"
+      "punpckhwd  %[src1],         %[src_lo],            %[src_hi]   \n\t"
+      "paddw      %[dest3],        %[src0],              %[src1]     \n\t"
+      "psrlw      %[dest3],        %[dest3],             %[eight]    \n\t"
+
+      "packsswh   %[src_lo],       %[dest0],             %[dest1]    \n\t"
+      "packsswh   %[src_hi],       %[dest2],             %[dest3]    \n\t"
+      "packushb   %[dest0],        %[src_lo],            %[src_hi]   \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_y])                    \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_y])                    \n\t"
+
+      "daddiu     %[src_argb4444], %[src_argb4444],      0x10        \n\t"
+      "daddiu     %[dst_y],        %[dst_y],             0x08        \n\t"
+      "daddiu     %[width],        %[width],            -0x08        \n\t"
+      "bgtz       %[width],        1b                                \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]),
+        [src_hi] "=&f"(ftmp[3]), [b] "=&f"(ftmp[4]), [g] "=&f"(ftmp[5]),
+        [r] "=&f"(ftmp[6]), [dest0] "=&f"(ftmp[7]), [dest1] "=&f"(ftmp[8]),
+        [dest2] "=&f"(ftmp[9]), [dest3] "=&f"(ftmp[10])
+      : [src_argb4444] "r"(src_argb4444), [dst_y] "r"(dst_y),
+        [width] "r"(width), [value] "f"(value), [mask] "f"(mask), [c0] "f"(c0),
+        [c1] "f"(c1), [eight] "f"(0x08), [four] "f"(0x04)
+      : "memory");
+}
+
+void RGB565ToUVRow_MMI(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  uint64_t ftmp[13];
+  uint64_t value = 0x2020202020202020;
+  uint64_t mask_u = 0x0026004a00700002;
+  uint64_t mask_v = 0x00020070005e0012;
+  uint64_t mask = 0x93;
+  uint64_t c0 = 0x001f001f001f001f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t c2 = 0x0007000700070007;
+  __asm__ volatile(
+      "daddu      %[next_rgb565], %[src_rgb565],       %[next_rgb565]   \n\t"
+      "1:                                                               \n\t"
+      "gsldrc1    %[src0],        0x00(%[src_rgb565])                   \n\t"
+      "gsldlc1    %[src0],        0x07(%[src_rgb565])                   \n\t"
+      "gsldrc1    %[src1],        0x00(%[next_rgb565])                  \n\t"
+      "gsldlc1    %[src1],        0x07(%[next_rgb565])                  \n\t"
+      "psrlh      %[dest0_u],     %[src0],             %[eight]         \n\t"
+      "and        %[b0],          %[src0],             %[c0]            \n\t"
+      "and        %[src0],        %[src0],             %[c1]            \n\t"
+      "psrlh      %[src0],        %[src0],             %[five]          \n\t"
+      "and        %[g0],          %[dest0_u],          %[c2]            \n\t"
+      "psllh      %[g0],          %[g0],               %[three]         \n\t"
+      "or         %[g0],          %[src0],             %[g0]            \n\t"
+      "psrlh      %[r0],          %[dest0_u],          %[three]         \n\t"
+      "psrlh      %[src0],        %[src1],             %[eight]         \n\t"
+      "and        %[dest0_u],     %[src1],             %[c0]            \n\t"
+      "and        %[src1],        %[src1],             %[c1]            \n\t"
+      "psrlh      %[src1],        %[src1],             %[five]          \n\t"
+      "and        %[dest0_v],     %[src0],             %[c2]            \n\t"
+      "psllh      %[dest0_v],     %[dest0_v],          %[three]         \n\t"
+      "or         %[dest0_v],     %[src1],             %[dest0_v]       \n\t"
+      "psrlh      %[src0],        %[src0],             %[three]         \n\t"
+      "paddh      %[b0],          %[b0],               %[dest0_u]       \n\t"
+      "paddh      %[g0],          %[g0],               %[dest0_v]       \n\t"
+      "paddh      %[r0],          %[r0],               %[src0]          \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[r0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[r0]            \n\t"
+      "punpcklwd  %[dest0_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest0_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[src0],        %[dest0_u],          %[dest0_v]       \n\t"
+      "psrlh      %[b0],          %[src0],             %[six]           \n\t"
+      "psllh      %[r0],          %[src0],             %[one]           \n\t"
+      "or         %[b0],          %[b0],               %[r0]            \n\t"
+      "punpcklhw  %[src0],        %[g0],               %[value]         \n\t"
+      "punpckhhw  %[src1],        %[g0],               %[value]         \n\t"
+      "punpcklwd  %[dest0_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest0_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[g0],          %[dest0_u],          %[dest0_v]       \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[g0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[g0]            \n\t"
+
+      "pmaddhw    %[dest0_v],     %[src0],             %[mask_v]        \n\t"
+      "pshufh     %[dest0_u],     %[src0],             %[mask]          \n\t"
+      "pmaddhw    %[dest0_u],     %[dest0_u],          %[mask_u]        \n\t"
+      "pmaddhw    %[g0],          %[src1],             %[mask_v]        \n\t"
+      "pshufh     %[b0],          %[src1],             %[mask]          \n\t"
+      "pmaddhw    %[b0],          %[b0],               %[mask_u]        \n\t"
+
+      "punpcklwd  %[src0],        %[dest0_u],          %[b0]            \n\t"
+      "punpckhwd  %[src1],        %[dest0_u],          %[b0]            \n\t"
+      "psubw      %[dest0_u],     %[src0],             %[src1]          \n\t"
+      "psraw      %[dest0_u],     %[dest0_u],          %[eight]         \n\t"
+      "punpcklwd  %[src0],        %[dest0_v],          %[g0]            \n\t"
+      "punpckhwd  %[src1],        %[dest0_v],          %[g0]            \n\t"
+      "psubw      %[dest0_v],     %[src1],             %[src0]          \n\t"
+      "psraw      %[dest0_v],     %[dest0_v],          %[eight]         \n\t"
+
+      "gsldrc1    %[src0],        0x08(%[src_rgb565])                   \n\t"
+      "gsldlc1    %[src0],        0x0f(%[src_rgb565])                   \n\t"
+      "gsldrc1    %[src1],        0x08(%[next_rgb565])                  \n\t"
+      "gsldlc1    %[src1],        0x0f(%[next_rgb565])                  \n\t"
+      "psrlh      %[dest1_u],     %[src0],             %[eight]         \n\t"
+      "and        %[b0],          %[src0],             %[c0]            \n\t"
+      "and        %[src0],        %[src0],             %[c1]            \n\t"
+      "psrlh      %[src0],        %[src0],             %[five]          \n\t"
+      "and        %[g0],          %[dest1_u],          %[c2]            \n\t"
+      "psllh      %[g0],          %[g0],               %[three]         \n\t"
+      "or         %[g0],          %[src0],             %[g0]            \n\t"
+      "psrlh      %[r0],          %[dest1_u],          %[three]         \n\t"
+      "psrlh      %[src0],        %[src1],             %[eight]         \n\t"
+      "and        %[dest1_u],     %[src1],             %[c0]            \n\t"
+      "and        %[src1],        %[src1],             %[c1]            \n\t"
+      "psrlh      %[src1],        %[src1],             %[five]          \n\t"
+      "and        %[dest1_v],     %[src0],             %[c2]            \n\t"
+      "psllh      %[dest1_v],     %[dest1_v],          %[three]         \n\t"
+      "or         %[dest1_v],     %[src1],             %[dest1_v]       \n\t"
+      "psrlh      %[src0],        %[src0],             %[three]         \n\t"
+      "paddh      %[b0],          %[b0],               %[dest1_u]       \n\t"
+      "paddh      %[g0],          %[g0],               %[dest1_v]       \n\t"
+      "paddh      %[r0],          %[r0],               %[src0]          \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[r0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[r0]            \n\t"
+      "punpcklwd  %[dest1_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest1_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[src0],        %[dest1_u],          %[dest1_v]       \n\t"
+      "psrlh      %[b0],          %[src0],             %[six]           \n\t"
+      "psllh      %[r0],          %[src0],             %[one]           \n\t"
+      "or         %[b0],          %[b0],               %[r0]            \n\t"
+      "punpcklhw  %[src0],        %[g0],               %[value]         \n\t"
+      "punpckhhw  %[src1],        %[g0],               %[value]         \n\t"
+      "punpcklwd  %[dest1_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest1_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[g0],          %[dest1_u],          %[dest1_v]       \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[g0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[g0]            \n\t"
+
+      "pmaddhw    %[dest1_v],     %[src0],             %[mask_v]        \n\t"
+      "pshufh     %[dest1_u],     %[src0],             %[mask]          \n\t"
+      "pmaddhw    %[dest1_u],     %[dest1_u],          %[mask_u]        \n\t"
+      "pmaddhw    %[g0],          %[src1],             %[mask_v]        \n\t"
+      "pshufh     %[b0],          %[src1],             %[mask]          \n\t"
+      "pmaddhw    %[b0],          %[b0],               %[mask_u]        \n\t"
+
+      "punpcklwd  %[src0],        %[dest1_u],          %[b0]            \n\t"
+      "punpckhwd  %[src1],        %[dest1_u],          %[b0]            \n\t"
+      "psubw      %[dest1_u],     %[src0],             %[src1]          \n\t"
+      "psraw      %[dest1_u],     %[dest1_u],          %[eight]         \n\t"
+      "punpcklwd  %[src0],        %[dest1_v],          %[g0]            \n\t"
+      "punpckhwd  %[src1],        %[dest1_v],          %[g0]            \n\t"
+      "psubw      %[dest1_v],     %[src1],             %[src0]          \n\t"
+      "psraw      %[dest1_v],     %[dest1_v],          %[eight]         \n\t"
+
+      "gsldrc1    %[src0],        0x10(%[src_rgb565])                   \n\t"
+      "gsldlc1    %[src0],        0x17(%[src_rgb565])                   \n\t"
+      "gsldrc1    %[src1],        0x10(%[next_rgb565])                  \n\t"
+      "gsldlc1    %[src1],        0x17(%[next_rgb565])                  \n\t"
+      "psrlh      %[dest2_u],     %[src0],             %[eight]         \n\t"
+      "and        %[b0],          %[src0],             %[c0]            \n\t"
+      "and        %[src0],        %[src0],             %[c1]            \n\t"
+      "psrlh      %[src0],        %[src0],             %[five]          \n\t"
+      "and        %[g0],          %[dest2_u],          %[c2]            \n\t"
+      "psllh      %[g0],          %[g0],               %[three]         \n\t"
+      "or         %[g0],          %[src0],             %[g0]            \n\t"
+      "psrlh      %[r0],          %[dest2_u],          %[three]         \n\t"
+      "psrlh      %[src0],        %[src1],             %[eight]         \n\t"
+      "and        %[dest2_u],     %[src1],             %[c0]            \n\t"
+      "and        %[src1],        %[src1],             %[c1]            \n\t"
+      "psrlh      %[src1],        %[src1],             %[five]          \n\t"
+      "and        %[dest2_v],     %[src0],             %[c2]            \n\t"
+      "psllh      %[dest2_v],     %[dest2_v],          %[three]         \n\t"
+      "or         %[dest2_v],     %[src1],             %[dest2_v]       \n\t"
+      "psrlh      %[src0],        %[src0],             %[three]         \n\t"
+      "paddh      %[b0],          %[b0],               %[dest2_u]       \n\t"
+      "paddh      %[g0],          %[g0],               %[dest2_v]       \n\t"
+      "paddh      %[r0],          %[r0],               %[src0]          \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[r0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[r0]            \n\t"
+      "punpcklwd  %[dest2_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest2_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[src0],        %[dest2_u],          %[dest2_v]       \n\t"
+      "psrlh      %[b0],          %[src0],             %[six]           \n\t"
+      "psllh      %[r0],          %[src0],             %[one]           \n\t"
+      "or         %[b0],          %[b0],               %[r0]            \n\t"
+      "punpcklhw  %[src0],        %[g0],               %[value]         \n\t"
+      "punpckhhw  %[src1],        %[g0],               %[value]         \n\t"
+      "punpcklwd  %[dest2_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest2_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[g0],          %[dest2_u],          %[dest2_v]       \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[g0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[g0]            \n\t"
+
+      "pmaddhw    %[dest2_v],     %[src0],             %[mask_v]        \n\t"
+      "pshufh     %[dest2_u],     %[src0],             %[mask]          \n\t"
+      "pmaddhw    %[dest2_u],     %[dest2_u],          %[mask_u]        \n\t"
+      "pmaddhw    %[g0],          %[src1],             %[mask_v]        \n\t"
+      "pshufh     %[b0],          %[src1],             %[mask]          \n\t"
+      "pmaddhw    %[b0],          %[b0],               %[mask_u]        \n\t"
+
+      "punpcklwd  %[src0],        %[dest2_u],          %[b0]            \n\t"
+      "punpckhwd  %[src1],        %[dest2_u],          %[b0]            \n\t"
+      "psubw      %[dest2_u],     %[src0],             %[src1]          \n\t"
+      "psraw      %[dest2_u],     %[dest2_u],          %[eight]         \n\t"
+      "punpcklwd  %[src0],        %[dest2_v],          %[g0]            \n\t"
+      "punpckhwd  %[src1],        %[dest2_v],          %[g0]            \n\t"
+      "psubw      %[dest2_v],     %[src1],             %[src0]          \n\t"
+      "psraw      %[dest2_v],     %[dest2_v],          %[eight]         \n\t"
+
+      "gsldrc1    %[src0],        0x18(%[src_rgb565])                   \n\t"
+      "gsldlc1    %[src0],        0x1f(%[src_rgb565])                   \n\t"
+      "gsldrc1    %[src1],        0x18(%[next_rgb565])                  \n\t"
+      "gsldlc1    %[src1],        0x1f(%[next_rgb565])                  \n\t"
+      "psrlh      %[dest3_u],     %[src0],             %[eight]         \n\t"
+      "and        %[b0],          %[src0],             %[c0]            \n\t"
+      "and        %[src0],        %[src0],             %[c1]            \n\t"
+      "psrlh      %[src0],        %[src0],             %[five]          \n\t"
+      "and        %[g0],          %[dest3_u],          %[c2]            \n\t"
+      "psllh      %[g0],          %[g0],               %[three]         \n\t"
+      "or         %[g0],          %[src0],             %[g0]            \n\t"
+      "psrlh      %[r0],          %[dest3_u],          %[three]         \n\t"
+      "psrlh      %[src0],        %[src1],             %[eight]         \n\t"
+      "and        %[dest3_u],     %[src1],             %[c0]            \n\t"
+      "and        %[src1],        %[src1],             %[c1]            \n\t"
+      "psrlh      %[src1],        %[src1],             %[five]          \n\t"
+      "and        %[dest3_v],     %[src0],             %[c2]            \n\t"
+      "psllh      %[dest3_v],     %[dest3_v],          %[three]         \n\t"
+      "or         %[dest3_v],     %[src1],             %[dest3_v]       \n\t"
+      "psrlh      %[src0],        %[src0],             %[three]         \n\t"
+      "paddh      %[b0],          %[b0],               %[dest3_u]       \n\t"
+      "paddh      %[g0],          %[g0],               %[dest3_v]       \n\t"
+      "paddh      %[r0],          %[r0],               %[src0]          \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[r0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[r0]            \n\t"
+      "punpcklwd  %[dest3_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest3_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[src0],        %[dest3_u],          %[dest3_v]       \n\t"
+      "psrlh      %[b0],          %[src0],             %[six]           \n\t"
+      "psllh      %[r0],          %[src0],             %[one]           \n\t"
+      "or         %[b0],          %[b0],               %[r0]            \n\t"
+      "punpcklhw  %[src0],        %[g0],               %[value]         \n\t"
+      "punpckhhw  %[src1],        %[g0],               %[value]         \n\t"
+      "punpcklwd  %[dest3_u],     %[src0],             %[src1]          \n\t"
+      "punpckhwd  %[dest3_v],     %[src0],             %[src1]          \n\t"
+      "paddh      %[g0],          %[dest3_u],          %[dest3_v]       \n\t"
+      "punpcklhw  %[src0],        %[b0],               %[g0]            \n\t"
+      "punpckhhw  %[src1],        %[b0],               %[g0]            \n\t"
+
+      "pmaddhw    %[dest3_v],     %[src0],             %[mask_v]        \n\t"
+      "pshufh     %[dest3_u],     %[src0],             %[mask]          \n\t"
+      "pmaddhw    %[dest3_u],     %[dest3_u],          %[mask_u]        \n\t"
+      "pmaddhw    %[g0],          %[src1],             %[mask_v]        \n\t"
+      "pshufh     %[b0],          %[src1],             %[mask]          \n\t"
+      "pmaddhw    %[b0],          %[b0],               %[mask_u]        \n\t"
+
+      "punpcklwd  %[src0],        %[dest3_u],          %[b0]            \n\t"
+      "punpckhwd  %[src1],        %[dest3_u],          %[b0]            \n\t"
+      "psubw      %[dest3_u],     %[src0],             %[src1]          \n\t"
+      "psraw      %[dest3_u],     %[dest3_u],          %[eight]         \n\t"
+      "punpcklwd  %[src0],        %[dest3_v],          %[g0]            \n\t"
+      "punpckhwd  %[src1],        %[dest3_v],          %[g0]            \n\t"
+      "psubw      %[dest3_v],     %[src1],             %[src0]          \n\t"
+      "psraw      %[dest3_v],     %[dest3_v],          %[eight]         \n\t"
+
+      "packsswh   %[src0],        %[dest0_u],          %[dest1_u]       \n\t"
+      "packsswh   %[src1],        %[dest2_u],          %[dest3_u]       \n\t"
+      "packushb   %[dest0_u],     %[src0],             %[src1]          \n\t"
+      "gssdlc1    %[dest0_u],     0x07(%[dst_u])                        \n\t"
+      "gssdrc1    %[dest0_u],     0x00(%[dst_u])                        \n\t"
+      "packsswh   %[src0],        %[dest0_v],          %[dest1_v]       \n\t"
+      "packsswh   %[src1],        %[dest2_v],          %[dest3_v]       \n\t"
+      "packushb   %[dest0_v],     %[src0],             %[src1]          \n\t"
+      "gssdlc1    %[dest0_v],     0x07(%[dst_v])                        \n\t"
+      "gssdrc1    %[dest0_v],     0x00(%[dst_v])                        \n\t"
+
+      "daddiu    %[src_rgb565],   %[src_rgb565],       0x20             \n\t"
+      "daddiu    %[next_rgb565],  %[next_rgb565],      0x20             \n\t"
+      "daddiu    %[dst_u],        %[dst_u],            0x08             \n\t"
+      "daddiu    %[dst_v],        %[dst_v],            0x08             \n\t"
+      "daddiu    %[width],        %[width],           -0x10             \n\t"
+      "bgtz      %[width],        1b                                    \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [b0] "=&f"(ftmp[2]),
+        [g0] "=&f"(ftmp[3]), [r0] "=&f"(ftmp[4]), [dest0_u] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest2_u] "=&f"(ftmp[7]),
+        [dest3_u] "=&f"(ftmp[8]), [dest0_v] "=&f"(ftmp[9]),
+        [dest1_v] "=&f"(ftmp[10]), [dest2_v] "=&f"(ftmp[11]),
+        [dest3_v] "=&f"(ftmp[12])
+      : [src_rgb565] "r"(src_rgb565), [next_rgb565] "r"(src_stride_rgb565),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [value] "f"(value), [c0] "f"(c0), [c1] "f"(c1), [c2] "f"(c2),
+        [mask] "f"(mask), [mask_u] "f"(mask_u), [mask_v] "f"(mask_v),
+        [eight] "f"(0x08), [six] "f"(0x06), [five] "f"(0x05), [three] "f"(0x03),
+        [one] "f"(0x01)
+      : "memory");
+}
+
+void ARGB1555ToUVRow_MMI(const uint8_t* src_argb1555,
+                         int src_stride_argb1555,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  uint64_t ftmp[11];
+  uint64_t value = 0x2020202020202020;
+  uint64_t mask_u = 0x0026004a00700002;
+  uint64_t mask_v = 0x00020070005e0012;
+  uint64_t mask = 0x93;
+  uint64_t c0 = 0x001f001f001f001f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t c2 = 0x0003000300030003;
+  uint64_t c3 = 0x007c007c007c007c;
+  __asm__ volatile(
+      "daddu      %[next_argb1555], %[src_argb1555],      %[next_argb1555] \n\t"
+      "1:                                                                  \n\t"
+      "gsldrc1    %[src0],          0x00(%[src_argb1555])                  \n\t"
+      "gsldlc1    %[src0],          0x07(%[src_argb1555])                  \n\t"
+      "gsldrc1    %[src1],          0x00(%[next_argb1555])                 \n\t"
+      "gsldlc1    %[src1],          0x07(%[next_argb1555])                 \n\t"
+      "psrlh      %[dest0_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[five]         \n\t"
+      "and        %[g0],            %[dest0_u],            %[c2]           \n\t"
+      "psllh      %[g0],            %[g0],                 %[three]        \n\t"
+      "or         %[g0],            %[src0],               %[g0]           \n\t"
+      "and        %[r0],            %[dest0_u],            %[c3]           \n\t"
+      "psrlh      %[r0],            %[r0],                 %[two]          \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest0_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[src1],          %[src1],               %[five]         \n\t"
+      "and        %[dest0_v],       %[src0],               %[c2]           \n\t"
+      "psllh      %[dest0_v],       %[dest0_v],            %[three]        \n\t"
+      "or         %[dest0_v],       %[src1],               %[dest0_v]      \n\t"
+      "and        %[src0],          %[src0],               %[c3]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[two]          \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest0_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest0_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest0_u],            %[dest0_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[six]          \n\t"
+      "psllh      %[r0],            %[src0],               %[one]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[six]          \n\t"
+      "psllh      %[g0],            %[g0],                 %[one]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest0_u],            %[dest0_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest0_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest0_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest0_u],       %[dest0_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest0_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest0_u],            %[b0]           \n\t"
+      "psubw      %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest0_u],       %[dest0_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest0_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest0_v],            %[g0]           \n\t"
+      "psubw      %[dest0_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest0_v],       %[dest0_v],            %[eight]        \n\t"
+
+      "gsldrc1    %[src0],          0x08(%[src_argb1555])                  \n\t"
+      "gsldlc1    %[src0],          0x0f(%[src_argb1555])                  \n\t"
+      "gsldrc1    %[src1],          0x08(%[next_argb1555])                 \n\t"
+      "gsldlc1    %[src1],          0x0f(%[next_argb1555])                 \n\t"
+      "psrlh      %[dest1_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[five]         \n\t"
+      "and        %[g0],            %[dest1_u],            %[c2]           \n\t"
+      "psllh      %[g0],            %[g0],                 %[three]        \n\t"
+      "or         %[g0],            %[src0],               %[g0]           \n\t"
+      "and        %[r0],            %[dest1_u],            %[c3]           \n\t"
+      "psrlh      %[r0],            %[r0],                 %[two]          \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest1_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[src1],          %[src1],               %[five]         \n\t"
+      "and        %[dest1_v],       %[src0],               %[c2]           \n\t"
+      "psllh      %[dest1_v],       %[dest1_v],            %[three]        \n\t"
+      "or         %[dest1_v],       %[src1],               %[dest1_v]      \n\t"
+      "and        %[src0],          %[src0],               %[c3]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[two]          \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest1_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest1_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest1_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest1_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest1_u],            %[dest1_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[six]          \n\t"
+      "psllh      %[r0],            %[src0],               %[one]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[six]          \n\t"
+      "psllh      %[g0],            %[g0],                 %[one]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest1_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest1_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest1_u],            %[dest1_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest1_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest1_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest1_u],       %[dest1_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest1_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest1_u],            %[b0]           \n\t"
+      "psubw      %[dest1_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest1_u],       %[dest1_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest1_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest1_v],            %[g0]           \n\t"
+      "psubw      %[dest1_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest1_v],       %[dest1_v],            %[eight]        \n\t"
+
+      "packsswh   %[dest0_u],       %[dest0_u],            %[dest1_u]      \n\t"
+      "packsswh   %[dest1_u],       %[dest0_v],            %[dest1_v]      \n\t"
+
+      "gsldrc1    %[src0],          0x10(%[src_argb1555])                  \n\t"
+      "gsldlc1    %[src0],          0x17(%[src_argb1555])                  \n\t"
+      "gsldrc1    %[src1],          0x10(%[next_argb1555])                 \n\t"
+      "gsldlc1    %[src1],          0x17(%[next_argb1555])                 \n\t"
+      "psrlh      %[dest2_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[five]         \n\t"
+      "and        %[g0],            %[dest2_u],            %[c2]           \n\t"
+      "psllh      %[g0],            %[g0],                 %[three]        \n\t"
+      "or         %[g0],            %[src0],               %[g0]           \n\t"
+      "and        %[r0],            %[dest2_u],            %[c3]           \n\t"
+      "psrlh      %[r0],            %[r0],                 %[two]          \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest2_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[src1],          %[src1],               %[five]         \n\t"
+      "and        %[dest0_v],       %[src0],               %[c2]           \n\t"
+      "psllh      %[dest0_v],       %[dest0_v],            %[three]        \n\t"
+      "or         %[dest0_v],       %[src1],               %[dest0_v]      \n\t"
+      "and        %[src0],          %[src0],               %[c3]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[two]          \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest2_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest0_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest2_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest2_u],            %[dest0_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[six]          \n\t"
+      "psllh      %[r0],            %[src0],               %[one]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[six]          \n\t"
+      "psllh      %[g0],            %[g0],                 %[one]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest2_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest2_u],            %[dest0_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest0_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest2_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest2_u],       %[dest2_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest2_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest2_u],            %[b0]           \n\t"
+      "psubw      %[dest2_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest2_u],       %[dest2_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest0_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest0_v],            %[g0]           \n\t"
+      "psubw      %[dest0_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest0_v],       %[dest0_v],            %[eight]        \n\t"
+
+      "gsldrc1    %[src0],          0x18(%[src_argb1555])                  \n\t"
+      "gsldlc1    %[src0],          0x1f(%[src_argb1555])                  \n\t"
+      "gsldrc1    %[src1],          0x18(%[next_argb1555])                 \n\t"
+      "gsldlc1    %[src1],          0x1f(%[next_argb1555])                 \n\t"
+      "psrlh      %[dest3_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[five]         \n\t"
+      "and        %[g0],            %[dest3_u],            %[c2]           \n\t"
+      "psllh      %[g0],            %[g0],                 %[three]        \n\t"
+      "or         %[g0],            %[src0],               %[g0]           \n\t"
+      "and        %[r0],            %[dest3_u],            %[c3]           \n\t"
+      "psrlh      %[r0],            %[r0],                 %[two]          \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest3_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[src1],          %[src1],               %[five]         \n\t"
+      "and        %[dest1_v],       %[src0],               %[c2]           \n\t"
+      "psllh      %[dest1_v],       %[dest1_v],            %[three]        \n\t"
+      "or         %[dest1_v],       %[src1],               %[dest1_v]      \n\t"
+      "and        %[src0],          %[src0],               %[c3]           \n\t"
+      "psrlh      %[src0],          %[src0],               %[two]          \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest3_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest1_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest3_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest1_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest3_u],            %[dest1_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[six]          \n\t"
+      "psllh      %[r0],            %[src0],               %[one]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[six]          \n\t"
+      "psllh      %[g0],            %[g0],                 %[one]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest3_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest1_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest3_u],            %[dest1_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest1_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest3_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest3_u],       %[dest3_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest3_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest3_u],            %[b0]           \n\t"
+      "psubw      %[dest3_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest3_u],       %[dest3_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest1_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest1_v],            %[g0]           \n\t"
+      "psubw      %[dest1_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest1_v],       %[dest1_v],            %[eight]        \n\t"
+
+      "packsswh   %[src1],          %[dest2_u],            %[dest3_u]      \n\t"
+      "packushb   %[dest0_u],       %[dest0_u],            %[src1]         \n\t"
+      "gssdlc1    %[dest0_u],       0x07(%[dst_u])                         \n\t"
+      "gssdrc1    %[dest0_u],       0x00(%[dst_u])                         \n\t"
+      "packsswh   %[src1],          %[dest0_v],            %[dest1_v]      \n\t"
+      "packushb   %[dest0_v],       %[dest1_u],            %[src1]         \n\t"
+      "gssdlc1    %[dest0_v],       0x07(%[dst_v])                         \n\t"
+      "gssdrc1    %[dest0_v],       0x00(%[dst_v])                         \n\t"
+
+      "daddiu    %[src_argb1555],   %[src_argb1555],       0x20            \n\t"
+      "daddiu    %[next_argb1555],  %[next_argb1555],      0x20            \n\t"
+      "daddiu    %[dst_u],          %[dst_u],              0x08            \n\t"
+      "daddiu    %[dst_v],          %[dst_v],              0x08            \n\t"
+      "daddiu    %[width],          %[width],             -0x10            \n\t"
+      "bgtz      %[width],          1b                                     \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [b0] "=&f"(ftmp[2]),
+        [g0] "=&f"(ftmp[3]), [r0] "=&f"(ftmp[4]), [dest0_u] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest2_u] "=&f"(ftmp[7]),
+        [dest3_u] "=&f"(ftmp[8]), [dest0_v] "=&f"(ftmp[9]),
+        [dest1_v] "=&f"(ftmp[10])
+      : [src_argb1555] "r"(src_argb1555),
+        [next_argb1555] "r"(src_stride_argb1555), [dst_u] "r"(dst_u),
+        [dst_v] "r"(dst_v), [width] "r"(width), [value] "f"(value),
+        [c0] "f"(c0), [c1] "f"(c1), [c2] "f"(c2), [c3] "f"(c3),
+        [mask] "f"(mask), [mask_u] "f"(mask_u), [mask_v] "f"(mask_v),
+        [eight] "f"(0x08), [six] "f"(0x06), [five] "f"(0x05), [three] "f"(0x03),
+        [two] "f"(0x02), [one] "f"(0x01)
+      : "memory");
+}
+
+void ARGB4444ToUVRow_MMI(const uint8_t* src_argb4444,
+                         int src_stride_argb4444,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  uint64_t ftmp[13];
+  uint64_t value = 0x2020202020202020;
+  uint64_t mask_u = 0x0026004a00700002;
+  uint64_t mask_v = 0x00020070005e0012;
+  uint64_t mask = 0x93;
+  uint64_t c0 = 0x000f000f000f000f;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  __asm__ volatile(
+      "daddu      %[next_argb4444], %[src_argb4444],      %[next_argb4444] \n\t"
+      "1:                                                                  \n\t"
+      "gsldrc1    %[src0],          0x00(%[src_argb4444])                  \n\t"
+      "gsldlc1    %[src0],          0x07(%[src_argb4444])                  \n\t"
+      "gsldrc1    %[src1],          0x00(%[next_argb4444])                 \n\t"
+      "gsldlc1    %[src1],          0x07(%[next_argb4444])                 \n\t"
+      "psrlh      %[dest0_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[g0],            %[src0],               %[four]         \n\t"
+      "and        %[r0],            %[dest0_u],            %[c0]           \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest0_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[dest0_v],       %[src1],               %[four]         \n\t"
+      "and        %[src0],          %[src0],               %[c0]           \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest0_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest0_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest0_u],            %[dest0_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[four]         \n\t"
+      "psllh      %[r0],            %[src0],               %[two]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[four]         \n\t"
+      "psllh      %[g0],            %[g0],                 %[two]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest0_u],            %[dest0_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest0_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest0_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest0_u],       %[dest0_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest0_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest0_u],            %[b0]           \n\t"
+      "psubw      %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest0_u],       %[dest0_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest0_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest0_v],            %[g0]           \n\t"
+      "psubw      %[dest0_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest0_v],       %[dest0_v],            %[eight]        \n\t"
+
+      "gsldrc1    %[src0],          0x08(%[src_argb4444])                  \n\t"
+      "gsldlc1    %[src0],          0x0f(%[src_argb4444])                  \n\t"
+      "gsldrc1    %[src1],          0x08(%[next_argb4444])                 \n\t"
+      "gsldlc1    %[src1],          0x0f(%[next_argb4444])                 \n\t"
+      "psrlh      %[dest1_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[g0],            %[src0],               %[four]         \n\t"
+      "and        %[r0],            %[dest1_u],            %[c0]           \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest1_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[dest1_v],       %[src1],               %[four]         \n\t"
+      "and        %[src0],          %[src0],               %[c0]           \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest1_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest1_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest1_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest1_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest1_u],            %[dest1_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[four]         \n\t"
+      "psllh      %[r0],            %[src0],               %[two]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[four]         \n\t"
+      "psllh      %[g0],            %[g0],                 %[two]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest1_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest1_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest1_u],            %[dest1_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest1_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest1_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest1_u],       %[dest1_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest1_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest1_u],            %[b0]           \n\t"
+      "psubw      %[dest1_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest1_u],       %[dest1_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest1_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest1_v],            %[g0]           \n\t"
+      "psubw      %[dest1_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest1_v],       %[dest1_v],            %[eight]        \n\t"
+
+      "gsldrc1    %[src0],          0x10(%[src_argb4444])                  \n\t"
+      "gsldlc1    %[src0],          0x17(%[src_argb4444])                  \n\t"
+      "gsldrc1    %[src1],          0x10(%[next_argb4444])                 \n\t"
+      "gsldlc1    %[src1],          0x17(%[next_argb4444])                 \n\t"
+      "psrlh      %[dest2_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[g0],            %[src0],               %[four]         \n\t"
+      "and        %[r0],            %[dest2_u],            %[c0]           \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest2_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[dest2_v],       %[src1],               %[four]         \n\t"
+      "and        %[src0],          %[src0],               %[c0]           \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest2_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest2_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest2_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest2_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest2_u],            %[dest2_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[four]         \n\t"
+      "psllh      %[r0],            %[src0],               %[two]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[four]         \n\t"
+      "psllh      %[g0],            %[g0],                 %[two]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest2_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest2_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest2_u],            %[dest2_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest2_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest2_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest2_u],       %[dest2_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest2_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest2_u],            %[b0]           \n\t"
+      "psubw      %[dest2_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest2_u],       %[dest2_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest2_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest2_v],            %[g0]           \n\t"
+      "psubw      %[dest2_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest2_v],       %[dest2_v],            %[eight]        \n\t"
+
+      "gsldrc1    %[src0],          0x18(%[src_argb4444])                  \n\t"
+      "gsldlc1    %[src0],          0x1f(%[src_argb4444])                  \n\t"
+      "gsldrc1    %[src1],          0x18(%[next_argb4444])                 \n\t"
+      "gsldlc1    %[src1],          0x1f(%[next_argb4444])                 \n\t"
+      "psrlh      %[dest3_u],       %[src0],               %[eight]        \n\t"
+      "and        %[b0],            %[src0],               %[c0]           \n\t"
+      "and        %[src0],          %[src0],               %[c1]           \n\t"
+      "psrlh      %[g0],            %[src0],               %[four]         \n\t"
+      "and        %[r0],            %[dest3_u],            %[c0]           \n\t"
+      "psrlh      %[src0],          %[src1],               %[eight]        \n\t"
+      "and        %[dest3_u],       %[src1],               %[c0]           \n\t"
+      "and        %[src1],          %[src1],               %[c1]           \n\t"
+      "psrlh      %[dest3_v],       %[src1],               %[four]         \n\t"
+      "and        %[src0],          %[src0],               %[c0]           \n\t"
+      "paddh      %[b0],            %[b0],                 %[dest3_u]      \n\t"
+      "paddh      %[g0],            %[g0],                 %[dest3_v]      \n\t"
+      "paddh      %[r0],            %[r0],                 %[src0]         \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[r0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[r0]           \n\t"
+      "punpcklwd  %[dest3_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest3_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[src0],          %[dest3_u],            %[dest3_v]      \n\t"
+      "psrlh      %[b0],            %[src0],               %[four]         \n\t"
+      "psllh      %[r0],            %[src0],               %[two]          \n\t"
+      "or         %[b0],            %[b0],                 %[r0]           \n\t"
+      "psrlh      %[r0],            %[g0],                 %[four]         \n\t"
+      "psllh      %[g0],            %[g0],                 %[two]          \n\t"
+      "or         %[g0],            %[g0],                 %[r0]           \n\t"
+      "punpcklhw  %[src0],          %[g0],                 %[value]        \n\t"
+      "punpckhhw  %[src1],          %[g0],                 %[value]        \n\t"
+      "punpcklwd  %[dest3_u],       %[src0],               %[src1]         \n\t"
+      "punpckhwd  %[dest3_v],       %[src0],               %[src1]         \n\t"
+      "paddh      %[g0],            %[dest3_u],            %[dest3_v]      \n\t"
+      "punpcklhw  %[src0],          %[b0],                 %[g0]           \n\t"
+      "punpckhhw  %[src1],          %[b0],                 %[g0]           \n\t"
+
+      "pmaddhw    %[dest3_v],       %[src0],               %[mask_v]       \n\t"
+      "pshufh     %[dest3_u],       %[src0],               %[mask]         \n\t"
+      "pmaddhw    %[dest3_u],       %[dest3_u],            %[mask_u]       \n\t"
+      "pmaddhw    %[g0],            %[src1],               %[mask_v]       \n\t"
+      "pshufh     %[b0],            %[src1],               %[mask]         \n\t"
+      "pmaddhw    %[b0],            %[b0],                 %[mask_u]       \n\t"
+
+      "punpcklwd  %[src0],          %[dest3_u],            %[b0]           \n\t"
+      "punpckhwd  %[src1],          %[dest3_u],            %[b0]           \n\t"
+      "psubw      %[dest3_u],       %[src0],               %[src1]         \n\t"
+      "psraw      %[dest3_u],       %[dest3_u],            %[eight]        \n\t"
+      "punpcklwd  %[src0],          %[dest3_v],            %[g0]           \n\t"
+      "punpckhwd  %[src1],          %[dest3_v],            %[g0]           \n\t"
+      "psubw      %[dest3_v],       %[src1],               %[src0]         \n\t"
+      "psraw      %[dest3_v],       %[dest3_v],            %[eight]        \n\t"
+
+      "packsswh   %[src0],          %[dest0_u],            %[dest1_u]      \n\t"
+      "packsswh   %[src1],          %[dest2_u],            %[dest3_u]      \n\t"
+      "packushb   %[dest0_u],       %[src0],               %[src1]         \n\t"
+      "gssdlc1    %[dest0_u],       0x07(%[dst_u])                         \n\t"
+      "gssdrc1    %[dest0_u],       0x00(%[dst_u])                         \n\t"
+      "packsswh   %[src0],          %[dest0_v],            %[dest1_v]      \n\t"
+      "packsswh   %[src1],          %[dest2_v],            %[dest3_v]      \n\t"
+      "packushb   %[dest0_v],       %[src0],               %[src1]         \n\t"
+      "gssdlc1    %[dest0_v],       0x07(%[dst_v])                         \n\t"
+      "gssdrc1    %[dest0_v],       0x00(%[dst_v])                         \n\t"
+
+      "daddiu    %[src_argb4444],   %[src_argb4444],       0x20            \n\t"
+      "daddiu    %[next_argb4444],  %[next_argb4444],      0x20            \n\t"
+      "daddiu    %[dst_u],          %[dst_u],              0x08            \n\t"
+      "daddiu    %[dst_v],          %[dst_v],              0x08            \n\t"
+      "daddiu    %[width],          %[width],             -0x10            \n\t"
+      "bgtz      %[width],          1b                                     \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [b0] "=&f"(ftmp[2]),
+        [g0] "=&f"(ftmp[3]), [r0] "=&f"(ftmp[4]), [dest0_u] "=&f"(ftmp[5]),
+        [dest1_u] "=&f"(ftmp[6]), [dest2_u] "=&f"(ftmp[7]),
+        [dest3_u] "=&f"(ftmp[8]), [dest0_v] "=&f"(ftmp[9]),
+        [dest1_v] "=&f"(ftmp[10]), [dest2_v] "=&f"(ftmp[11]),
+        [dest3_v] "=&f"(ftmp[12])
+      : [src_argb4444] "r"(src_argb4444),
+        [next_argb4444] "r"(src_stride_argb4444), [dst_u] "r"(dst_u),
+        [dst_v] "r"(dst_v), [width] "r"(width), [value] "f"(value),
+        [c0] "f"(c0), [c1] "f"(c1), [mask] "f"(mask), [mask_u] "f"(mask_u),
+        [mask_v] "f"(mask_v), [eight] "f"(0x08), [four] "f"(0x04),
+        [two] "f"(0x02)
+      : "memory");
+}
+
+void ARGBToUV444Row_MMI(const uint8_t* src_argb,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  uint64_t ftmp[12];
+  const uint64_t value = 0x4040;
+  const uint64_t mask_u = 0x0026004a00700002;
+  const uint64_t mask_v = 0x00020070005e0012;
+
+  __asm__ volatile(
+      "1:                                                               \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_argb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_argb])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "dsll       %[dest0_u],      %[src_lo],         %[sixteen]        \n\t"
+      "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest0_v],      %[src_lo],         %[value]          \n\t"
+      "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
+
+      "dsll       %[src_lo],       %[src_hi],         %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest0_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_u],        %[src_lo]         \n\t"
+      "psubw      %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest0_u],      %[dest0_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest0_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest0_v],        %[src_hi]         \n\t"
+      "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x08(%[src_argb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_argb])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "dsll       %[dest1_u],      %[src_lo],         %[sixteen]        \n\t"
+      "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest1_v],      %[src_lo],         %[value]          \n\t"
+      "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
+      "dsll       %[src_lo],       %[src_hi],         %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest1_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_u],        %[src_lo]         \n\t"
+      "psubw      %[dest1_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest1_u],      %[dest1_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest1_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest1_v],        %[src_hi]         \n\t"
+      "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_argb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_argb])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "dsll       %[dest2_u],      %[src_lo],         %[sixteen]        \n\t"
+      "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest2_v],      %[src_lo],         %[value]          \n\t"
+      "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
+      "dsll       %[src_lo],       %[src_hi],         %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest2_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_u],        %[src_lo]         \n\t"
+      "psubw      %[dest2_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest2_u],      %[dest2_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest2_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest2_v],        %[src_hi]         \n\t"
+      "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
+
+      "gsldrc1    %[src0],         0x18(%[src_argb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_argb])                    \n\t"
+      "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
+      "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
+      "dsll       %[dest3_u],      %[src_lo],         %[sixteen]        \n\t"
+      "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
+      "pinsrh_3   %[dest3_v],      %[src_lo],         %[value]          \n\t"
+      "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
+      "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
+      "dsll       %[src_lo],       %[src_hi],         %[sixteen]        \n\t"
+      "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
+      "pmaddhw    %[src_lo],       %[src_lo],         %[mask_u]         \n\t"
+      "pmaddhw    %[src_hi],       %[src_hi],         %[mask_v]         \n\t"
+
+      "punpcklwd  %[src0],         %[dest3_u],        %[src_lo]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_u],        %[src_lo]         \n\t"
+      "psubw      %[dest3_u],      %[src0],           %[src1]           \n\t"
+      "psraw      %[dest3_u],      %[dest3_u],        %[eight]          \n\t"
+      "punpcklwd  %[src0],         %[dest3_v],        %[src_hi]         \n\t"
+      "punpckhwd  %[src1],         %[dest3_v],        %[src_hi]         \n\t"
+      "psubw      %[dest3_v],      %[src1],           %[src0]           \n\t"
+      "psraw      %[dest3_v],      %[dest3_v],        %[eight]          \n\t"
+
+      "packsswh   %[src0],         %[dest0_u],        %[dest1_u]        \n\t"
+      "packsswh   %[src1],         %[dest2_u],        %[dest3_u]        \n\t"
+      "packushb   %[dest0_u],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_u],      0x07(%[dst_u])                       \n\t"
+      "gssdrc1    %[dest0_u],      0x00(%[dst_u])                       \n\t"
+
+      "packsswh   %[src0],         %[dest0_v],        %[dest1_v]        \n\t"
+      "packsswh   %[src1],         %[dest2_v],        %[dest3_v]        \n\t"
+      "packushb   %[dest0_v],      %[src0],           %[src1]           \n\t"
+      "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
+      "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
+
+      "daddiu     %[src_argb],     %[src_argb],       0x20              \n\t"
+      "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
+      "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
+      "daddi      %[width],        %[width],         -0x08              \n\t"
+      "bgtz       %[width],        1b                                   \n\t"
+      : [src0] "=&f"(ftmp[0]), [src1] "=&f"(ftmp[1]), [src_lo] "=&f"(ftmp[2]),
+        [src_hi] "=&f"(ftmp[3]), [dest0_u] "=&f"(ftmp[4]),
+        [dest0_v] "=&f"(ftmp[5]), [dest1_u] "=&f"(ftmp[6]),
+        [dest1_v] "=&f"(ftmp[7]), [dest2_u] "=&f"(ftmp[8]),
+        [dest2_v] "=&f"(ftmp[9]), [dest3_u] "=&f"(ftmp[10]),
+        [dest3_v] "=&f"(ftmp[11])
+      : [src_argb] "r"(src_argb), [dst_u] "r"(dst_u), [dst_v] "r"(dst_v),
+        [width] "r"(width), [mask_u] "f"(mask_u), [mask_v] "f"(mask_v),
+        [value] "f"(value), [zero] "f"(0x00), [sixteen] "f"(0x10),
+        [eight] "f"(0x08)
+      : "memory");
+}
+
+void ARGBGrayRow_MMI(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  uint64_t src, src_lo, src_hi, src37, dest, dest_lo, dest_hi;
+  uint64_t tmp0, tmp1;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x01;
+  const uint64_t mask2 = 0x0080004D0096001DULL;
+  const uint64_t mask3 = 0xFF000000FF000000ULL;
+  const uint64_t mask4 = ~mask3;
+  const uint64_t shift = 0x08;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+
+      "and        %[src37],        %[src],            %[mask3]      \n\t"
+
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_lo],       %[src_lo],         %[mask1]      \n\t"
+      "pmaddhw    %[dest_lo],      %[src_lo],         %[mask2]      \n\t"
+      "punpcklwd  %[tmp0],         %[dest_lo],        %[dest_lo]    \n\t"
+      "punpckhwd  %[tmp1],         %[dest_lo],        %[dest_lo]    \n\t"
+      "paddw      %[dest_lo],      %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest_lo],      %[dest_lo],        %[shift]      \n\t"
+      "packsswh   %[dest_lo],      %[dest_lo],        %[dest_lo]    \n\t"
+
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+      "pinsrh_3   %[src_hi],       %[src_hi],         %[mask1]      \n\t"
+      "pmaddhw    %[dest_hi],      %[src_hi],         %[mask2]      \n\t"
+      "punpcklwd  %[tmp0],         %[dest_hi],        %[dest_hi]    \n\t"
+      "punpckhwd  %[tmp1],         %[dest_hi],        %[dest_hi]    \n\t"
+      "paddw      %[dest_hi],      %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest_hi],      %[dest_hi],        %[shift]      \n\t"
+      "packsswh   %[dest_hi],      %[dest_hi],        %[dest_hi]    \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "and        %[dest],         %[dest],           %[mask4]      \n\t"
+      "or         %[dest],         %[dest],           %[src37]      \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo), [tmp0] "=&f"(tmp0),
+        [tmp1] "=&f"(tmp1), [src] "=&f"(src), [dest] "=&f"(dest),
+        [src37] "=&f"(src37)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb), [width] "r"(width),
+        [shift] "f"(shift), [mask0] "f"(mask0), [mask1] "f"(mask1),
+        [mask2] "f"(mask2), [mask3] "f"(mask3), [mask4] "f"(mask4)
+      : "memory");
+}
+
+// Convert a row of image to Sepia tone.
+void ARGBSepiaRow_MMI(uint8_t* dst_argb, int width) {
+  uint64_t dest, dest_lo, dest_hi, dest37, dest0, dest1, dest2;
+  uint64_t tmp0, tmp1;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x002300440011ULL;
+  const uint64_t mask2 = 0x002D00580016ULL;
+  const uint64_t mask3 = 0x003200620018ULL;
+  const uint64_t mask4 = 0xFF000000FF000000ULL;
+  const uint64_t shift = 0x07;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gsldrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "and        %[dest37],       %[dest],           %[mask4]      \n\t"
+
+      "punpcklbh  %[dest_lo],      %[dest],           %[mask0]      \n\t"
+      "pmaddhw    %[dest0],        %[dest_lo],        %[mask1]      \n\t"
+      "pmaddhw    %[dest1],        %[dest_lo],        %[mask2]      \n\t"
+      "pmaddhw    %[dest2],        %[dest_lo],        %[mask3]      \n\t"
+      "punpcklwd  %[tmp0],         %[dest0],          %[dest1]      \n\t"
+      "punpckhwd  %[tmp1],         %[dest0],          %[dest1]      \n\t"
+      "paddw      %[dest0],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[shift]      \n\t"
+      "punpcklwd  %[tmp0],         %[dest2],          %[mask0]      \n\t"
+      "punpckhwd  %[tmp1],         %[dest2],          %[mask0]      \n\t"
+      "paddw      %[dest1],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[shift]      \n\t"
+      "packsswh   %[dest_lo],      %[dest0],          %[dest1]      \n\t"
+
+      "punpckhbh  %[dest_hi],      %[dest],           %[mask0]      \n\t"
+      "pmaddhw    %[dest0],        %[dest_hi],        %[mask1]      \n\t"
+      "pmaddhw    %[dest1],        %[dest_hi],        %[mask2]      \n\t"
+      "pmaddhw    %[dest2],        %[dest_hi],        %[mask3]      \n\t"
+      "punpcklwd  %[tmp0],         %[dest0],          %[dest1]      \n\t"
+      "punpckhwd  %[tmp1],         %[dest0],          %[dest1]      \n\t"
+      "paddw      %[dest0],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[shift]      \n\t"
+      "punpcklwd  %[tmp0],         %[dest2],          %[mask0]      \n\t"
+      "punpckhwd  %[tmp1],         %[dest2],          %[mask0]      \n\t"
+      "paddw      %[dest1],        %[tmp0],           %[tmp1]       \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[shift]      \n\t"
+      "packsswh   %[dest_hi],      %[dest0],          %[dest1]      \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "or         %[dest],         %[dest],           %[dest37]     \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest37] "=&f"(dest37), [tmp0] "=&f"(tmp0), [tmp1] "=&f"(tmp1),
+        [dest] "=&f"(dest)
+      : [dst_ptr] "r"(dst_argb), [width] "r"(width), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [mask2] "f"(mask2), [mask3] "f"(mask3),
+        [mask4] "f"(mask4), [shift] "f"(shift)
+      : "memory");
+}
+
+// Apply color matrix to a row of image. Matrix is signed.
+// TODO(fbarchard): Consider adding rounding (+32).
+void ARGBColorMatrixRow_MMI(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const int8_t* matrix_argb,
+                            int width) {
+  uint64_t src, src_hi, src_lo, dest, dest_lo, dest_hi, dest0, dest1, dest2,
+      dest3;
+  uint64_t matrix, matrix_hi, matrix_lo;
+  uint64_t tmp0, tmp1;
+  const uint64_t shift0 = 0x06;
+  const uint64_t shift1 = 0x08;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x08;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+
+      "gsldlc1    %[matrix],       0x07(%[matrix_ptr])              \n\t"
+      "gsldrc1    %[matrix],       0x00(%[matrix_ptr])              \n\t"
+      "punpcklbh  %[matrix_lo],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "psrah      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "punpckhbh  %[matrix_hi],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "psrah      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "pmaddhw    %[dest_lo],      %[src_lo],         %[matrix_lo]  \n\t"
+      "pmaddhw    %[dest_hi],      %[src_lo],         %[matrix_hi]  \n\t"
+      "punpcklwd  %[tmp0],         %[dest_lo],        %[dest_hi]    \n\t"
+      "punpckhwd  %[tmp1],         %[dest_lo],        %[dest_hi]    \n\t"
+      "paddw      %[dest0],        %[tmp0],           %[tmp1]       \n\t"
+      "psraw      %[dest0],        %[dest0],          %[shift0]     \n\t"
+
+      "gsldlc1    %[matrix],       0x0f(%[matrix_ptr])              \n\t"
+      "gsldrc1    %[matrix],       0x08(%[matrix_ptr])              \n\t"
+      "punpcklbh  %[matrix_lo],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "psrah      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "punpckhbh  %[matrix_hi],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "psrah      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "pmaddhw    %[dest_lo],      %[src_lo],         %[matrix_lo]  \n\t"
+      "pmaddhw    %[dest_hi],      %[src_lo],         %[matrix_hi]  \n\t"
+      "punpcklwd  %[tmp0],         %[dest_lo],        %[dest_hi]    \n\t"
+      "punpckhwd  %[tmp1],         %[dest_lo],        %[dest_hi]    \n\t"
+      "paddw      %[dest1],        %[tmp0],           %[tmp1]       \n\t"
+      "psraw      %[dest1],        %[dest1],          %[shift0]     \n\t"
+
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+
+      "gsldlc1    %[matrix],       0x07(%[matrix_ptr])              \n\t"
+      "gsldrc1    %[matrix],       0x00(%[matrix_ptr])              \n\t"
+      "punpcklbh  %[matrix_lo],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "psrah      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "punpckhbh  %[matrix_hi],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "psrah      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "pmaddhw    %[dest_lo],      %[src_hi],         %[matrix_lo]  \n\t"
+      "pmaddhw    %[dest_hi],      %[src_hi],         %[matrix_hi]  \n\t"
+      "punpcklwd  %[tmp0],         %[dest_lo],        %[dest_hi]    \n\t"
+      "punpckhwd  %[tmp1],         %[dest_lo],        %[dest_hi]    \n\t"
+      "paddw      %[dest2],        %[tmp0],           %[tmp1]       \n\t"
+      "psraw      %[dest2],        %[dest2],          %[shift0]     \n\t"
+
+      "gsldlc1    %[matrix],       0x0f(%[matrix_ptr])              \n\t"
+      "gsldrc1    %[matrix],       0x08(%[matrix_ptr])              \n\t"
+      "punpcklbh  %[matrix_lo],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "psrah      %[matrix_lo],    %[matrix_lo],      %[shift1]     \n\t"
+      "punpckhbh  %[matrix_hi],    %[matrix],         %[mask0]      \n\t"
+      "psllh      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "psrah      %[matrix_hi],    %[matrix_hi],      %[shift1]     \n\t"
+      "pmaddhw    %[dest_lo],      %[src_hi],         %[matrix_lo]  \n\t"
+      "pmaddhw    %[dest_hi],      %[src_hi],         %[matrix_hi]  \n\t"
+      "punpcklwd  %[tmp0],         %[dest_lo],        %[dest_hi]    \n\t"
+      "punpckhwd  %[tmp1],         %[dest_lo],        %[dest_hi]    \n\t"
+      "paddw      %[dest3],        %[tmp0],           %[tmp1]       \n\t"
+      "psraw      %[dest3],        %[dest3],          %[shift0]     \n\t"
+
+      "packsswh   %[tmp0],         %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[tmp1],         %[dest2],          %[dest3]      \n\t"
+      "packushb   %[dest],         %[tmp0],           %[tmp1]       \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3), [src] "=&f"(src), [dest] "=&f"(dest),
+        [tmp0] "=&f"(tmp0), [tmp1] "=&f"(tmp1), [matrix_hi] "=&f"(matrix_hi),
+        [matrix_lo] "=&f"(matrix_lo), [matrix] "=&f"(matrix)
+      : [src_ptr] "r"(src_argb), [matrix_ptr] "r"(matrix_argb),
+        [dst_ptr] "r"(dst_argb), [width] "r"(width), [shift0] "f"(shift0),
+        [shift1] "f"(shift1), [mask0] "f"(mask0), [mask1] "f"(mask1)
+      : "memory");
+}
+
+void ARGBShadeRow_MMI(const uint8_t* src_argb,
+                      uint8_t* dst_argb,
+                      int width,
+                      uint32_t value) {
+  uint64_t src, src_hi, src_lo, dest, dest_lo, dest_hi;
+  const uint64_t shift = 0x08;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[src]        \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[src]        \n\t"
+
+      "punpcklbh  %[value],        %[value],          %[value]      \n\t"
+
+      "pmulhuh    %[dest_lo],      %[src_lo],         %[value]      \n\t"
+      "psrlh      %[dest_lo],      %[dest_lo],        %[shift]      \n\t"
+      "pmulhuh    %[dest_hi],      %[src_hi],         %[value]      \n\t"
+      "psrlh      %[dest_hi],      %[dest_hi],        %[shift]      \n\t"
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo), [src] "=&f"(src),
+        [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb), [width] "r"(width),
+        [value] "f"(value), [shift] "f"(shift)
+      : "memory");
+}
+
+void ARGBMultiplyRow_MMI(const uint8_t* src_argb,
+                         const uint8_t* src_argb1,
+                         uint8_t* dst_argb,
+                         int width) {
+  uint64_t src0, src0_hi, src0_lo, src1, src1_hi, src1_lo;
+  uint64_t dest, dest_lo, dest_hi;
+  const uint64_t mask = 0x0;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src0],         0x07(%[src0_ptr])                \n\t"
+      "gsldrc1    %[src0],         0x00(%[src0_ptr])                \n\t"
+      "punpcklbh  %[src0_lo],      %[src0],           %[src0]       \n\t"
+      "punpckhbh  %[src0_hi],      %[src0],           %[src0]       \n\t"
+
+      "gsldlc1    %[src1],         0x07(%[src1_ptr])                \n\t"
+      "gsldrc1    %[src1],         0x00(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src1_lo],      %[src1],           %[mask]       \n\t"
+      "punpckhbh  %[src1_hi],      %[src1],           %[mask]       \n\t"
+
+      "pmulhuh    %[dest_lo],      %[src0_lo],        %[src1_lo]    \n\t"
+      "pmulhuh    %[dest_hi],      %[src0_hi],        %[src1_hi]    \n\t"
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src0_ptr],     %[src0_ptr],       0x08          \n\t"
+      "daddiu     %[src1_ptr],     %[src1_ptr],       0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0_hi] "=&f"(src0_hi), [src0_lo] "=&f"(src0_lo),
+        [src1_hi] "=&f"(src1_hi), [src1_lo] "=&f"(src1_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo), [src0] "=&f"(src0),
+        [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
+        [dst_ptr] "r"(dst_argb), [width] "r"(width), [mask] "f"(mask)
+      : "memory");
+}
+
+void ARGBAddRow_MMI(const uint8_t* src_argb,
+                    const uint8_t* src_argb1,
+                    uint8_t* dst_argb,
+                    int width) {
+  uint64_t src0, src1, dest;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src0],         0x07(%[src0_ptr])                \n\t"
+      "gsldrc1    %[src0],         0x00(%[src0_ptr])                \n\t"
+      "gsldlc1    %[src1],         0x07(%[src1_ptr])                \n\t"
+      "gsldrc1    %[src1],         0x00(%[src1_ptr])                \n\t"
+      "paddusb    %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src0_ptr],     %[src0_ptr],       0x08          \n\t"
+      "daddiu     %[src1_ptr],     %[src1_ptr],       0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
+        [dst_ptr] "r"(dst_argb), [width] "r"(width)
+      : "memory");
+}
+
+void ARGBSubtractRow_MMI(const uint8_t* src_argb,
+                         const uint8_t* src_argb1,
+                         uint8_t* dst_argb,
+                         int width) {
+  uint64_t src0, src1, dest;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src0],         0x07(%[src0_ptr])                \n\t"
+      "gsldrc1    %[src0],         0x00(%[src0_ptr])                \n\t"
+      "gsldlc1    %[src1],         0x07(%[src1_ptr])                \n\t"
+      "gsldrc1    %[src1],         0x00(%[src1_ptr])                \n\t"
+      "psubusb    %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src0_ptr],     %[src0_ptr],       0x08          \n\t"
+      "daddiu     %[src1_ptr],     %[src1_ptr],       0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
+        [dst_ptr] "r"(dst_argb), [width] "r"(width)
+      : "memory");
+}
+
+// Sobel functions which mimics SSSE3.
+void SobelXRow_MMI(const uint8_t* src_y0,
+                   const uint8_t* src_y1,
+                   const uint8_t* src_y2,
+                   uint8_t* dst_sobelx,
+                   int width) {
+  uint64_t y00 = 0, y10 = 0, y20 = 0;
+  uint64_t y02 = 0, y12 = 0, y22 = 0;
+  uint64_t zero = 0x0;
+  uint64_t sobel = 0x0;
+  __asm__ volatile(
+      "1:	                                         \n\t"
+      "gsldlc1   %[y00],        0x07(%[src_y0])          \n\t"  // a=src_y0[i]
+      "gsldrc1   %[y00],        0x00(%[src_y0])          \n\t"
+      "gsldlc1   %[y02],        0x09(%[src_y0])          \n\t"  // a_sub=src_y0[i+2]
+      "gsldrc1   %[y02],        0x02(%[src_y0])          \n\t"
+
+      "gsldlc1   %[y10],        0x07(%[src_y1])          \n\t"  // b=src_y1[i]
+      "gsldrc1   %[y10],        0x00(%[src_y1])          \n\t"
+      "gsldlc1   %[y12],        0x09(%[src_y1])          \n\t"  // b_sub=src_y1[i+2]
+      "gsldrc1   %[y12],        0x02(%[src_y1])          \n\t"
+
+      "gsldlc1   %[y20],        0x07(%[src_y2])          \n\t"  // c=src_y2[i]
+      "gsldrc1   %[y20],        0x00(%[src_y2])          \n\t"
+      "gsldlc1   %[y22],        0x09(%[src_y2])          \n\t"  // c_sub=src_y2[i+2]
+      "gsldrc1   %[y22],        0x02(%[src_y2])          \n\t"
+
+      "punpcklbh %[y00],        %[y00],          %[zero] \n\t"
+      "punpcklbh %[y10],        %[y10],          %[zero] \n\t"
+      "punpcklbh %[y20],        %[y20],          %[zero] \n\t"
+
+      "punpcklbh %[y02],        %[y02],          %[zero] \n\t"
+      "punpcklbh %[y12],        %[y12],          %[zero] \n\t"
+      "punpcklbh %[y22],        %[y22],          %[zero] \n\t"
+
+      "paddh     %[y00],        %[y00],          %[y10]  \n\t"  // a+b
+      "paddh     %[y20],        %[y20],          %[y10]  \n\t"  // c+b
+      "paddh     %[y00],        %[y00],          %[y20]  \n\t"  // a+2b+c
+
+      "paddh     %[y02],        %[y02],          %[y12]  \n\t"  // a_sub+b_sub
+      "paddh     %[y22],        %[y22],          %[y12]  \n\t"  // c_sub+b_sub
+      "paddh     %[y02],        %[y02],          %[y22]  \n\t"  // a_sub+2b_sub+c_sub
+
+      "pmaxsh    %[y10],        %[y00],          %[y02]  \n\t"
+      "pminsh    %[y20],        %[y00],          %[y02]  \n\t"
+      "psubh     %[sobel],      %[y10],          %[y20]  \n\t"  // Abs
+
+      "gsldlc1   %[y00],        0x0B(%[src_y0])          \n\t"
+      "gsldrc1   %[y00],        0x04(%[src_y0])          \n\t"
+      "gsldlc1   %[y02],        0x0D(%[src_y0])          \n\t"
+      "gsldrc1   %[y02],        0x06(%[src_y0])          \n\t"
+
+      "gsldlc1   %[y10],        0x0B(%[src_y1])          \n\t"
+      "gsldrc1   %[y10],        0x04(%[src_y1])          \n\t"
+      "gsldlc1   %[y12],        0x0D(%[src_y1])          \n\t"
+      "gsldrc1   %[y12],        0x06(%[src_y1])          \n\t"
+
+      "gsldlc1   %[y20],        0x0B(%[src_y2])          \n\t"
+      "gsldrc1   %[y20],        0x04(%[src_y2])          \n\t"
+      "gsldlc1   %[y22],        0x0D(%[src_y2])          \n\t"
+      "gsldrc1   %[y22],        0x06(%[src_y2])          \n\t"
+
+      "punpcklbh %[y00],        %[y00],          %[zero] \n\t"
+      "punpcklbh %[y10],        %[y10],          %[zero] \n\t"
+      "punpcklbh %[y20],        %[y20],          %[zero] \n\t"
+
+      "punpcklbh %[y02],        %[y02],          %[zero] \n\t"
+      "punpcklbh %[y12],        %[y12],          %[zero] \n\t"
+      "punpcklbh %[y22],        %[y22],          %[zero] \n\t"
+
+      "paddh     %[y00],        %[y00],          %[y10]  \n\t"
+      "paddh     %[y20],        %[y20],          %[y10]  \n\t"
+      "paddh     %[y00],        %[y00],          %[y20]  \n\t"
+
+      "paddh     %[y02],        %[y02],          %[y12]  \n\t"
+      "paddh     %[y22],        %[y22],          %[y12]  \n\t"
+      "paddh     %[y02],        %[y02],          %[y22]  \n\t"
+
+      "pmaxsh    %[y10],        %[y00],          %[y02]  \n\t"
+      "pminsh    %[y20],        %[y00],          %[y02]  \n\t"
+      "psubh     %[y00],        %[y10],          %[y20]  \n\t"
+
+      "packushb  %[sobel],      %[sobel],        %[y00]  \n\t"  // clamp255
+      "gssdrc1   %[sobel],      0(%[dst_sobelx])         \n\t"
+      "gssdlc1   %[sobel],      7(%[dst_sobelx])         \n\t"
+
+      "daddiu    %[src_y0],     %[src_y0],      8        \n\t"
+      "daddiu    %[src_y1],     %[src_y1],      8        \n\t"
+      "daddiu    %[src_y2],     %[src_y2],      8        \n\t"
+      "daddiu    %[dst_sobelx], %[dst_sobelx],  8        \n\t"
+      "daddiu    %[width],      %[width],      -8        \n\t"
+      "bgtz      %[width],      1b                       \n\t"
+      "nop                                               \n\t"
+      : [sobel] "=&f"(sobel), [y00] "=&f"(y00), [y10] "=&f"(y10),
+        [y20] "=&f"(y20), [y02] "=&f"(y02), [y12] "=&f"(y12), [y22] "=&f"(y22)
+      : [src_y0] "r"(src_y0), [src_y1] "r"(src_y1), [src_y2] "r"(src_y2),
+        [dst_sobelx] "r"(dst_sobelx), [width] "r"(width), [zero] "f"(zero)
+      : "memory");
+}
+
+void SobelYRow_MMI(const uint8_t* src_y0,
+                   const uint8_t* src_y1,
+                   uint8_t* dst_sobely,
+                   int width) {
+  uint64_t y00 = 0, y01 = 0, y02 = 0;
+  uint64_t y10 = 0, y11 = 0, y12 = 0;
+  uint64_t zero = 0x0;
+  uint64_t sobel = 0x0;
+  __asm__ volatile(
+      "1:	                                        \n\t"
+      "gsldlc1   %[y00],        0x07(%[src_y0])         \n\t"  // a=src_y0[i]
+      "gsldrc1   %[y00],        0x00(%[src_y0])         \n\t"
+      "gsldlc1   %[y01],        0x08(%[src_y0])         \n\t"  // b=src_y0[i+1]
+      "gsldrc1   %[y01],        0x01(%[src_y0])         \n\t"
+      "gsldlc1   %[y02],        0x09(%[src_y0])         \n\t"  // c=src_y0[i+2]
+      "gsldrc1   %[y02],        0x02(%[src_y0])         \n\t"
+
+      "gsldlc1   %[y10],        0x07(%[src_y1])         \n\t"  // a_sub=src_y1[i]
+      "gsldrc1   %[y10],        0x00(%[src_y1])         \n\t"
+      "gsldlc1   %[y11],        0x08(%[src_y1])         \n\t"  // b_sub=src_y1[i+1]
+      "gsldrc1   %[y11],        0x01(%[src_y1])         \n\t"
+      "gsldlc1   %[y12],        0x09(%[src_y1])         \n\t"  // c_sub=src_y1[i+2]
+      "gsldrc1   %[y12],        0x02(%[src_y1])         \n\t"
+
+      "punpcklbh %[y00],        %[y00],         %[zero] \n\t"
+      "punpcklbh %[y01],        %[y01],         %[zero] \n\t"
+      "punpcklbh %[y02],        %[y02],         %[zero] \n\t"
+
+      "punpcklbh %[y10],        %[y10],         %[zero] \n\t"
+      "punpcklbh %[y11],        %[y11],         %[zero] \n\t"
+      "punpcklbh %[y12],        %[y12],         %[zero] \n\t"
+
+      "paddh     %[y00],        %[y00],         %[y01]  \n\t"  // a+b
+      "paddh     %[y02],        %[y02],         %[y01]  \n\t"  // c+b
+      "paddh     %[y00],        %[y00],         %[y02]  \n\t"  // a+2b+c
+
+      "paddh     %[y10],        %[y10],         %[y11]  \n\t"  // a_sub+b_sub
+      "paddh     %[y12],        %[y12],         %[y11]  \n\t"  // c_sub+b_sub
+      "paddh     %[y10],        %[y10],         %[y12]  \n\t"  // a_sub+2b_sub+c_sub
+
+      "pmaxsh    %[y02],        %[y00],         %[y10]  \n\t"
+      "pminsh    %[y12],        %[y00],         %[y10]  \n\t"
+      "psubh     %[sobel],      %[y02],         %[y12]  \n\t"  // Abs
+
+      "gsldlc1   %[y00],        0x0B(%[src_y0])         \n\t"
+      "gsldrc1   %[y00],        0x04(%[src_y0])         \n\t"
+      "gsldlc1   %[y01],        0x0C(%[src_y0])         \n\t"
+      "gsldrc1   %[y01],        0x05(%[src_y0])         \n\t"
+      "gsldlc1   %[y02],        0x0D(%[src_y0])         \n\t"
+      "gsldrc1   %[y02],        0x06(%[src_y0])         \n\t"
+
+      "gsldlc1   %[y10],        0x0B(%[src_y1])         \n\t"
+      "gsldrc1   %[y10],        0x04(%[src_y1])         \n\t"
+      "gsldlc1   %[y11],        0x0C(%[src_y1])         \n\t"
+      "gsldrc1   %[y11],        0x05(%[src_y1])         \n\t"
+      "gsldlc1   %[y12],        0x0D(%[src_y1])         \n\t"
+      "gsldrc1   %[y12],        0x06(%[src_y1])         \n\t"
+
+      "punpcklbh %[y00],        %[y00],         %[zero] \n\t"
+      "punpcklbh %[y01],        %[y01],         %[zero] \n\t"
+      "punpcklbh %[y02],        %[y02],         %[zero] \n\t"
+
+      "punpcklbh %[y10],        %[y10],         %[zero] \n\t"
+      "punpcklbh %[y11],        %[y11],         %[zero] \n\t"
+      "punpcklbh %[y12],        %[y12],         %[zero] \n\t"
+
+      "paddh     %[y00],        %[y00],         %[y01]  \n\t"
+      "paddh     %[y02],        %[y02],         %[y01]  \n\t"
+      "paddh     %[y00],        %[y00],         %[y02]  \n\t"
+
+      "paddh     %[y10],        %[y10],         %[y11]  \n\t"
+      "paddh     %[y12],        %[y12],         %[y11]  \n\t"
+      "paddh     %[y10],        %[y10],         %[y12]  \n\t"
+
+      "pmaxsh    %[y02],        %[y00],         %[y10]  \n\t"
+      "pminsh    %[y12],        %[y00],         %[y10]  \n\t"
+      "psubh     %[y00],        %[y02],         %[y12]  \n\t"
+
+      "packushb  %[sobel],      %[sobel],       %[y00]  \n\t"  // clamp255
+      "gssdrc1   %[sobel],      0(%[dst_sobely])        \n\t"
+      "gssdlc1   %[sobel],      7(%[dst_sobely])        \n\t"
+
+      "daddiu    %[src_y0],     %[src_y0],      8       \n\t"
+      "daddiu    %[src_y1],     %[src_y1],      8       \n\t"
+      "daddiu    %[dst_sobely], %[dst_sobely],  8       \n\t"
+      "daddiu    %[width],      %[width],      -8       \n\t"
+      "bgtz      %[width],      1b                      \n\t"
+      "nop                                              \n\t"
+      : [sobel] "=&f"(sobel), [y00] "=&f"(y00), [y01] "=&f"(y01),
+        [y02] "=&f"(y02), [y10] "=&f"(y10), [y11] "=&f"(y11), [y12] "=&f"(y12)
+      : [src_y0] "r"(src_y0), [src_y1] "r"(src_y1),
+        [dst_sobely] "r"(dst_sobely), [width] "r"(width), [zero] "f"(zero)
+      : "memory");
+}
+
+void SobelRow_MMI(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width) {
+  double temp[3];
+  uint64_t c1 = 0xff000000ff000000;
+  __asm__ volatile(
+      "1:	                                          \n\t"
+      "gsldlc1   %[t0],         0x07(%[src_sobelx])       \n\t"  // a=src_sobelx[i]
+      "gsldrc1   %[t0],         0x00(%[src_sobelx])       \n\t"
+      "gsldlc1   %[t1],         0x07(%[src_sobely])       \n\t"  // b=src_sobely[i]
+      "gsldrc1   %[t1],         0x00(%[src_sobely])       \n\t"
+      // s7 s6 s5 s4 s3 s2 s1 s0 = a+b
+      "paddusb   %[t2] ,        %[t0],              %[t1] \n\t"
+
+      // s3 s2 s1 s0->s3 s3 s2 s2 s1 s1 s0 s0
+      "punpcklbh %[t0],         %[t2],              %[t2] \n\t"
+
+      // s1 s1 s0 s0->s1 s2 s1 s1 s0 s0 s0 s0
+      "punpcklbh %[t1],         %[t0],              %[t0] \n\t"
+      "or        %[t1],         %[t1],              %[c1] \n\t"
+      // 255 s1 s1 s1 s55 s0 s0 s0
+      "gssdrc1   %[t1],         0x00(%[dst_argb])	  \n\t"
+      "gssdlc1   %[t1],         0x07(%[dst_argb])         \n\t"
+
+      // s3 s3 s2 s2->s3 s3 s3 s3 s2 s2 s2 s2
+      "punpckhbh %[t1],         %[t0],              %[t0] \n\t"
+      "or        %[t1],         %[t1],              %[c1] \n\t"
+      // 255 s3 s3 s3 255 s2 s2 s2
+      "gssdrc1   %[t1],         0x08(%[dst_argb])	  \n\t"
+      "gssdlc1   %[t1],         0x0f(%[dst_argb])         \n\t"
+
+      // s7 s6 s5 s4->s7 s7 s6 s6 s5 s5 s4 s4
+      "punpckhbh %[t0],         %[t2],              %[t2] \n\t"
+
+      // s5 s5 s4 s4->s5 s5 s5 s5 s4 s4 s4 s4
+      "punpcklbh %[t1],         %[t0],              %[t0] \n\t"
+      "or        %[t1],         %[t1],              %[c1] \n\t"
+      "gssdrc1   %[t1],         0x10(%[dst_argb])	  \n\t"
+      "gssdlc1   %[t1],         0x17(%[dst_argb])         \n\t"
+
+      // s7 s7 s6 s6->s7 s7 s7 s7 s6 s6 s6 s6
+      "punpckhbh %[t1],         %[t0],              %[t0] \n\t"
+      "or        %[t1],         %[t1],              %[c1] \n\t"
+      "gssdrc1   %[t1],         0x18(%[dst_argb])	  \n\t"
+      "gssdlc1   %[t1],         0x1f(%[dst_argb])         \n\t"
+
+      "daddiu    %[dst_argb],   %[dst_argb],        32    \n\t"
+      "daddiu    %[src_sobelx], %[src_sobelx],      8     \n\t"
+      "daddiu    %[src_sobely], %[src_sobely],      8     \n\t"
+      "daddiu    %[width],      %[width],          -8     \n\t"
+      "bgtz      %[width],      1b                        \n\t"
+      "nop                                                \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [t2] "=&f"(temp[2])
+      : [src_sobelx] "r"(src_sobelx), [src_sobely] "r"(src_sobely),
+        [dst_argb] "r"(dst_argb), [width] "r"(width), [c1] "f"(c1)
+      : "memory");
+}
+
+void SobelToPlaneRow_MMI(const uint8_t* src_sobelx,
+                         const uint8_t* src_sobely,
+                         uint8_t* dst_y,
+                         int width) {
+  uint64_t tr = 0;
+  uint64_t tb = 0;
+  __asm__ volatile(
+      "1:	                                       \n\t"
+      "gsldrc1 %[tr],         0x0(%[src_sobelx])       \n\t"
+      "gsldlc1 %[tr],         0x7(%[src_sobelx])       \n\t"  // r=src_sobelx[i]
+      "gsldrc1 %[tb],         0x0(%[src_sobely])       \n\t"
+      "gsldlc1 %[tb],         0x7(%[src_sobely])       \n\t"  // b=src_sobely[i]
+      "paddusb %[tr],         %[tr],             %[tb] \n\t"  // g
+      "gssdrc1 %[tr],         0x0(%[dst_y])	       \n\t"
+      "gssdlc1 %[tr],         0x7(%[dst_y])            \n\t"
+
+      "daddiu  %[dst_y],      %[dst_y],          8     \n\t"
+      "daddiu  %[src_sobelx], %[src_sobelx],     8     \n\t"
+      "daddiu  %[src_sobely], %[src_sobely],     8     \n\t"
+      "daddiu  %[width],      %[width],         -8     \n\t"
+      "bgtz    %[width],      1b                       \n\t"
+      "nop                                             \n\t"
+      : [tr] "=&f"(tr), [tb] "=&f"(tb)
+      : [src_sobelx] "r"(src_sobelx), [src_sobely] "r"(src_sobely),
+        [dst_y] "r"(dst_y), [width] "r"(width)
+      : "memory");
+}
+
+void SobelXYRow_MMI(const uint8_t* src_sobelx,
+                    const uint8_t* src_sobely,
+                    uint8_t* dst_argb,
+                    int width) {
+  uint64_t temp[3];
+  uint64_t result = 0;
+  uint64_t gb = 0;
+  uint64_t cr = 0;
+  uint64_t c1 = 0xffffffffffffffff;
+  __asm__ volatile(
+      "1:	                                          \n\t"
+      "gsldlc1   %[tr],         0x07(%[src_sobelx])       \n\t"  // r=src_sobelx[i]
+      "gsldrc1   %[tr],         0x00(%[src_sobelx])       \n\t"
+      "gsldlc1   %[tb],         0x07(%[src_sobely])       \n\t"  // b=src_sobely[i]
+      "gsldrc1   %[tb],         0x00(%[src_sobely])       \n\t"
+      "paddusb   %[tg] ,        %[tr],              %[tb] \n\t"  // g
+
+      // g3 b3 g2 b2 g1 b1 g0 b0
+      "punpcklbh %[gb],         %[tb],              %[tg] \n\t"
+      // c3 r3 r2 r2 c1 r1 c0 r0
+      "punpcklbh %[cr],         %[tr],              %[c1] \n\t"
+      // c1 r1 g1 b1 c0 r0 g0 b0
+      "punpcklhw %[result],     %[gb],              %[cr] \n\t"
+      "gssdrc1   %[result],     0x00(%[dst_argb])	  \n\t"
+      "gssdlc1   %[result],     0x07(%[dst_argb])         \n\t"
+      // c3 r3 g3 b3 c2 r2 g2 b2
+      "punpckhhw %[result],     %[gb],              %[cr] \n\t"
+      "gssdrc1   %[result],     0x08(%[dst_argb])	  \n\t"
+      "gssdlc1   %[result],     0x0f(%[dst_argb])         \n\t"
+
+      // g7 b7 g6 b6 g5 b5 g4 b4
+      "punpckhbh %[gb],         %[tb],              %[tg] \n\t"
+      // c7 r7 c6 r6 c5 r5 c4 r4
+      "punpckhbh %[cr],         %[tr],              %[c1] \n\t"
+      // c5 r5 g5 b5 c4 r4 g4 b4
+      "punpcklhw %[result],     %[gb],              %[cr] \n\t"
+      "gssdrc1   %[result],     0x10(%[dst_argb])	  \n\t"
+      "gssdlc1   %[result],     0x17(%[dst_argb])         \n\t"
+      // c7 r7 g7 b7 c6 r6 g6 b6
+      "punpckhhw %[result],     %[gb],              %[cr] \n\t"
+      "gssdrc1   %[result],     0x18(%[dst_argb])	  \n\t"
+      "gssdlc1   %[result],     0x1f(%[dst_argb])         \n\t"
+
+      "daddiu    %[dst_argb],   %[dst_argb],        32    \n\t"
+      "daddiu    %[src_sobelx], %[src_sobelx],      8     \n\t"
+      "daddiu    %[src_sobely], %[src_sobely],      8     \n\t"
+      "daddiu    %[width],      %[width],          -8     \n\t"
+      "bgtz      %[width],      1b                        \n\t"
+      "nop                                                \n\t"
+      : [tr] "=&f"(temp[0]), [tb] "=&f"(temp[1]), [tg] "=&f"(temp[2]),
+        [gb] "=&f"(gb), [cr] "=&f"(cr), [result] "=&f"(result)
+      : [src_sobelx] "r"(src_sobelx), [src_sobely] "r"(src_sobely),
+        [dst_argb] "r"(dst_argb), [width] "r"(width), [c1] "f"(c1)
+      : "memory");
+}
+
+void J400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  // Copy a Y to RGB.
+  uint64_t src, dest;
+  const uint64_t mask0 = 0x00ffffff00ffffffULL;
+  const uint64_t mask1 = ~mask0;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gslwlc1    %[src],          0x03(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src],          %[src],            %[src]        \n\t"
+      "punpcklhw  %[dest],         %[src],            %[src]        \n\t"
+      "and        %[dest],         %[dest],           %[mask0]      \n\t"
+      "or         %[dest],         %[dest],           %[mask1]      \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+
+      "punpckhhw  %[dest],         %[src],            %[src]        \n\t"
+      "and        %[dest],         %[dest],           %[mask0]      \n\t"
+      "or         %[dest],         %[dest],           %[mask1]      \n\t"
+      "gssdrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+      "gssdlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x04          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_y), [dst_ptr] "r"(dst_argb), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [width] "r"(width)
+      : "memory");
+}
+
+// TODO - respect YuvConstants
+void I400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* rgb_buf,
+                       const struct YuvConstants*, int width) {
+  uint64_t src, src_lo, src_hi, dest, dest_lo, dest_hi;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x55;
+  const uint64_t mask2 = 0xAA;
+  const uint64_t mask3 = 0xFF;
+  const uint64_t mask4 = 0x4A354A354A354A35ULL;
+  const uint64_t mask5 = 0x0488048804880488ULL;
+  const uint64_t shift0 = 0x08;
+  const uint64_t shift1 = 0x06;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t"
+
+      "pshufh     %[src],          %[src_lo],         %[mask0]      \n\t"
+      "psllh      %[dest_lo],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[src]        \n\t"
+      "pmulhuh    %[dest_lo],      %[dest_lo],        %[mask4]      \n\t"
+      "psubh      %[dest_lo],      %[dest_lo],        %[mask5]      \n\t"
+      "psrah      %[dest_lo],      %[dest_lo],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_lo],      %[dest_lo],        %[mask3]      \n\t"
+      "pshufh     %[src],          %[src_lo],         %[mask1]      \n\t"
+      "psllh      %[dest_hi],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[src]        \n\t"
+      "pmulhuh    %[dest_hi],      %[dest_hi],        %[mask4]      \n\t"
+      "psubh      %[dest_hi],      %[dest_hi],        %[mask5]      \n\t"
+      "psrah      %[dest_hi],      %[dest_hi],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_hi],      %[dest_hi],        %[mask3]      \n\t"
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "pshufh     %[src],          %[src_lo],         %[mask2]      \n\t"
+      "psllh      %[dest_lo],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[src]        \n\t"
+      "pmulhuh    %[dest_lo],      %[dest_lo],        %[mask4]      \n\t"
+      "psubh      %[dest_lo],      %[dest_lo],        %[mask5]      \n\t"
+      "psrah      %[dest_lo],      %[dest_lo],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_lo],      %[dest_lo],        %[mask3]      \n\t"
+      "pshufh     %[src],          %[src_lo],         %[mask3]      \n\t"
+      "psllh      %[dest_hi],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[src]        \n\t"
+      "pmulhuh    %[dest_hi],      %[dest_hi],        %[mask4]      \n\t"
+      "psubh      %[dest_hi],      %[dest_hi],        %[mask5]      \n\t"
+      "psrah      %[dest_hi],      %[dest_hi],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_hi],      %[dest_hi],        %[mask3]      \n\t"
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+
+      "pshufh     %[src],          %[src_hi],         %[mask0]      \n\t"
+      "psllh      %[dest_lo],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[src]        \n\t"
+      "pmulhuh    %[dest_lo],      %[dest_lo],        %[mask4]      \n\t"
+      "psubh      %[dest_lo],      %[dest_lo],        %[mask5]      \n\t"
+      "psrah      %[dest_lo],      %[dest_lo],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_lo],      %[dest_lo],        %[mask3]      \n\t"
+      "pshufh     %[src],          %[src_hi],         %[mask1]      \n\t"
+      "psllh      %[dest_hi],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[src]        \n\t"
+      "pmulhuh    %[dest_hi],      %[dest_hi],        %[mask4]      \n\t"
+      "psubh      %[dest_hi],      %[dest_hi],        %[mask5]      \n\t"
+      "psrah      %[dest_hi],      %[dest_hi],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_hi],      %[dest_hi],        %[mask3]      \n\t"
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x17(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x10(%[dst_ptr])                 \n\t"
+
+      "pshufh     %[src],          %[src_hi],         %[mask2]      \n\t"
+      "psllh      %[dest_lo],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[src]        \n\t"
+      "pmulhuh    %[dest_lo],      %[dest_lo],        %[mask4]      \n\t"
+      "psubh      %[dest_lo],      %[dest_lo],        %[mask5]      \n\t"
+      "psrah      %[dest_lo],      %[dest_lo],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_lo],      %[dest_lo],        %[mask3]      \n\t"
+      "pshufh     %[src],          %[src_hi],         %[mask3]      \n\t"
+      "psllh      %[dest_hi],      %[src],            %[shift0]     \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[src]        \n\t"
+      "pmulhuh    %[dest_hi],      %[dest_hi],        %[mask4]      \n\t"
+      "psubh      %[dest_hi],      %[dest_hi],        %[mask5]      \n\t"
+      "psrah      %[dest_hi],      %[dest_hi],        %[shift1]     \n\t"
+      "pinsrh_3   %[dest_hi],      %[dest_hi],        %[mask3]      \n\t"
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x1f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x18(%[dst_ptr])                 \n\t"
+
+      "daddi      %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x20          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest), [src_hi] "=&f"(src_hi),
+        [src_lo] "=&f"(src_lo), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo)
+      : [src_ptr] "r"(src_y), [dst_ptr] "r"(rgb_buf), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [mask2] "f"(mask2), [mask3] "f"(mask3),
+        [mask4] "f"(mask4), [mask5] "f"(mask5), [shift0] "f"(shift0),
+        [shift1] "f"(shift1), [width] "r"(width)
+      : "memory");
+}
+
+void MirrorRow_MMI(const uint8_t* src, uint8_t* dst, int width) {
+  uint64_t source, src0, src1, dest;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x1b;
+
+  src += width - 1;
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[source],       0(%[src_ptr])                    \n\t"
+      "gsldrc1    %[source],       -7(%[src_ptr])                   \n\t"
+      "punpcklbh  %[src0],         %[source],         %[mask0]      \n\t"
+      "pshufh     %[src0],         %[src0],           %[mask1]      \n\t"
+      "punpckhbh  %[src1],         %[source],         %[mask0]      \n\t"
+      "pshufh     %[src1],         %[src1],           %[mask1]      \n\t"
+      "packushb   %[dest],         %[src1],           %[src0]       \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddi      %[src_ptr],      %[src_ptr],       -0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [source] "=&f"(source), [dest] "=&f"(dest), [src0] "=&f"(src0),
+        [src1] "=&f"(src1)
+      : [src_ptr] "r"(src), [dst_ptr] "r"(dst), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [width] "r"(width)
+      : "memory");
+}
+
+void MirrorSplitUVRow_MMI(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  uint64_t src0, src1, dest0, dest1;
+  const uint64_t mask0 = 0x00ff00ff00ff00ffULL;
+  const uint64_t mask1 = 0x1b;
+  const uint64_t shift = 0x08;
+
+  src_uv += (width - 1) << 1;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src0],         1(%[src_ptr])                    \n\t"
+      "gsldrc1    %[src0],         -6(%[src_ptr])                   \n\t"
+      "gsldlc1    %[src1],         -7(%[src_ptr])                   \n\t"
+      "gsldrc1    %[src1],         -14(%[src_ptr])                  \n\t"
+
+      "and        %[dest0],        %[src0],           %[mask0]      \n\t"
+      "pshufh     %[dest0],        %[dest0],          %[mask1]      \n\t"
+      "and        %[dest1],        %[src1],           %[mask0]      \n\t"
+      "pshufh     %[dest1],        %[dest1],          %[mask1]      \n\t"
+      "packushb   %[dest0],        %[dest0],          %[dest1]      \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dstu_ptr])                \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dstu_ptr])                \n\t"
+
+      "psrlh      %[dest0],        %[src0],           %[shift]      \n\t"
+      "pshufh     %[dest0],        %[dest0],          %[mask1]      \n\t"
+      "psrlh      %[dest1],        %[src1],           %[shift]      \n\t"
+      "pshufh     %[dest1],        %[dest1],          %[mask1]      \n\t"
+      "packushb   %[dest0],        %[dest0],          %[dest1]      \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dstv_ptr])                \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dstv_ptr])                \n\t"
+
+      "daddi      %[src_ptr],      %[src_ptr],       -0x10          \n\t"
+      "daddiu     %[dstu_ptr],     %[dstu_ptr],       0x08          \n\t"
+      "daddiu     %[dstv_ptr],     %[dstv_ptr],       0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [src0] "=&f"(src0),
+        [src1] "=&f"(src1)
+      : [src_ptr] "r"(src_uv), [dstu_ptr] "r"(dst_u), [dstv_ptr] "r"(dst_v),
+        [width] "r"(width), [mask0] "f"(mask0), [mask1] "f"(mask1),
+        [shift] "f"(shift)
+      : "memory");
+}
+
+void ARGBMirrorRow_MMI(const uint8_t* src, uint8_t* dst, int width) {
+  src += (width - 1) * 4;
+  uint64_t temp = 0x0;
+  uint64_t shuff = 0x4e;  // 01 00 11 10
+  __asm__ volatile(
+      "1:                                      \n\t"
+      "gsldlc1 %[temp],  3(%[src])     	       \n\t"
+      "gsldrc1 %[temp], -4(%[src])     	       \n\t"
+      "pshufh  %[temp],  %[temp],    %[shuff]  \n\t"
+      "gssdrc1 %[temp],  0x0(%[dst])           \n\t"
+      "gssdlc1 %[temp],  0x7(%[dst])           \n\t"
+
+      "daddiu  %[src],   %[src],    -0x08      \n\t"
+      "daddiu  %[dst],   %[dst],     0x08      \n\t"
+      "daddiu  %[width], %[width],  -0x02      \n\t"
+      "bnez    %[width], 1b                    \n\t"
+      : [temp] "=&f"(temp)
+      : [src] "r"(src), [dst] "r"(dst), [width] "r"(width), [shuff] "f"(shuff)
+      : "memory");
+}
+
+void SplitUVRow_MMI(const uint8_t* src_uv,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  uint64_t c0 = 0x00ff00ff00ff00ff;
+  uint64_t temp[4];
+  uint64_t shift = 0x08;
+  __asm__ volatile(
+      "1:	                                    \n\t"
+      "gsldrc1  %[t0],     0x00(%[src_uv])          \n\t"
+      "gsldlc1  %[t0],     0x07(%[src_uv])          \n\t"
+      "gsldrc1  %[t1],     0x08(%[src_uv])          \n\t"
+      "gsldlc1  %[t1],     0x0f(%[src_uv])          \n\t"
+
+      "and      %[t2],     %[t0],          %[c0]    \n\t"
+      "and      %[t3],     %[t1],          %[c0]    \n\t"
+      "packushb %[t2],     %[t2],          %[t3]    \n\t"
+      "gssdrc1  %[t2],     0x0(%[dst_u])	    \n\t"
+      "gssdlc1  %[t2],     0x7(%[dst_u])            \n\t"
+
+      "psrlh    %[t2],     %[t0],          %[shift] \n\t"
+      "psrlh    %[t3],     %[t1],          %[shift] \n\t"
+      "packushb %[t2],     %[t2],          %[t3]    \n\t"
+      "gssdrc1  %[t2],     0x0(%[dst_v])            \n\t"
+      "gssdlc1  %[t2],     0x7(%[dst_v])            \n\t"
+
+      "daddiu   %[src_uv], %[src_uv],      16       \n\t"
+      "daddiu   %[dst_u],  %[dst_u],       8        \n\t"
+      "daddiu   %[dst_v],  %[dst_v],       8        \n\t"
+      "daddiu   %[width],  %[width],      -8        \n\t"
+      "bgtz     %[width],  1b                       \n\t"
+      "nop                                          \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [t2] "=&f"(temp[2]),
+        [t3] "=&f"(temp[3])
+      : [src_uv] "r"(src_uv), [dst_u] "r"(dst_u), [dst_v] "r"(dst_v),
+        [width] "r"(width), [c0] "f"(c0), [shift] "f"(shift)
+      : "memory");
+}
+
+void MergeUVRow_MMI(const uint8_t* src_u,
+                    const uint8_t* src_v,
+                    uint8_t* dst_uv,
+                    int width) {
+  uint64_t temp[3];
+  __asm__ volatile(
+      "1:	                                 \n\t"
+      "gsldrc1   %[t0],     0x0(%[src_u])        \n\t"
+      "gsldlc1   %[t0],     0x7(%[src_u])        \n\t"
+      "gsldrc1   %[t1],     0x0(%[src_v])        \n\t"
+      "gsldlc1   %[t1],     0x7(%[src_v])        \n\t"
+      "punpcklbh %[t2],     %[t0],         %[t1] \n\t"
+      "gssdrc1   %[t2],     0x0(%[dst_uv])	 \n\t"
+      "gssdlc1   %[t2],     0x7(%[dst_uv])       \n\t"
+      "punpckhbh %[t2],     %[t0],         %[t1] \n\t"
+      "gssdrc1   %[t2],     0x8(%[dst_uv])	 \n\t"
+      "gssdlc1   %[t2],     0xf(%[dst_uv])       \n\t"
+
+      "daddiu    %[src_u],  %[src_u],      8     \n\t"
+      "daddiu    %[src_v],  %[src_v],      8     \n\t"
+      "daddiu    %[dst_uv], %[dst_uv],     16    \n\t"
+      "daddiu    %[width],  %[width],     -8     \n\t"
+      "bgtz      %[width],  1b                   \n\t"
+      "nop                                       \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [t2] "=&f"(temp[2])
+      : [dst_uv] "r"(dst_uv), [src_u] "r"(src_u), [src_v] "r"(src_v),
+        [width] "r"(width)
+      : "memory");
+}
+
+void SplitRGBRow_MMI(const uint8_t* src_rgb,
+                     uint8_t* dst_r,
+                     uint8_t* dst_g,
+                     uint8_t* dst_b,
+                     int width) {
+  uint64_t src[4];
+  uint64_t dest_hi, dest_lo, dest;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gslwlc1    %[src0],         0x03(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gslwlc1    %[src1],         0x06(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src1],         0x03(%[src_ptr])                 \n\t"
+      "punpcklbh  %[dest_lo],      %[src0],           %[src1]       \n\t"
+      "gslwlc1    %[src2],         0x09(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src2],         0x06(%[src_ptr])                 \n\t"
+      "gslwlc1    %[src3],         0x0c(%[src_ptr])                 \n\t"
+      "gslwrc1    %[src3],         0x09(%[src_ptr])                 \n\t"
+      "punpcklbh  %[dest_hi],      %[src2],           %[src3]       \n\t"
+
+      "punpcklhw  %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gsswlc1    %[dest],         0x03(%[dstr_ptr])                \n\t"
+      "gsswrc1    %[dest],         0x00(%[dstr_ptr])                \n\t"
+      "punpckhwd  %[dest],         %[dest],           %[dest]       \n\t"
+      "gsswlc1    %[dest],         0x03(%[dstg_ptr])                \n\t"
+      "gsswrc1    %[dest],         0x00(%[dstg_ptr])                \n\t"
+      "punpckhhw  %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gsswlc1    %[dest],         0x03(%[dstb_ptr])                \n\t"
+      "gsswrc1    %[dest],         0x00(%[dstb_ptr])                \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x0c          \n\t"
+      "daddiu     %[dstr_ptr],     %[dstr_ptr],       0x04          \n\t"
+      "daddiu     %[dstg_ptr],     %[dstg_ptr],       0x04          \n\t"
+      "daddiu     %[dstb_ptr],     %[dstb_ptr],       0x04          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src[0]), [src1] "=&f"(src[1]), [src2] "=&f"(src[2]),
+        [src3] "=&f"(src[3]), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_rgb), [dstr_ptr] "r"(dst_r), [dstg_ptr] "r"(dst_g),
+        [dstb_ptr] "r"(dst_b), [width] "r"(width)
+      : "memory");
+}
+
+void MergeRGBRow_MMI(const uint8_t* src_r,
+                     const uint8_t* src_g,
+                     const uint8_t* src_b,
+                     uint8_t* dst_rgb,
+                     int width) {
+  uint64_t srcr, srcg, srcb, dest;
+  uint64_t srcrg_hi, srcrg_lo, srcbz_hi, srcbz_lo;
+  const uint64_t temp = 0x0;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[srcr],         0x07(%[srcr_ptr])                \n\t"
+      "gsldrc1    %[srcr],         0x00(%[srcr_ptr])                \n\t"
+      "gsldlc1    %[srcg],         0x07(%[srcg_ptr])                \n\t"
+      "gsldrc1    %[srcg],         0x00(%[srcg_ptr])                \n\t"
+      "punpcklbh  %[srcrg_lo],     %[srcr],           %[srcg]       \n\t"
+      "punpckhbh  %[srcrg_hi],     %[srcr],           %[srcg]       \n\t"
+
+      "gsldlc1    %[srcb],         0x07(%[srcb_ptr])                \n\t"
+      "gsldrc1    %[srcb],         0x00(%[srcb_ptr])                \n\t"
+      "punpcklbh  %[srcbz_lo],     %[srcb],           %[temp]       \n\t"
+      "punpckhbh  %[srcbz_hi],     %[srcb],           %[temp]       \n\t"
+
+      "punpcklhw  %[dest],         %[srcrg_lo],       %[srcbz_lo]   \n\t"
+      "gsswlc1    %[dest],         0x03(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+      "punpckhwd  %[dest],         %[dest],           %[dest]       \n\t"
+      "gsswlc1    %[dest],         0x06(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x03(%[dst_ptr])                 \n\t"
+      "punpckhhw  %[dest],         %[srcrg_lo],       %[srcbz_lo]   \n\t"
+      "gsswlc1    %[dest],         0x09(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x06(%[dst_ptr])                 \n\t"
+      "punpckhwd  %[dest],         %[dest],           %[dest]       \n\t"
+      "gsswlc1    %[dest],         0x0c(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x09(%[dst_ptr])                 \n\t"
+      "punpcklhw  %[dest],         %[srcrg_hi],       %[srcbz_hi]   \n\t"
+      "gsswlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x0c(%[dst_ptr])                 \n\t"
+      "punpckhwd  %[dest],         %[dest],           %[dest]       \n\t"
+      "gsswlc1    %[dest],         0x12(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "punpckhhw  %[dest],         %[srcrg_hi],       %[srcbz_hi]   \n\t"
+      "gsswlc1    %[dest],         0x15(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x12(%[dst_ptr])                 \n\t"
+      "punpckhwd  %[dest],         %[dest],           %[dest]       \n\t"
+      "gsswlc1    %[dest],         0x18(%[dst_ptr])                 \n\t"
+      "gsswrc1    %[dest],         0x15(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[srcr_ptr],     %[srcr_ptr],       0x08          \n\t"
+      "daddiu     %[srcg_ptr],     %[srcg_ptr],       0x08          \n\t"
+      "daddiu     %[srcb_ptr],     %[srcb_ptr],       0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x18          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [srcr] "=&f"(srcr), [srcg] "=&f"(srcg), [srcb] "=&f"(srcb),
+        [dest] "=&f"(dest), [srcrg_hi] "=&f"(srcrg_hi),
+        [srcrg_lo] "=&f"(srcrg_lo), [srcbz_hi] "=&f"(srcbz_hi),
+        [srcbz_lo] "=&f"(srcbz_lo)
+      : [srcr_ptr] "r"(src_r), [srcg_ptr] "r"(src_g), [srcb_ptr] "r"(src_b),
+        [dst_ptr] "r"(dst_rgb), [width] "r"(width), [temp] "f"(temp)
+      : "memory");
+}
+
+// Filter 2 rows of YUY2 UV's (422) into U and V (420).
+void YUY2ToUVRow_MMI(const uint8_t* src_yuy2,
+                     int src_stride_yuy2,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  uint64_t c0 = 0xff00ff00ff00ff00;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t temp[3];
+  uint64_t data[4];
+  uint64_t shift = 0x08;
+  uint64_t src_stride = 0x0;
+  __asm__ volatile(
+      "1:	                                                     \n\t"
+      "gsldrc1  %[t0],         0x00(%[src_yuy2])                     \n\t"
+      "gsldlc1  %[t0],         0x07(%[src_yuy2])                     \n\t"
+      "daddu    %[src_stride], %[src_yuy2],       %[src_stride_yuy2] \n\t"
+      "gsldrc1  %[t1],         0x00(%[src_stride])                   \n\t"
+      "gsldlc1  %[t1],         0x07(%[src_stride])                   \n\t"
+      "pavgb    %[t0],         %[t0],             %[t1]              \n\t"
+
+      "gsldrc1  %[t2],         0x08(%[src_yuy2])                     \n\t"
+      "gsldlc1  %[t2],         0x0f(%[src_yuy2])                     \n\t"
+      "gsldrc1  %[t1],         0x08(%[src_stride])                   \n\t"
+      "gsldlc1  %[t1],         0x0f(%[src_stride])                   \n\t"
+      "pavgb    %[t1],         %[t2],             %[t1]              \n\t"
+
+      "and      %[t0],         %[t0],             %[c0]              \n\t"
+      "and      %[t1],         %[t1],             %[c0]              \n\t"
+      "psrlh    %[t0],         %[t0],             %[shift]           \n\t"
+      "psrlh    %[t1],         %[t1],             %[shift]           \n\t"
+      "packushb %[t0],         %[t0],             %[t1]              \n\t"
+      "mov.s    %[t1],         %[t0]                                 \n\t"
+      "and      %[d0],         %[t0],             %[c1]              \n\t"
+      "psrlh    %[d1],         %[t1],             %[shift]           \n\t"
+
+      "gsldrc1  %[t0],         0x10(%[src_yuy2])                     \n\t"
+      "gsldlc1  %[t0],         0x17(%[src_yuy2])                     \n\t"
+      "gsldrc1  %[t1],         0x10(%[src_stride])                   \n\t"
+      "gsldlc1  %[t1],         0x17(%[src_stride])                   \n\t"
+      "pavgb    %[t0],         %[t0],              %[t1]             \n\t"
+
+      "gsldrc1  %[t2],         0x18(%[src_yuy2])                     \n\t"
+      "gsldlc1  %[t2],         0x1f(%[src_yuy2])                     \n\t"
+      "gsldrc1  %[t1],         0x18(%[src_stride])                   \n\t"
+      "gsldlc1  %[t1],         0x1f(%[src_stride])                   \n\t"
+      "pavgb    %[t1],         %[t2],              %[t1]             \n\t"
+
+      "and      %[t0],         %[t0],              %[c0]             \n\t"
+      "and      %[t1],         %[t1],              %[c0]             \n\t"
+      "psrlh    %[t0],         %[t0],              %[shift]          \n\t"
+      "psrlh    %[t1],         %[t1],              %[shift]          \n\t"
+      "packushb %[t0],         %[t0],              %[t1]             \n\t"
+      "mov.s    %[t1],         %[t0]                                 \n\t"
+      "and      %[d2],         %[t0],              %[c1]             \n\t"
+      "psrlh    %[d3],         %[t1],              %[shift]          \n\t"
+
+      "packushb %[d0],         %[d0],              %[d2]             \n\t"
+      "packushb %[d1],         %[d1],              %[d3]             \n\t"
+      "gssdrc1  %[d0],         0x0(%[dst_u])	                     \n\t"
+      "gssdlc1  %[d0],         0x7(%[dst_u])                         \n\t"
+      "gssdrc1  %[d1],         0x0(%[dst_v])	                     \n\t"
+      "gssdlc1  %[d1],         0x7(%[dst_v])                         \n\t"
+      "daddiu   %[src_yuy2],   %[src_yuy2],        32                \n\t"
+      "daddiu   %[dst_u],      %[dst_u],           8                 \n\t"
+      "daddiu   %[dst_v],      %[dst_v],           8                 \n\t"
+      "daddiu   %[width],      %[width],          -16                \n\t"
+      "bgtz     %[width],      1b                                    \n\t"
+      "nop                                                           \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [t2] "=&f"(temp[2]),
+        [d0] "=&f"(data[0]), [d1] "=&f"(data[1]), [d2] "=&f"(data[2]),
+        [d3] "=&f"(data[3]), [src_stride] "=&r"(src_stride)
+      : [src_yuy2] "r"(src_yuy2), [src_stride_yuy2] "r"(src_stride_yuy2),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [c0] "f"(c0), [c1] "f"(c1), [shift] "f"(shift)
+      : "memory");
+}
+
+// Copy row of YUY2 UV's (422) into U and V (422).
+void YUY2ToUV422Row_MMI(const uint8_t* src_yuy2,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  uint64_t c0 = 0xff00ff00ff00ff00;
+  uint64_t c1 = 0x00ff00ff00ff00ff;
+  uint64_t temp[2];
+  uint64_t data[4];
+  uint64_t shift = 0x08;
+  __asm__ volatile(
+      "1:	                                        \n\t"
+      "gsldrc1  %[t0],       0x00(%[src_yuy2])          \n\t"
+      "gsldlc1  %[t0],       0x07(%[src_yuy2])          \n\t"
+      "gsldrc1  %[t1],       0x08(%[src_yuy2])          \n\t"
+      "gsldlc1  %[t1],       0x0f(%[src_yuy2])          \n\t"
+      "and      %[t0],       %[t0],            %[c0]    \n\t"
+      "and      %[t1],       %[t1],            %[c0]    \n\t"
+      "psrlh    %[t0],       %[t0],            %[shift] \n\t"
+      "psrlh    %[t1],       %[t1],            %[shift] \n\t"
+      "packushb %[t0],       %[t0],            %[t1]    \n\t"
+      "mov.s    %[t1],       %[t0]                      \n\t"
+      "and      %[d0],       %[t0],            %[c1]    \n\t"
+      "psrlh    %[d1],       %[t1],            %[shift] \n\t"
+
+      "gsldrc1  %[t0],       0x10(%[src_yuy2])          \n\t"
+      "gsldlc1  %[t0],       0x17(%[src_yuy2])          \n\t"
+      "gsldrc1  %[t1],       0x18(%[src_yuy2])          \n\t"
+      "gsldlc1  %[t1],       0x1f(%[src_yuy2])          \n\t"
+      "and      %[t0],       %[t0],            %[c0]    \n\t"
+      "and      %[t1],       %[t1],            %[c0]    \n\t"
+      "psrlh    %[t0],       %[t0],            %[shift] \n\t"
+      "psrlh    %[t1],       %[t1],            %[shift] \n\t"
+      "packushb %[t0],       %[t0],            %[t1]    \n\t"
+      "mov.s    %[t1],       %[t0]                      \n\t"
+      "and      %[d2],       %[t0],            %[c1]    \n\t"
+      "psrlh    %[d3],       %[t1],            %[shift] \n\t"
+
+      "packushb %[d0],       %[d0],            %[d2]    \n\t"
+      "packushb %[d1],       %[d1],            %[d3]    \n\t"
+      "gssdrc1  %[d0],       0x0(%[dst_u])	        \n\t"
+      "gssdlc1  %[d0],       0x7(%[dst_u])              \n\t"
+      "gssdrc1  %[d1],       0x0(%[dst_v])	        \n\t"
+      "gssdlc1  %[d1],       0x7(%[dst_v])              \n\t"
+      "daddiu   %[src_yuy2], %[src_yuy2],      32       \n\t"
+      "daddiu   %[dst_u],    %[dst_u],         8        \n\t"
+      "daddiu   %[dst_v],    %[dst_v],         8        \n\t"
+      "daddiu   %[width],    %[width],        -16       \n\t"
+      "bgtz     %[width],    1b                         \n\t"
+      "nop                                              \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [d0] "=&f"(data[0]),
+        [d1] "=&f"(data[1]), [d2] "=&f"(data[2]), [d3] "=&f"(data[3])
+      : [src_yuy2] "r"(src_yuy2), [dst_u] "r"(dst_u), [dst_v] "r"(dst_v),
+        [width] "r"(width), [c0] "f"(c0), [c1] "f"(c1), [shift] "f"(shift)
+      : "memory");
+}
+
+// Copy row of YUY2 Y's (422) into Y (420/422).
+void YUY2ToYRow_MMI(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  uint64_t c0 = 0x00ff00ff00ff00ff;
+  uint64_t temp[2];
+  __asm__ volatile(
+      "1:	                                     \n\t"
+      "gsldrc1  %[t0],       0x00(%[src_yuy2])       \n\t"
+      "gsldlc1  %[t0],       0x07(%[src_yuy2])       \n\t"
+      "gsldrc1  %[t1],       0x08(%[src_yuy2])       \n\t"
+      "gsldlc1  %[t1],       0x0f(%[src_yuy2])       \n\t"
+      "and      %[t0],       %[t0],            %[c0] \n\t"
+      "and      %[t1],       %[t1],            %[c0] \n\t"
+      "packushb %[t0],       %[t0],            %[t1] \n\t"
+      "gssdrc1  %[t0],       0x0(%[dst_y])	     \n\t"
+      "gssdlc1  %[t0],       0x7(%[dst_y])           \n\t"
+      "daddiu   %[src_yuy2], %[src_yuy2],      16    \n\t"
+      "daddiu   %[dst_y],    %[dst_y],         8     \n\t"
+      "daddiu   %[width],    %[width],        -8     \n\t"
+      "bgtz     %[width],    1b                      \n\t"
+      "nop                                           \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1])
+      : [src_yuy2] "r"(src_yuy2), [dst_y] "r"(dst_y), [width] "r"(width),
+        [c0] "f"(c0)
+      : "memory");
+}
+
+// Filter 2 rows of UYVY UV's (422) into U and V (420).
+void UYVYToUVRow_MMI(const uint8_t* src_uyvy,
+                     int src_stride_uyvy,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  // Output a row of UV values.
+  uint64_t c0 = 0x00ff00ff00ff00ff;
+  uint64_t temp[3];
+  uint64_t data[4];
+  uint64_t shift = 0x08;
+  uint64_t src_stride = 0x0;
+  __asm__ volatile(
+      "1:	                                                      \n\t"
+      "gsldrc1  %[t0],         0x00(%[src_uyvy])                      \n\t"
+      "gsldlc1  %[t0],         0x07(%[src_uyvy])                      \n\t"
+      "daddu    %[src_stride], %[src_uyvy],        %[src_stride_uyvy] \n\t"
+      "gsldrc1  %[t1],         0x00(%[src_stride])                    \n\t"
+      "gsldlc1  %[t1],         0x07(%[src_stride])                    \n\t"
+      "pavgb    %[t0],         %[t0],              %[t1]              \n\t"
+
+      "gsldrc1  %[t2],         0x08(%[src_uyvy])                      \n\t"
+      "gsldlc1  %[t2],         0x0f(%[src_uyvy])                      \n\t"
+      "gsldrc1  %[t1],         0x08(%[src_stride])                    \n\t"
+      "gsldlc1  %[t1],         0x0f(%[src_stride])                    \n\t"
+      "pavgb    %[t1],         %[t2],              %[t1]              \n\t"
+
+      "and      %[t0],         %[t0],              %[c0]              \n\t"
+      "and      %[t1],         %[t1],              %[c0]              \n\t"
+      "packushb %[t0],         %[t0],              %[t1]              \n\t"
+      "mov.s    %[t1],         %[t0]                                  \n\t"
+      "and      %[d0],         %[t0],              %[c0]              \n\t"
+      "psrlh    %[d1],         %[t1],              %[shift]           \n\t"
+
+      "gsldrc1  %[t0],         0x10(%[src_uyvy])                      \n\t"
+      "gsldlc1  %[t0],         0x17(%[src_uyvy])                      \n\t"
+      "gsldrc1  %[t1],         0x10(%[src_stride])                    \n\t"
+      "gsldlc1  %[t1],         0x17(%[src_stride])                    \n\t"
+      "pavgb    %[t0],         %[t0],              %[t1]              \n\t"
+
+      "gsldrc1  %[t2],         0x18(%[src_uyvy])                      \n\t"
+      "gsldlc1  %[t2],         0x1f(%[src_uyvy])                      \n\t"
+      "gsldrc1  %[t1],         0x18(%[src_stride])                    \n\t"
+      "gsldlc1  %[t1],         0x1f(%[src_stride])                    \n\t"
+      "pavgb    %[t1],         %[t2],              %[t1]              \n\t"
+
+      "and      %[t0],         %[t0],              %[c0]              \n\t"
+      "and      %[t1],         %[t1],              %[c0]              \n\t"
+      "packushb %[t0],         %[t0],              %[t1]              \n\t"
+      "mov.s    %[t1],         %[t0]                                  \n\t"
+      "and      %[d2],         %[t0],              %[c0]              \n\t"
+      "psrlh    %[d3],         %[t1],              %[shift]           \n\t"
+
+      "packushb %[d0],         %[d0],              %[d2]              \n\t"
+      "packushb %[d1],         %[d1],              %[d3]              \n\t"
+      "gssdrc1  %[d0],         0x0(%[dst_u])	                      \n\t"
+      "gssdlc1  %[d0],         0x7(%[dst_u])                          \n\t"
+      "gssdrc1  %[d1],         0x0(%[dst_v])	                      \n\t"
+      "gssdlc1  %[d1],         0x7(%[dst_v])                          \n\t"
+      "daddiu   %[src_uyvy],   %[src_uyvy],        32                 \n\t"
+      "daddiu   %[dst_u],      %[dst_u],           8                  \n\t"
+      "daddiu   %[dst_v],      %[dst_v],           8                  \n\t"
+      "daddiu   %[width],      %[width],          -16                 \n\t"
+      "bgtz     %[width],      1b                                     \n\t"
+      "nop                                                            \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [t2] "=&f"(temp[2]),
+        [d0] "=&f"(data[0]), [d1] "=&f"(data[1]), [d2] "=&f"(data[2]),
+        [d3] "=&f"(data[3]), [src_stride] "=&r"(src_stride)
+      : [src_uyvy] "r"(src_uyvy), [src_stride_uyvy] "r"(src_stride_uyvy),
+        [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
+        [c0] "f"(c0), [shift] "f"(shift)
+      : "memory");
+}
+
+// Copy row of UYVY UV's (422) into U and V (422).
+void UYVYToUV422Row_MMI(const uint8_t* src_uyvy,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  // Output a row of UV values.
+  uint64_t c0 = 0x00ff00ff00ff00ff;
+  uint64_t temp[2];
+  uint64_t data[4];
+  uint64_t shift = 0x08;
+  __asm__ volatile(
+      "1:	                                        \n\t"
+      "gsldrc1  %[t0],       0x00(%[src_uyvy])          \n\t"
+      "gsldlc1  %[t0],       0x07(%[src_uyvy])          \n\t"
+      "gsldrc1  %[t1],       0x08(%[src_uyvy])          \n\t"
+      "gsldlc1  %[t1],       0x0f(%[src_uyvy])          \n\t"
+      "and      %[t0],       %[t0],            %[c0]    \n\t"
+      "and      %[t1],       %[t1],            %[c0]    \n\t"
+      "packushb %[t0],       %[t0],            %[t1]    \n\t"
+      "mov.s    %[t1],       %[t0]                      \n\t"
+      "and      %[d0],       %[t0],            %[c0]    \n\t"
+      "psrlh    %[d1],       %[t1],            %[shift] \n\t"
+
+      "gsldrc1  %[t0],       0x10(%[src_uyvy])          \n\t"
+      "gsldlc1  %[t0],       0x17(%[src_uyvy])          \n\t"
+      "gsldrc1  %[t1],       0x18(%[src_uyvy])          \n\t"
+      "gsldlc1  %[t1],       0x1f(%[src_uyvy])          \n\t"
+      "and      %[t0],       %[t0],            %[c0]    \n\t"
+      "and      %[t1],       %[t1],            %[c0]    \n\t"
+      "packushb %[t0],       %[t0],            %[t1]    \n\t"
+      "mov.s    %[t1],       %[t0]                      \n\t"
+      "and      %[d2],       %[t0],            %[c0]    \n\t"
+      "psrlh    %[d3],       %[t1],            %[shift] \n\t"
+
+      "packushb %[d0],       %[d0],            %[d2]    \n\t"
+      "packushb %[d1],       %[d1],            %[d3]    \n\t"
+      "gssdrc1  %[d0],       0x0(%[dst_u])	        \n\t"
+      "gssdlc1  %[d0],       0x7(%[dst_u])              \n\t"
+      "gssdrc1  %[d1],       0x0(%[dst_v])	        \n\t"
+      "gssdlc1  %[d1],       0x7(%[dst_v])              \n\t"
+      "daddiu   %[src_uyvy], %[src_uyvy],      32       \n\t"
+      "daddiu   %[dst_u],    %[dst_u],         8        \n\t"
+      "daddiu   %[dst_v],    %[dst_v],         8        \n\t"
+      "daddiu   %[width],    %[width],        -16       \n\t"
+      "bgtz     %[width],    1b                         \n\t"
+      "nop                                              \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1]), [d0] "=&f"(data[0]),
+        [d1] "=&f"(data[1]), [d2] "=&f"(data[2]), [d3] "=&f"(data[3])
+      : [src_uyvy] "r"(src_uyvy), [dst_u] "r"(dst_u), [dst_v] "r"(dst_v),
+        [width] "r"(width), [c0] "f"(c0), [shift] "f"(shift)
+      : "memory");
+}
+
+// Copy row of UYVY Y's (422) into Y (420/422).
+void UYVYToYRow_MMI(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  // Output a row of Y values.
+  uint64_t c0 = 0x00ff00ff00ff00ff;
+  uint64_t shift = 0x08;
+  uint64_t temp[2];
+  __asm__ volatile(
+      "1:	                                        \n\t"
+      "gsldrc1  %[t0],       0x00(%[src_uyvy])          \n\t"
+      "gsldlc1  %[t0],       0x07(%[src_uyvy])          \n\t"
+      "gsldrc1  %[t1],       0x08(%[src_uyvy])          \n\t"
+      "gsldlc1  %[t1],       0x0f(%[src_uyvy])          \n\t"
+      "dsrl     %[t0],       %[t0],            %[shift] \n\t"
+      "dsrl     %[t1],       %[t1],            %[shift] \n\t"
+      "and      %[t0],       %[t0],            %[c0]    \n\t"
+      "and      %[t1],       %[t1],            %[c0]    \n\t"
+      "and      %[t1],       %[t1],            %[c0]    \n\t"
+      "packushb %[t0],       %[t0],            %[t1]    \n\t"
+      "gssdrc1  %[t0],       0x0(%[dst_y])	        \n\t"
+      "gssdlc1  %[t0],       0x7(%[dst_y])              \n\t"
+      "daddiu   %[src_uyvy], %[src_uyvy],      16       \n\t"
+      "daddiu   %[dst_y],    %[dst_y],         8        \n\t"
+      "daddiu   %[width],    %[width],        -8        \n\t"
+      "bgtz     %[width],    1b                         \n\t"
+      "nop                                              \n\t"
+      : [t0] "=&f"(temp[0]), [t1] "=&f"(temp[1])
+      : [src_uyvy] "r"(src_uyvy), [dst_y] "r"(dst_y), [width] "r"(width),
+        [c0] "f"(c0), [shift] "f"(shift)
+      : "memory");
+}
+
+// Blend src_argb over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb or src_argb1.
+// This code mimics the SSSE3 version for better testability.
+void ARGBBlendRow_MMI(const uint8_t* src_argb,
+                      const uint8_t* src_argb1,
+                      uint8_t* dst_argb,
+                      int width) {
+  uint64_t src0, src1, dest, alpha, src0_hi, src0_lo, src1_hi, src1_lo, dest_hi,
+      dest_lo;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x00FFFFFF00FFFFFFULL;
+  const uint64_t mask2 = 0x00FF00FF00FF00FFULL;
+  const uint64_t mask3 = 0xFF;
+  const uint64_t mask4 = ~mask1;
+  const uint64_t shift = 0x08;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src0],         0x07(%[src0_ptr])                \n\t"
+      "gsldrc1    %[src0],         0x00(%[src0_ptr])                \n\t"
+      "punpcklbh  %[src0_lo],      %[src0],           %[mask0]      \n\t"
+
+      "gsldlc1    %[src1],         0x07(%[src1_ptr])                \n\t"
+      "gsldrc1    %[src1],         0x00(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src1_lo],      %[src1],           %[mask0]      \n\t"
+
+      "psubush    %[alpha],        %[mask2],          %[src0_lo]    \n\t"
+      "pshufh     %[alpha],        %[alpha],          %[mask3]      \n\t"
+      "pmullh     %[dest_lo],      %[src1_lo],        %[alpha]      \n\t"
+      "psrlh      %[dest_lo],      %[dest_lo],        %[shift]      \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[src0_lo]    \n\t"
+
+      "punpckhbh  %[src0_hi],      %[src0],           %[mask0]      \n\t"
+      "punpckhbh  %[src1_hi],      %[src1],           %[mask0]      \n\t"
+
+      "psubush    %[alpha],        %[mask2],          %[src0_hi]    \n\t"
+      "pshufh     %[alpha],        %[alpha],          %[mask3]      \n\t"
+      "pmullh     %[dest_hi],      %[src1_hi],        %[alpha]      \n\t"
+      "psrlh      %[dest_hi],      %[dest_hi],        %[shift]      \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[src0_hi]    \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "and        %[dest],         %[dest],           %[mask1]      \n\t"
+      "or         %[dest],         %[dest],           %[mask4]      \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src0_ptr],     %[src0_ptr],       0x08          \n\t"
+      "daddiu     %[src1_ptr],     %[src1_ptr],       0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [alpha] "=&f"(alpha),
+        [dest] "=&f"(dest), [src0_hi] "=&f"(src0_hi), [src0_lo] "=&f"(src0_lo),
+        [src1_hi] "=&f"(src1_hi), [src1_lo] "=&f"(src1_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo)
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
+        [dst_ptr] "r"(dst_argb), [mask0] "f"(mask0), [mask1] "f"(mask1),
+        [mask2] "f"(mask2), [mask3] "f"(mask3), [mask4] "f"(mask4),
+        [shift] "f"(shift), [width] "r"(width)
+      : "memory");
+}
+
+void BlendPlaneRow_MMI(const uint8_t* src0,
+                       const uint8_t* src1,
+                       const uint8_t* alpha,
+                       uint8_t* dst,
+                       int width) {
+  uint64_t source0, source1, dest, alph;
+  uint64_t src0_hi, src0_lo, src1_hi, src1_lo, alpha_hi, alpha_lo, dest_hi,
+      dest_lo;
+  uint64_t alpha_rev, alpha_rev_lo, alpha_rev_hi;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0xFFFFFFFFFFFFFFFFULL;
+  const uint64_t mask2 = 0x00FF00FF00FF00FFULL;
+  const uint64_t shift = 0x08;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src0],         0x07(%[src0_ptr])                \n\t"
+      "gsldrc1    %[src0],         0x00(%[src0_ptr])                \n\t"
+      "punpcklbh  %[src0_lo],      %[src0],           %[mask0]      \n\t"
+      "punpckhbh  %[src0_hi],      %[src0],           %[mask0]      \n\t"
+
+      "gsldlc1    %[src1],         0x07(%[src1_ptr])                \n\t"
+      "gsldrc1    %[src1],         0x00(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src1_lo],      %[src1],           %[mask0]      \n\t"
+      "punpckhbh  %[src1_hi],      %[src1],           %[mask0]      \n\t"
+
+      "gsldlc1    %[alpha],        0x07(%[alpha_ptr])               \n\t"
+      "gsldrc1    %[alpha],        0x00(%[alpha_ptr])               \n\t"
+      "psubusb    %[alpha_r],      %[mask1],          %[alpha]      \n\t"
+      "punpcklbh  %[alpha_lo],     %[alpha],          %[mask0]      \n\t"
+      "punpckhbh  %[alpha_hi],     %[alpha],          %[mask0]      \n\t"
+      "punpcklbh  %[alpha_rlo],    %[alpha_r],        %[mask0]      \n\t"
+      "punpckhbh  %[alpha_rhi],    %[alpha_r],        %[mask0]      \n\t"
+
+      "pmullh     %[dest_lo],      %[src0_lo],        %[alpha_lo]   \n\t"
+      "pmullh     %[dest],         %[src1_lo],        %[alpha_rlo]  \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[dest]       \n\t"
+      "paddush    %[dest_lo],      %[dest_lo],        %[mask2]      \n\t"
+      "psrlh      %[dest_lo],      %[dest_lo],        %[shift]      \n\t"
+
+      "pmullh     %[dest_hi],      %[src0_hi],        %[alpha_hi]   \n\t"
+      "pmullh     %[dest],         %[src1_hi],        %[alpha_rhi]  \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[dest]       \n\t"
+      "paddush    %[dest_hi],      %[dest_hi],        %[mask2]      \n\t"
+      "psrlh      %[dest_hi],      %[dest_hi],        %[shift]      \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src0_ptr],     %[src0_ptr],       0x08          \n\t"
+      "daddiu     %[src1_ptr],     %[src1_ptr],       0x08          \n\t"
+      "daddiu     %[alpha_ptr],    %[alpha_ptr],      0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(source0), [src1] "=&f"(source1), [alpha] "=&f"(alph),
+        [dest] "=&f"(dest), [src0_hi] "=&f"(src0_hi), [src0_lo] "=&f"(src0_lo),
+        [src1_hi] "=&f"(src1_hi), [src1_lo] "=&f"(src1_lo),
+        [alpha_hi] "=&f"(alpha_hi), [alpha_lo] "=&f"(alpha_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [alpha_rlo] "=&f"(alpha_rev_lo), [alpha_rhi] "=&f"(alpha_rev_hi),
+        [alpha_r] "=&f"(alpha_rev)
+      : [src0_ptr] "r"(src0), [src1_ptr] "r"(src1), [alpha_ptr] "r"(alpha),
+        [dst_ptr] "r"(dst), [mask0] "f"(mask0), [mask1] "f"(mask1),
+        [mask2] "f"(mask2), [shift] "f"(shift), [width] "r"(width)
+      : "memory");
+}
+
+// Multiply source RGB by alpha and store to destination.
+// This code mimics the SSSE3 version for better testability.
+void ARGBAttenuateRow_MMI(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          int width) {
+  uint64_t src, src_hi, src_lo, dest, dest_hi, dest_lo, alpha;
+  const uint64_t mask0 = 0xFF;
+  const uint64_t mask1 = 0xFF000000FF000000ULL;
+  const uint64_t mask2 = ~mask1;
+  const uint64_t shift = 0x08;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[src]        \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[src]        \n\t"
+
+      "pshufh     %[alpha],        %[src_lo],         %[mask0]      \n\t"
+      "pmulhuh    %[dest_lo],      %[alpha],          %[src_lo]     \n\t"
+      "psrlh      %[dest_lo],      %[dest_lo],        %[shift]      \n\t"
+      "pshufh     %[alpha],        %[src_hi],         %[mask0]      \n\t"
+      "pmulhuh    %[dest_hi],      %[alpha],          %[src_hi]     \n\t"
+      "psrlh      %[dest_hi],      %[dest_hi],        %[shift]      \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "and        %[dest],         %[dest],           %[mask2]      \n\t"
+      "and        %[src],          %[src],            %[mask1]      \n\t"
+      "or         %[dest],         %[dest],           %[src]        \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest), [src_hi] "=&f"(src_hi),
+        [src_lo] "=&f"(src_lo), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [alpha] "=&f"(alpha)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [mask2] "f"(mask2), [shift] "f"(shift),
+        [width] "r"(width)
+      : "memory");
+}
+
+void ComputeCumulativeSumRow_MMI(const uint8_t* row,
+                                 int32_t* cumsum,
+                                 const int32_t* previous_cumsum,
+                                 int width) {
+  int64_t row_sum[2] = {0, 0};
+  uint64_t src, dest0, dest1, presrc0, presrc1, dest;
+  const uint64_t mask = 0x0;
+
+  __asm__ volatile(
+      "xor        %[row_sum0],     %[row_sum0],       %[row_sum0]   \n\t"
+      "xor        %[row_sum1],     %[row_sum1],       %[row_sum1]   \n\t"
+
+      "1:                                                           \n\t"
+      "gslwlc1    %[src],          0x03(%[row_ptr])                 \n\t"
+      "gslwrc1    %[src],          0x00(%[row_ptr])                 \n\t"
+
+      "punpcklbh  %[src],          %[src],            %[mask]       \n\t"
+      "punpcklhw  %[dest0],        %[src],            %[mask]       \n\t"
+      "punpckhhw  %[dest1],        %[src],            %[mask]       \n\t"
+
+      "paddw      %[row_sum0],     %[row_sum0],       %[dest0]      \n\t"
+      "paddw      %[row_sum1],     %[row_sum1],       %[dest1]      \n\t"
+
+      "gsldlc1    %[presrc0],      0x07(%[pre_ptr])                 \n\t"
+      "gsldrc1    %[presrc0],      0x00(%[pre_ptr])                 \n\t"
+      "gsldlc1    %[presrc1],      0x0f(%[pre_ptr])                 \n\t"
+      "gsldrc1    %[presrc1],      0x08(%[pre_ptr])                 \n\t"
+
+      "paddw      %[dest0],        %[row_sum0],       %[presrc0]    \n\t"
+      "paddw      %[dest1],        %[row_sum1],       %[presrc1]    \n\t"
+
+      "gssdlc1    %[dest0],        0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_ptr])                 \n\t"
+      "gssdlc1    %[dest1],        0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest1],        0x08(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[row_ptr],      %[row_ptr],        0x04          \n\t"
+      "daddiu     %[pre_ptr],      %[pre_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x01          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest), [dest0] "=&f"(dest0),
+        [dest1] "=&f"(dest1), [row_sum0] "+&f"(row_sum[0]),
+        [row_sum1] "+&f"(row_sum[1]), [presrc0] "=&f"(presrc0),
+        [presrc1] "=&f"(presrc1)
+      : [row_ptr] "r"(row), [pre_ptr] "r"(previous_cumsum),
+        [dst_ptr] "r"(cumsum), [width] "r"(width), [mask] "f"(mask)
+      : "memory");
+}
+
+// C version 2x2 -> 2x1.
+void InterpolateRow_MMI(uint8_t* dst_ptr,
+                        const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int source_y_fraction) {
+  if (source_y_fraction == 0) {
+    __asm__ volatile(
+        "1:	                              \n\t"
+        "ld     $t0,        0x0(%[src_ptr])   \n\t"
+        "sd     $t0,        0x0(%[dst_ptr])   \n\t"
+        "daddiu %[src_ptr], %[src_ptr],     8 \n\t"
+        "daddiu %[dst_ptr], %[dst_ptr],     8 \n\t"
+        "daddiu %[width],   %[width],      -8 \n\t"
+        "bgtz   %[width],   1b                \n\t"
+        "nop                                  \n\t"
+        :
+        : [dst_ptr] "r"(dst_ptr), [src_ptr] "r"(src_ptr), [width] "r"(width)
+        : "memory");
+    return;
+  }
+  if (source_y_fraction == 128) {
+    uint64_t uv = 0x0;
+    uint64_t uv_stride = 0x0;
+    __asm__ volatile(
+        "1:	                                            \n\t"
+        "gsldrc1 %[uv],        0x0(%[src_ptr])              \n\t"
+        "gsldlc1 %[uv],        0x7(%[src_ptr])              \n\t"
+        "daddu   $t0,          %[src_ptr],     %[stride]    \n\t"
+        "gsldrc1 %[uv_stride], 0x0($t0)                     \n\t"
+        "gsldlc1 %[uv_stride], 0x7($t0)                     \n\t"
+
+        "pavgb   %[uv],        %[uv],          %[uv_stride] \n\t"
+        "gssdrc1 %[uv],        0x0(%[dst_ptr])              \n\t"
+        "gssdlc1 %[uv],        0x7(%[dst_ptr])              \n\t"
+
+        "daddiu  %[src_ptr],   %[src_ptr],     8            \n\t"
+        "daddiu  %[dst_ptr],   %[dst_ptr],     8            \n\t"
+        "daddiu  %[width],     %[width],      -8            \n\t"
+        "bgtz    %[width],     1b                           \n\t"
+        "nop                                                \n\t"
+        : [uv] "=&f"(uv), [uv_stride] "=&f"(uv_stride)
+        : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst_ptr), [width] "r"(width),
+          [stride] "r"((int64_t)src_stride)
+        : "memory");
+    return;
+  }
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint64_t temp;
+  uint64_t data[4];
+  uint64_t zero = 0x0;
+  uint64_t c0 = 0x0080008000800080;
+  uint64_t fy0 = 0x0100010001000100;
+  uint64_t shift = 0x8;
+  __asm__ volatile(
+      "pshufh    %[fy1],      %[fy1],          %[zero]  \n\t"
+      "psubh     %[fy0],      %[fy0],          %[fy1]   \n\t"
+      "1:	                                        \n\t"
+      "gsldrc1   %[t0],       0x0(%[src_ptr])           \n\t"
+      "gsldlc1   %[t0],       0x7(%[src_ptr])           \n\t"
+      "punpcklbh %[d0],       %[t0],           %[zero]  \n\t"
+      "punpckhbh %[d1],       %[t0],           %[zero]  \n\t"
+      "gsldrc1   %[t0],       0x0(%[src_ptr1])          \n\t"
+      "gsldlc1   %[t0],       0x7(%[src_ptr1])          \n\t"
+      "punpcklbh %[d2],       %[t0],           %[zero]  \n\t"
+      "punpckhbh %[d3],       %[t0],           %[zero]  \n\t"
+
+      "pmullh    %[d0],       %[d0],           %[fy0]   \n\t"
+      "pmullh    %[d2],       %[d2],           %[fy1]   \n\t"
+      "paddh     %[d0],       %[d0],           %[d2]    \n\t"
+      "paddh     %[d0],       %[d0],           %[c0]    \n\t"
+      "psrlh     %[d0],       %[d0],           %[shift] \n\t"
+
+      "pmullh    %[d1],       %[d1],           %[fy0]   \n\t"
+      "pmullh    %[d3],       %[d3],           %[fy1]   \n\t"
+      "paddh     %[d1],       %[d1],           %[d3]    \n\t"
+      "paddh     %[d1],       %[d1],           %[c0]    \n\t"
+      "psrlh     %[d1],       %[d1],           %[shift] \n\t"
+
+      "packushb  %[d0],       %[d0],           %[d1]    \n\t"
+      "gssdrc1   %[d0],       0x0(%[dst_ptr])           \n\t"
+      "gssdlc1   %[d0],       0x7(%[dst_ptr])           \n\t"
+      "daddiu    %[src_ptr],  %[src_ptr],      8        \n\t"
+      "daddiu    %[src_ptr1], %[src_ptr1],     8        \n\t"
+      "daddiu    %[dst_ptr],  %[dst_ptr],      8        \n\t"
+      "daddiu    %[width],    %[width],       -8        \n\t"
+      "bgtz      %[width],    1b                        \n\t"
+      "nop                                              \n\t"
+      : [t0] "=&f"(temp), [d0] "=&f"(data[0]), [d1] "=&f"(data[1]),
+        [d2] "=&f"(data[2]), [d3] "=&f"(data[3])
+      : [src_ptr] "r"(src_ptr), [src_ptr1] "r"(src_ptr1),
+        [dst_ptr] "r"(dst_ptr), [width] "r"(width),
+        [fy1] "f"(source_y_fraction), [fy0] "f"(fy0), [c0] "f"(c0),
+        [shift] "f"(shift), [zero] "f"(zero)
+      : "memory");
+}
+
+// Use first 4 shuffler values to reorder ARGB channels.
+void ARGBShuffleRow_MMI(const uint8_t* src_argb,
+                        uint8_t* dst_argb,
+                        const uint8_t* shuffler,
+                        int width) {
+  uint64_t source, dest0, dest1, dest;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = (shuffler[0] & 0x03) | ((shuffler[1] & 0x03) << 2) |
+                         ((shuffler[2] & 0x03) << 4) |
+                         ((shuffler[3] & 0x03) << 6);
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+
+      "punpcklbh  %[dest0],        %[src],            %[mask0]      \n\t"
+      "pshufh     %[dest0],        %[dest0],          %[mask1]      \n\t"
+      "punpckhbh  %[dest1],        %[src],            %[mask0]      \n\t"
+      "pshufh     %[dest1],        %[dest1],          %[mask1]      \n\t"
+      "packushb   %[dest],         %[dest0],          %[dest1]      \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(source), [dest] "=&f"(dest), [dest0] "=&f"(dest0),
+        [dest1] "=&f"(dest1)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [width] "r"(width)
+      : "memory");
+}
+
+void I422ToYUY2Row_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_frame,
+                       int width) {
+  uint64_t temp[3];
+  uint64_t vu = 0x0;
+  __asm__ volatile(
+      "1:	                                        \n\t"
+      "gsldlc1   %[ty],        0x7(%[src_y])            \n\t"  // r=src_sobelx[i]
+      "gsldrc1   %[ty],        0x0(%[src_y])            \n\t"  // r=src_sobelx[i]
+      "gslwlc1   %[tu],        0x3(%[src_u])            \n\t"  // b=src_sobely[i]
+      "gslwrc1   %[tu],        0x0(%[src_u])            \n\t"  // b=src_sobely[i]
+      "gslwlc1   %[tv],        0x3(%[src_v])            \n\t"  // b=src_sobely[i]
+      "gslwrc1   %[tv],        0x0(%[src_v])            \n\t"  // b=src_sobely[i]
+      "punpcklbh %[vu],        %[tu],             %[tv]	\n\t"  // g
+      "punpcklbh %[tu],        %[ty],             %[vu]	\n\t"  // g
+      "gssdlc1   %[tu],        0x7(%[dst_frame])        \n\t"
+      "gssdrc1   %[tu],        0x0(%[dst_frame])        \n\t"
+      "punpckhbh %[tu],        %[ty],             %[vu]	\n\t"  // g
+      "gssdlc1   %[tu],        0x0F(%[dst_frame])       \n\t"
+      "gssdrc1   %[tu],        0x08(%[dst_frame])       \n\t"
+      "daddiu    %[src_y],     %[src_y],          8     \n\t"
+      "daddiu    %[src_u],     %[src_u],          4     \n\t"
+      "daddiu    %[src_v],     %[src_v],          4     \n\t"
+      "daddiu    %[dst_frame], %[dst_frame],      16    \n\t"
+      "daddiu    %[width],     %[width],         -8     \n\t"
+      "bgtz      %[width],     1b                       \n\t"
+      "nop                                              \n\t"
+      : [ty] "=&f"(temp[1]), [tu] "=&f"(temp[1]), [tv] "=&f"(temp[1]),
+        [vu] "=&f"(vu)
+      : [src_y] "r"(src_y), [src_u] "r"(src_u), [src_v] "r"(src_v),
+        [dst_frame] "r"(dst_frame), [width] "r"(width)
+      : "memory");
+}
+
+void I422ToUYVYRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_frame,
+                       int width) {
+  uint64_t temp[3];
+  uint64_t vu = 0x0;
+  __asm__ volatile(
+      "1:	                                        \n\t"
+      "gsldlc1   %[ty],        0x7(%[src_y])            \n\t"  // r=src_sobelx[i]
+      "gsldrc1   %[ty],        0x0(%[src_y])            \n\t"  // r=src_sobelx[i]
+      "gslwlc1   %[tu],        0x3(%[src_u])            \n\t"  // b=src_sobely[i]
+      "gslwrc1   %[tu],        0x0(%[src_u])            \n\t"  // b=src_sobely[i]
+      "gslwlc1   %[tv],        0x3(%[src_v])            \n\t"  // b=src_sobely[i]
+      "gslwrc1   %[tv],        0x0(%[src_v])            \n\t"  // b=src_sobely[i]
+      "punpcklbh %[vu],        %[tu],             %[tv]	\n\t"  // g
+      "punpcklbh %[tu],        %[vu],             %[ty]	\n\t"  // g
+      "gssdlc1   %[tu],        0x7(%[dst_frame])        \n\t"
+      "gssdrc1   %[tu],        0x0(%[dst_frame])        \n\t"
+      "punpckhbh %[tu],        %[vu],             %[ty]	\n\t"  // g
+      "gssdlc1   %[tu],        0x0F(%[dst_frame])       \n\t"
+      "gssdrc1   %[tu],        0x08(%[dst_frame])       \n\t"
+      "daddiu    %[src_y],     %[src_y],          8     \n\t"
+      "daddiu    %[src_u],     %[src_u],          4     \n\t"
+      "daddiu    %[src_v],     %[src_v],          4     \n\t"
+      "daddiu    %[dst_frame], %[dst_frame],      16    \n\t"
+      "daddiu    %[width],     %[width],         -8     \n\t"
+      "bgtz      %[width],     1b                       \n\t"
+      "nop                                              \n\t"
+      : [ty] "=&f"(temp[1]), [tu] "=&f"(temp[1]), [tv] "=&f"(temp[1]),
+        [vu] "=&f"(vu)
+      : [src_y] "r"(src_y), [src_u] "r"(src_u), [src_v] "r"(src_v),
+        [dst_frame] "r"(dst_frame), [width] "r"(width)
+      : "memory");
+}
+
+void ARGBCopyAlphaRow_MMI(const uint8_t* src, uint8_t* dst, int width) {
+  uint64_t source, dest;
+  const uint64_t mask0 = 0xff000000ff000000ULL;
+  const uint64_t mask1 = ~mask0;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gsldrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "and        %[src],          %[src],            %[mask0]      \n\t"
+      "and        %[dest],         %[dest],           %[mask1]      \n\t"
+      "or         %[dest],         %[src],            %[dest]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(source), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src), [dst_ptr] "r"(dst), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [width] "r"(width)
+      : "memory");
+}
+
+void ARGBExtractAlphaRow_MMI(const uint8_t* src_argb,
+                             uint8_t* dst_a,
+                             int width) {
+  uint64_t src, dest0, dest1, dest_lo, dest_hi, dest;
+  const uint64_t mask = 0xff000000ff000000ULL;
+  const uint64_t shift = 0x18;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "and        %[dest0],        %[src],            %[mask]       \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[shift]      \n\t"
+      "gsldlc1    %[src],          0x0f(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x08(%[src_ptr])                 \n\t"
+      "and        %[dest1],        %[src],            %[mask]       \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[shift]      \n\t"
+      "packsswh   %[dest_lo],      %[dest0],          %[dest1]      \n\t"
+
+      "gsldlc1    %[src],          0x17(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x10(%[src_ptr])                 \n\t"
+      "and        %[dest0],        %[src],            %[mask]       \n\t"
+      "psrlw      %[dest0],        %[dest0],          %[shift]      \n\t"
+      "gsldlc1    %[src],          0x1f(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x18(%[src_ptr])                 \n\t"
+      "and        %[dest1],        %[src],            %[mask]       \n\t"
+      "psrlw      %[dest1],        %[dest1],          %[shift]      \n\t"
+      "packsswh   %[dest_hi],      %[dest0],          %[dest1]      \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x20          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest), [dest0] "=&f"(dest0),
+        [dest1] "=&f"(dest1), [dest_lo] "=&f"(dest_lo), [dest_hi] "=&f"(dest_hi)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_a), [mask] "f"(mask),
+        [shift] "f"(shift), [width] "r"(width)
+      : "memory");
+}
+
+void ARGBCopyYToAlphaRow_MMI(const uint8_t* src, uint8_t* dst, int width) {
+  uint64_t source, dest0, dest1, dest;
+  const uint64_t mask0 = 0x0;
+  const uint64_t mask1 = 0x00ffffff00ffffffULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+
+      "punpcklbh  %[dest0],        %[mask0],          %[src]        \n\t"
+      "punpcklhw  %[dest1],        %[mask0],          %[dest0]      \n\t"
+      "gsldlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gsldrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+      "and        %[dest],         %[dest],           %[mask1]      \n\t"
+      "or         %[dest],         %[dest],           %[dest1]      \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+      "punpckhhw  %[dest1],        %[mask0],          %[dest0]      \n\t"
+      "gsldlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gsldrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+      "and        %[dest],         %[dest],           %[mask1]      \n\t"
+      "or         %[dest],         %[dest],           %[dest1]      \n\t"
+      "gssdlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+
+      "punpckhbh  %[dest0],        %[mask0],          %[src]        \n\t"
+      "punpcklhw  %[dest1],        %[mask0],          %[dest0]      \n\t"
+      "gsldlc1    %[dest],         0x17(%[dst_ptr])                 \n\t"
+      "gsldrc1    %[dest],         0x10(%[dst_ptr])                 \n\t"
+      "and        %[dest],         %[dest],           %[mask1]      \n\t"
+      "or         %[dest],         %[dest],           %[dest1]      \n\t"
+      "gssdlc1    %[dest],         0x17(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x10(%[dst_ptr])                 \n\t"
+      "punpckhhw  %[dest1],        %[mask0],          %[dest0]      \n\t"
+      "gsldlc1    %[dest],         0x1f(%[dst_ptr])                 \n\t"
+      "gsldrc1    %[dest],         0x18(%[dst_ptr])                 \n\t"
+      "and        %[dest],         %[dest],           %[mask1]      \n\t"
+      "or         %[dest],         %[dest],           %[dest1]      \n\t"
+      "gssdlc1    %[dest],         0x1f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x18(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x20          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(source), [dest] "=&f"(dest), [dest0] "=&f"(dest0),
+        [dest1] "=&f"(dest1)
+      : [src_ptr] "r"(src), [dst_ptr] "r"(dst), [mask0] "f"(mask0),
+        [mask1] "f"(mask1), [width] "r"(width)
+      : "memory");
+}
+
+void I444ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+  __asm__ volatile (
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"//yg
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"//bb
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"//ub
+    "or         %[ub],           %[ub],             %[mask]       \n\t"//must sign extension
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"//bg
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"//ug
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"//vg
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"//br
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"//vr
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"//sign extension
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"//u
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"//v
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"//u*ug
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"//v*vg
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"//v*vr
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//rrrrbbbb
+    "packushb   %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"//ffffgggg
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//gbgbgbgb
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//frfrfrfr
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x04          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+    : [y]"=&f"(y),
+      [u]"=&f"(u),                         [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [alpha]"f"(-1),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask)
+    : "memory"
+  );
+}
+
+// Also used for 420
+void I422ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"//yg
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"//bb
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"//ub
+    "or         %[ub],           %[ub],             %[mask]       \n\t"//must sign extension
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"//bg
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"//ug
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"//vg
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"//br
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"//vr
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"//sign extension
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"//v
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"//u*ug
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"//v*vg
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"//v*vr
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//rrrrbbbb
+    "packushb   %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"//ffffgggg
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//gbgbgbgb
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//frfrfrfr
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),
+      [u]"=&f"(u),                         [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [alpha]"f"(-1),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask)
+    : "memory"
+  );
+}
+
+// 10 bit YUV to ARGB
+void I210ToARGBRow_MMI(const uint16_t* src_y,
+                       const uint16_t* src_u,
+                       const uint16_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gsldlc1    %[y],            0x07(%[y_ptr])                   \n\t"
+    "gsldrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "psllh      %[y],            %[y],              %[six]        \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "punpcklhw  %[u],            %[u],              %[u]          \n\t"
+    "psrah      %[u],            %[u],              %[two]        \n\t"
+    "punpcklhw  %[v],            %[v],              %[v]          \n\t"
+    "psrah      %[v],            %[v],              %[two]        \n\t"
+    "pminsh     %[u],            %[u],              %[mask1]      \n\t"
+    "pminsh     %[v],            %[v],              %[mask1]      \n\t"
+
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "packushb   %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x08          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x04          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),
+      [u]"=&f"(u),                         [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [alpha]"f"(-1),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask),                     [two]"f"(0x02),
+      [mask1]"f"(0x00ff00ff00ff00ff)
+    : "memory"
+  );
+}
+
+void I422AlphaToARGBRow_MMI(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            const uint8_t* src_a,
+                            uint8_t* rgb_buf,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  uint64_t y,u,v,a;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+    "gslwlc1    %[a],            0x03(%[a_ptr])                   \n\t"
+    "gslwrc1    %[a],            0x00(%[a_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//rrrrbbbb
+    "packushb   %[g_vec0],       %[g_vec0],         %[a]          \n\t"
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[a]          \n\t"//aaaagggg
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[a_ptr],        %[a_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),                         [a]"=&f"(a),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [a_ptr]"r"(src_a),                   [zero]"f"(0x00),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask)
+    : "memory"
+  );
+}
+
+void I422ToRGB24Row_MMI(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "packushb   %[g_vec0],       %[g_vec0],         %[zero]       \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"
+
+    "punpckhwd  %[r_vec0],       %[g_vec0],         %[g_vec0]     \n\t"
+    "psllw      %[r_vec1],       %[r_vec0],         %[lmove1]     \n\t"
+    "or         %[g_vec0],       %[g_vec0],         %[r_vec1]     \n\t"
+    "psrlw      %[r_vec1],       %[r_vec0],         %[rmove1]     \n\t"
+    "pextrh     %[r_vec1],       %[r_vec1],         %[zero]       \n\t"
+    "pinsrh_2   %[g_vec0],       %[g_vec0],         %[r_vec1]     \n\t"
+    "pextrh     %[r_vec1],       %[g_vec1],         %[zero]       \n\t"
+    "pinsrh_3   %[g_vec0],       %[g_vec0],         %[r_vec1]     \n\t"
+    "pextrh     %[r_vec1],       %[g_vec1],         %[one]        \n\t"
+    "punpckhwd  %[g_vec1],       %[g_vec1],         %[g_vec1]     \n\t"
+    "psllw      %[g_vec1],       %[g_vec1],         %[rmove1]     \n\t"
+    "or         %[g_vec1],       %[g_vec1],         %[r_vec1]     \n\t"
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gsswlc1    %[g_vec1],       0x0b(%[rgbbuf_ptr])              \n\t"
+    "gsswrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x0c          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask]"f"(mask),
+      [lmove1]"f"(0x18),                   [rmove1]"f"(0x8),
+      [one]"f"(0x1)
+    : "memory"
+  );
+}
+
+void I422ToARGB4444Row_MMI(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb4444,
+                           const struct YuvConstants* yuvconstants,
+                           int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "and        %[g_vec],        %[g_vec],          %[mask1]      \n\t"
+    "psrlw      %[g_vec],        %[g_vec],          %[four]       \n\t"
+    "psrlw      %[r_vec],        %[g_vec],          %[four]       \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "punpcklbh  %[r_vec],        %[alpha],          %[zero]       \n\t"
+    "and        %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+
+    "and        %[b_vec],        %[b_vec],          %[mask1]      \n\t"
+    "psrlw      %[b_vec],        %[b_vec],          %[four]       \n\t"
+    "psrlw      %[r_vec],        %[b_vec],          %[four]       \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpcklbh  %[r_vec],        %[alpha],          %[zero]       \n\t"
+    "and        %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_argb4444])            \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_argb4444])            \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[dst_argb4444], %[dst_argb4444],   0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [dst_argb4444]"r"(dst_argb4444),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask]"f"(0xff00ff00ff00ff00),
+      [four]"f"(0x4),                      [mask1]"f"(0xf0f0f0f0f0f0f0f0),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void I422ToARGB1555Row_MMI(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb1555,
+                           const struct YuvConstants* yuvconstants,
+                           int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "psrlw      %[temp],         %[g_vec],          %[three]      \n\t"
+    "and        %[g_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "or         %[g_vec],        %[g_vec],          %[mask3]      \n\t"
+
+    "psrlw      %[temp],         %[b_vec],          %[three]      \n\t"
+    "and        %[b_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "or         %[b_vec],        %[b_vec],          %[mask3]      \n\t"
+
+    "punpcklhw  %[r_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[r_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_argb1555])            \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_argb1555])            \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[dst_argb1555], %[dst_argb1555],   0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [dst_argb1555]"r"(dst_argb1555),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [three]"f"(0x3),                     [mask2]"f"(0x1f0000001f),
+      [eight]"f"(0x8),                     [mask3]"f"(0x800000008000),
+      [lmove5]"f"(0x5)
+    : "memory"
+  );
+}
+
+void I422ToRGB565Row_MMI(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[g_vec],          %[three]      \n\t"
+    "and        %[g_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[b_vec],          %[three]      \n\t"
+    "and        %[b_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpcklhw  %[r_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[r_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_rgb565])             \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_rgb565])             \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[dst_rgb565],   %[dst_rgb565],     0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [dst_rgb565]"r"(dst_rgb565),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [three]"f"(0x3),                     [mask2]"f"(0x1f0000001f),
+      [eight]"f"(0x8),                     [seven]"f"(0x7),
+      [lmove5]"f"(0x5)
+    : "memory"
+  );
+}
+
+void NV12ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[uv_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[uv_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[vshu]       \n\t"
+    "pshufh     %[u],            %[u],              %[ushu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],       0x07(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[g_vec],       0x00(%[rgbbuf_ptr])               \n\t"
+    "gssdlc1    %[b_vec],       0x0f(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[b_vec],       0x08(%[rgbbuf_ptr])               \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[uv_ptr],       %[uv_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [uv_ptr]"r"(src_uv),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void NV21ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[vu_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[vu_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[ushu]       \n\t"
+    "pshufh     %[u],            %[u],              %[vshu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],       0x07(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[g_vec],       0x00(%[rgbbuf_ptr])               \n\t"
+    "gssdlc1    %[b_vec],       0x0f(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[b_vec],       0x08(%[rgbbuf_ptr])               \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[vu_ptr],       %[vu_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [vu_ptr]"r"(src_vu),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void NV12ToRGB24Row_MMI(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[uv_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[uv_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[vshu]       \n\t"
+    "pshufh     %[u],            %[u],              %[ushu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpckhwd  %[r_vec],        %[g_vec],          %[g_vec]      \n\t"
+    "psllw      %[temp],         %[r_vec],          %[lmove1]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrlw      %[temp],         %[r_vec],          %[rmove1]     \n\t"
+    "pextrh     %[temp],         %[temp],           %[zero]       \n\t"
+    "pinsrh_2   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[zero]       \n\t"
+    "pinsrh_3   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[one]        \n\t"
+    "punpckhwd  %[b_vec],        %[b_vec],          %[b_vec]      \n\t"
+    "psllw      %[b_vec],        %[b_vec],          %[rmove1]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gsswlc1    %[b_vec],        0x0b(%[rgbbuf_ptr])              \n\t"
+    "gsswrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[uv_ptr],       %[uv_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x0C          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [uv_ptr]"r"(src_uv),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1),                      [lmove1]"f"(0x18),
+      [one]"f"(0x1),                       [rmove1]"f"(0x8)
+    : "memory"
+  );
+}
+
+void NV21ToRGB24Row_MMI(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[vu_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[vu_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[ushu]       \n\t"
+    "pshufh     %[u],            %[u],              %[vshu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpckhwd  %[r_vec],        %[g_vec],          %[g_vec]      \n\t"
+    "psllw      %[temp],         %[r_vec],          %[lmove1]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrlw      %[temp],         %[r_vec],          %[rmove1]     \n\t"
+    "pextrh     %[temp],         %[temp],           %[zero]       \n\t"
+    "pinsrh_2   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[zero]       \n\t"
+    "pinsrh_3   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[one]        \n\t"
+    "punpckhwd  %[b_vec],        %[b_vec],          %[b_vec]      \n\t"
+    "psllw      %[b_vec],        %[b_vec],          %[rmove1]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gsswlc1    %[b_vec],        0x0b(%[rgbbuf_ptr])              \n\t"
+    "gsswrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[vu_ptr],       %[vu_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x0C          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [vu_ptr]"r"(src_vu),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [lmove1]"f"(0x18),                   [rmove1]"f"(0x8),
+      [one]"f"(0x1)
+    : "memory"
+  );
+}
+
+void NV12ToRGB565Row_MMI(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[uv_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[uv_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[vshu]       \n\t"
+    "pshufh     %[u],            %[u],              %[ushu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[g_vec],          %[three]      \n\t"
+    "and        %[g_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psubb      %[y],            %[eight],          %[three]      \n\t"//5
+    "psllw      %[r_vec],        %[r_vec],          %[y]          \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[b_vec],          %[three]      \n\t"
+    "and        %[b_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psubb      %[y],            %[eight],          %[three]      \n\t"//5
+    "psllw      %[r_vec],        %[r_vec],          %[y]          \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpcklhw  %[r_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[r_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_rgb565])             \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_rgb565])             \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+	"daddiu     %[uv_ptr],       %[uv_ptr],         0x04          \n\t"
+    "daddiu     %[dst_rgb565],   %[dst_rgb565],     0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [uv_ptr]"r"(src_uv),
+      [dst_rgb565]"r"(dst_rgb565),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [three]"f"(0x3),                     [mask2]"f"(0x1f0000001f),
+      [eight]"f"(0x8),                     [seven]"f"(0x7)
+    : "memory"
+  );
+}
+
+void YUY2ToARGBRow_MMI(const uint8_t* src_yuy2,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gsldlc1    %[y],            0x07(%[yuy2_ptr])                \n\t"
+    "gsldrc1    %[y],            0x00(%[yuy2_ptr])                \n\t"
+    "psrlh      %[temp],         %[y],              %[eight]      \n\t"
+    "pshufh     %[u],            %[temp],           %[ushu]       \n\t"
+    "pshufh     %[v],            %[temp],           %[vshu]       \n\t"
+
+    "psrlh      %[temp],         %[mask1],          %[eight]      \n\t"
+    "and        %[y],            %[y],              %[temp]       \n\t"
+    "psllh      %[temp],         %[y],              %[eight]      \n\t"
+    "or         %[y],            %[y],              %[temp]       \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[b_vec],        0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[yuy2_ptr],     %[yuy2_ptr],       0x08          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [yuy2_ptr]"r"(src_yuy2),             [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1),                      [eight]"f"(0x8)
+    : "memory"
+  );
+}
+
+void UYVYToARGBRow_MMI(const uint8_t* src_uyvy,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gsldlc1    %[y],            0x07(%[uyvy_ptr])                \n\t"
+    "gsldrc1    %[y],            0x00(%[uyvy_ptr])                \n\t"
+    "psrlh      %[temp],         %[mask1],          %[eight]      \n\t"
+    "and        %[temp],         %[y],              %[temp]       \n\t"
+    "pshufh     %[u],            %[temp],           %[ushu]       \n\t"
+    "pshufh     %[v],            %[temp],           %[vshu]       \n\t"
+
+    "psrlh      %[y],            %[y],              %[eight]      \n\t"
+    "psllh      %[temp],         %[y],              %[eight]      \n\t"
+    "or         %[y],            %[y],              %[temp]       \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[b_vec],        0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[uyvy_ptr],     %[uyvy_ptr],       0x08          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [uyvy_ptr]"r"(src_uyvy),             [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1),                      [eight]"f"(0x8)
+    : "memory"
+  );
+}
+
+void I422ToRGBARow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[alpha],          %[g_vec]      \n\t"
+    "punpcklbh  %[b_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],       0x07(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[g_vec],       0x00(%[rgbbuf_ptr])               \n\t"
+    "gssdlc1    %[b_vec],       0x0f(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[b_vec],       0x08(%[rgbbuf_ptr])               \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void ARGBSetRow_MMI(uint8_t* dst_argb, uint32_t v32, int width) {
+  __asm__ volatile (
+    "punpcklwd  %[v32],          %[v32],            %[v32]        \n\t"
+    "1:                                                           \n\t"
+    "gssdlc1    %[v32],          0x07(%[dst_ptr])                 \n\t"
+    "gssdrc1    %[v32],          0x00(%[dst_ptr])                 \n\t"
+    "gssdlc1    %[v32],          0x0f(%[dst_ptr])                 \n\t"
+    "gssdrc1    %[v32],          0x08(%[dst_ptr])                 \n\t"
+
+    "daddi      %[width],        %[width],         -0x04          \n\t"
+    "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+    "bnez       %[width],        1b                               \n\t"
+    : [v32]"+&f"(v32)
+    : [dst_ptr]"r"(dst_argb),           [width]"r"(width)
+    : "memory"
+  );
+}
+// clang-format on
+
+// 10 bit YUV to ARGB
+#endif  // !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/row_msa.cc b/media/libyuv/source/row_msa.cc
new file mode 100644
index 0000000000..b7d5bb5ecf
--- /dev/null
+++ b/media/libyuv/source/row_msa.cc
@@ -0,0 +1,3597 @@
+/*
+ *  Copyright 2016 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string.h>
+
+#include "libyuv/row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ALPHA_VAL (-1)
+
+// Fill YUV -> RGB conversion constants into vectors
+#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, yg, yb) \
+  {                                                      \
+    ub = __msa_fill_w(yuvconst->kUVToB[0]);              \
+    vr = __msa_fill_w(yuvconst->kUVToR[1]);              \
+    ug = __msa_fill_w(yuvconst->kUVToG[0]);              \
+    vg = __msa_fill_w(yuvconst->kUVToG[1]);              \
+    yg = __msa_fill_w(yuvconst->kYToRgb[0]);             \
+    yb = __msa_fill_w(yuvconst->kYBiasToRgb[0]);         \
+  }
+
+// Load YUV 422 pixel data
+#define READYUV422(psrc_y, psrc_u, psrc_v, out_y, out_u, out_v)    \
+  {                                                                \
+    uint64_t y_m;                                                  \
+    uint32_t u_m, v_m;                                             \
+    v4i32 zero_m = {0};                                            \
+    y_m = LD(psrc_y);                                              \
+    u_m = LW(psrc_u);                                              \
+    v_m = LW(psrc_v);                                              \
+    out_y = (v16u8)__msa_insert_d((v2i64)zero_m, 0, (int64_t)y_m); \
+    out_u = (v16u8)__msa_insert_w(zero_m, 0, (int32_t)u_m);        \
+    out_v = (v16u8)__msa_insert_w(zero_m, 0, (int32_t)v_m);        \
+  }
+
+// Clip input vector elements between 0 to 255
+#define CLIP_0TO255(in0, in1, in2, in3, in4, in5) \
+  {                                               \
+    v4i32 max_m = __msa_ldi_w(0xFF);              \
+                                                  \
+    in0 = __msa_maxi_s_w(in0, 0);                 \
+    in1 = __msa_maxi_s_w(in1, 0);                 \
+    in2 = __msa_maxi_s_w(in2, 0);                 \
+    in3 = __msa_maxi_s_w(in3, 0);                 \
+    in4 = __msa_maxi_s_w(in4, 0);                 \
+    in5 = __msa_maxi_s_w(in5, 0);                 \
+    in0 = __msa_min_s_w(max_m, in0);              \
+    in1 = __msa_min_s_w(max_m, in1);              \
+    in2 = __msa_min_s_w(max_m, in2);              \
+    in3 = __msa_min_s_w(max_m, in3);              \
+    in4 = __msa_min_s_w(max_m, in4);              \
+    in5 = __msa_min_s_w(max_m, in5);              \
+  }
+
+// Convert 8 pixels of YUV 420 to RGB.
+#define YUVTORGB(in_y, in_uv, ubvr, ugvg, yg, yb, out_b, out_g, out_r) \
+  {                                                                    \
+    v8i16 vec0_m, vec1_m;                                              \
+    v4i32 reg0_m, reg1_m, reg2_m, reg3_m, reg4_m;                      \
+    v4i32 reg5_m, reg6_m, reg7_m;                                      \
+    v16i8 temp_m, zero_m = {0};                                        \
+                                                                       \
+    vec0_m = (v8i16)__msa_ilvr_b((v16i8)in_y, (v16i8)in_y);            \
+    vec1_m = (v8i16)__msa_ilvr_b((v16i8)zero_m, (v16i8)in_uv);         \
+    reg0_m = (v4i32)__msa_ilvr_h((v8i16)zero_m, (v8i16)vec0_m);        \
+    reg1_m = (v4i32)__msa_ilvl_h((v8i16)zero_m, (v8i16)vec0_m);        \
+    vec1_m = (v8i16)__msa_subv_h(vec1_m, const_0x80);                  \
+    temp_m = (v16i8)__msa_clti_s_h(vec1_m, 0);                         \
+    reg2_m = (v4i32)__msa_ilvr_h((v8i16)temp_m, (v8i16)vec1_m);        \
+    reg3_m = (v4i32)__msa_ilvl_h((v8i16)temp_m, (v8i16)vec1_m);        \
+    reg0_m *= yg;                                                      \
+    reg1_m *= yg;                                                      \
+    reg2_m *= ubvr;                                                    \
+    reg3_m *= ubvr;                                                    \
+    reg0_m = __msa_srai_w(reg0_m, 16);                                 \
+    reg1_m = __msa_srai_w(reg1_m, 16);                                 \
+    reg0_m += yb;                                                      \
+    reg1_m += yb;                                                      \
+    reg4_m = __msa_dotp_s_w((v8i16)vec1_m, (v8i16)ugvg);               \
+    reg5_m = __msa_ilvev_w(reg2_m, reg2_m);                            \
+    reg6_m = __msa_ilvev_w(reg3_m, reg3_m);                            \
+    reg7_m = __msa_ilvr_w(reg4_m, reg4_m);                             \
+    reg2_m = __msa_ilvod_w(reg2_m, reg2_m);                            \
+    reg3_m = __msa_ilvod_w(reg3_m, reg3_m);                            \
+    reg4_m = __msa_ilvl_w(reg4_m, reg4_m);                             \
+    reg5_m = reg0_m + reg5_m;                                          \
+    reg6_m = reg1_m + reg6_m;                                          \
+    reg2_m = reg0_m + reg2_m;                                          \
+    reg3_m = reg1_m + reg3_m;                                          \
+    reg7_m = reg0_m - reg7_m;                                          \
+    reg4_m = reg1_m - reg4_m;                                          \
+    reg5_m = __msa_srai_w(reg5_m, 6);                                  \
+    reg6_m = __msa_srai_w(reg6_m, 6);                                  \
+    reg7_m = __msa_srai_w(reg7_m, 6);                                  \
+    reg4_m = __msa_srai_w(reg4_m, 6);                                  \
+    reg2_m = __msa_srai_w(reg2_m, 6);                                  \
+    reg3_m = __msa_srai_w(reg3_m, 6);                                  \
+    CLIP_0TO255(reg5_m, reg6_m, reg7_m, reg4_m, reg2_m, reg3_m);       \
+    out_b = __msa_pckev_h((v8i16)reg6_m, (v8i16)reg5_m);               \
+    out_g = __msa_pckev_h((v8i16)reg4_m, (v8i16)reg7_m);               \
+    out_r = __msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);               \
+  }
+
+// Pack and Store 8 ARGB values.
+#define STOREARGB(in0, in1, in2, in3, pdst_argb)           \
+  {                                                        \
+    v8i16 vec0_m, vec1_m;                                  \
+    v16u8 dst0_m, dst1_m;                                  \
+    vec0_m = (v8i16)__msa_ilvev_b((v16i8)in1, (v16i8)in0); \
+    vec1_m = (v8i16)__msa_ilvev_b((v16i8)in3, (v16i8)in2); \
+    dst0_m = (v16u8)__msa_ilvr_h(vec1_m, vec0_m);          \
+    dst1_m = (v16u8)__msa_ilvl_h(vec1_m, vec0_m);          \
+    ST_UB2(dst0_m, dst1_m, pdst_argb, 16);                 \
+  }
+
+// Takes ARGB input and calculates Y.
+#define ARGBTOY(argb0, argb1, argb2, argb3, const0, const1, const2, shift, \
+                y_out)                                                     \
+  {                                                                        \
+    v16u8 vec0_m, vec1_m, vec2_m, vec3_m;                                  \
+    v8u16 reg0_m, reg1_m;                                                  \
+                                                                           \
+    vec0_m = (v16u8)__msa_pckev_h((v8i16)argb1, (v8i16)argb0);             \
+    vec1_m = (v16u8)__msa_pckev_h((v8i16)argb3, (v8i16)argb2);             \
+    vec2_m = (v16u8)__msa_pckod_h((v8i16)argb1, (v8i16)argb0);             \
+    vec3_m = (v16u8)__msa_pckod_h((v8i16)argb3, (v8i16)argb2);             \
+    reg0_m = __msa_dotp_u_h(vec0_m, const0);                               \
+    reg1_m = __msa_dotp_u_h(vec1_m, const0);                               \
+    reg0_m = __msa_dpadd_u_h(reg0_m, vec2_m, const1);                      \
+    reg1_m = __msa_dpadd_u_h(reg1_m, vec3_m, const1);                      \
+    reg0_m += const2;                                                      \
+    reg1_m += const2;                                                      \
+    reg0_m = (v8u16)__msa_srai_h((v8i16)reg0_m, shift);                    \
+    reg1_m = (v8u16)__msa_srai_h((v8i16)reg1_m, shift);                    \
+    y_out = (v16u8)__msa_pckev_b((v16i8)reg1_m, (v16i8)reg0_m);            \
+  }
+
+// Loads current and next row of ARGB input and averages it to calculate U and V
+#define READ_ARGB(s_ptr, t_ptr, argb0, argb1, argb2, argb3, const_0x0101) \
+  {                                                                       \
+    v16u8 src0_m, src1_m, src2_m, src3_m, src4_m, src5_m, src6_m, src7_m; \
+    v16u8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m; \
+    v8u16 reg0_m, reg1_m, reg2_m, reg3_m, reg4_m, reg5_m, reg6_m, reg7_m; \
+    v8u16 reg8_m, reg9_m;                                                 \
+                                                                          \
+    src0_m = (v16u8)__msa_ld_b((void*)s, 0);                              \
+    src1_m = (v16u8)__msa_ld_b((void*)s, 16);                             \
+    src2_m = (v16u8)__msa_ld_b((void*)s, 32);                             \
+    src3_m = (v16u8)__msa_ld_b((void*)s, 48);                             \
+    src4_m = (v16u8)__msa_ld_b((void*)t, 0);                              \
+    src5_m = (v16u8)__msa_ld_b((void*)t, 16);                             \
+    src6_m = (v16u8)__msa_ld_b((void*)t, 32);                             \
+    src7_m = (v16u8)__msa_ld_b((void*)t, 48);                             \
+    vec0_m = (v16u8)__msa_ilvr_b((v16i8)src0_m, (v16i8)src4_m);           \
+    vec1_m = (v16u8)__msa_ilvr_b((v16i8)src1_m, (v16i8)src5_m);           \
+    vec2_m = (v16u8)__msa_ilvr_b((v16i8)src2_m, (v16i8)src6_m);           \
+    vec3_m = (v16u8)__msa_ilvr_b((v16i8)src3_m, (v16i8)src7_m);           \
+    vec4_m = (v16u8)__msa_ilvl_b((v16i8)src0_m, (v16i8)src4_m);           \
+    vec5_m = (v16u8)__msa_ilvl_b((v16i8)src1_m, (v16i8)src5_m);           \
+    vec6_m = (v16u8)__msa_ilvl_b((v16i8)src2_m, (v16i8)src6_m);           \
+    vec7_m = (v16u8)__msa_ilvl_b((v16i8)src3_m, (v16i8)src7_m);           \
+    reg0_m = __msa_hadd_u_h(vec0_m, vec0_m);                              \
+    reg1_m = __msa_hadd_u_h(vec1_m, vec1_m);                              \
+    reg2_m = __msa_hadd_u_h(vec2_m, vec2_m);                              \
+    reg3_m = __msa_hadd_u_h(vec3_m, vec3_m);                              \
+    reg4_m = __msa_hadd_u_h(vec4_m, vec4_m);                              \
+    reg5_m = __msa_hadd_u_h(vec5_m, vec5_m);                              \
+    reg6_m = __msa_hadd_u_h(vec6_m, vec6_m);                              \
+    reg7_m = __msa_hadd_u_h(vec7_m, vec7_m);                              \
+    reg8_m = (v8u16)__msa_pckev_d((v2i64)reg4_m, (v2i64)reg0_m);          \
+    reg9_m = (v8u16)__msa_pckev_d((v2i64)reg5_m, (v2i64)reg1_m);          \
+    reg8_m += (v8u16)__msa_pckod_d((v2i64)reg4_m, (v2i64)reg0_m);         \
+    reg9_m += (v8u16)__msa_pckod_d((v2i64)reg5_m, (v2i64)reg1_m);         \
+    reg0_m = (v8u16)__msa_pckev_d((v2i64)reg6_m, (v2i64)reg2_m);          \
+    reg1_m = (v8u16)__msa_pckev_d((v2i64)reg7_m, (v2i64)reg3_m);          \
+    reg0_m += (v8u16)__msa_pckod_d((v2i64)reg6_m, (v2i64)reg2_m);         \
+    reg1_m += (v8u16)__msa_pckod_d((v2i64)reg7_m, (v2i64)reg3_m);         \
+    reg8_m += const_0x0101;                                               \
+    reg9_m += const_0x0101;                                               \
+    reg0_m += const_0x0101;                                               \
+    reg1_m += const_0x0101;                                               \
+    argb0 = (v8u16)__msa_srai_h((v8i16)reg8_m, 1);                        \
+    argb1 = (v8u16)__msa_srai_h((v8i16)reg9_m, 1);                        \
+    argb2 = (v8u16)__msa_srai_h((v8i16)reg0_m, 1);                        \
+    argb3 = (v8u16)__msa_srai_h((v8i16)reg1_m, 1);                        \
+  }
+
+#define ARGBTOUV(argb0, argb1, argb2, argb3, const0, const1, const2, const3, \
+                 shf0, shf1, shf2, shf3, shift, u_out, v_out)                \
+  {                                                                          \
+    v8u16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;    \
+    v4u32 reg0_m, reg1_m, reg2_m, reg3_m;                                    \
+                                                                             \
+    vec0_m = (v8u16)__msa_vshf_h(shf0, (v16i8)argb1, (v16i8)argb0);          \
+    vec1_m = (v8u16)__msa_vshf_h(shf0, (v16i8)argb3, (v16i8)argb2);          \
+    vec2_m = (v8u16)__msa_vshf_h(shf1, (v16i8)argb1, (v16i8)argb0);          \
+    vec3_m = (v8u16)__msa_vshf_h(shf1, (v16i8)argb3, (v16i8)argb2);          \
+    vec4_m = (v8u16)__msa_vshf_h(shf2, (v16i8)argb1, (v16i8)argb0);          \
+    vec5_m = (v8u16)__msa_vshf_h(shf2, (v16i8)argb3, (v16i8)argb2);          \
+    vec6_m = (v8u16)__msa_vshf_h(shf3, (v16i8)argb1, (v16i8)argb0);          \
+    vec7_m = (v8u16)__msa_vshf_h(shf3, (v16i8)argb3, (v16i8)argb2);          \
+    reg0_m = __msa_dotp_u_w(vec0_m, const0);                                 \
+    reg1_m = __msa_dotp_u_w(vec1_m, const0);                                 \
+    reg2_m = __msa_dotp_u_w(vec4_m, const0);                                 \
+    reg3_m = __msa_dotp_u_w(vec5_m, const0);                                 \
+    reg0_m += const1;                                                        \
+    reg1_m += const1;                                                        \
+    reg2_m += const1;                                                        \
+    reg3_m += const1;                                                        \
+    reg0_m -= (v4u32)__msa_dotp_u_w(vec2_m, const2);                         \
+    reg1_m -= (v4u32)__msa_dotp_u_w(vec3_m, const2);                         \
+    reg2_m -= (v4u32)__msa_dotp_u_w(vec6_m, const3);                         \
+    reg3_m -= (v4u32)__msa_dotp_u_w(vec7_m, const3);                         \
+    reg0_m = __msa_srl_w(reg0_m, shift);                                     \
+    reg1_m = __msa_srl_w(reg1_m, shift);                                     \
+    reg2_m = __msa_srl_w(reg2_m, shift);                                     \
+    reg3_m = __msa_srl_w(reg3_m, shift);                                     \
+    u_out = (v8u16)__msa_pckev_h((v8i16)reg1_m, (v8i16)reg0_m);              \
+    v_out = (v8u16)__msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);              \
+  }
+
+// Takes ARGB input and calculates U and V.
+#define ARGBTOUV_H(argb0, argb1, argb2, argb3, const0, const1, const2, const3, \
+                   shf0, shf1, shf2, shf3, v_out, u_out)                       \
+  {                                                                            \
+    v8u16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;      \
+    v4u32 reg0_m, reg1_m, reg2_m, reg3_m;                                      \
+                                                                               \
+    vec0_m = __msa_vshf_h(shf0, (v16i8)argb1, (v16i8)argb0);                   \
+    vec1_m = __msa_vshf_h(shf0, (v16i8)argb3, (v16i8)argb2);                   \
+    vec2_m = __msa_vshf_h(shf1, (v16i8)argb1, (v16i8)argb0);                   \
+    vec3_m = __msa_vshf_h(shf1, (v16i8)argb3, (v16i8)argb2);                   \
+    vec4_m = __msa_vshf_h(shf2, (v16i8)argb1, (v16i8)argb0);                   \
+    vec5_m = __msa_vshf_h(shf2, (v16i8)argb3, (v16i8)argb2);                   \
+    vec6_m = __msa_vshf_h(shf3, (v16i8)argb1, (v16i8)argb0);                   \
+    vec7_m = __msa_vshf_h(shf3, (v16i8)argb3, (v16i8)argb2);                   \
+    reg0_m = __msa_dotp_u_w(vec0_m, const1);                                   \
+    reg1_m = __msa_dotp_u_w(vec1_m, const1);                                   \
+    reg2_m = __msa_dotp_u_w(vec4_m, const1);                                   \
+    reg3_m = __msa_dotp_u_w(vec5_m, const1);                                   \
+    reg0_m += (v4u32)const3;                                                   \
+    reg1_m += (v4u32)const3;                                                   \
+    reg2_m += (v4u32)const3;                                                   \
+    reg3_m += (v4u32)const3;                                                   \
+    reg0_m -= __msa_dotp_u_w(vec2_m, const0);                                  \
+    reg1_m -= __msa_dotp_u_w(vec3_m, const0);                                  \
+    reg2_m -= __msa_dotp_u_w(vec6_m, const2);                                  \
+    reg3_m -= __msa_dotp_u_w(vec7_m, const2);                                  \
+    u_out = (v16u8)__msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);                \
+    v_out = (v16u8)__msa_pckev_h((v8i16)reg1_m, (v8i16)reg0_m);                \
+    u_out = (v16u8)__msa_pckod_b((v16i8)u_out, (v16i8)u_out);                  \
+    v_out = (v16u8)__msa_pckod_b((v16i8)v_out, (v16i8)v_out);                  \
+  }
+
+// Load I444 pixel data
+#define READI444(psrc_y, psrc_u, psrc_v, out_y, out_u, out_v) \
+  {                                                           \
+    uint64_t y_m, u_m, v_m;                                   \
+    v2i64 zero_m = {0};                                       \
+    y_m = LD(psrc_y);                                         \
+    u_m = LD(psrc_u);                                         \
+    v_m = LD(psrc_v);                                         \
+    out_y = (v16u8)__msa_insert_d(zero_m, 0, (int64_t)y_m);   \
+    out_u = (v16u8)__msa_insert_d(zero_m, 0, (int64_t)u_m);   \
+    out_v = (v16u8)__msa_insert_d(zero_m, 0, (int64_t)v_m);   \
+  }
+
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _dst0) \
+  {                                                              \
+    v16u8 _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5;              \
+    v8i16 _reg0, _reg1, _reg2, _reg3, _reg4, _reg5;              \
+    _tmp0 = (v16u8)__msa_ilvev_b(_tmpb, _nexb);                  \
+    _tmp1 = (v16u8)__msa_ilvod_b(_tmpb, _nexb);                  \
+    _tmp2 = (v16u8)__msa_ilvev_b(_tmpg, _nexg);                  \
+    _tmp3 = (v16u8)__msa_ilvod_b(_tmpg, _nexg);                  \
+    _tmp4 = (v16u8)__msa_ilvev_b(_tmpr, _nexr);                  \
+    _tmp5 = (v16u8)__msa_ilvod_b(_tmpr, _nexr);                  \
+    _reg0 = (v8i16)__msa_hadd_u_h(_tmp0, _tmp0);                 \
+    _reg1 = (v8i16)__msa_hadd_u_h(_tmp1, _tmp1);                 \
+    _reg2 = (v8i16)__msa_hadd_u_h(_tmp2, _tmp2);                 \
+    _reg3 = (v8i16)__msa_hadd_u_h(_tmp3, _tmp3);                 \
+    _reg4 = (v8i16)__msa_hadd_u_h(_tmp4, _tmp4);                 \
+    _reg5 = (v8i16)__msa_hadd_u_h(_tmp5, _tmp5);                 \
+    _reg0 = (v8i16)__msa_aver_u_h(_reg0, _reg1);                 \
+    _reg2 = (v8i16)__msa_aver_u_h(_reg2, _reg3);                 \
+    _reg4 = (v8i16)__msa_aver_u_h(_reg4, _reg5);                 \
+    _reg1 = const_8080 + const_112 * _reg0;                      \
+    _reg3 = const_8080 + const_112 * _reg4;                      \
+    _reg1 = (v8i16)__msa_msubv_h(_reg1, const_74, _reg2);        \
+    _reg3 = (v8i16)__msa_msubv_h(_reg3, const_94, _reg2);        \
+    _reg1 = (v8i16)__msa_msubv_h(_reg1, const_38, _reg4);        \
+    _reg3 = (v8i16)__msa_msubv_h(_reg3, const_18, _reg0);        \
+    _dst0 = (v16u8)__msa_pckod_b(_reg3, _reg1);                  \
+  }
+
+void MirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16i8 shuffler = {15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0};
+  src += width - 64;
+
+  for (x = 0; x < width; x += 64) {
+    LD_UB4(src, 16, src3, src2, src1, src0);
+    VSHF_B2_UB(src3, src3, src2, src2, shuffler, shuffler, dst3, dst2);
+    VSHF_B2_UB(src1, src1, src0, src0, shuffler, shuffler, dst1, dst0);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, 16);
+    dst += 64;
+    src -= 64;
+  }
+}
+
+void MirrorUVRow_MSA(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  int x;
+  v8u16 src, dst;
+  v8u16 shuffler = {7, 6, 5, 4, 3, 2, 1, 0};
+  src_uv += (width - 8) << 1;
+  for (x = 0; x < width; x += 8) {
+    src = LD_UH(src_uv);
+    dst = __msa_vshf_h(shuffler, src, src);
+    ST_UH(dst, dst_uv);
+    src_uv -= 16;
+    dst_uv += 16;
+  }
+}
+
+void ARGBMirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16i8 shuffler = {12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3};
+  src += width * 4 - 64;
+
+  for (x = 0; x < width; x += 16) {
+    LD_UB4(src, 16, src3, src2, src1, src0);
+    VSHF_B2_UB(src3, src3, src2, src2, shuffler, shuffler, dst3, dst2);
+    VSHF_B2_UB(src1, src1, src0, src0, shuffler, shuffler, dst1, dst0);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, 16);
+    dst += 64;
+    src -= 64;
+  }
+}
+
+void I422ToYUY2Row_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  int x;
+  v16u8 src_u0, src_v0, src_y0, src_y1, vec_uv0, vec_uv1;
+  v16u8 dst_yuy2_0, dst_yuy2_1, dst_yuy2_2, dst_yuy2_3;
+
+  for (x = 0; x < width; x += 32) {
+    src_u0 = LD_UB(src_u);
+    src_v0 = LD_UB(src_v);
+    LD_UB2(src_y, 16, src_y0, src_y1);
+    ILVRL_B2_UB(src_v0, src_u0, vec_uv0, vec_uv1);
+    ILVRL_B2_UB(vec_uv0, src_y0, dst_yuy2_0, dst_yuy2_1);
+    ILVRL_B2_UB(vec_uv1, src_y1, dst_yuy2_2, dst_yuy2_3);
+    ST_UB4(dst_yuy2_0, dst_yuy2_1, dst_yuy2_2, dst_yuy2_3, dst_yuy2, 16);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_yuy2 += 64;
+  }
+}
+
+void I422ToUYVYRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_uyvy,
+                       int width) {
+  int x;
+  v16u8 src_u0, src_v0, src_y0, src_y1, vec_uv0, vec_uv1;
+  v16u8 dst_uyvy0, dst_uyvy1, dst_uyvy2, dst_uyvy3;
+
+  for (x = 0; x < width; x += 32) {
+    src_u0 = LD_UB(src_u);
+    src_v0 = LD_UB(src_v);
+    LD_UB2(src_y, 16, src_y0, src_y1);
+    ILVRL_B2_UB(src_v0, src_u0, vec_uv0, vec_uv1);
+    ILVRL_B2_UB(src_y0, vec_uv0, dst_uyvy0, dst_uyvy1);
+    ILVRL_B2_UB(src_y1, vec_uv1, dst_uyvy2, dst_uyvy3);
+    ST_UB4(dst_uyvy0, dst_uyvy1, dst_uyvy2, dst_uyvy3, dst_uyvy, 16);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_uyvy += 64;
+  }
+}
+
+void I422ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  v16u8 src0, src1, src2;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    READYUV422(src_y, src_u, src_v, src0, src1, src2);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    STOREARGB(vec0, vec1, vec2, alpha, dst_argb);
+    src_y += 8;
+    src_u += 4;
+    src_v += 4;
+    dst_argb += 32;
+  }
+}
+
+void I422ToRGBARow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  v16u8 src0, src1, src2;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    READYUV422(src_y, src_u, src_v, src0, src1, src2);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    STOREARGB(alpha, vec0, vec1, vec2, dst_argb);
+    src_y += 8;
+    src_u += 4;
+    src_v += 4;
+    dst_argb += 32;
+  }
+}
+
+void I422AlphaToARGBRow_MSA(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            const uint8_t* src_a,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  int x;
+  int64_t data_a;
+  v16u8 src0, src1, src2, src3;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v4i32 zero = {0};
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    data_a = LD(src_a);
+    READYUV422(src_y, src_u, src_v, src0, src1, src2);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    src3 = (v16u8)__msa_insert_d((v2i64)zero, 0, data_a);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    src3 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src3);
+    STOREARGB(vec0, vec1, vec2, src3, dst_argb);
+    src_y += 8;
+    src_u += 4;
+    src_v += 4;
+    src_a += 8;
+    dst_argb += 32;
+  }
+}
+
+void I422ToRGB24Row_MSA(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int32_t width) {
+  int x;
+  int64_t data_u, data_v;
+  v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v16u8 reg0, reg1, reg2, reg3;
+  v2i64 zero = {0};
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+  v16i8 shuffler0 = {0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, 8, 9, 20, 10};
+  v16i8 shuffler1 = {0, 21, 1, 2, 22, 3, 4, 23, 5, 6, 24, 7, 8, 25, 9, 10};
+  v16i8 shuffler2 = {26, 6,  7,  27, 8,  9,  28, 10,
+                     11, 29, 12, 13, 30, 14, 15, 31};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_y, 0);
+    data_u = LD(src_u);
+    data_v = LD(src_v);
+    src1 = (v16u8)__msa_insert_d(zero, 0, data_u);
+    src2 = (v16u8)__msa_insert_d(zero, 0, data_v);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    src3 = (v16u8)__msa_sldi_b((v16i8)src0, (v16i8)src0, 8);
+    src4 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src1, 8);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    YUVTORGB(src3, src4, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec3, vec4, vec5);
+    reg0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
+    reg2 = (v16u8)__msa_ilvev_b((v16i8)vec4, (v16i8)vec3);
+    reg3 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec2);
+    reg1 = (v16u8)__msa_sldi_b((v16i8)reg2, (v16i8)reg0, 11);
+    dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)reg3, (v16i8)reg0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)reg3, (v16i8)reg1);
+    dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)reg3, (v16i8)reg2);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    ST_UB(dst2, (dst_argb + 32));
+    src_y += 16;
+    src_u += 8;
+    src_v += 8;
+    dst_argb += 48;
+  }
+}
+
+// TODO(fbarchard): Consider AND instead of shift to isolate 5 upper bits of R.
+void I422ToRGB565Row_MSA(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  int x;
+  v16u8 src0, src1, src2, dst0;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    READYUV422(src_y, src_u, src_v, src0, src1, src2);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    vec0 = __msa_srli_h(vec0, 3);
+    vec1 = __msa_srli_h(vec1, 2);
+    vec2 = __msa_srli_h(vec2, 3);
+    vec2 = __msa_slli_h(vec2, 11);
+    vec1 = __msa_slli_h(vec1, 5);
+    vec0 |= vec1;
+    dst0 = (v16u8)(vec2 | vec0);
+    ST_UB(dst0, dst_rgb565);
+    src_y += 8;
+    src_u += 4;
+    src_v += 4;
+    dst_rgb565 += 16;
+  }
+}
+
+// TODO(fbarchard): Consider AND instead of shift to isolate 4 upper bits of G.
+void I422ToARGB4444Row_MSA(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb4444,
+                           const struct YuvConstants* yuvconstants,
+                           int width) {
+  int x;
+  v16u8 src0, src1, src2, dst0;
+  v8i16 vec0, vec1, vec2;
+  v8u16 reg0, reg1, reg2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v8u16 const_0xF000 = (v8u16)__msa_fill_h(0xF000);
+  v8u16 mask = (v8u16)__msa_fill_h(0x00F0);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    READYUV422(src_y, src_u, src_v, src0, src1, src2);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    reg0 = (v8u16)__msa_srli_h(vec0, 4);
+    reg2 = (v8u16)__msa_srli_h(vec2, 4);
+    reg1 = (v8u16)__msa_and_v(vec1, mask);
+    reg2 = (v8u16)__msa_slli_h(reg2, 8);
+    reg1 |= const_0xF000;
+    reg0 |= reg2;
+    dst0 = (v16u8)(reg1 | reg0);
+    ST_UB(dst0, dst_argb4444);
+    src_y += 8;
+    src_u += 4;
+    src_v += 4;
+    dst_argb4444 += 16;
+  }
+}
+
+void I422ToARGB1555Row_MSA(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb1555,
+                           const struct YuvConstants* yuvconstants,
+                           int width) {
+  int x;
+  v16u8 src0, src1, src2, dst0;
+  v8i16 vec0, vec1, vec2;
+  v8u16 reg0, reg1, reg2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v8u16 const_0x8000 = (v8u16)__msa_fill_h(0x8000);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    READYUV422(src_y, src_u, src_v, src0, src1, src2);
+    src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    reg0 = (v8u16)__msa_srli_h(vec0, 3);
+    reg1 = (v8u16)__msa_srli_h(vec1, 3);
+    reg2 = (v8u16)__msa_srli_h(vec2, 3);
+    reg1 = (v8u16)__msa_slli_h((v8i16)reg1, 5);
+    reg2 = (v8u16)__msa_slli_h((v8i16)reg2, 10);
+    reg1 |= const_0x8000;
+    reg0 |= reg2;
+    dst0 = (v16u8)(reg1 | reg0);
+    ST_UB(dst0, dst_argb1555);
+    src_y += 8;
+    src_u += 4;
+    src_v += 4;
+    dst_argb1555 += 16;
+  }
+}
+
+void YUY2ToYRow_MSA(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    LD_UB4(src_yuy2, 16, src0, src1, src2, src3);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst_y, 16);
+    src_yuy2 += 64;
+    dst_y += 32;
+  }
+}
+
+void YUY2ToUVRow_MSA(const uint8_t* src_yuy2,
+                     int src_stride_yuy2,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  const uint8_t* src_yuy2_next = src_yuy2 + src_stride_yuy2;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 vec0, vec1, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    LD_UB4(src_yuy2, 16, src0, src1, src2, src3);
+    LD_UB4(src_yuy2_next, 16, src4, src5, src6, src7);
+    src0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    src1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    src2 = (v16u8)__msa_pckod_b((v16i8)src5, (v16i8)src4);
+    src3 = (v16u8)__msa_pckod_b((v16i8)src7, (v16i8)src6);
+    vec0 = __msa_aver_u_b(src0, src2);
+    vec1 = __msa_aver_u_b(src1, src3);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_yuy2 += 64;
+    src_yuy2_next += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void YUY2ToUV422Row_MSA(const uint8_t* src_yuy2,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    LD_UB4(src_yuy2, 16, src0, src1, src2, src3);
+    src0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    src1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_yuy2 += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void UYVYToYRow_MSA(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    LD_UB4(src_uyvy, 16, src0, src1, src2, src3);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst_y, 16);
+    src_uyvy += 64;
+    dst_y += 32;
+  }
+}
+
+void UYVYToUVRow_MSA(const uint8_t* src_uyvy,
+                     int src_stride_uyvy,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  const uint8_t* src_uyvy_next = src_uyvy + src_stride_uyvy;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 vec0, vec1, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    LD_UB4(src_uyvy, 16, src0, src1, src2, src3);
+    LD_UB4(src_uyvy_next, 16, src4, src5, src6, src7);
+    src0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    src1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    src2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4);
+    src3 = (v16u8)__msa_pckev_b((v16i8)src7, (v16i8)src6);
+    vec0 = __msa_aver_u_b(src0, src2);
+    vec1 = __msa_aver_u_b(src1, src3);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_uyvy += 64;
+    src_uyvy_next += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void UYVYToUV422Row_MSA(const uint8_t* src_uyvy,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    LD_UB4(src_uyvy, 16, src0, src1, src2, src3);
+    src0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    src1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_uyvy += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0;
+  v8u16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v16i8 zero = {0};
+  v8u16 const_0x19 = (v8u16)__msa_ldi_h(0x19);
+  v8u16 const_0x81 = (v8u16)__msa_ldi_h(0x81);
+  v8u16 const_0x42 = (v8u16)__msa_ldi_h(0x42);
+  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    vec2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    reg0 = (v8u16)__msa_ilvev_b(zero, (v16i8)vec0);
+    reg1 = (v8u16)__msa_ilvev_b(zero, (v16i8)vec1);
+    reg2 = (v8u16)__msa_ilvev_b(zero, (v16i8)vec2);
+    reg3 = (v8u16)__msa_ilvev_b(zero, (v16i8)vec3);
+    reg4 = (v8u16)__msa_ilvod_b(zero, (v16i8)vec0);
+    reg5 = (v8u16)__msa_ilvod_b(zero, (v16i8)vec1);
+    reg0 *= const_0x19;
+    reg1 *= const_0x19;
+    reg2 *= const_0x81;
+    reg3 *= const_0x81;
+    reg4 *= const_0x42;
+    reg5 *= const_0x42;
+    reg0 += reg2;
+    reg1 += reg3;
+    reg0 += reg4;
+    reg1 += reg5;
+    reg0 += const_0x1080;
+    reg1 += const_0x1080;
+    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 8);
+    reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    ST_UB(dst0, dst_y);
+    src_argb += 64;
+    dst_y += 16;
+  }
+}
+
+void ARGBToUVRow_MSA(const uint8_t* src_argb,
+                     int src_stride_argb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* src_argb_next = src_argb + src_stride_argb;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v8u16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, reg8, reg9;
+  v16u8 dst0, dst1;
+  v8u16 const_0x70 = (v8u16)__msa_ldi_h(0x38);
+  v8u16 const_0x4A = (v8u16)__msa_ldi_h(0x25);
+  v8u16 const_0x26 = (v8u16)__msa_ldi_h(0x13);
+  v8u16 const_0x5E = (v8u16)__msa_ldi_h(0x2f);
+  v8u16 const_0x12 = (v8u16)__msa_ldi_h(0x09);
+  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
+
+  for (x = 0; x < width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb, 48);
+    src4 = (v16u8)__msa_ld_b((v16u8*)src_argb, 64);
+    src5 = (v16u8)__msa_ld_b((v16u8*)src_argb, 80);
+    src6 = (v16u8)__msa_ld_b((v16u8*)src_argb, 96);
+    src7 = (v16u8)__msa_ld_b((v16u8*)src_argb, 112);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    vec2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4);
+    vec3 = (v16u8)__msa_pckev_b((v16i8)src7, (v16i8)src6);
+    vec4 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec5 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    vec6 = (v16u8)__msa_pckod_b((v16i8)src5, (v16i8)src4);
+    vec7 = (v16u8)__msa_pckod_b((v16i8)src7, (v16i8)src6);
+    vec8 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    vec9 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    vec4 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4);
+    vec5 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6);
+    vec0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)vec3, (v16i8)vec2);
+    reg0 = __msa_hadd_u_h(vec8, vec8);
+    reg1 = __msa_hadd_u_h(vec9, vec9);
+    reg2 = __msa_hadd_u_h(vec4, vec4);
+    reg3 = __msa_hadd_u_h(vec5, vec5);
+    reg4 = __msa_hadd_u_h(vec0, vec0);
+    reg5 = __msa_hadd_u_h(vec1, vec1);
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 16);
+    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 32);
+    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 48);
+    src4 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 64);
+    src5 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 80);
+    src6 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 96);
+    src7 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 112);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    vec2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4);
+    vec3 = (v16u8)__msa_pckev_b((v16i8)src7, (v16i8)src6);
+    vec4 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec5 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    vec6 = (v16u8)__msa_pckod_b((v16i8)src5, (v16i8)src4);
+    vec7 = (v16u8)__msa_pckod_b((v16i8)src7, (v16i8)src6);
+    vec8 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    vec9 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    vec4 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4);
+    vec5 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6);
+    vec0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)vec3, (v16i8)vec2);
+    reg0 += __msa_hadd_u_h(vec8, vec8);
+    reg1 += __msa_hadd_u_h(vec9, vec9);
+    reg2 += __msa_hadd_u_h(vec4, vec4);
+    reg3 += __msa_hadd_u_h(vec5, vec5);
+    reg4 += __msa_hadd_u_h(vec0, vec0);
+    reg5 += __msa_hadd_u_h(vec1, vec1);
+    reg0 += const_0x0001;
+    reg1 += const_0x0001;
+    reg2 += const_0x0001;
+    reg3 += const_0x0001;
+    reg4 += const_0x0001;
+    reg5 += const_0x0001;
+    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 1);
+    reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 1);
+    reg2 = (v8u16)__msa_srai_h((v8i16)reg2, 1);
+    reg3 = (v8u16)__msa_srai_h((v8i16)reg3, 1);
+    reg4 = (v8u16)__msa_srai_h((v8i16)reg4, 1);
+    reg5 = (v8u16)__msa_srai_h((v8i16)reg5, 1);
+    reg6 = reg0 * const_0x70;
+    reg7 = reg1 * const_0x70;
+    reg8 = reg2 * const_0x4A;
+    reg9 = reg3 * const_0x4A;
+    reg6 += const_0x8080;
+    reg7 += const_0x8080;
+    reg8 += reg4 * const_0x26;
+    reg9 += reg5 * const_0x26;
+    reg0 *= const_0x12;
+    reg1 *= const_0x12;
+    reg2 *= const_0x5E;
+    reg3 *= const_0x5E;
+    reg4 *= const_0x70;
+    reg5 *= const_0x70;
+    reg2 += reg0;
+    reg3 += reg1;
+    reg4 += const_0x8080;
+    reg5 += const_0x8080;
+    reg6 -= reg8;
+    reg7 -= reg9;
+    reg4 -= reg2;
+    reg5 -= reg3;
+    reg6 = (v8u16)__msa_srai_h((v8i16)reg6, 8);
+    reg7 = (v8u16)__msa_srai_h((v8i16)reg7, 8);
+    reg4 = (v8u16)__msa_srai_h((v8i16)reg4, 8);
+    reg5 = (v8u16)__msa_srai_h((v8i16)reg5, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg7, (v16i8)reg6);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_argb += 128;
+    src_argb_next += 128;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBToRGB24Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2;
+  v16i8 shuffler0 = {0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20};
+  v16i8 shuffler1 = {5,  6,  8,  9,  10, 12, 13, 14,
+                     16, 17, 18, 20, 21, 22, 24, 25};
+  v16i8 shuffler2 = {10, 12, 13, 14, 16, 17, 18, 20,
+                     21, 22, 24, 25, 26, 28, 29, 30};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src2, (v16i8)src1);
+    dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst_rgb, 16);
+    ST_UB(dst2, (dst_rgb + 32));
+    src_argb += 64;
+    dst_rgb += 48;
+  }
+}
+
+void ARGBToRAWRow_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2;
+  v16i8 shuffler0 = {2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, 18, 17, 16, 22};
+  v16i8 shuffler1 = {5,  4,  10, 9,  8,  14, 13, 12,
+                     18, 17, 16, 22, 21, 20, 26, 25};
+  v16i8 shuffler2 = {8,  14, 13, 12, 18, 17, 16, 22,
+                     21, 20, 26, 25, 24, 30, 29, 28};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src2, (v16i8)src1);
+    dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst_rgb, 16);
+    ST_UB(dst2, (dst_rgb + 32));
+    src_argb += 64;
+    dst_rgb += 48;
+  }
+}
+
+void ARGBToRGB565Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    vec0 = (v16u8)__msa_srai_b((v16i8)src0, 3);
+    vec1 = (v16u8)__msa_slli_b((v16i8)src0, 3);
+    vec2 = (v16u8)__msa_srai_b((v16i8)src0, 5);
+    vec4 = (v16u8)__msa_srai_b((v16i8)src1, 3);
+    vec5 = (v16u8)__msa_slli_b((v16i8)src1, 3);
+    vec6 = (v16u8)__msa_srai_b((v16i8)src1, 5);
+    vec1 = (v16u8)__msa_sldi_b(zero, (v16i8)vec1, 1);
+    vec2 = (v16u8)__msa_sldi_b(zero, (v16i8)vec2, 1);
+    vec5 = (v16u8)__msa_sldi_b(zero, (v16i8)vec5, 1);
+    vec6 = (v16u8)__msa_sldi_b(zero, (v16i8)vec6, 1);
+    vec3 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 2);
+    vec7 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 2);
+    vec0 = __msa_binsli_b(vec0, vec1, 2);
+    vec1 = __msa_binsli_b(vec2, vec3, 4);
+    vec4 = __msa_binsli_b(vec4, vec5, 2);
+    vec5 = __msa_binsli_b(vec6, vec7, 4);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
+    vec4 = (v16u8)__msa_ilvev_b((v16i8)vec5, (v16i8)vec4);
+    dst0 = (v16u8)__msa_pckev_h((v8i16)vec4, (v8i16)vec0);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBToARGB1555Row_MSA(const uint8_t* src_argb,
+                           uint8_t* dst_rgb,
+                           int width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    vec0 = (v16u8)__msa_srai_b((v16i8)src0, 3);
+    vec1 = (v16u8)__msa_slli_b((v16i8)src0, 2);
+    vec2 = (v16u8)__msa_srai_b((v16i8)vec0, 3);
+    vec1 = (v16u8)__msa_sldi_b(zero, (v16i8)vec1, 1);
+    vec2 = (v16u8)__msa_sldi_b(zero, (v16i8)vec2, 1);
+    vec3 = (v16u8)__msa_srai_b((v16i8)src0, 1);
+    vec5 = (v16u8)__msa_srai_b((v16i8)src1, 3);
+    vec6 = (v16u8)__msa_slli_b((v16i8)src1, 2);
+    vec7 = (v16u8)__msa_srai_b((v16i8)vec5, 3);
+    vec6 = (v16u8)__msa_sldi_b(zero, (v16i8)vec6, 1);
+    vec7 = (v16u8)__msa_sldi_b(zero, (v16i8)vec7, 1);
+    vec8 = (v16u8)__msa_srai_b((v16i8)src1, 1);
+    vec3 = (v16u8)__msa_sldi_b(zero, (v16i8)vec3, 2);
+    vec8 = (v16u8)__msa_sldi_b(zero, (v16i8)vec8, 2);
+    vec4 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 3);
+    vec9 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 3);
+    vec0 = __msa_binsli_b(vec0, vec1, 2);
+    vec5 = __msa_binsli_b(vec5, vec6, 2);
+    vec1 = __msa_binsli_b(vec2, vec3, 5);
+    vec6 = __msa_binsli_b(vec7, vec8, 5);
+    vec1 = __msa_binsli_b(vec1, vec4, 0);
+    vec6 = __msa_binsli_b(vec6, vec9, 0);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
+    vec1 = (v16u8)__msa_ilvev_b((v16i8)vec6, (v16i8)vec5);
+    dst0 = (v16u8)__msa_pckev_h((v8i16)vec1, (v8i16)vec0);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBToARGB4444Row_MSA(const uint8_t* src_argb,
+                           uint8_t* dst_rgb,
+                           int width) {
+  int x;
+  v16u8 src0, src1;
+  v16u8 vec0, vec1;
+  v16u8 dst0;
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    vec0 = (v16u8)__msa_srai_b((v16i8)src0, 4);
+    vec1 = (v16u8)__msa_srai_b((v16i8)src1, 4);
+    src0 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 1);
+    src1 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 1);
+    vec0 = __msa_binsli_b(vec0, src0, 3);
+    vec1 = __msa_binsli_b(vec1, src1, 3);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBToUV444Row_MSA(const uint8_t* src_argb,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int32_t width) {
+  int32_t x;
+  v16u8 src0, src1, src2, src3, reg0, reg1, reg2, reg3, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 vec8, vec9, vec10, vec11;
+  v8u16 const_112 = (v8u16)__msa_ldi_h(112);
+  v8u16 const_74 = (v8u16)__msa_ldi_h(74);
+  v8u16 const_38 = (v8u16)__msa_ldi_h(38);
+  v8u16 const_94 = (v8u16)__msa_ldi_h(94);
+  v8u16 const_18 = (v8u16)__msa_ldi_h(18);
+  v8u16 const_32896 = (v8u16)__msa_fill_h(32896);
+  v16i8 zero = {0};
+
+  for (x = width; x > 0; x -= 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    reg0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    reg1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    reg2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    reg3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    src0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    src1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    src2 = (v16u8)__msa_pckod_b((v16i8)reg1, (v16i8)reg0);
+    vec0 = (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b(zero, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b(zero, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b(zero, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b(zero, (v16i8)src2);
+    vec5 = (v8u16)__msa_ilvl_b(zero, (v16i8)src2);
+    vec10 = vec0 * const_18;
+    vec11 = vec1 * const_18;
+    vec8 = vec2 * const_94;
+    vec9 = vec3 * const_94;
+    vec6 = vec4 * const_112;
+    vec7 = vec5 * const_112;
+    vec0 *= const_112;
+    vec1 *= const_112;
+    vec2 *= const_74;
+    vec3 *= const_74;
+    vec4 *= const_38;
+    vec5 *= const_38;
+    vec8 += vec10;
+    vec9 += vec11;
+    vec6 += const_32896;
+    vec7 += const_32896;
+    vec0 += const_32896;
+    vec1 += const_32896;
+    vec2 += vec4;
+    vec3 += vec5;
+    vec0 -= vec2;
+    vec1 -= vec3;
+    vec6 -= vec8;
+    vec7 -= vec9;
+    vec0 = (v8u16)__msa_srai_h((v8i16)vec0, 8);
+    vec1 = (v8u16)__msa_srai_h((v8i16)vec1, 8);
+    vec6 = (v8u16)__msa_srai_h((v8i16)vec6, 8);
+    vec7 = (v8u16)__msa_srai_h((v8i16)vec7, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_argb += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBMultiplyRow_MSA(const uint8_t* src_argb,
+                         const uint8_t* src_argb1,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v4u32 reg0, reg1, reg2, reg3;
+  v8i16 zero = {0};
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src0, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src1);
+    reg0 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec0);
+    reg1 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec0);
+    reg2 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec1);
+    reg3 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec1);
+    reg0 *= (v4u32)__msa_ilvr_h(zero, (v8i16)vec2);
+    reg1 *= (v4u32)__msa_ilvl_h(zero, (v8i16)vec2);
+    reg2 *= (v4u32)__msa_ilvr_h(zero, (v8i16)vec3);
+    reg3 *= (v4u32)__msa_ilvl_h(zero, (v8i16)vec3);
+    reg0 = (v4u32)__msa_srai_w((v4i32)reg0, 16);
+    reg1 = (v4u32)__msa_srai_w((v4i32)reg1, 16);
+    reg2 = (v4u32)__msa_srai_w((v4i32)reg2, 16);
+    reg3 = (v4u32)__msa_srai_w((v4i32)reg3, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_argb);
+    src_argb += 16;
+    src_argb1 += 16;
+    dst_argb += 16;
+  }
+}
+
+void ARGBAddRow_MSA(const uint8_t* src_argb,
+                    const uint8_t* src_argb1,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb1, 16);
+    dst0 = __msa_adds_u_b(src0, src2);
+    dst1 = __msa_adds_u_b(src1, src3);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBSubtractRow_MSA(const uint8_t* src_argb,
+                         const uint8_t* src_argb1,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb1, 16);
+    dst0 = __msa_subs_u_b(src0, src2);
+    dst1 = __msa_subs_u_b(src1, src3);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBAttenuateRow_MSA(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  v16u8 src0, src1, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v4u32 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v8i16 zero = {0};
+  v16u8 mask = {0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src0, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src1, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src1, (v16i8)src1);
+    vec4 = (v8u16)__msa_fill_h(vec0[3]);
+    vec5 = (v8u16)__msa_fill_h(vec0[7]);
+    vec6 = (v8u16)__msa_fill_h(vec1[3]);
+    vec7 = (v8u16)__msa_fill_h(vec1[7]);
+    vec4 = (v8u16)__msa_pckev_d((v2i64)vec5, (v2i64)vec4);
+    vec5 = (v8u16)__msa_pckev_d((v2i64)vec7, (v2i64)vec6);
+    vec6 = (v8u16)__msa_fill_h(vec2[3]);
+    vec7 = (v8u16)__msa_fill_h(vec2[7]);
+    vec8 = (v8u16)__msa_fill_h(vec3[3]);
+    vec9 = (v8u16)__msa_fill_h(vec3[7]);
+    vec6 = (v8u16)__msa_pckev_d((v2i64)vec7, (v2i64)vec6);
+    vec7 = (v8u16)__msa_pckev_d((v2i64)vec9, (v2i64)vec8);
+    reg0 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec4);
+    reg1 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec4);
+    reg2 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec5);
+    reg3 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec5);
+    reg4 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec6);
+    reg5 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec6);
+    reg6 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec7);
+    reg7 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec7);
+    reg0 *= (v4u32)__msa_ilvr_h(zero, (v8i16)vec0);
+    reg1 *= (v4u32)__msa_ilvl_h(zero, (v8i16)vec0);
+    reg2 *= (v4u32)__msa_ilvr_h(zero, (v8i16)vec1);
+    reg3 *= (v4u32)__msa_ilvl_h(zero, (v8i16)vec1);
+    reg4 *= (v4u32)__msa_ilvr_h(zero, (v8i16)vec2);
+    reg5 *= (v4u32)__msa_ilvl_h(zero, (v8i16)vec2);
+    reg6 *= (v4u32)__msa_ilvr_h(zero, (v8i16)vec3);
+    reg7 *= (v4u32)__msa_ilvl_h(zero, (v8i16)vec3);
+    reg0 = (v4u32)__msa_srai_w((v4i32)reg0, 24);
+    reg1 = (v4u32)__msa_srai_w((v4i32)reg1, 24);
+    reg2 = (v4u32)__msa_srai_w((v4i32)reg2, 24);
+    reg3 = (v4u32)__msa_srai_w((v4i32)reg3, 24);
+    reg4 = (v4u32)__msa_srai_w((v4i32)reg4, 24);
+    reg5 = (v4u32)__msa_srai_w((v4i32)reg5, 24);
+    reg6 = (v4u32)__msa_srai_w((v4i32)reg6, 24);
+    reg7 = (v4u32)__msa_srai_w((v4i32)reg7, 24);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec2 = (v8u16)__msa_pckev_h((v8i16)reg5, (v8i16)reg4);
+    vec3 = (v8u16)__msa_pckev_h((v8i16)reg7, (v8i16)reg6);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    dst0 = __msa_bmnz_v(dst0, src0, mask);
+    dst1 = __msa_bmnz_v(dst1, src1, mask);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBToRGB565DitherRow_MSA(const uint8_t* src_argb,
+                               uint8_t* dst_rgb,
+                               uint32_t dither4,
+                               int width) {
+  int x;
+  v16u8 src0, src1, dst0, vec0, vec1;
+  v8i16 vec_d0;
+  v8i16 reg0, reg1, reg2;
+  v16i8 zero = {0};
+  v8i16 max = __msa_ldi_h(0xFF);
+
+  vec_d0 = (v8i16)__msa_fill_w(dither4);
+  vec_d0 = (v8i16)__msa_ilvr_b(zero, (v16i8)vec_d0);
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    reg0 = (v8i16)__msa_ilvev_b(zero, (v16i8)vec0);
+    reg1 = (v8i16)__msa_ilvev_b(zero, (v16i8)vec1);
+    reg2 = (v8i16)__msa_ilvod_b(zero, (v16i8)vec0);
+    reg0 += vec_d0;
+    reg1 += vec_d0;
+    reg2 += vec_d0;
+    reg0 = __msa_maxi_s_h((v8i16)reg0, 0);
+    reg1 = __msa_maxi_s_h((v8i16)reg1, 0);
+    reg2 = __msa_maxi_s_h((v8i16)reg2, 0);
+    reg0 = __msa_min_s_h((v8i16)max, (v8i16)reg0);
+    reg1 = __msa_min_s_h((v8i16)max, (v8i16)reg1);
+    reg2 = __msa_min_s_h((v8i16)max, (v8i16)reg2);
+    reg0 = __msa_srai_h(reg0, 3);
+    reg2 = __msa_srai_h(reg2, 3);
+    reg1 = __msa_srai_h(reg1, 2);
+    reg2 = __msa_slli_h(reg2, 11);
+    reg1 = __msa_slli_h(reg1, 5);
+    reg0 |= reg1;
+    dst0 = (v16u8)(reg0 | reg2);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBShuffleRow_MSA(const uint8_t* src_argb,
+                        uint8_t* dst_argb,
+                        const uint8_t* shuffler,
+                        int width) {
+  int x;
+  v16u8 src0, src1, dst0, dst1;
+  v16i8 vec0;
+  v16i8 shuffler_vec = {0, 0, 0, 0, 4, 4, 4, 4, 8, 8, 8, 8, 12, 12, 12, 12};
+  int32_t val = LW((int32_t*)shuffler);
+
+  vec0 = (v16i8)__msa_fill_w(val);
+  shuffler_vec += vec0;
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
+    dst0 = (v16u8)__msa_vshf_b(shuffler_vec, (v16i8)src0, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler_vec, (v16i8)src1, (v16i8)src1);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBShadeRow_MSA(const uint8_t* src_argb,
+                      uint8_t* dst_argb,
+                      int width,
+                      uint32_t value) {
+  int x;
+  v16u8 src0, dst0;
+  v8u16 vec0, vec1;
+  v4u32 reg0, reg1, reg2, reg3, rgba_scale;
+  v8i16 zero = {0};
+
+  rgba_scale[0] = value;
+  rgba_scale = (v4u32)__msa_ilvr_b((v16i8)rgba_scale, (v16i8)rgba_scale);
+  rgba_scale = (v4u32)__msa_ilvr_h(zero, (v8i16)rgba_scale);
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src0, (v16i8)src0);
+    reg0 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec0);
+    reg1 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec0);
+    reg2 = (v4u32)__msa_ilvr_h(zero, (v8i16)vec1);
+    reg3 = (v4u32)__msa_ilvl_h(zero, (v8i16)vec1);
+    reg0 *= rgba_scale;
+    reg1 *= rgba_scale;
+    reg2 *= rgba_scale;
+    reg3 *= rgba_scale;
+    reg0 = (v4u32)__msa_srai_w((v4i32)reg0, 24);
+    reg1 = (v4u32)__msa_srai_w((v4i32)reg1, 24);
+    reg2 = (v4u32)__msa_srai_w((v4i32)reg2, 24);
+    reg3 = (v4u32)__msa_srai_w((v4i32)reg3, 24);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_argb);
+    src_argb += 16;
+    dst_argb += 16;
+  }
+}
+
+void ARGBGrayRow_MSA(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  int x;
+  v16u8 src0, src1, vec0, vec1, dst0, dst1;
+  v8u16 reg0;
+  v16u8 const_0x4D = (v16u8)__msa_ldi_h(0x4D);
+  v16u8 const_0x961D = (v16u8)__msa_fill_h(0x961D);
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
+    vec0 = (v16u8)__msa_pckev_h((v8i16)src1, (v8i16)src0);
+    vec1 = (v16u8)__msa_pckod_h((v8i16)src1, (v8i16)src0);
+    reg0 = __msa_dotp_u_h(vec0, const_0x961D);
+    reg0 = __msa_dpadd_u_h(reg0, vec1, const_0x4D);
+    reg0 = (v8u16)__msa_srari_h((v8i16)reg0, 8);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)reg0, (v16i8)reg0);
+    vec1 = (v16u8)__msa_ilvod_b((v16i8)vec1, (v16i8)vec0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBSepiaRow_MSA(uint8_t* dst_argb, int width) {
+  int x;
+  v16u8 src0, src1, dst0, dst1, vec0, vec1, vec2, vec3, vec4, vec5;
+  v8u16 reg0, reg1, reg2;
+  v16u8 const_0x4411 = (v16u8)__msa_fill_h(0x4411);
+  v16u8 const_0x23 = (v16u8)__msa_ldi_h(0x23);
+  v16u8 const_0x5816 = (v16u8)__msa_fill_h(0x5816);
+  v16u8 const_0x2D = (v16u8)__msa_ldi_h(0x2D);
+  v16u8 const_0x6218 = (v16u8)__msa_fill_h(0x6218);
+  v16u8 const_0x32 = (v16u8)__msa_ldi_h(0x32);
+  v8u16 const_0xFF = (v8u16)__msa_ldi_h(0xFF);
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)dst_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)dst_argb, 16);
+    vec0 = (v16u8)__msa_pckev_h((v8i16)src1, (v8i16)src0);
+    vec1 = (v16u8)__msa_pckod_h((v8i16)src1, (v8i16)src0);
+    vec3 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec1);
+    reg0 = (v8u16)__msa_dotp_u_h(vec0, const_0x4411);
+    reg1 = (v8u16)__msa_dotp_u_h(vec0, const_0x5816);
+    reg2 = (v8u16)__msa_dotp_u_h(vec0, const_0x6218);
+    reg0 = (v8u16)__msa_dpadd_u_h(reg0, vec1, const_0x23);
+    reg1 = (v8u16)__msa_dpadd_u_h(reg1, vec1, const_0x2D);
+    reg2 = (v8u16)__msa_dpadd_u_h(reg2, vec1, const_0x32);
+    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 7);
+    reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 7);
+    reg2 = (v8u16)__msa_srai_h((v8i16)reg2, 7);
+    reg1 = (v8u16)__msa_min_u_h((v8u16)reg1, const_0xFF);
+    reg2 = (v8u16)__msa_min_u_h((v8u16)reg2, const_0xFF);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)reg0, (v16i8)reg0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg1);
+    vec2 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg2);
+    vec4 = (v16u8)__msa_ilvr_b((v16i8)vec2, (v16i8)vec0);
+    vec5 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec1);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)vec5, (v16i8)vec4);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)vec5, (v16i8)vec4);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    dst_argb += 32;
+  }
+}
+
+void ARGB4444ToARGBRow_MSA(const uint8_t* src_argb4444,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  v16u8 src0, src1;
+  v8u16 vec0, vec1, vec2, vec3;
+  v16u8 dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb4444, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb4444, 16);
+    vec0 = (v8u16)__msa_andi_b(src0, 0x0F);
+    vec1 = (v8u16)__msa_andi_b(src1, 0x0F);
+    vec2 = (v8u16)__msa_andi_b(src0, 0xF0);
+    vec3 = (v8u16)__msa_andi_b(src1, 0xF0);
+    vec0 |= (v8u16)__msa_slli_b((v16i8)vec0, 4);
+    vec1 |= (v8u16)__msa_slli_b((v16i8)vec1, 4);
+    vec2 |= (v8u16)__msa_srli_b((v16i8)vec2, 4);
+    vec3 |= (v8u16)__msa_srli_b((v16i8)vec3, 4);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)vec2, (v16i8)vec0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)vec2, (v16i8)vec0);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec1);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)vec3, (v16i8)vec1);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_argb4444 += 32;
+    dst_argb += 64;
+  }
+}
+
+void ARGB1555ToARGBRow_MSA(const uint8_t* src_argb1555,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  v8u16 src0, src1;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6;
+  v16u8 dst0, dst1, dst2, dst3;
+  v8u16 const_0x1F = (v8u16)__msa_ldi_h(0x1F);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v8u16)__msa_ld_h((void*)src_argb1555, 0);
+    src1 = (v8u16)__msa_ld_h((void*)src_argb1555, 16);
+    vec0 = src0 & const_0x1F;
+    vec1 = src1 & const_0x1F;
+    src0 = (v8u16)__msa_srli_h((v8i16)src0, 5);
+    src1 = (v8u16)__msa_srli_h((v8i16)src1, 5);
+    vec2 = src0 & const_0x1F;
+    vec3 = src1 & const_0x1F;
+    src0 = (v8u16)__msa_srli_h((v8i16)src0, 5);
+    src1 = (v8u16)__msa_srli_h((v8i16)src1, 5);
+    vec4 = src0 & const_0x1F;
+    vec5 = src1 & const_0x1F;
+    src0 = (v8u16)__msa_srli_h((v8i16)src0, 5);
+    src1 = (v8u16)__msa_srli_h((v8i16)src1, 5);
+    reg0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    reg1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    reg2 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4);
+    reg3 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    reg4 = (v16u8)__msa_slli_b((v16i8)reg0, 3);
+    reg5 = (v16u8)__msa_slli_b((v16i8)reg1, 3);
+    reg6 = (v16u8)__msa_slli_b((v16i8)reg2, 3);
+    reg4 |= (v16u8)__msa_srai_b((v16i8)reg0, 2);
+    reg5 |= (v16u8)__msa_srai_b((v16i8)reg1, 2);
+    reg6 |= (v16u8)__msa_srai_b((v16i8)reg2, 2);
+    reg3 = -reg3;
+    reg0 = (v16u8)__msa_ilvr_b((v16i8)reg6, (v16i8)reg4);
+    reg1 = (v16u8)__msa_ilvl_b((v16i8)reg6, (v16i8)reg4);
+    reg2 = (v16u8)__msa_ilvr_b((v16i8)reg3, (v16i8)reg5);
+    reg3 = (v16u8)__msa_ilvl_b((v16i8)reg3, (v16i8)reg5);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)reg2, (v16i8)reg0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)reg2, (v16i8)reg0);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)reg3, (v16i8)reg1);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)reg3, (v16i8)reg1);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_argb1555 += 32;
+    dst_argb += 64;
+  }
+}
+
+void RGB565ToARGBRow_MSA(const uint8_t* src_rgb565,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  v8u16 src0, src1, vec0, vec1, vec2, vec3, vec4, vec5;
+  v8u16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v16u8 res0, res1, res2, res3, dst0, dst1, dst2, dst3;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8u16 const_0x1F = (v8u16)__msa_ldi_h(0x1F);
+  v8u16 const_0x7E0 = (v8u16)__msa_fill_h(0x7E0);
+  v8u16 const_0xF800 = (v8u16)__msa_fill_h(0xF800);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v8u16)__msa_ld_h((void*)src_rgb565, 0);
+    src1 = (v8u16)__msa_ld_h((void*)src_rgb565, 16);
+    vec0 = src0 & const_0x1F;
+    vec1 = src0 & const_0x7E0;
+    vec2 = src0 & const_0xF800;
+    vec3 = src1 & const_0x1F;
+    vec4 = src1 & const_0x7E0;
+    vec5 = src1 & const_0xF800;
+    reg0 = (v8u16)__msa_slli_h((v8i16)vec0, 3);
+    reg1 = (v8u16)__msa_srli_h((v8i16)vec1, 3);
+    reg2 = (v8u16)__msa_srli_h((v8i16)vec2, 8);
+    reg3 = (v8u16)__msa_slli_h((v8i16)vec3, 3);
+    reg4 = (v8u16)__msa_srli_h((v8i16)vec4, 3);
+    reg5 = (v8u16)__msa_srli_h((v8i16)vec5, 8);
+    reg0 |= (v8u16)__msa_srli_h((v8i16)vec0, 2);
+    reg1 |= (v8u16)__msa_srli_h((v8i16)vec1, 9);
+    reg2 |= (v8u16)__msa_srli_h((v8i16)vec2, 13);
+    reg3 |= (v8u16)__msa_srli_h((v8i16)vec3, 2);
+    reg4 |= (v8u16)__msa_srli_h((v8i16)vec4, 9);
+    reg5 |= (v8u16)__msa_srli_h((v8i16)vec5, 13);
+    res0 = (v16u8)__msa_ilvev_b((v16i8)reg2, (v16i8)reg0);
+    res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)reg1);
+    res2 = (v16u8)__msa_ilvev_b((v16i8)reg5, (v16i8)reg3);
+    res3 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)reg4);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)res1, (v16i8)res0);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)res3, (v16i8)res2);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)res3, (v16i8)res2);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_rgb565 += 32;
+    dst_argb += 64;
+  }
+}
+
+void RGB24ToARGBRow_MSA(const uint8_t* src_rgb24,
+                        uint8_t* dst_argb,
+                        int width) {
+  int x;
+  v16u8 src0, src1, src2;
+  v16u8 vec0, vec1, vec2;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v16i8 shuffler = {0, 1, 2, 16, 3, 4, 5, 17, 6, 7, 8, 18, 9, 10, 11, 19};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_rgb24, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_rgb24, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_rgb24, 32);
+    vec0 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 12);
+    vec1 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8);
+    vec2 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src2, 4);
+    dst0 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)vec0);
+    dst2 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)vec1);
+    dst3 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)vec2);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_rgb24 += 48;
+    dst_argb += 64;
+  }
+}
+
+void RAWToARGBRow_MSA(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  int x;
+  v16u8 src0, src1, src2;
+  v16u8 vec0, vec1, vec2;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v16i8 mask = {2, 1, 0, 16, 5, 4, 3, 17, 8, 7, 6, 18, 11, 10, 9, 19};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_raw, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_raw, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_raw, 32);
+    vec0 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 12);
+    vec1 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8);
+    vec2 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src2, 4);
+    dst0 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)vec0);
+    dst2 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)vec1);
+    dst3 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)vec2);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_raw += 48;
+    dst_argb += 64;
+  }
+}
+
+void ARGB1555ToYRow_MSA(const uint8_t* src_argb1555,
+                        uint8_t* dst_y,
+                        int width) {
+  int x;
+  v16u8 src0, src1, tmp0, tmp1, tmpb, tmpg, tmpr;
+  v16u8 reg0, reg1, reg2, dst;
+  v8i16 tmpr_l, tmpr_r, tmpg_l, tmpg_r, tmpb_l, tmpb_r;
+  v8i16 res0, res1;
+  v8i16 const_66 = (v8i16)__msa_ldi_h(66);
+  v8i16 const_129 = (v8i16)__msa_ldi_h(129);
+  v8i16 const_25 = (v8i16)__msa_ldi_h(25);
+  v8u16 const_1080 = (v8u16)__msa_fill_h(0x1080);
+  v16u8 zero = (v16u8)__msa_ldi_b(0);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb1555, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb1555, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    tmpg = (v16u8)__msa_srli_b(tmp0, 5);
+    reg0 = (v16u8)__msa_andi_b(tmp1, 0x03);
+    reg0 = (v16u8)__msa_slli_b(reg0, 3);
+    tmpg = (v16u8)__msa_or_v(tmpg, reg0);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x7C);
+    tmpr = (v16u8)__msa_srli_b(reg1, 2);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_slli_b(tmpg, 3);
+    reg2 = (v16u8)__msa_slli_b(tmpr, 3);
+    tmpb = (v16u8)__msa_srli_b(tmpb, 2);
+    tmpg = (v16u8)__msa_srli_b(tmpg, 2);
+    tmpr = (v16u8)__msa_srli_b(tmpr, 2);
+    tmpb = (v16u8)__msa_or_v(reg0, tmpb);
+    tmpg = (v16u8)__msa_or_v(reg1, tmpg);
+    tmpr = (v16u8)__msa_or_v(reg2, tmpr);
+    tmpb_r = (v8i16)__msa_ilvr_b(zero, tmpb);
+    tmpb_l = (v8i16)__msa_ilvl_b(zero, tmpb);
+    tmpg_r = (v8i16)__msa_ilvr_b(zero, tmpg);
+    tmpg_l = (v8i16)__msa_ilvl_b(zero, tmpg);
+    tmpr_r = (v8i16)__msa_ilvr_b(zero, tmpr);
+    tmpr_l = (v8i16)__msa_ilvl_b(zero, tmpr);
+    res0 = const_1080 + const_25 * tmpb_r;
+    res1 = const_1080 + const_25 * tmpb_l;
+    res0 += const_129 * tmpg_r;
+    res1 += const_129 * tmpg_l;
+    res0 += const_66 * tmpr_r;
+    res1 += const_66 * tmpr_l;
+    dst = (v16u8)__msa_pckod_b(res1, res0);
+    ST_UB(dst, dst_y);
+    src_argb1555 += 32;
+    dst_y += 16;
+  }
+}
+
+void RGB565ToYRow_MSA(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, tmp0, tmp1, tmpb, tmpg, tmpr;
+  v16u8 reg0, reg1, dst;
+  v8i16 tmpr_l, tmpr_r, tmpg_l, tmpg_r, tmpb_l, tmpb_r;
+  v8i16 res0, res1;
+  v8i16 const_66 = (v8i16)__msa_ldi_h(66);
+  v8i16 const_129 = (v8i16)__msa_ldi_h(129);
+  v8i16 const_25 = (v8i16)__msa_ldi_h(25);
+  v8i16 const_1080 = (v8i16)__msa_fill_h(0x1080);
+  v16u8 zero = __msa_ldi_b(0);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_rgb565, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_rgb565, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    tmpr = (v16u8)__msa_andi_b(tmp1, 0xF8);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x07);
+    reg0 = (v16u8)__msa_srli_b(tmp0, 5);
+    reg1 = (v16u8)__msa_slli_b(reg1, 3);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_srli_b(tmpb, 2);
+    tmpb = (v16u8)__msa_or_v(reg1, reg0);
+    reg0 = (v16u8)__msa_slli_b(tmpg, 2);
+    reg1 = (v16u8)__msa_srli_b(tmpg, 4);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    reg0 = (v16u8)__msa_srli_b(tmpr, 5);
+    tmpr = (v16u8)__msa_or_v(tmpr, reg0);
+    tmpb_r = (v8i16)__msa_ilvr_b(zero, tmpb);
+    tmpb_l = (v8i16)__msa_ilvl_b(zero, tmpb);
+    tmpg_r = (v8i16)__msa_ilvr_b(zero, tmpg);
+    tmpg_l = (v8i16)__msa_ilvl_b(zero, tmpg);
+    tmpr_r = (v8i16)__msa_ilvr_b(zero, tmpr);
+    tmpr_l = (v8i16)__msa_ilvl_b(zero, tmpr);
+    res0 = const_1080 + const_25 * tmpb_r;
+    res1 = const_1080 + const_25 * tmpb_l;
+    res0 += const_129 * tmpg_r;
+    res1 += const_129 * tmpg_l;
+    res0 += const_66 * tmpr_r;
+    res1 += const_66 * tmpr_l;
+    dst = (v16u8)__msa_pckod_b(res1, res0);
+    ST_UB(dst, dst_y);
+    src_rgb565 += 32;
+    dst_y += 16;
+  }
+}
+
+void RGB24ToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, reg0, reg1, reg2, reg3, dst0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 const_0x8119 = (v8u16)__msa_fill_h(0x8119);
+  v8u16 const_0x42 = (v8u16)__msa_fill_h(0x42);
+  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+  v16i8 mask0 = {0, 1, 2, 3, 3, 4, 5, 6, 6, 7, 8, 9, 9, 10, 11, 12};
+  v16i8 mask1 = {12, 13, 14, 15, 15, 16, 17, 18,
+                 18, 19, 20, 21, 21, 22, 23, 24};
+  v16i8 mask2 = {8, 9, 10, 11, 11, 12, 13, 14, 14, 15, 16, 17, 17, 18, 19, 20};
+  v16i8 mask3 = {4, 5, 6, 7, 7, 8, 9, 10, 10, 11, 12, 13, 13, 14, 15, 16};
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    reg0 = (v16u8)__msa_vshf_b(mask0, zero, (v16i8)src0);
+    reg1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    reg2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src2, (v16i8)src1);
+    reg3 = (v16u8)__msa_vshf_b(mask3, zero, (v16i8)src2);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec2 = (v8u16)__msa_pckod_h((v8i16)reg1, (v8i16)reg0);
+    vec3 = (v8u16)__msa_pckod_h((v8i16)reg3, (v8i16)reg2);
+    vec0 = __msa_dotp_u_h((v16u8)vec0, (v16u8)const_0x8119);
+    vec1 = __msa_dotp_u_h((v16u8)vec1, (v16u8)const_0x8119);
+    vec0 = __msa_dpadd_u_h(vec0, (v16u8)vec2, (v16u8)const_0x42);
+    vec1 = __msa_dpadd_u_h(vec1, (v16u8)vec3, (v16u8)const_0x42);
+    vec0 += const_0x1080;
+    vec1 += const_0x1080;
+    vec0 = (v8u16)__msa_srai_h((v8i16)vec0, 8);
+    vec1 = (v8u16)__msa_srai_h((v8i16)vec1, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_y);
+    src_argb += 48;
+    dst_y += 16;
+  }
+}
+
+void RAWToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, reg0, reg1, reg2, reg3, dst0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 const_0x8142 = (v8u16)__msa_fill_h(0x8142);
+  v8u16 const_0x19 = (v8u16)__msa_fill_h(0x19);
+  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+  v16i8 mask0 = {0, 1, 2, 3, 3, 4, 5, 6, 6, 7, 8, 9, 9, 10, 11, 12};
+  v16i8 mask1 = {12, 13, 14, 15, 15, 16, 17, 18,
+                 18, 19, 20, 21, 21, 22, 23, 24};
+  v16i8 mask2 = {8, 9, 10, 11, 11, 12, 13, 14, 14, 15, 16, 17, 17, 18, 19, 20};
+  v16i8 mask3 = {4, 5, 6, 7, 7, 8, 9, 10, 10, 11, 12, 13, 13, 14, 15, 16};
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    reg0 = (v16u8)__msa_vshf_b(mask0, zero, (v16i8)src0);
+    reg1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    reg2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src2, (v16i8)src1);
+    reg3 = (v16u8)__msa_vshf_b(mask3, zero, (v16i8)src2);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec2 = (v8u16)__msa_pckod_h((v8i16)reg1, (v8i16)reg0);
+    vec3 = (v8u16)__msa_pckod_h((v8i16)reg3, (v8i16)reg2);
+    vec0 = __msa_dotp_u_h((v16u8)vec0, (v16u8)const_0x8142);
+    vec1 = __msa_dotp_u_h((v16u8)vec1, (v16u8)const_0x8142);
+    vec0 = __msa_dpadd_u_h(vec0, (v16u8)vec2, (v16u8)const_0x19);
+    vec1 = __msa_dpadd_u_h(vec1, (v16u8)vec3, (v16u8)const_0x19);
+    vec0 += const_0x1080;
+    vec1 += const_0x1080;
+    vec0 = (v8u16)__msa_srai_h((v8i16)vec0, 8);
+    vec1 = (v8u16)__msa_srai_h((v8i16)vec1, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_y);
+    src_argb += 48;
+    dst_y += 16;
+  }
+}
+
+void ARGB1555ToUVRow_MSA(const uint8_t* src_argb1555,
+                         int src_stride_argb1555,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  const uint16_t* s = (const uint16_t*)src_argb1555;
+  const uint16_t* t = (const uint16_t*)(src_argb1555 + src_stride_argb1555);
+  int64_t res0, res1;
+  v16u8 src0, src1, src2, src3, dst;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v16u8 reg0, reg1, reg2, reg3;
+  v16u8 tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  v8i16 const_112 = (v8i16)__msa_ldi_h(0x38);
+  v8i16 const_74 = (v8i16)__msa_ldi_h(0x25);
+  v8i16 const_38 = (v8i16)__msa_ldi_h(0x13);
+  v8i16 const_94 = (v8i16)__msa_ldi_h(0x2F);
+  v8i16 const_18 = (v8i16)__msa_ldi_h(0x09);
+  v8u16 const_8080 = (v8u16)__msa_fill_h(0x8080);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v8u16)__msa_ld_b((void*)s, 0);
+    src1 = (v8u16)__msa_ld_b((void*)s, 16);
+    src2 = (v8u16)__msa_ld_b((void*)t, 0);
+    src3 = (v8u16)__msa_ld_b((void*)t, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmp2 = (v16u8)__msa_pckev_b(src3, src2);
+    tmp3 = (v16u8)__msa_pckod_b(src3, src2);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    nexb = (v16u8)__msa_andi_b(tmp2, 0x1F);
+    tmpg = (v16u8)__msa_srli_b(tmp0, 5);
+    nexg = (v16u8)__msa_srli_b(tmp2, 5);
+    reg0 = (v16u8)__msa_andi_b(tmp1, 0x03);
+    reg2 = (v16u8)__msa_andi_b(tmp3, 0x03);
+    reg0 = (v16u8)__msa_slli_b(reg0, 3);
+    reg2 = (v16u8)__msa_slli_b(reg2, 3);
+    tmpg = (v16u8)__msa_or_v(tmpg, reg0);
+    nexg = (v16u8)__msa_or_v(nexg, reg2);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x7C);
+    reg3 = (v16u8)__msa_andi_b(tmp3, 0x7C);
+    tmpr = (v16u8)__msa_srli_b(reg1, 2);
+    nexr = (v16u8)__msa_srli_b(reg3, 2);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_slli_b(tmpg, 3);
+    reg2 = (v16u8)__msa_slli_b(tmpr, 3);
+    tmpb = (v16u8)__msa_srli_b(tmpb, 2);
+    tmpg = (v16u8)__msa_srli_b(tmpg, 2);
+    tmpr = (v16u8)__msa_srli_b(tmpr, 2);
+    tmpb = (v16u8)__msa_or_v(reg0, tmpb);
+    tmpg = (v16u8)__msa_or_v(reg1, tmpg);
+    tmpr = (v16u8)__msa_or_v(reg2, tmpr);
+    reg0 = (v16u8)__msa_slli_b(nexb, 3);
+    reg1 = (v16u8)__msa_slli_b(nexg, 3);
+    reg2 = (v16u8)__msa_slli_b(nexr, 3);
+    nexb = (v16u8)__msa_srli_b(nexb, 2);
+    nexg = (v16u8)__msa_srli_b(nexg, 2);
+    nexr = (v16u8)__msa_srli_b(nexr, 2);
+    nexb = (v16u8)__msa_or_v(reg0, nexb);
+    nexg = (v16u8)__msa_or_v(reg1, nexg);
+    nexr = (v16u8)__msa_or_v(reg2, nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst);
+    res0 = __msa_copy_u_d((v2i64)dst, 0);
+    res1 = __msa_copy_u_d((v2i64)dst, 1);
+    SD(res0, dst_u);
+    SD(res1, dst_v);
+    s += 16;
+    t += 16;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void RGB565ToUVRow_MSA(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  const uint16_t* s = (const uint16_t*)src_rgb565;
+  const uint16_t* t = (const uint16_t*)(src_rgb565 + src_stride_rgb565);
+  int64_t res0, res1;
+  v16u8 src0, src1, src2, src3, dst;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v16u8 reg0, reg1, reg2, reg3;
+  v16u8 tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  v8i16 const_112 = (v8i16)__msa_ldi_h(0x38);
+  v8i16 const_74 = (v8i16)__msa_ldi_h(0x25);
+  v8i16 const_38 = (v8i16)__msa_ldi_h(0x13);
+  v8i16 const_94 = (v8i16)__msa_ldi_h(0x2F);
+  v8i16 const_18 = (v8i16)__msa_ldi_h(0x09);
+  v8u16 const_8080 = (v8u16)__msa_fill_h(0x8080);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)s, 0);
+    src1 = (v16u8)__msa_ld_b((void*)s, 16);
+    src2 = (v16u8)__msa_ld_b((void*)t, 0);
+    src3 = (v16u8)__msa_ld_b((void*)t, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmp2 = (v16u8)__msa_pckev_b(src3, src2);
+    tmp3 = (v16u8)__msa_pckod_b(src3, src2);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    tmpr = (v16u8)__msa_andi_b(tmp1, 0xF8);
+    nexb = (v16u8)__msa_andi_b(tmp2, 0x1F);
+    nexr = (v16u8)__msa_andi_b(tmp3, 0xF8);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x07);
+    reg3 = (v16u8)__msa_andi_b(tmp3, 0x07);
+    reg0 = (v16u8)__msa_srli_b(tmp0, 5);
+    reg1 = (v16u8)__msa_slli_b(reg1, 3);
+    reg2 = (v16u8)__msa_srli_b(tmp2, 5);
+    reg3 = (v16u8)__msa_slli_b(reg3, 3);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    nexg = (v16u8)__msa_or_v(reg2, reg3);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_srli_b(tmpb, 2);
+    reg2 = (v16u8)__msa_slli_b(nexb, 3);
+    reg3 = (v16u8)__msa_srli_b(nexb, 2);
+    tmpb = (v16u8)__msa_or_v(reg1, reg0);
+    nexb = (v16u8)__msa_or_v(reg2, reg3);
+    reg0 = (v16u8)__msa_slli_b(tmpg, 2);
+    reg1 = (v16u8)__msa_srli_b(tmpg, 4);
+    reg2 = (v16u8)__msa_slli_b(nexg, 2);
+    reg3 = (v16u8)__msa_srli_b(nexg, 4);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    nexg = (v16u8)__msa_or_v(reg2, reg3);
+    reg0 = (v16u8)__msa_srli_b(tmpr, 5);
+    reg2 = (v16u8)__msa_srli_b(nexr, 5);
+    tmpr = (v16u8)__msa_or_v(tmpr, reg0);
+    nexr = (v16u8)__msa_or_v(nexr, reg2);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst);
+    res0 = __msa_copy_u_d((v2i64)dst, 0);
+    res1 = __msa_copy_u_d((v2i64)dst, 1);
+    SD(res0, dst_u);
+    SD(res1, dst_v);
+    s += 16;
+    t += 16;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void RGB24ToUVRow_MSA(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  int64_t res0, res1;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 inp0, inp1, inp2, inp3, inp4, inp5;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 reg0, reg1, reg2, reg3;
+  v16u8 dst0;
+  v8u16 const_0x70 = (v8u16)__msa_fill_h(0x38);
+  v8u16 const_0x4A = (v8u16)__msa_fill_h(0x25);
+  v8u16 const_0x26 = (v8u16)__msa_fill_h(0x13);
+  v8u16 const_0x5E = (v8u16)__msa_fill_h(0x2f);
+  v8u16 const_0x12 = (v8u16)__msa_fill_h(0x09);
+  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
+  v16i8 mask = {0, 1, 2, 16, 3, 4, 5, 17, 6, 7, 8, 18, 9, 10, 11, 19};
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 16) {
+    inp0 = (v16u8)__msa_ld_b((void*)s, 0);
+    inp1 = (v16u8)__msa_ld_b((void*)s, 16);
+    inp2 = (v16u8)__msa_ld_b((void*)s, 32);
+    inp3 = (v16u8)__msa_ld_b((void*)t, 0);
+    inp4 = (v16u8)__msa_ld_b((void*)t, 16);
+    inp5 = (v16u8)__msa_ld_b((void*)t, 32);
+    src1 = (v16u8)__msa_sldi_b((v16i8)inp1, (v16i8)inp0, 12);
+    src5 = (v16u8)__msa_sldi_b((v16i8)inp4, (v16i8)inp3, 12);
+    src2 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp1, 8);
+    src6 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp4, 8);
+    src3 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp2, 4);
+    src7 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp5, 4);
+    src0 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp0);
+    src1 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src1);
+    src2 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src2);
+    src3 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src3);
+    src4 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp3);
+    src5 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src5);
+    src6 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src6);
+    src7 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src7);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src4, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src4, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src5, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src5, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b((v16i8)src6, (v16i8)src2);
+    vec5 = (v8u16)__msa_ilvl_b((v16i8)src6, (v16i8)src2);
+    vec6 = (v8u16)__msa_ilvr_b((v16i8)src7, (v16i8)src3);
+    vec7 = (v8u16)__msa_ilvl_b((v16i8)src7, (v16i8)src3);
+    vec0 = (v8u16)__msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = (v8u16)__msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = (v8u16)__msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = (v8u16)__msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = (v8u16)__msa_hadd_u_h((v16u8)vec4, (v16u8)vec4);
+    vec5 = (v8u16)__msa_hadd_u_h((v16u8)vec5, (v16u8)vec5);
+    vec6 = (v8u16)__msa_hadd_u_h((v16u8)vec6, (v16u8)vec6);
+    vec7 = (v8u16)__msa_hadd_u_h((v16u8)vec7, (v16u8)vec7);
+    reg0 = (v8i16)__msa_pckev_d((v2i64)vec1, (v2i64)vec0);
+    reg1 = (v8i16)__msa_pckev_d((v2i64)vec3, (v2i64)vec2);
+    reg2 = (v8i16)__msa_pckev_d((v2i64)vec5, (v2i64)vec4);
+    reg3 = (v8i16)__msa_pckev_d((v2i64)vec7, (v2i64)vec6);
+    reg0 += (v8i16)__msa_pckod_d((v2i64)vec1, (v2i64)vec0);
+    reg1 += (v8i16)__msa_pckod_d((v2i64)vec3, (v2i64)vec2);
+    reg2 += (v8i16)__msa_pckod_d((v2i64)vec5, (v2i64)vec4);
+    reg3 += (v8i16)__msa_pckod_d((v2i64)vec7, (v2i64)vec6);
+    reg0 += const_0x0001;
+    reg1 += const_0x0001;
+    reg2 += const_0x0001;
+    reg3 += const_0x0001;
+    reg0 = __msa_srai_h((v8i16)reg0, 1);
+    reg1 = __msa_srai_h((v8i16)reg1, 1);
+    reg2 = __msa_srai_h((v8i16)reg2, 1);
+    reg3 = __msa_srai_h((v8i16)reg3, 1);
+    vec4 = (v8u16)__msa_pckev_h(reg1, reg0);
+    vec5 = (v8u16)__msa_pckev_h(reg3, reg2);
+    vec6 = (v8u16)__msa_pckod_h(reg1, reg0);
+    vec7 = (v8u16)__msa_pckod_h(reg3, reg2);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)vec5, (v8i16)vec4);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)vec7, (v8i16)vec6);
+    vec2 = (v8u16)__msa_pckod_h((v8i16)vec5, (v8i16)vec4);
+    vec3 = vec0 * const_0x70;
+    vec4 = vec1 * const_0x4A;
+    vec5 = vec2 * const_0x26;
+    vec2 *= const_0x70;
+    vec1 *= const_0x5E;
+    vec0 *= const_0x12;
+    reg0 = __msa_subv_h((v8i16)vec3, (v8i16)vec4);
+    reg1 = __msa_subv_h((v8i16)const_0x8080, (v8i16)vec5);
+    reg2 = __msa_subv_h((v8i16)vec2, (v8i16)vec1);
+    reg3 = __msa_subv_h((v8i16)const_0x8080, (v8i16)vec0);
+    reg0 += reg1;
+    reg2 += reg3;
+    reg0 = __msa_srai_h(reg0, 8);
+    reg2 = __msa_srai_h(reg2, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0);
+    res0 = __msa_copy_u_d((v2i64)dst0, 0);
+    res1 = __msa_copy_u_d((v2i64)dst0, 1);
+    SD(res0, dst_u);
+    SD(res1, dst_v);
+    t += 48;
+    s += 48;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void RAWToUVRow_MSA(const uint8_t* src_rgb,
+                    int src_stride_rgb,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  int x;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  int64_t res0, res1;
+  v16u8 inp0, inp1, inp2, inp3, inp4, inp5;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 reg0, reg1, reg2, reg3;
+  v16u8 dst0;
+  v8u16 const_0x70 = (v8u16)__msa_fill_h(0x38);
+  v8u16 const_0x4A = (v8u16)__msa_fill_h(0x25);
+  v8u16 const_0x26 = (v8u16)__msa_fill_h(0x13);
+  v8u16 const_0x5E = (v8u16)__msa_fill_h(0x2f);
+  v8u16 const_0x12 = (v8u16)__msa_fill_h(0x09);
+  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
+  v16i8 mask = {0, 1, 2, 16, 3, 4, 5, 17, 6, 7, 8, 18, 9, 10, 11, 19};
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 16) {
+    inp0 = (v16u8)__msa_ld_b((void*)s, 0);
+    inp1 = (v16u8)__msa_ld_b((void*)s, 16);
+    inp2 = (v16u8)__msa_ld_b((void*)s, 32);
+    inp3 = (v16u8)__msa_ld_b((void*)t, 0);
+    inp4 = (v16u8)__msa_ld_b((void*)t, 16);
+    inp5 = (v16u8)__msa_ld_b((void*)t, 32);
+    src1 = (v16u8)__msa_sldi_b((v16i8)inp1, (v16i8)inp0, 12);
+    src5 = (v16u8)__msa_sldi_b((v16i8)inp4, (v16i8)inp3, 12);
+    src2 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp1, 8);
+    src6 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp4, 8);
+    src3 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp2, 4);
+    src7 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp5, 4);
+    src0 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp0);
+    src1 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src1);
+    src2 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src2);
+    src3 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src3);
+    src4 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp3);
+    src5 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src5);
+    src6 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src6);
+    src7 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src7);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src4, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src4, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src5, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src5, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b((v16i8)src6, (v16i8)src2);
+    vec5 = (v8u16)__msa_ilvl_b((v16i8)src6, (v16i8)src2);
+    vec6 = (v8u16)__msa_ilvr_b((v16i8)src7, (v16i8)src3);
+    vec7 = (v8u16)__msa_ilvl_b((v16i8)src7, (v16i8)src3);
+    vec0 = (v8u16)__msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = (v8u16)__msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = (v8u16)__msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = (v8u16)__msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = (v8u16)__msa_hadd_u_h((v16u8)vec4, (v16u8)vec4);
+    vec5 = (v8u16)__msa_hadd_u_h((v16u8)vec5, (v16u8)vec5);
+    vec6 = (v8u16)__msa_hadd_u_h((v16u8)vec6, (v16u8)vec6);
+    vec7 = (v8u16)__msa_hadd_u_h((v16u8)vec7, (v16u8)vec7);
+    reg0 = (v8i16)__msa_pckev_d((v2i64)vec1, (v2i64)vec0);
+    reg1 = (v8i16)__msa_pckev_d((v2i64)vec3, (v2i64)vec2);
+    reg2 = (v8i16)__msa_pckev_d((v2i64)vec5, (v2i64)vec4);
+    reg3 = (v8i16)__msa_pckev_d((v2i64)vec7, (v2i64)vec6);
+    reg0 += (v8i16)__msa_pckod_d((v2i64)vec1, (v2i64)vec0);
+    reg1 += (v8i16)__msa_pckod_d((v2i64)vec3, (v2i64)vec2);
+    reg2 += (v8i16)__msa_pckod_d((v2i64)vec5, (v2i64)vec4);
+    reg3 += (v8i16)__msa_pckod_d((v2i64)vec7, (v2i64)vec6);
+    reg0 += const_0x0001;
+    reg1 += const_0x0001;
+    reg2 += const_0x0001;
+    reg3 += const_0x0001;
+    reg0 = __msa_srai_h(reg0, 1);
+    reg1 = __msa_srai_h(reg1, 1);
+    reg2 = __msa_srai_h(reg2, 1);
+    reg3 = __msa_srai_h(reg3, 1);
+    vec4 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec5 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec6 = (v8u16)__msa_pckod_h((v8i16)reg1, (v8i16)reg0);
+    vec7 = (v8u16)__msa_pckod_h((v8i16)reg3, (v8i16)reg2);
+    vec0 = (v8u16)__msa_pckod_h((v8i16)vec5, (v8i16)vec4);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)vec7, (v8i16)vec6);
+    vec2 = (v8u16)__msa_pckev_h((v8i16)vec5, (v8i16)vec4);
+    vec3 = vec0 * const_0x70;
+    vec4 = vec1 * const_0x4A;
+    vec5 = vec2 * const_0x26;
+    vec2 *= const_0x70;
+    vec1 *= const_0x5E;
+    vec0 *= const_0x12;
+    reg0 = __msa_subv_h((v8i16)vec3, (v8i16)vec4);
+    reg1 = __msa_subv_h((v8i16)const_0x8080, (v8i16)vec5);
+    reg2 = __msa_subv_h((v8i16)vec2, (v8i16)vec1);
+    reg3 = __msa_subv_h((v8i16)const_0x8080, (v8i16)vec0);
+    reg0 += reg1;
+    reg2 += reg3;
+    reg0 = __msa_srai_h(reg0, 8);
+    reg2 = __msa_srai_h(reg2, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0);
+    res0 = __msa_copy_u_d((v2i64)dst0, 0);
+    res1 = __msa_copy_u_d((v2i64)dst0, 1);
+    SD(res0, dst_u);
+    SD(res1, dst_v);
+    t += 48;
+    s += 48;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void NV12ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  uint64_t val0, val1;
+  v16u8 src0, src1, res0, res1, dst0, dst1;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v16u8 zero = {0};
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    val0 = LD(src_y);
+    val1 = LD(src_uv);
+    src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
+    src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    res0 = (v16u8)__msa_ilvev_b((v16i8)vec2, (v16i8)vec0);
+    res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)vec1);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)res1, (v16i8)res0);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_y += 8;
+    src_uv += 8;
+    dst_argb += 32;
+  }
+}
+
+void NV12ToRGB565Row_MSA(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  int x;
+  uint64_t val0, val1;
+  v16u8 src0, src1, dst0;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+  v16u8 zero = {0};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    val0 = LD(src_y);
+    val1 = LD(src_uv);
+    src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
+    src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    vec0 = vec0 >> 3;
+    vec1 = (vec1 >> 2) << 5;
+    vec2 = (vec2 >> 3) << 11;
+    dst0 = (v16u8)(vec0 | vec1 | vec2);
+    ST_UB(dst0, dst_rgb565);
+    src_y += 8;
+    src_uv += 8;
+    dst_rgb565 += 16;
+  }
+}
+
+void NV21ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  uint64_t val0, val1;
+  v16u8 src0, src1, res0, res1, dst0, dst1;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v16u8 zero = {0};
+  v16i8 shuffler = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    val0 = LD(src_y);
+    val1 = LD(src_vu);
+    src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
+    src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
+    src1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src1, (v16i8)src1);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    res0 = (v16u8)__msa_ilvev_b((v16i8)vec2, (v16i8)vec0);
+    res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)vec1);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)res1, (v16i8)res0);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_y += 8;
+    src_vu += 8;
+    dst_argb += 32;
+  }
+}
+
+void SobelRow_MSA(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width) {
+  int x;
+  v16u8 src0, src1, vec0, dst0, dst1, dst2, dst3;
+  v16i8 mask0 = {0, 0, 0, 16, 1, 1, 1, 16, 2, 2, 2, 16, 3, 3, 3, 16};
+  v16i8 const_0x4 = __msa_ldi_b(0x4);
+  v16i8 mask1 = mask0 + const_0x4;
+  v16i8 mask2 = mask1 + const_0x4;
+  v16i8 mask3 = mask2 + const_0x4;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_sobelx, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_sobely, 0);
+    vec0 = __msa_adds_u_b(src0, src1);
+    dst0 = (v16u8)__msa_vshf_b(mask0, (v16i8)alpha, (v16i8)vec0);
+    dst1 = (v16u8)__msa_vshf_b(mask1, (v16i8)alpha, (v16i8)vec0);
+    dst2 = (v16u8)__msa_vshf_b(mask2, (v16i8)alpha, (v16i8)vec0);
+    dst3 = (v16u8)__msa_vshf_b(mask3, (v16i8)alpha, (v16i8)vec0);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void SobelToPlaneRow_MSA(const uint8_t* src_sobelx,
+                         const uint8_t* src_sobely,
+                         uint8_t* dst_y,
+                         int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((void*)src_sobelx, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_sobelx, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_sobely, 0);
+    src3 = (v16u8)__msa_ld_b((void*)src_sobely, 16);
+    dst0 = __msa_adds_u_b(src0, src2);
+    dst1 = __msa_adds_u_b(src1, src3);
+    ST_UB2(dst0, dst1, dst_y, 16);
+    src_sobelx += 32;
+    src_sobely += 32;
+    dst_y += 32;
+  }
+}
+
+void SobelXYRow_MSA(const uint8_t* src_sobelx,
+                    const uint8_t* src_sobely,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  v16u8 src0, src1, vec0, vec1, vec2;
+  v16u8 reg0, reg1, dst0, dst1, dst2, dst3;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_sobelx, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_sobely, 0);
+    vec0 = __msa_adds_u_b(src0, src1);
+    vec1 = (v16u8)__msa_ilvr_b((v16i8)src0, (v16i8)src1);
+    vec2 = (v16u8)__msa_ilvl_b((v16i8)src0, (v16i8)src1);
+    reg0 = (v16u8)__msa_ilvr_b((v16i8)alpha, (v16i8)vec0);
+    reg1 = (v16u8)__msa_ilvl_b((v16i8)alpha, (v16i8)vec0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)reg0, (v16i8)vec1);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)reg0, (v16i8)vec1);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)reg1, (v16i8)vec2);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)reg1, (v16i8)vec2);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void ARGBToYJRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0;
+  v16u8 const_0x961D = (v16u8)__msa_fill_h(0x961D);
+  v16u8 const_0x4D = (v16u8)__msa_fill_h(0x4D);
+  v8u16 const_0x80 = (v8u16)__msa_fill_h(0x80);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    ARGBTOY(src0, src1, src2, src3, const_0x961D, const_0x4D, const_0x80, 8,
+            dst0);
+    ST_UB(dst0, dst_y);
+    src_argb += 64;
+    dst_y += 16;
+  }
+}
+
+void BGRAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0;
+  v16u8 const_0x4200 = (v16u8)__msa_fill_h(0x4200);
+  v16u8 const_0x1981 = (v16u8)__msa_fill_h(0x1981);
+  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    ARGBTOY(src0, src1, src2, src3, const_0x4200, const_0x1981, const_0x1080, 8,
+            dst0);
+    ST_UB(dst0, dst_y);
+    src_argb += 64;
+    dst_y += 16;
+  }
+}
+
+void ABGRToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0;
+  v16u8 const_0x8142 = (v16u8)__msa_fill_h(0x8142);
+  v16u8 const_0x19 = (v16u8)__msa_fill_h(0x19);
+  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    ARGBTOY(src0, src1, src2, src3, const_0x8142, const_0x19, const_0x1080, 8,
+            dst0);
+    ST_UB(dst0, dst_y);
+    src_argb += 64;
+    dst_y += 16;
+  }
+}
+
+void RGBAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0;
+  v16u8 const_0x1900 = (v16u8)__msa_fill_h(0x1900);
+  v16u8 const_0x4281 = (v16u8)__msa_fill_h(0x4281);
+  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    ARGBTOY(src0, src1, src2, src3, const_0x1900, const_0x4281, const_0x1080, 8,
+            dst0);
+    ST_UB(dst0, dst_y);
+    src_argb += 64;
+    dst_y += 16;
+  }
+}
+
+void ARGBToUVJRow_MSA(const uint8_t* src_rgb,
+                      int src_stride_rgb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  v8u16 src0, src1, src2, src3, src4, src5, src6, src7;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 dst0, dst1, dst2, dst3;
+  v16u8 zero = {0};
+  v8i16 shuffler0 = {0, 3, 4, 7, 8, 11, 12, 15};
+  v8i16 shuffler1 = {1, 2, 5, 6, 9, 10, 13, 14};
+  v8i16 shuffler2 = {2, 3, 6, 7, 10, 11, 14, 15};
+  v8i16 shuffler3 = {0, 1, 4, 5, 8, 9, 12, 13};
+  v8u16 const_0x0000003f = (v8u16)__msa_fill_w(0x0000003f);
+  v4u32 const_0x00008080 = (v8u16)__msa_fill_w(0x00008080);
+  v8u16 const_0x0015002a = (v8u16)__msa_fill_w(0x0015002a);
+  v8u16 const_0x0035000a = (v8u16)__msa_fill_w(0x0035000a);
+  v4i32 shift = __msa_fill_w(0x00000008);
+
+  for (x = 0; x < width; x += 32) {
+    src1 = __msa_ld_b((void*)s, 0);
+    src3 = __msa_ld_b((void*)s, 16);
+    src5 = __msa_ld_b((void*)t, 0);
+    src7 = __msa_ld_b((void*)t, 16);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec0 = __msa_aver_u_h(src4, src5);
+    vec1 = __msa_aver_u_h(src6, src7);
+
+    src1 = __msa_ld_b((void*)s, 32);
+    src3 = __msa_ld_b((void*)s, 48);
+    src5 = __msa_ld_b((void*)t, 32);
+    src7 = __msa_ld_b((void*)t, 48);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec2 = __msa_aver_u_h(src4, src5);
+    vec3 = __msa_aver_u_h(src6, src7);
+    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x0000003f, const_0x00008080,
+             const_0x0015002a, const_0x0035000a, shuffler0, shuffler1,
+             shuffler2, shuffler3, shift, dst0, dst1);
+
+    src1 = __msa_ld_b((void*)s, 64);
+    src3 = __msa_ld_b((void*)s, 80);
+    src5 = __msa_ld_b((void*)t, 64);
+    src7 = __msa_ld_b((void*)t, 80);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec0 = __msa_aver_u_h(src4, src5);
+    vec1 = __msa_aver_u_h(src6, src7);
+
+    src1 = __msa_ld_b((void*)s, 96);
+    src3 = __msa_ld_b((void*)s, 112);
+    src5 = __msa_ld_b((void*)t, 96);
+    src7 = __msa_ld_b((void*)t, 112);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec2 = __msa_aver_u_h(src4, src5);
+    vec3 = __msa_aver_u_h(src6, src7);
+    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x0000003f, const_0x00008080,
+             const_0x0015002a, const_0x0035000a, shuffler0, shuffler1,
+             shuffler2, shuffler3, shift, dst2, dst3);
+
+    dst0 = (v8u16)__msa_pckev_b(dst2, dst0);
+    dst1 = (v8u16)__msa_pckev_b(dst3, dst1);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    s += 128;
+    t += 128;
+    dst_v += 16;
+    dst_u += 16;
+  }
+}
+
+void BGRAToUVRow_MSA(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  const uint8_t unused = 0xf;
+  v8u16 src0, src1, src2, src3;
+  v16u8 dst0, dst1;
+  v8i16 shuffler0 = {1, unused, 5, unused, 9, unused, 13, unused};
+  v8i16 shuffler1 = {2, 3, 6, 7, 10, 11, 14, 15};
+  v8i16 shuffler2 = {3, unused, 7, unused, 11, unused, 15, unused};
+  v8i16 shuffler3 = {1, 2, 5, 6, 9, 10, 13, 14};
+  v8u16 const_0x09002f = (v8u16)__msa_fill_w(0x09002f);
+  v8u16 const_0x000038 = (v8u16)__msa_fill_w(0x0038);
+  v8u16 const_0x250013 = (v8u16)__msa_fill_w(0x250013);
+  v4u32 const_0x008080 = (v4u32)__msa_fill_w(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
+
+  for (x = 0; x < width; x += 16) {
+    READ_ARGB(s, t, src0, src1, src2, src3, const_0x0001);
+    ARGBTOUV_H(src0, src1, src2, src3, const_0x09002f, const_0x000038,
+               const_0x250013, const_0x008080, shuffler0, shuffler1, shuffler2,
+               shuffler3, dst0, dst1);
+    *((uint64_t*)dst_v) = __msa_copy_u_d((v2i64)dst0, 0);
+    *((uint64_t*)dst_u) = __msa_copy_u_d((v2i64)dst1, 0);
+    s += 64;
+    t += 64;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void ABGRToUVRow_MSA(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  const uint8_t unused = 0xf;
+  v8u16 src0, src1, src2, src3;
+  v16u8 dst0, dst1;
+  v8i16 shuffler0 = {0, unused, 4, unused, 8, unused, 12, unused};
+  v8i16 shuffler1 = {1, 2, 5, 6, 9, 10, 13, 14};
+  v8i16 shuffler2 = {2, unused, 6, unused, 10, unused, 14, unused};
+  v8i16 shuffler3 = {0, 1, 4, 5, 8, 9, 12, 13};
+  v8u16 const_0x09002f = (v8u16)__msa_fill_w(0x09002f);
+  v8u16 const_0x000038 = (v8u16)__msa_fill_w(0x0038);
+  v8u16 const_0x250013 = (v8u16)__msa_fill_w(0x250013);
+  v4u32 const_0x008080 = (v4u32)__msa_fill_w(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
+
+  for (x = 0; x < width; x += 16) {
+    READ_ARGB(s, t, src0, src1, src2, src3, const_0x0001);
+    ARGBTOUV_H(src0, src1, src2, src3, const_0x09002f, const_0x000038,
+               const_0x250013, const_0x008080, shuffler0, shuffler1, shuffler2,
+               shuffler3, dst0, dst1);
+    *((uint64_t*)dst_v) = __msa_copy_u_d((v2i64)dst0, 0);
+    *((uint64_t*)dst_u) = __msa_copy_u_d((v2i64)dst1, 0);
+    s += 64;
+    t += 64;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void RGBAToUVRow_MSA(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  const uint8_t unused = 0xf;
+  v8u16 src0, src1, src2, src3;
+  v16u8 dst0, dst1;
+  v8i16 shuffler0 = {3, unused, 7, unused, 11, unused, 15, unused};
+  v8i16 shuffler1 = {2, 1, 6, 5, 10, 9, 14, 13};
+  v8i16 shuffler2 = {1, unused, 5, unused, 9, unused, 13, unused};
+  v8i16 shuffler3 = {3, 2, 7, 6, 11, 10, 15, 14};
+  v8u16 const_0x09002f = (v8u16)__msa_fill_w(0x09002f);
+  v8u16 const_0x000038 = (v8u16)__msa_fill_w(0x0038);
+  v8u16 const_0x250013 = (v8u16)__msa_fill_w(0x250013);
+  v4u32 const_0x008080 = (v4u32)__msa_fill_w(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
+
+  for (x = 0; x < width; x += 16) {
+    READ_ARGB(s, t, src0, src1, src2, src3, const_0x0001);
+    ARGBTOUV_H(src0, src1, src2, src3, const_0x09002f, const_0x000038,
+               const_0x250013, const_0x008080, shuffler0, shuffler1, shuffler2,
+               shuffler3, dst0, dst1);
+    *((uint64_t*)dst_v) = __msa_copy_u_d((v2i64)dst0, 0);
+    *((uint64_t*)dst_u) = __msa_copy_u_d((v2i64)dst1, 0);
+    s += 64;
+    t += 64;
+    dst_u += 8;
+    dst_v += 8;
+  }
+}
+
+void I444ToARGBRow_MSA(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  v16u8 src0, src1, src2, dst0, dst1;
+  v8i16 vec0, vec1, vec2;
+  v4i32 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, reg8, reg9;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 zero = {0};
+  v4i32 const_0x80 = __msa_fill_w(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+
+  for (x = 0; x < width; x += 8) {
+    READI444(src_y, src_u, src_v, src0, src1, src2);
+    vec0 = (v8i16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    reg0 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
+    reg1 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
+    reg0 *= vec_yg;
+    reg1 *= vec_yg;
+    reg0 = __msa_srai_w(reg0, 16);
+    reg1 = __msa_srai_w(reg1, 16);
+    reg0 += vec_yb;
+    reg1 += vec_yb;
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
+    vec1 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src2);
+    reg6 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
+    reg7 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
+    reg8 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec1);
+    reg9 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec1);
+    reg6 -= const_0x80;
+    reg7 -= const_0x80;
+    reg8 -= const_0x80;
+    reg9 -= const_0x80;
+    tmp0 = reg0 + reg6 * vec_ub;
+    tmp1 = reg1 + reg7 * vec_ub;
+    tmp2 = reg0 + reg8 * vec_vr;
+    tmp3 = reg1 + reg9 * vec_vr;
+    tmp4 = reg6 * vec_ug;
+    tmp5 = reg7 * vec_ug;
+    tmp4 += reg8 * vec_vg;
+    tmp5 += reg9 * vec_vg;
+    tmp4 = reg0 - tmp4;
+    tmp5 = reg1 - tmp5;
+    reg0 = __msa_srai_w(tmp0, 6);
+    reg1 = __msa_srai_w(tmp1, 6);
+    reg2 = __msa_srai_w(tmp2, 6);
+    reg3 = __msa_srai_w(tmp3, 6);
+    reg4 = __msa_srai_w(tmp4, 6);
+    reg5 = __msa_srai_w(tmp5, 6);
+    CLIP_0TO255(reg0, reg1, reg2, reg3, reg4, reg5);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg5, (v8i16)reg4);
+    vec2 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec0 = (v8u16)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
+    vec1 = (v8u16)__msa_ilvev_b((v16i8)alpha, (v16i8)vec2);
+    dst0 = (v16u8)__msa_ilvr_h((v8i16)vec1, (v8i16)vec0);
+    dst1 = (v16u8)__msa_ilvl_h((v8i16)vec1, (v8i16)vec0);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_y += 8;
+    src_u += 8;
+    src_v += 8;
+    dst_argb += 32;
+  }
+}
+
+// TODO - respect YuvConstants
+void I400ToARGBRow_MSA(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+#if defined(__aarch64__) || defined(__arm__)
+  int ygb = yuvconstants->kUVBiasBGR[3];
+  int yg = yuvconstants->kYToRgb[1];
+#else
+  int ygb = yuvconstants->kYBiasToRgb[0];
+  int yg = yuvconstants->kYToRgb[0];
+#endif
+  v16u8 src0, res0, res1, res2, res3, res4, dst0, dst1, dst2, dst3;
+  v8i16 vec0, vec1;
+  v4i32 reg0, reg1, reg2, reg3;
+  v4i32 vec_yg = __msa_fill_w(yg);
+  v8i16 vec_ygb = __msa_fill_h(ygb);
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 max = __msa_ldi_h(0xFF);
+  v8i16 zero = {0};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_y, 0);
+    vec0 = (v8i16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8i16)__msa_ilvl_b((v16i8)src0, (v16i8)src0);
+    reg0 = (v4i32)__msa_ilvr_h(zero, vec0);
+    reg1 = (v4i32)__msa_ilvl_h(zero, vec0);
+    reg2 = (v4i32)__msa_ilvr_h(zero, vec1);
+    reg3 = (v4i32)__msa_ilvl_h(zero, vec1);
+    reg0 *= vec_yg;
+    reg1 *= vec_yg;
+    reg2 *= vec_yg;
+    reg3 *= vec_yg;
+    reg0 = __msa_srai_w(reg0, 16);
+    reg1 = __msa_srai_w(reg1, 16);
+    reg2 = __msa_srai_w(reg2, 16);
+    reg3 = __msa_srai_w(reg3, 16);
+    vec0 = (v8i16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8i16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec0 += vec_ygb;
+    vec1 += vec_ygb;
+    vec0 = __msa_srai_h(vec0, 6);
+    vec1 = __msa_srai_h(vec1, 6);
+    vec0 = __msa_maxi_s_h(vec0, 0);
+    vec1 = __msa_maxi_s_h(vec1, 0);
+    vec0 = __msa_min_s_h(max, vec0);
+    vec1 = __msa_min_s_h(max, vec1);
+    res0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    res1 = (v16u8)__msa_ilvr_b((v16i8)res0, (v16i8)res0);
+    res2 = (v16u8)__msa_ilvl_b((v16i8)res0, (v16i8)res0);
+    res3 = (v16u8)__msa_ilvr_b((v16i8)alpha, (v16i8)res0);
+    res4 = (v16u8)__msa_ilvl_b((v16i8)alpha, (v16i8)res0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)res3, (v16i8)res1);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)res3, (v16i8)res1);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)res4, (v16i8)res2);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)res4, (v16i8)res2);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_y += 16;
+    dst_argb += 64;
+  }
+}
+
+void J400ToARGBRow_MSA(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  int x;
+  v16u8 src0, vec0, vec1, vec2, vec3, dst0, dst1, dst2, dst3;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_y, 0);
+    vec0 = (v16u8)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_ilvl_b((v16i8)src0, (v16i8)src0);
+    vec2 = (v16u8)__msa_ilvr_b((v16i8)alpha, (v16i8)src0);
+    vec3 = (v16u8)__msa_ilvl_b((v16i8)alpha, (v16i8)src0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)vec2, (v16i8)vec0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)vec2, (v16i8)vec0);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec1);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)vec3, (v16i8)vec1);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_y += 16;
+    dst_argb += 64;
+  }
+}
+
+void YUY2ToARGBRow_MSA(const uint8_t* src_yuy2,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  v16u8 src0, src1, src2;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_yuy2, 0);
+    src1 = (v16u8)__msa_pckev_b((v16i8)src0, (v16i8)src0);
+    src2 = (v16u8)__msa_pckod_b((v16i8)src0, (v16i8)src0);
+    YUVTORGB(src1, src2, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    STOREARGB(vec0, vec1, vec2, alpha, dst_argb);
+    src_yuy2 += 16;
+    dst_argb += 32;
+  }
+}
+
+void UYVYToARGBRow_MSA(const uint8_t* src_uyvy,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  v16u8 src0, src1, src2;
+  v8i16 vec0, vec1, vec2;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
+  v4i32 vec_ubvr, vec_ugvg;
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
+  v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_uyvy, 0);
+    src1 = (v16u8)__msa_pckod_b((v16i8)src0, (v16i8)src0);
+    src2 = (v16u8)__msa_pckev_b((v16i8)src0, (v16i8)src0);
+    YUVTORGB(src1, src2, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    STOREARGB(vec0, vec1, vec2, alpha, dst_argb);
+    src_uyvy += 16;
+    dst_argb += 32;
+  }
+}
+
+void InterpolateRow_MSA(uint8_t* dst_ptr,
+                        const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int32_t source_y_fraction) {
+  int32_t y1_fraction = source_y_fraction;
+  int32_t y0_fraction = 256 - y1_fraction;
+  uint16_t y_fractions;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3, y_frac;
+
+  if (0 == y1_fraction) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+
+  if (128 == y1_fraction) {
+    for (x = 0; x < width; x += 32) {
+      src0 = (v16u8)__msa_ld_b((void*)s, 0);
+      src1 = (v16u8)__msa_ld_b((void*)s, 16);
+      src2 = (v16u8)__msa_ld_b((void*)t, 0);
+      src3 = (v16u8)__msa_ld_b((void*)t, 16);
+      dst0 = __msa_aver_u_b(src0, src2);
+      dst1 = __msa_aver_u_b(src1, src3);
+      ST_UB2(dst0, dst1, dst_ptr, 16);
+      s += 32;
+      t += 32;
+      dst_ptr += 32;
+    }
+    return;
+  }
+
+  y_fractions = (uint16_t)(y0_fraction + (y1_fraction << 8));
+  y_frac = (v8u16)__msa_fill_h(y_fractions);
+
+  for (x = 0; x < width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((void*)s, 0);
+    src1 = (v16u8)__msa_ld_b((void*)s, 16);
+    src2 = (v16u8)__msa_ld_b((void*)t, 0);
+    src3 = (v16u8)__msa_ld_b((void*)t, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = (v8u16)__msa_dotp_u_h((v16u8)vec0, (v16u8)y_frac);
+    vec1 = (v8u16)__msa_dotp_u_h((v16u8)vec1, (v16u8)y_frac);
+    vec2 = (v8u16)__msa_dotp_u_h((v16u8)vec2, (v16u8)y_frac);
+    vec3 = (v8u16)__msa_dotp_u_h((v16u8)vec3, (v16u8)y_frac);
+    vec0 = (v8u16)__msa_srari_h((v8i16)vec0, 8);
+    vec1 = (v8u16)__msa_srari_h((v8i16)vec1, 8);
+    vec2 = (v8u16)__msa_srari_h((v8i16)vec2, 8);
+    vec3 = (v8u16)__msa_srari_h((v8i16)vec3, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    ST_UB2(dst0, dst1, dst_ptr, 16);
+    s += 32;
+    t += 32;
+    dst_ptr += 32;
+  }
+}
+
+void ARGBSetRow_MSA(uint8_t* dst_argb, uint32_t v32, int width) {
+  int x;
+  v4i32 dst0 = __builtin_msa_fill_w(v32);
+
+  for (x = 0; x < width; x += 4) {
+    ST_UB(dst0, dst_argb);
+    dst_argb += 16;
+  }
+}
+
+void RAWToRGB24Row_MSA(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2;
+  v16i8 shuffler0 = {2, 1, 0, 5, 4, 3, 8, 7, 6, 11, 10, 9, 14, 13, 12, 17};
+  v16i8 shuffler1 = {8,  7,  12, 11, 10, 15, 14, 13,
+                     18, 17, 16, 21, 20, 19, 24, 23};
+  v16i8 shuffler2 = {14, 19, 18, 17, 22, 21, 20, 25,
+                     24, 23, 28, 27, 26, 31, 30, 29};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_raw, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_raw, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_raw, 32);
+    src3 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 8);
+    src4 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8);
+    dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src4, (v16i8)src3);
+    dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src2, (v16i8)src1);
+    ST_UB2(dst0, dst1, dst_rgb24, 16);
+    ST_UB(dst2, (dst_rgb24 + 32));
+    src_raw += 48;
+    dst_rgb24 += 48;
+  }
+}
+
+void MergeUVRow_MSA(const uint8_t* src_u,
+                    const uint8_t* src_v,
+                    uint8_t* dst_uv,
+                    int width) {
+  int x;
+  v16u8 src0, src1, dst0, dst1;
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_u, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_v, 0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)src1, (v16i8)src0);
+    ST_UB2(dst0, dst1, dst_uv, 16);
+    src_u += 16;
+    src_v += 16;
+    dst_uv += 32;
+  }
+}
+
+void ARGBExtractAlphaRow_MSA(const uint8_t* src_argb,
+                             uint8_t* dst_a,
+                             int width) {
+  int i;
+  v16u8 src0, src1, src2, src3, vec0, vec1, dst0;
+
+  for (i = 0; i < width; i += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    vec0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_a);
+    src_argb += 64;
+    dst_a += 16;
+  }
+}
+
+void ARGBBlendRow_MSA(const uint8_t* src_argb,
+                      const uint8_t* src_argb1,
+                      uint8_t* dst_argb,
+                      int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 vec8, vec9, vec10, vec11, vec12, vec13;
+  v8u16 const_256 = (v8u16)__msa_ldi_h(256);
+  v16u8 const_255 = (v16u8)__msa_ldi_b(255);
+  v16u8 mask = {0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255};
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb1, 16);
+    vec0 = (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b(zero, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b(zero, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b(zero, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b(zero, (v16i8)src2);
+    vec5 = (v8u16)__msa_ilvl_b(zero, (v16i8)src2);
+    vec6 = (v8u16)__msa_ilvr_b(zero, (v16i8)src3);
+    vec7 = (v8u16)__msa_ilvl_b(zero, (v16i8)src3);
+    vec8 = (v8u16)__msa_fill_h(vec0[3]);
+    vec9 = (v8u16)__msa_fill_h(vec0[7]);
+    vec10 = (v8u16)__msa_fill_h(vec1[3]);
+    vec11 = (v8u16)__msa_fill_h(vec1[7]);
+    vec8 = (v8u16)__msa_pckev_d((v2i64)vec9, (v2i64)vec8);
+    vec9 = (v8u16)__msa_pckev_d((v2i64)vec11, (v2i64)vec10);
+    vec10 = (v8u16)__msa_fill_h(vec2[3]);
+    vec11 = (v8u16)__msa_fill_h(vec2[7]);
+    vec12 = (v8u16)__msa_fill_h(vec3[3]);
+    vec13 = (v8u16)__msa_fill_h(vec3[7]);
+    vec10 = (v8u16)__msa_pckev_d((v2i64)vec11, (v2i64)vec10);
+    vec11 = (v8u16)__msa_pckev_d((v2i64)vec13, (v2i64)vec12);
+    vec8 = const_256 - vec8;
+    vec9 = const_256 - vec9;
+    vec10 = const_256 - vec10;
+    vec11 = const_256 - vec11;
+    vec8 *= vec4;
+    vec9 *= vec5;
+    vec10 *= vec6;
+    vec11 *= vec7;
+    vec8 = (v8u16)__msa_srai_h((v8i16)vec8, 8);
+    vec9 = (v8u16)__msa_srai_h((v8i16)vec9, 8);
+    vec10 = (v8u16)__msa_srai_h((v8i16)vec10, 8);
+    vec11 = (v8u16)__msa_srai_h((v8i16)vec11, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)vec9, (v16i8)vec8);
+    dst3 = (v16u8)__msa_pckev_b((v16i8)vec11, (v16i8)vec10);
+    dst0 = (v16u8)__msa_adds_u_b(dst0, dst2);
+    dst1 = (v16u8)__msa_adds_u_b(dst1, dst3);
+    dst0 = __msa_bmnz_v(dst0, const_255, mask);
+    dst1 = __msa_bmnz_v(dst1, const_255, mask);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBQuantizeRow_MSA(uint8_t* dst_argb,
+                         int scale,
+                         int interval_size,
+                         int interval_offset,
+                         int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v4i32 tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  v4i32 vec_scale = __msa_fill_w(scale);
+  v16u8 vec_int_sz = (v16u8)__msa_fill_b(interval_size);
+  v16u8 vec_int_ofst = (v16u8)__msa_fill_b(interval_offset);
+  v16i8 mask = {0, 1, 2, 19, 4, 5, 6, 23, 8, 9, 10, 27, 12, 13, 14, 31};
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)dst_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)dst_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)dst_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)dst_argb, 48);
+    vec0 = (v8i16)__msa_ilvr_b(zero, (v16i8)src0);
+    vec1 = (v8i16)__msa_ilvl_b(zero, (v16i8)src0);
+    vec2 = (v8i16)__msa_ilvr_b(zero, (v16i8)src1);
+    vec3 = (v8i16)__msa_ilvl_b(zero, (v16i8)src1);
+    vec4 = (v8i16)__msa_ilvr_b(zero, (v16i8)src2);
+    vec5 = (v8i16)__msa_ilvl_b(zero, (v16i8)src2);
+    vec6 = (v8i16)__msa_ilvr_b(zero, (v16i8)src3);
+    vec7 = (v8i16)__msa_ilvl_b(zero, (v16i8)src3);
+    tmp0 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
+    tmp1 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
+    tmp2 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec1);
+    tmp3 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec1);
+    tmp4 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec2);
+    tmp5 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec2);
+    tmp6 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec3);
+    tmp7 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec3);
+    tmp8 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec4);
+    tmp9 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec4);
+    tmp10 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec5);
+    tmp11 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec5);
+    tmp12 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec6);
+    tmp13 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec6);
+    tmp14 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec7);
+    tmp15 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec7);
+    tmp0 *= vec_scale;
+    tmp1 *= vec_scale;
+    tmp2 *= vec_scale;
+    tmp3 *= vec_scale;
+    tmp4 *= vec_scale;
+    tmp5 *= vec_scale;
+    tmp6 *= vec_scale;
+    tmp7 *= vec_scale;
+    tmp8 *= vec_scale;
+    tmp9 *= vec_scale;
+    tmp10 *= vec_scale;
+    tmp11 *= vec_scale;
+    tmp12 *= vec_scale;
+    tmp13 *= vec_scale;
+    tmp14 *= vec_scale;
+    tmp15 *= vec_scale;
+    tmp0 >>= 16;
+    tmp1 >>= 16;
+    tmp2 >>= 16;
+    tmp3 >>= 16;
+    tmp4 >>= 16;
+    tmp5 >>= 16;
+    tmp6 >>= 16;
+    tmp7 >>= 16;
+    tmp8 >>= 16;
+    tmp9 >>= 16;
+    tmp10 >>= 16;
+    tmp11 >>= 16;
+    tmp12 >>= 16;
+    tmp13 >>= 16;
+    tmp14 >>= 16;
+    tmp15 >>= 16;
+    vec0 = (v8i16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8i16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    vec2 = (v8i16)__msa_pckev_h((v8i16)tmp5, (v8i16)tmp4);
+    vec3 = (v8i16)__msa_pckev_h((v8i16)tmp7, (v8i16)tmp6);
+    vec4 = (v8i16)__msa_pckev_h((v8i16)tmp9, (v8i16)tmp8);
+    vec5 = (v8i16)__msa_pckev_h((v8i16)tmp11, (v8i16)tmp10);
+    vec6 = (v8i16)__msa_pckev_h((v8i16)tmp13, (v8i16)tmp12);
+    vec7 = (v8i16)__msa_pckev_h((v8i16)tmp15, (v8i16)tmp14);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4);
+    dst3 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6);
+    dst0 *= vec_int_sz;
+    dst1 *= vec_int_sz;
+    dst2 *= vec_int_sz;
+    dst3 *= vec_int_sz;
+    dst0 += vec_int_ofst;
+    dst1 += vec_int_ofst;
+    dst2 += vec_int_ofst;
+    dst3 += vec_int_ofst;
+    dst0 = (v16u8)__msa_vshf_b(mask, (v16i8)src0, (v16i8)dst0);
+    dst1 = (v16u8)__msa_vshf_b(mask, (v16i8)src1, (v16i8)dst1);
+    dst2 = (v16u8)__msa_vshf_b(mask, (v16i8)src2, (v16i8)dst2);
+    dst3 = (v16u8)__msa_vshf_b(mask, (v16i8)src3, (v16i8)dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    dst_argb += 64;
+  }
+}
+
+void ARGBColorMatrixRow_MSA(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const int8_t* matrix_argb,
+                            int width) {
+  int32_t x;
+  v16i8 src0;
+  v16u8 src1, src2, dst0, dst1;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v8i16 vec10, vec11, vec12, vec13, vec14, vec15, vec16, vec17;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v4i32 tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  v16i8 zero = {0};
+  v8i16 max = __msa_ldi_h(255);
+
+  src0 = __msa_ld_b((void*)matrix_argb, 0);
+  vec0 = (v8i16)__msa_ilvr_b(zero, src0);
+  vec1 = (v8i16)__msa_ilvl_b(zero, src0);
+
+  for (x = 0; x < width; x += 8) {
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    vec2 = (v8i16)__msa_ilvr_b(zero, (v16i8)src1);
+    vec3 = (v8i16)__msa_ilvl_b(zero, (v16i8)src1);
+    vec4 = (v8i16)__msa_ilvr_b(zero, (v16i8)src2);
+    vec5 = (v8i16)__msa_ilvl_b(zero, (v16i8)src2);
+    vec6 = (v8i16)__msa_pckod_d((v2i64)vec2, (v2i64)vec2);
+    vec7 = (v8i16)__msa_pckod_d((v2i64)vec3, (v2i64)vec3);
+    vec8 = (v8i16)__msa_pckod_d((v2i64)vec4, (v2i64)vec4);
+    vec9 = (v8i16)__msa_pckod_d((v2i64)vec5, (v2i64)vec5);
+    vec2 = (v8i16)__msa_pckev_d((v2i64)vec2, (v2i64)vec2);
+    vec3 = (v8i16)__msa_pckev_d((v2i64)vec3, (v2i64)vec3);
+    vec4 = (v8i16)__msa_pckev_d((v2i64)vec4, (v2i64)vec4);
+    vec5 = (v8i16)__msa_pckev_d((v2i64)vec5, (v2i64)vec5);
+    vec10 = vec2 * vec0;
+    vec11 = vec2 * vec1;
+    vec12 = vec6 * vec0;
+    vec13 = vec6 * vec1;
+    tmp0 = __msa_hadd_s_w(vec10, vec10);
+    tmp1 = __msa_hadd_s_w(vec11, vec11);
+    tmp2 = __msa_hadd_s_w(vec12, vec12);
+    tmp3 = __msa_hadd_s_w(vec13, vec13);
+    vec14 = vec3 * vec0;
+    vec15 = vec3 * vec1;
+    vec16 = vec7 * vec0;
+    vec17 = vec7 * vec1;
+    tmp4 = __msa_hadd_s_w(vec14, vec14);
+    tmp5 = __msa_hadd_s_w(vec15, vec15);
+    tmp6 = __msa_hadd_s_w(vec16, vec16);
+    tmp7 = __msa_hadd_s_w(vec17, vec17);
+    vec10 = __msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec11 = __msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    vec12 = __msa_pckev_h((v8i16)tmp5, (v8i16)tmp4);
+    vec13 = __msa_pckev_h((v8i16)tmp7, (v8i16)tmp6);
+    tmp0 = __msa_hadd_s_w(vec10, vec10);
+    tmp1 = __msa_hadd_s_w(vec11, vec11);
+    tmp2 = __msa_hadd_s_w(vec12, vec12);
+    tmp3 = __msa_hadd_s_w(vec13, vec13);
+    tmp0 = __msa_srai_w(tmp0, 6);
+    tmp1 = __msa_srai_w(tmp1, 6);
+    tmp2 = __msa_srai_w(tmp2, 6);
+    tmp3 = __msa_srai_w(tmp3, 6);
+    vec2 = vec4 * vec0;
+    vec6 = vec4 * vec1;
+    vec3 = vec8 * vec0;
+    vec7 = vec8 * vec1;
+    tmp8 = __msa_hadd_s_w(vec2, vec2);
+    tmp9 = __msa_hadd_s_w(vec6, vec6);
+    tmp10 = __msa_hadd_s_w(vec3, vec3);
+    tmp11 = __msa_hadd_s_w(vec7, vec7);
+    vec4 = vec5 * vec0;
+    vec8 = vec5 * vec1;
+    vec5 = vec9 * vec0;
+    vec9 = vec9 * vec1;
+    tmp12 = __msa_hadd_s_w(vec4, vec4);
+    tmp13 = __msa_hadd_s_w(vec8, vec8);
+    tmp14 = __msa_hadd_s_w(vec5, vec5);
+    tmp15 = __msa_hadd_s_w(vec9, vec9);
+    vec14 = __msa_pckev_h((v8i16)tmp9, (v8i16)tmp8);
+    vec15 = __msa_pckev_h((v8i16)tmp11, (v8i16)tmp10);
+    vec16 = __msa_pckev_h((v8i16)tmp13, (v8i16)tmp12);
+    vec17 = __msa_pckev_h((v8i16)tmp15, (v8i16)tmp14);
+    tmp4 = __msa_hadd_s_w(vec14, vec14);
+    tmp5 = __msa_hadd_s_w(vec15, vec15);
+    tmp6 = __msa_hadd_s_w(vec16, vec16);
+    tmp7 = __msa_hadd_s_w(vec17, vec17);
+    tmp4 = __msa_srai_w(tmp4, 6);
+    tmp5 = __msa_srai_w(tmp5, 6);
+    tmp6 = __msa_srai_w(tmp6, 6);
+    tmp7 = __msa_srai_w(tmp7, 6);
+    vec10 = __msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec11 = __msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    vec12 = __msa_pckev_h((v8i16)tmp5, (v8i16)tmp4);
+    vec13 = __msa_pckev_h((v8i16)tmp7, (v8i16)tmp6);
+    vec10 = __msa_maxi_s_h(vec10, 0);
+    vec11 = __msa_maxi_s_h(vec11, 0);
+    vec12 = __msa_maxi_s_h(vec12, 0);
+    vec13 = __msa_maxi_s_h(vec13, 0);
+    vec10 = __msa_min_s_h(vec10, max);
+    vec11 = __msa_min_s_h(vec11, max);
+    vec12 = __msa_min_s_h(vec12, max);
+    vec13 = __msa_min_s_h(vec13, max);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec11, (v16i8)vec10);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec13, (v16i8)vec12);
+    ST_UB2(dst0, dst1, dst_argb, 16);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void SplitUVRow_MSA(const uint8_t* src_uv,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((void*)src_uv, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_uv, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_uv, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_uv, 48);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    dst2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    dst3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst_u, 16);
+    ST_UB2(dst2, dst3, dst_v, 16);
+    src_uv += 64;
+    dst_u += 32;
+    dst_v += 32;
+  }
+}
+
+void SetRow_MSA(uint8_t* dst, uint8_t v8, int width) {
+  int x;
+  v16u8 dst0 = (v16u8)__msa_fill_b(v8);
+
+  for (x = 0; x < width; x += 16) {
+    ST_UB(dst0, dst);
+    dst += 16;
+  }
+}
+
+void MirrorSplitUVRow_MSA(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  int x;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16i8 mask0 = {30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0};
+  v16i8 mask1 = {31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11, 9, 7, 5, 3, 1};
+
+  src_uv += (2 * width);
+
+  for (x = 0; x < width; x += 32) {
+    src_uv -= 64;
+    src2 = (v16u8)__msa_ld_b((void*)src_uv, 0);
+    src3 = (v16u8)__msa_ld_b((void*)src_uv, 16);
+    src0 = (v16u8)__msa_ld_b((void*)src_uv, 32);
+    src1 = (v16u8)__msa_ld_b((void*)src_uv, 48);
+    dst0 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    dst2 = (v16u8)__msa_vshf_b(mask0, (v16i8)src1, (v16i8)src0);
+    dst3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst_v, 16);
+    ST_UB2(dst2, dst3, dst_u, 16);
+    dst_u += 32;
+    dst_v += 32;
+  }
+}
+
+void SobelXRow_MSA(const uint8_t* src_y0,
+                   const uint8_t* src_y1,
+                   const uint8_t* src_y2,
+                   uint8_t* dst_sobelx,
+                   int32_t width) {
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, dst0;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16i8 mask0 = {0, 2, 1, 3, 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7, 9};
+  v16i8 tmp = __msa_ldi_b(8);
+  v16i8 mask1 = mask0 + tmp;
+  v8i16 zero = {0};
+  v8i16 max = __msa_ldi_h(255);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_y0, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_y0, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_y1, 0);
+    src3 = (v16u8)__msa_ld_b((void*)src_y1, 16);
+    src4 = (v16u8)__msa_ld_b((void*)src_y2, 0);
+    src5 = (v16u8)__msa_ld_b((void*)src_y2, 16);
+    vec0 = (v8i16)__msa_vshf_b(mask0, (v16i8)src1, (v16i8)src0);
+    vec1 = (v8i16)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v8i16)__msa_vshf_b(mask0, (v16i8)src3, (v16i8)src2);
+    vec3 = (v8i16)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec4 = (v8i16)__msa_vshf_b(mask0, (v16i8)src5, (v16i8)src4);
+    vec5 = (v8i16)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec0 = (v8i16)__msa_hsub_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = (v8i16)__msa_hsub_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = (v8i16)__msa_hsub_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = (v8i16)__msa_hsub_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = (v8i16)__msa_hsub_u_h((v16u8)vec4, (v16u8)vec4);
+    vec5 = (v8i16)__msa_hsub_u_h((v16u8)vec5, (v16u8)vec5);
+    vec0 += vec2;
+    vec1 += vec3;
+    vec4 += vec2;
+    vec5 += vec3;
+    vec0 += vec4;
+    vec1 += vec5;
+    vec0 = __msa_add_a_h(zero, vec0);
+    vec1 = __msa_add_a_h(zero, vec1);
+    vec0 = __msa_maxi_s_h(vec0, 0);
+    vec1 = __msa_maxi_s_h(vec1, 0);
+    vec0 = __msa_min_s_h(max, vec0);
+    vec1 = __msa_min_s_h(max, vec1);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_sobelx);
+    src_y0 += 16;
+    src_y1 += 16;
+    src_y2 += 16;
+    dst_sobelx += 16;
+  }
+}
+
+void SobelYRow_MSA(const uint8_t* src_y0,
+                   const uint8_t* src_y1,
+                   uint8_t* dst_sobely,
+                   int32_t width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6;
+  v8i16 zero = {0};
+  v8i16 max = __msa_ldi_h(255);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((void*)src_y0, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_y1, 0);
+    vec0 = (v8i16)__msa_ilvr_b((v16i8)zero, (v16i8)src0);
+    vec1 = (v8i16)__msa_ilvl_b((v16i8)zero, (v16i8)src0);
+    vec2 = (v8i16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
+    vec3 = (v8i16)__msa_ilvl_b((v16i8)zero, (v16i8)src1);
+    vec0 -= vec2;
+    vec1 -= vec3;
+    vec6[0] = src_y0[16] - src_y1[16];
+    vec6[1] = src_y0[17] - src_y1[17];
+    vec2 = (v8i16)__msa_sldi_b((v16i8)vec1, (v16i8)vec0, 2);
+    vec3 = (v8i16)__msa_sldi_b((v16i8)vec6, (v16i8)vec1, 2);
+    vec4 = (v8i16)__msa_sldi_b((v16i8)vec1, (v16i8)vec0, 4);
+    vec5 = (v8i16)__msa_sldi_b((v16i8)vec6, (v16i8)vec1, 4);
+    vec0 += vec2;
+    vec1 += vec3;
+    vec4 += vec2;
+    vec5 += vec3;
+    vec0 += vec4;
+    vec1 += vec5;
+    vec0 = __msa_add_a_h(zero, vec0);
+    vec1 = __msa_add_a_h(zero, vec1);
+    vec0 = __msa_maxi_s_h(vec0, 0);
+    vec1 = __msa_maxi_s_h(vec1, 0);
+    vec0 = __msa_min_s_h(max, vec0);
+    vec1 = __msa_min_s_h(max, vec1);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_sobely);
+    src_y0 += 16;
+    src_y1 += 16;
+    dst_sobely += 16;
+  }
+}
+
+void HalfFloatRow_MSA(const uint16_t* src,
+                      uint16_t* dst,
+                      float scale,
+                      int width) {
+  int i;
+  v8u16 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  v4u32 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4f32 fvec0, fvec1, fvec2, fvec3, fvec4, fvec5, fvec6, fvec7;
+  v4f32 mult_vec;
+  v8i16 zero = {0};
+  mult_vec[0] = 1.9259299444e-34f * scale;
+  mult_vec = (v4f32)__msa_splati_w((v4i32)mult_vec, 0);
+
+  for (i = 0; i < width; i += 32) {
+    src0 = (v8u16)__msa_ld_h((void*)src, 0);
+    src1 = (v8u16)__msa_ld_h((void*)src, 16);
+    src2 = (v8u16)__msa_ld_h((void*)src, 32);
+    src3 = (v8u16)__msa_ld_h((void*)src, 48);
+    vec0 = (v4u32)__msa_ilvr_h(zero, (v8i16)src0);
+    vec1 = (v4u32)__msa_ilvl_h(zero, (v8i16)src0);
+    vec2 = (v4u32)__msa_ilvr_h(zero, (v8i16)src1);
+    vec3 = (v4u32)__msa_ilvl_h(zero, (v8i16)src1);
+    vec4 = (v4u32)__msa_ilvr_h(zero, (v8i16)src2);
+    vec5 = (v4u32)__msa_ilvl_h(zero, (v8i16)src2);
+    vec6 = (v4u32)__msa_ilvr_h(zero, (v8i16)src3);
+    vec7 = (v4u32)__msa_ilvl_h(zero, (v8i16)src3);
+    fvec0 = __msa_ffint_u_w(vec0);
+    fvec1 = __msa_ffint_u_w(vec1);
+    fvec2 = __msa_ffint_u_w(vec2);
+    fvec3 = __msa_ffint_u_w(vec3);
+    fvec4 = __msa_ffint_u_w(vec4);
+    fvec5 = __msa_ffint_u_w(vec5);
+    fvec6 = __msa_ffint_u_w(vec6);
+    fvec7 = __msa_ffint_u_w(vec7);
+    fvec0 *= mult_vec;
+    fvec1 *= mult_vec;
+    fvec2 *= mult_vec;
+    fvec3 *= mult_vec;
+    fvec4 *= mult_vec;
+    fvec5 *= mult_vec;
+    fvec6 *= mult_vec;
+    fvec7 *= mult_vec;
+    vec0 = ((v4u32)fvec0) >> 13;
+    vec1 = ((v4u32)fvec1) >> 13;
+    vec2 = ((v4u32)fvec2) >> 13;
+    vec3 = ((v4u32)fvec3) >> 13;
+    vec4 = ((v4u32)fvec4) >> 13;
+    vec5 = ((v4u32)fvec5) >> 13;
+    vec6 = ((v4u32)fvec6) >> 13;
+    vec7 = ((v4u32)fvec7) >> 13;
+    dst0 = (v8u16)__msa_pckev_h((v8i16)vec1, (v8i16)vec0);
+    dst1 = (v8u16)__msa_pckev_h((v8i16)vec3, (v8i16)vec2);
+    dst2 = (v8u16)__msa_pckev_h((v8i16)vec5, (v8i16)vec4);
+    dst3 = (v8u16)__msa_pckev_h((v8i16)vec7, (v8i16)vec6);
+    ST_UH2(dst0, dst1, dst, 8);
+    ST_UH2(dst2, dst3, dst + 16, 8);
+    src += 32;
+    dst += 32;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
diff --git a/media/libyuv/source/row_neon.cc b/media/libyuv/source/row_neon.cc
index 909df060c6..37f6db0cd9 100644
--- a/media/libyuv/source/row_neon.cc
+++ b/media/libyuv/source/row_neon.cc
@@ -19,1482 +19,1913 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
     !defined(__aarch64__)
 
+// d8-d15, r4-r11,r14(lr) need to be preserved if used. r13(sp),r15(pc) are
+// reserved.
+
+// q0: Y uint16x8_t
+// d2: U uint8x8_t
+// d3: V uint8x8_t
+
 // Read 8 Y, 4 U and 4 V from 422
-#define READYUV422                                                             \
-    MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    MEMACCESS(1)                                                               \
-    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
-    MEMACCESS(2)                                                               \
-    "vld1.32    {d2[1]}, [%2]!                 \n"
-
-// Read 8 Y, 2 U and 2 V from 422
-#define READYUV411                                                             \
-    MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    MEMACCESS(1)                                                               \
-    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
-    MEMACCESS(2)                                                               \
-    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
-    "vmov.u8    d3, d2                         \n"                             \
-    "vzip.u8    d2, d3                         \n"
+#define READYUV422                               \
+  "vld1.8     {d0}, [%[src_y]]!              \n" \
+  "vld1.32    {d2[0]}, [%[src_u]]!           \n" \
+  "vld1.32    {d2[1]}, [%[src_v]]!           \n" \
+  "vmov.u8    d1, d0                         \n" \
+  "vmovl.u8   q1, d2                         \n" \
+  "vzip.u8    d0, d1                         \n" \
+  "vsli.u16   q1, q1, #8                     \n"
 
 // Read 8 Y, 8 U and 8 V from 444
-#define READYUV444                                                             \
-    MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    MEMACCESS(1)                                                               \
-    "vld1.8     {d2}, [%1]!                    \n"                             \
-    MEMACCESS(2)                                                               \
-    "vld1.8     {d3}, [%2]!                    \n"                             \
-    "vpaddl.u8  q1, q1                         \n"                             \
-    "vrshrn.u16 d2, q1, #1                     \n"
+#define READYUV444                               \
+  "vld1.8     {d0}, [%[src_y]]!              \n" \
+  "vld1.8     {d2}, [%[src_u]]!              \n" \
+  "vmovl.u8   q0, d0                         \n" \
+  "vld1.8     {d3}, [%[src_v]]!              \n" \
+  "vsli.u16   q0, q0, #8                     \n"
 
 // Read 8 Y, and set 4 U and 4 V to 128
-#define READYUV400                                                             \
-    MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    "vmov.u8    d2, #128                       \n"
+#define READYUV400                               \
+  "vld1.8     {d0}, [%[src_y]]!              \n" \
+  "vmov.u8    q1, #128                       \n" \
+  "vmovl.u8   q0, d0                         \n" \
+  "vsli.u16   q0, q0, #8                     \n"
 
 // Read 8 Y and 4 UV from NV12
-#define READNV12                                                               \
-    MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    MEMACCESS(1)                                                               \
-    "vld1.8     {d2}, [%1]!                    \n"                             \
-    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
-    "vuzp.u8    d2, d3                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
+#define READNV12                                                              \
+  "vld1.8     {d0}, [%[src_y]]!              \n"                              \
+  "vld1.8     {d2}, [%[src_uv]]!             \n"                              \
+  "vmov.u8    d1, d0                         \n"                              \
+  "vmov.u8    d3, d2                         \n"                              \
+  "vzip.u8    d0, d1                         \n"                              \
+  "vsli.u16   d2, d2, #8                     \n" /* Duplicate low byte (U) */ \
+  "vsri.u16   d3, d3, #8                     \n" /* Duplicate high byte (V) */
 
 // Read 8 Y and 4 VU from NV21
 #define READNV21                                                               \
-    MEMACCESS(0)                                                               \
-    "vld1.8     {d0}, [%0]!                    \n"                             \
-    MEMACCESS(1)                                                               \
-    "vld1.8     {d2}, [%1]!                    \n"                             \
-    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
-    "vuzp.u8    d3, d2                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
+  "vld1.8     {d0}, [%[src_y]]!              \n"                               \
+  "vld1.8     {d2}, [%[src_vu]]!             \n"                               \
+  "vmov.u8    d1, d0                         \n"                               \
+  "vmov.u8    d3, d2                         \n"                               \
+  "vzip.u8    d0, d1                         \n"                               \
+  "vsri.u16   d2, d2, #8                     \n" /* Duplicate high byte (U) */ \
+  "vsli.u16   d3, d3, #8                     \n" /* Duplicate low byte (V) */
 
 // Read 8 YUY2
-#define READYUY2                                                               \
-    MEMACCESS(0)                                                               \
-    "vld2.8     {d0, d2}, [%0]!                \n"                             \
-    "vmov.u8    d3, d2                         \n"                             \
-    "vuzp.u8    d2, d3                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
+#define READYUY2                                 \
+  "vld2.8     {d0, d2}, [%[src_yuy2]]!       \n" \
+  "vmovl.u8   q0, d0                         \n" \
+  "vmov.u8    d3, d2                         \n" \
+  "vsli.u16   q0, q0, #8                     \n" \
+  "vsli.u16   d2, d2, #8                     \n" \
+  "vsri.u16   d3, d3, #8                     \n"
 
 // Read 8 UYVY
-#define READUYVY                                                               \
-    MEMACCESS(0)                                                               \
-    "vld2.8     {d2, d3}, [%0]!                \n"                             \
-    "vmov.u8    d0, d3                         \n"                             \
-    "vmov.u8    d3, d2                         \n"                             \
-    "vuzp.u8    d2, d3                         \n"                             \
-    "vtrn.u32   d2, d3                         \n"
-
-#define YUVTORGB_SETUP                                                         \
-    MEMACCESS([kUVToRB])                                                       \
-    "vld1.8     {d24}, [%[kUVToRB]]            \n"                             \
-    MEMACCESS([kUVToG])                                                        \
-    "vld1.8     {d25}, [%[kUVToG]]             \n"                             \
-    MEMACCESS([kUVBiasBGR])                                                    \
-    "vld1.16    {d26[], d27[]}, [%[kUVBiasBGR]]! \n"                           \
-    MEMACCESS([kUVBiasBGR])                                                    \
-    "vld1.16    {d8[], d9[]}, [%[kUVBiasBGR]]!   \n"                           \
-    MEMACCESS([kUVBiasBGR])                                                    \
-    "vld1.16    {d28[], d29[]}, [%[kUVBiasBGR]]  \n"                           \
-    MEMACCESS([kYToRgb])                                                       \
-    "vld1.32    {d30[], d31[]}, [%[kYToRgb]]     \n"
-
-#define YUVTORGB                                                               \
-    "vmull.u8   q8, d2, d24                    \n" /* u/v B/R component      */\
-    "vmull.u8   q9, d2, d25                    \n" /* u/v G component        */\
-    "vmovl.u8   q0, d0                         \n" /* Y                      */\
-    "vmovl.s16  q10, d1                        \n"                             \
-    "vmovl.s16  q0, d0                         \n"                             \
-    "vmul.s32   q10, q10, q15                  \n"                             \
-    "vmul.s32   q0, q0, q15                    \n"                             \
-    "vqshrun.s32 d0, q0, #16                   \n"                             \
-    "vqshrun.s32 d1, q10, #16                  \n" /* Y                      */\
-    "vadd.s16   d18, d19                       \n"                             \
-    "vshll.u16  q1, d16, #16                   \n" /* Replicate u * UB       */\
-    "vshll.u16  q10, d17, #16                  \n" /* Replicate v * VR       */\
-    "vshll.u16  q3, d18, #16                   \n" /* Replicate (v*VG + u*UG)*/\
-    "vaddw.u16  q1, q1, d16                    \n"                             \
-    "vaddw.u16  q10, q10, d17                  \n"                             \
-    "vaddw.u16  q3, q3, d18                    \n"                             \
-    "vqadd.s16  q8, q0, q13                    \n" /* B */                     \
-    "vqadd.s16  q9, q0, q14                    \n" /* R */                     \
-    "vqadd.s16  q0, q0, q4                     \n" /* G */                     \
-    "vqadd.s16  q8, q8, q1                     \n" /* B */                     \
-    "vqadd.s16  q9, q9, q10                    \n" /* R */                     \
-    "vqsub.s16  q0, q0, q3                     \n" /* G */                     \
-    "vqshrun.s16 d20, q8, #6                   \n" /* B */                     \
-    "vqshrun.s16 d22, q9, #6                   \n" /* R */                     \
-    "vqshrun.s16 d21, q0, #6                   \n" /* G */
-
-void I444ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+#define READUYVY                                 \
+  "vld2.8     {d2, d3}, [%[src_uyvy]]!       \n" \
+  "vmovl.u8   q0, d3                         \n" \
+  "vmov.u8    d3, d2                         \n" \
+  "vsli.u16   q0, q0, #8                     \n" \
+  "vsli.u16   d2, d2, #8                     \n" \
+  "vsri.u16   d3, d3, #8                     \n"
+
+#define YUVTORGB_SETUP                                        \
+  "vld4.8     {d26[], d27[], d28[], d29[]}, [%[kUVCoeff]] \n" \
+  "vld1.16    {d31[]}, [%[kRGBCoeffBias]]!   \n"              \
+  "vld1.16    {d20[], d21[]}, [%[kRGBCoeffBias]]! \n"         \
+  "vld1.16    {d22[], d23[]}, [%[kRGBCoeffBias]]! \n"         \
+  "vld1.16    {d24[], d25[]}, [%[kRGBCoeffBias]] \n"
+
+// q0: B uint16x8_t
+// q1: G uint16x8_t
+// q2: R uint16x8_t
+
+// Convert from YUV to 2.14 fixed point RGB
+#define YUVTORGB                                           \
+  "vmull.u16  q2, d1, d31                    \n"           \
+  "vmull.u8   q8, d3, d29                    \n" /* DGV */ \
+  "vmull.u16  q0, d0, d31                    \n"           \
+  "vmlal.u8   q8, d2, d28                    \n" /* DG */  \
+  "vqshrn.u32 d0, q0, #16                    \n"           \
+  "vqshrn.u32 d1, q2, #16                    \n" /* Y */   \
+  "vmull.u8   q9, d2, d26                    \n" /* DB */  \
+  "vmull.u8   q2, d3, d27                    \n" /* DR */  \
+  "vadd.u16   q4, q0, q11                    \n" /* G */   \
+  "vadd.u16   q2, q0, q2                     \n" /* R */   \
+  "vadd.u16   q0, q0, q9                     \n" /* B */   \
+  "vqsub.u16  q1, q4, q8                     \n" /* G */   \
+  "vqsub.u16  q0, q0, q10                    \n" /* B */   \
+  "vqsub.u16  q2, q2, q12                    \n" /* R */
+
+// Convert from 2.14 fixed point RGB To 8 bit RGB
+#define RGBTORGB8                                        \
+  "vqshrn.u16 d4, q2, #6                     \n" /* R */ \
+  "vqshrn.u16 d2, q1, #6                     \n" /* G */ \
+  "vqshrn.u16 d0, q0, #6                     \n" /* B */
+
+#define YUVTORGB_REGS \
+  "q0", "q1", "q2", "q4", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "d31"
+
+#define STORERGBA                                \
+  "vmov.u8    d1, d0                         \n" \
+  "vmov.u8    d3, d4                         \n" \
+  "vmov.u8    d0, d6                         \n" \
+  "vst4.8     {d0, d1, d2, d3}, [%[dst_rgba]]! \n"
+
+void I444ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READYUV444
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void I422ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void I444ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
+}
+
+void I422ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void I422AlphaToARGBRow_NEON(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             const uint8* src_a,
-                             uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void I444AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %5, %5, #8                     \n"
-    MEMACCESS(3)
-    "vld1.8     {d23}, [%3]!                   \n"
-    MEMACCESS(4)
-    "vst4.8     {d20, d21, d22, d23}, [%4]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(src_a),     // %3
-      "+r"(dst_argb),  // %4
-      "+r"(width)      // %5
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void I411ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READYUV411
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    MEMACCESS(3)
-    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void I422ToRGBARow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_rgba,
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "vld1.8      {d6}, [%[src_a]]!             \n"
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "vld1.8      {d6}, [%[src_a]]!             \n"
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void I422ToRGBARow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_rgba,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    "vmov.u8    d19, #255                      \n"  // d19 modified by YUVTORGB
-    MEMACCESS(3)
-    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_rgba),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void I422ToRGB24Row_NEON(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_rgb24,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUV422 YUVTORGB
+      RGBTORGB8 "subs        %[width], %[width], #8        \n" STORERGBA
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgba] "+r"(dst_rgba),                         // %[dst_rgba]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void I422ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    MEMACCESS(3)
-    "vst3.8     {d20, d21, d22}, [%3]!         \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),      // %0
-      "+r"(src_u),      // %1
-      "+r"(src_v),      // %2
-      "+r"(dst_rgb24),  // %3
-      "+r"(width)       // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-#define ARGBTORGB565                                                           \
-    "vshll.u8    q0, d22, #8                   \n"  /* R                    */ \
-    "vshll.u8    q8, d21, #8                   \n"  /* G                    */ \
-    "vshll.u8    q9, d20, #8                   \n"  /* B                    */ \
-    "vsri.16     q0, q8, #5                    \n"  /* RG                   */ \
-    "vsri.16     q0, q9, #11                   \n"  /* RGB                  */
-
-void I422ToRGB565Row_NEON(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          uint8* dst_rgb565,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
+}
+
+#define ARGBTORGB565                                                        \
+  "vshll.u8    q2, d4, #8                    \n" /* R                    */ \
+  "vshll.u8    q1, d2, #8                    \n" /* G                    */ \
+  "vshll.u8    q0, d0, #8                    \n" /* B                    */ \
+  "vsri.16     q2, q1, #5                    \n" /* RG                   */ \
+  "vsri.16     q2, q0, #11                   \n" /* RGB                  */
+
+void I422ToRGB565Row_NEON(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    ARGBTORGB565
-    MEMACCESS(3)
-    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_rgb565),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-#define ARGBTOARGB1555                                                         \
-    "vshll.u8    q0, d23, #8                   \n"  /* A                    */ \
-    "vshll.u8    q8, d22, #8                   \n"  /* R                    */ \
-    "vshll.u8    q9, d21, #8                   \n"  /* G                    */ \
-    "vshll.u8    q10, d20, #8                  \n"  /* B                    */ \
-    "vsri.16     q0, q8, #1                    \n"  /* AR                   */ \
-    "vsri.16     q0, q9, #6                    \n"  /* ARG                  */ \
-    "vsri.16     q0, q10, #11                  \n"  /* ARGB                 */
-
-void I422ToARGB1555Row_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb1555,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUV422 YUVTORGB
+      RGBTORGB8 "subs        %[width], %[width], #8        \n" ARGBTORGB565
+      "vst1.8      {q2}, [%[dst_rgb565]]!        \n"  // store 8 pixels RGB565.
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb565] "+r"(dst_rgb565),                     // %[dst_rgb565]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
+}
+
+#define ARGBTOARGB1555                                                      \
+  "vshll.u8    q3, d6, #8                    \n" /* A                    */ \
+  "vshll.u8    q2, d4, #8                    \n" /* R                    */ \
+  "vshll.u8    q1, d2, #8                    \n" /* G                    */ \
+  "vshll.u8    q0, d0, #8                    \n" /* B                    */ \
+  "vsri.16     q3, q2, #1                    \n" /* AR                   */ \
+  "vsri.16     q3, q1, #6                    \n" /* ARG                  */ \
+  "vsri.16     q3, q0, #11                   \n" /* ARGB                 */
+
+void I422ToARGB1555Row_NEON(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
                             const struct YuvConstants* yuvconstants,
                             int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
-    ARGBTOARGB1555
-    MEMACCESS(3)
-    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
-    "bgt        1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_argb1555),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-#define ARGBTOARGB4444                                                         \
-    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
-    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
-    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
-    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
-    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
-    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
-    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
-
-void I422ToARGB4444Row_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb4444,
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vmov.u8     d6, #0xff                     \n" ARGBTOARGB1555
+      "vst1.8      {q3}, [%[dst_argb1555]]!      \n"  // store 8 pixels RGB1555.
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb1555] "+r"(dst_argb1555),                 // %[dst_argb1555]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "q3");
+}
+
+#define ARGBTOARGB4444                                                      \
+  "vshr.u8    d0, d0, #4                     \n" /* B                    */ \
+  "vbic.32    d2, d2, d7                     \n" /* G                    */ \
+  "vshr.u8    d4, d4, #4                     \n" /* R                    */ \
+  "vbic.32    d6, d6, d7                     \n" /* A                    */ \
+  "vorr       d0, d0, d2                     \n" /* BG                   */ \
+  "vorr       d1, d4, d6                     \n" /* RA                   */ \
+  "vzip.u8    d0, d1                         \n" /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
                             const struct YuvConstants* yuvconstants,
                             int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB
-    "subs       %4, %4, #8                     \n"
-    "vmov.u8    d23, #255                      \n"
-    ARGBTOARGB4444
-    MEMACCESS(3)
-    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
-    "bgt        1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_argb4444),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void I400ToARGBRow_NEON(const uint8* src_y,
-                        uint8* dst_argb,
-                        int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READYUV400
-    YUVTORGB
-    "subs       %2, %2, #8                     \n"
-    MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&kYuvI601Constants.kUVToRB),
-      [kUVToG]"r"(&kYuvI601Constants.kUVToG),
-      [kUVBiasBGR]"r"(&kYuvI601Constants.kUVBiasBGR),
-      [kYToRgb]"r"(&kYuvI601Constants.kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void J400ToARGBRow_NEON(const uint8* src_y,
-                        uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "vmov.u8     d7, #0x0f                     \n"  // vbic bits to clear
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n" ARGBTOARGB4444
+      "vst1.8      {q0}, [%[dst_argb4444]]!      \n"  // store 8 pixels
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb4444] "+r"(dst_argb4444),                 // %[dst_argb4444]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "q3");
+}
+
+void I400ToARGBRow_NEON(const uint8_t* src_y,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d20}, [%0]!                   \n"
-    "vmov       d21, d20                       \n"
-    "vmov       d22, d20                       \n"
-    "subs       %2, %2, #8                     \n"
-    MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    :
-    : "cc", "memory", "d20", "d21", "d22", "d23"
-  );
-}
-
-void NV12ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_uv,
-                        uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUV400 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "vmov.u8     d23, #255                     \n"
+      "1:                                        \n"
+      "vld1.8      {d20}, [%0]!                  \n"
+      "vmov        d21, d20                      \n"
+      "vmov        d22, d20                      \n"
+      "subs        %2, %2, #8                    \n"
+      "vst4.8      {d20, d21, d22, d23}, [%1]!   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "d20", "d21", "d22", "d23");
+}
+
+void NV12ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READNV12
-    YUVTORGB
-    "subs       %3, %3, #8                     \n"
-    MEMACCESS(2)
-    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_uv),    // %1
-      "+r"(dst_argb),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void NV21ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_vu,
-                        uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_uv] "+r"(src_uv),                             // %[src_uv]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void NV21ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READNV21
-    YUVTORGB
-    "subs       %3, %3, #8                     \n"
-    MEMACCESS(2)
-    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_vu),    // %1
-      "+r"(dst_argb),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void NV12ToRGB565Row_NEON(const uint8* src_y,
-                          const uint8* src_uv,
-                          uint8* dst_rgb565,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV21 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_vu] "+r"(src_vu),                             // %[src_vu]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void NV12ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_uv] "+r"(src_uv),                             // %[src_uv]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
+}
+
+void NV21ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV21 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_vu] "+r"(src_vu),                             // %[src_vu]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
+}
+
+void NV12ToRGB565Row_NEON(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READNV12
-    YUVTORGB
-    "subs       %3, %3, #8                     \n"
-    ARGBTORGB565
-    MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_uv),    // %1
-      "+r"(dst_rgb565),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
-                        uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n" ARGBTORGB565
+      "vst1.8      {q2}, [%[dst_rgb565]]!        \n"  // store 8 pixels RGB565.
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_uv] "+r"(src_uv),                             // %[src_uv]
+        [dst_rgb565] "+r"(dst_rgb565),                     // %[dst_rgb565]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
+}
+
+void YUY2ToARGBRow_NEON(const uint8_t* src_yuy2,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READYUY2
-    YUVTORGB
-    "subs       %2, %2, #8                     \n"
-    MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_yuy2),  // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
-void UYVYToARGBRow_NEON(const uint8* src_uyvy,
-                        uint8* dst_argb,
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUY2 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_yuy2] "+r"(src_yuy2),                         // %[src_yuy2]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void UYVYToARGBRow_NEON(const uint8_t* src_uyvy,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "vmov.u8    d23, #255                      \n"
-  "1:                                          \n"
-    READUYVY
-    YUVTORGB
-    "subs       %2, %2, #8                     \n"
-    MEMACCESS(1)
-    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-    "bgt        1b                             \n"
-    : "+r"(src_uyvy),  // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
-      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READUYVY YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_uyvy] "+r"(src_uyvy),                         // %[src_uyvy]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 // Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
-void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store U
-    MEMACCESS(2)
-    "vst1.8     {q1}, [%2]!                    \n"  // store V
-    "bgt        1b                             \n"
-    : "+r"(src_uv),  // %0
-      "+r"(dst_u),   // %1
-      "+r"(dst_v),   // %2
-      "+r"(width)    // %3  // Output registers
-    :                       // Input registers
-    : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-
-// Reads 16 U's and V's and writes out 16 pairs of UV.
-void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void SplitUVRow_NEON(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load U
-    MEMACCESS(1)
-    "vld1.8     {q1}, [%1]!                    \n"  // load V
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    MEMACCESS(2)
-    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
-    "bgt        1b                             \n"
-    :
-      "+r"(src_u),   // %0
-      "+r"(src_v),   // %1
-      "+r"(dst_uv),  // %2
-      "+r"(width)    // %3  // Output registers
-    :                       // Input registers
-    : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-
-// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
-void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
-    "subs       %2, %2, #32                    \n"  // 32 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
-    "bgt        1b                             \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(count)  // %2  // Output registers
-  :                     // Input registers
-  : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-
-// SetRow writes 'count' bytes using an 8 bit value repeated.
-void SetRow_NEON(uint8* dst, uint8 v8, int count) {
-  asm volatile (
-    "vdup.8    q0, %2                          \n"  // duplicate 16 bytes
-  "1:                                          \n"
-    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
-    MEMACCESS(0)
-    "vst1.8    {q0}, [%0]!                     \n"  // store
-    "bgt       1b                              \n"
-  : "+r"(dst),   // %0
-    "+r"(count)  // %1
-  : "r"(v8)      // %2
-  : "cc", "memory", "q0"
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pairs of UV
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%1]!                   \n"  // store U
+      "vst1.8      {q1}, [%2]!                   \n"  // store V
+      "bgt         1b                            \n"
+      : "+r"(src_uv),               // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(width)                 // %3  // Output registers
+      :                             // Input registers
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 byte Y's from tile and writes out 16 Y's.
+// MM21 Y tiles are 16x32 so src_tile_stride = 512 bytes
+// MM21 UV tiles are 8x16 so src_tile_stride = 256 bytes
+// width measured in bytes so 8 UV = 16.
+void DetileRow_NEON(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0], %3                \n"  // load 16 bytes
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "pld         [%0, #1792]                   \n"
+      "vst1.8      {q0}, [%1]!                   \n"  // store 16 bytes
+      "bgt         1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "q0"  // Clobber List
+  );
+}
+
+// Reads 16 byte Y's of 16 bits from tile and writes out 16 Y's.
+void DetileRow_16_NEON(const uint16_t* src,
+                       ptrdiff_t src_tile_stride,
+                       uint16_t* dst,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {q0, q1}, [%0], %3            \n"  // load 16 pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "pld         [%0, #3584]                   \n"
+      "vst1.16     {q0, q1}, [%1]!               \n"  // store 16 pixels
+      "bgt         1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst),                  // %1
+        "+r"(width)                 // %2
+      : "r"(src_tile_stride * 2)    // %3
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Read 16 bytes of UV, detile, and write 8 bytes of U and 8 bytes of V.
+void DetileSplitUVRow_NEON(const uint8_t* src_uv,
+                           ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {d0, d1}, [%0], %4            \n"
+      "subs        %3, %3, #16                   \n"
+      "pld         [%0, #1792]                   \n"
+      "vst1.8      {d0}, [%1]!                   \n"
+      "vst1.8      {d1}, [%2]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_uv),               // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(width)                 // %3
+      : "r"(src_tile_stride)        // %4
+      : "cc", "memory", "d0", "d1"  // Clobber List
+  );
+}
+
+#if LIBYUV_USE_ST2
+// Read 16 Y, 8 UV, and write 8 YUYV.
+void DetileToYUY2_NEON(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0], %4                \n"  // Load 16 Y
+      "pld         [%0, #1792]                   \n"
+      "vld1.8      {q1}, [%1], %5                \n"  // Load 8 UV
+      "pld         [%1, #1792]                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vst2.8      {q0, q1}, [%2]!               \n"
+      "bgt         1b                            \n"
+      : "+r"(src_y),                            // %0
+        "+r"(src_uv),                           // %1
+        "+r"(dst_yuy2),                         // %2
+        "+r"(width)                             // %3
+      : "r"(src_y_tile_stride),                 // %4
+        "r"(src_uv_tile_stride)                 // %5
+      : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber list
+  );
+}
+#else
+// Read 16 Y, 8 UV, and write 8 YUYV.
+void DetileToYUY2_NEON(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0], %4                \n"  // Load 16 Y
+      "vld1.8      {q1}, [%1], %5                \n"  // Load 8 UV
+      "subs        %3, %3, #16                   \n"
+      "pld         [%0, #1792]                   \n"
+      "vzip.8      q0, q1                        \n"
+      "pld         [%1, #1792]                   \n"
+      "vst1.8      {q0, q1}, [%2]!               \n"
+      "bgt         1b                            \n"
+      : "+r"(src_y),                            // %0
+        "+r"(src_uv),                           // %1
+        "+r"(dst_yuy2),                         // %2
+        "+r"(width)                             // %3
+      : "r"(src_y_tile_stride),                 // %4
+        "r"(src_uv_tile_stride)                 // %5
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber list
   );
 }
+#endif
 
-// ARGBSetRow writes 'count' pixels using an 32 bit value repeated.
-void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
-  asm volatile (
-    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
-  "1:                                          \n"
-    "subs      %1, %1, #4                      \n"  // 4 pixels per loop
-    MEMACCESS(0)
-    "vst1.8    {q0}, [%0]!                     \n"  // store
-    "bgt       1b                              \n"
-  : "+r"(dst),   // %0
-    "+r"(count)  // %1
-  : "r"(v32)     // %2
-  : "cc", "memory", "q0"
-  );
+void UnpackMT2T_NEON(const uint8_t* src, uint16_t* dst, size_t size) {
+  const uint8_t* src_lower_bits = src;
+  const uint8_t* src_upper_bits = src + 16;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d1, d3, d5, d7}, [%1]!       \n"  // Load 32 bytes of upper
+                                                      // bits.
+      "vld1.8      {d6}, [%0]!                   \n"  // Load 8 bytes of lower
+                                                      // bits.
+      "vshl.u8     d4, d6, #2                    \n"  // Align lower bits.
+      "vshl.u8     d2, d6, #4                    \n"
+      "vshl.u8     d0, d6, #6                    \n"
+      "vzip.u8     d0, d1                        \n"  // Zip lower and upper
+                                                      // bits together.
+      "vzip.u8     d2, d3                        \n"
+      "vzip.u8     d4, d5                        \n"
+      "vzip.u8     d6, d7                        \n"
+      "vsri.u16    q0, q0, #10                   \n"  // Copy upper 6 bits into
+                                                      // lower 6 bits for better
+                                                      // accuracy in
+                                                      // conversions.
+      "vsri.u16    q1, q1, #10                   \n"
+      "vsri.u16    q2, q2, #10                   \n"
+      "vsri.u16    q3, q3, #10                   \n"
+      "vst4.16     {d0, d2, d4, d6}, [%2]!       \n"  // Store 32 pixels
+      "vst4.16     {d1, d3, d5, d7}, [%2]!       \n"
+      "vld4.8      {d1, d3, d5, d7}, [%1]!       \n"  // Process last 32 pixels
+                                                      // in the block
+      "vld1.8      {d6}, [%0]!                   \n"
+      "vshl.u8     d4, d6, #2                    \n"
+      "vshl.u8     d2, d6, #4                    \n"
+      "vshl.u8     d0, d6, #6                    \n"
+      "vzip.u8     d0, d1                        \n"
+      "vzip.u8     d2, d3                        \n"
+      "vzip.u8     d4, d5                        \n"
+      "vzip.u8     d6, d7                        \n"
+      "vsri.u16    q0, q0, #10                   \n"
+      "vsri.u16    q1, q1, #10                   \n"
+      "vsri.u16    q2, q2, #10                   \n"
+      "vsri.u16    q3, q3, #10                   \n"
+      "vst4.16     {d0, d2, d4, d6}, [%2]!       \n"
+      "vst4.16     {d1, d3, d5, d7}, [%2]!       \n"
+      "mov         %0, %1                        \n"
+      "add         %1, %0, #16                   \n"
+      "subs        %3, %3, #80                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_lower_bits),  // %0
+        "+r"(src_upper_bits),  // %1
+        "+r"(dst),             // %2
+        "+r"(size)             // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
-void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    // Start at end of source row.
-    "mov        r3, #-16                       \n"
-    "add        %0, %0, %2                     \n"
-    "sub        %0, #16                        \n"
-
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
-    "subs       %2, #16                        \n"  // 16 pixels per loop.
-    "vrev64.8   q0, q0                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  :
-  : "cc", "memory", "r3", "q0"
-  );
-}
-
-void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
+                     int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load U
+      "vld1.8      {q1}, [%1]!                   \n"  // load V
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop
+      "vst2.8      {q0, q1}, [%2]!               \n"  // store 16 pairs of UV
+      "bgt         1b                            \n"
+      : "+r"(src_u),                // %0
+        "+r"(src_v),                // %1
+        "+r"(dst_uv),               // %2
+        "+r"(width)                 // %3  // Output registers
+      :                             // Input registers
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 packed RGB and write to planar dst_r, dst_g, dst_b.
+void SplitRGBRow_NEON(const uint8_t* src_rgb,
+                      uint8_t* dst_r,
+                      uint8_t* dst_g,
+                      uint8_t* dst_b,
                       int width) {
-  asm volatile (
-    // Start at end of source row.
-    "mov        r12, #-16                      \n"
-    "add        %0, %0, %3, lsl #1             \n"
-    "sub        %0, #16                        \n"
-
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
-    "subs       %3, #8                         \n"  // 8 pixels per loop.
-    "vrev64.8   q0, q0                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
-    MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_uv),  // %0
-    "+r"(dst_u),   // %1
-    "+r"(dst_v),   // %2
-    "+r"(width)    // %3
-  :
-  : "cc", "memory", "r12", "q0"
-  );
-}
-
-void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    // Start at end of source row.
-    "mov        r3, #-16                       \n"
-    "add        %0, %0, %2, lsl #2             \n"
-    "sub        %0, #16                        \n"
-
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
-    "subs       %2, #4                         \n"  // 4 pixels per loop.
-    "vrev64.32  q0, q0                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  :
-  : "cc", "memory", "r3", "q0"
-  );
-}
-
-void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d4, #255                       \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    MEMACCESS(1)
-    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_rgb24),  // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)         // %2
-  :
-  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
-  );
-}
-
-void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d4, #255                       \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vswp.u8    d1, d3                         \n"  // swap R, B
-    MEMACCESS(1)
-    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_raw),   // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  :
-  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
-  );
-}
-
-void RAWToRGB24Row_NEON(const uint8* src_raw, uint8* dst_rgb24, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vswp.u8    d1, d3                         \n"  // swap R, B
-    MEMACCESS(1)
-    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
-    "bgt        1b                             \n"
-  : "+r"(src_raw),    // %0
-    "+r"(dst_rgb24),  // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "d1", "d2", "d3"  // Clobber List
-  );
-}
-
-#define RGB565TOARGB                                                           \
-    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
-    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
-    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
-    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
-    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
-    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
-    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
-    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
-    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
-    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
-
-void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d3, #255                       \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    RGB565TOARGB
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_rgb565),  // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
-  );
+  asm volatile(
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RGB
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // next 8 RGB
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%1]!                   \n"  // store R
+      "vst1.8      {q1}, [%2]!                   \n"  // store G
+      "vst1.8      {q2}, [%3]!                   \n"  // store B
+      "bgt         1b                            \n"
+      : "+r"(src_rgb),                    // %0
+        "+r"(dst_r),                      // %1
+        "+r"(dst_g),                      // %2
+        "+r"(dst_b),                      // %3
+        "+r"(width)                       // %4
+      :                                   // Input registers
+      : "cc", "memory", "d0", "d1", "d2"  // Clobber List
+  );
+}
+
+// Reads 16 planar R's, G's and B's and writes out 16 packed RGB at a time
+void MergeRGBRow_NEON(const uint8_t* src_r,
+                      const uint8_t* src_g,
+                      const uint8_t* src_b,
+                      uint8_t* dst_rgb,
+                      int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load R
+      "vld1.8      {q1}, [%1]!                   \n"  // load G
+      "vld1.8      {q2}, [%2]!                   \n"  // load B
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst3.8      {d0, d2, d4}, [%3]!           \n"  // store 8 RGB
+      "vst3.8      {d1, d3, d5}, [%3]!           \n"  // next 8 RGB
+      "bgt         1b                            \n"
+      : "+r"(src_r),                      // %0
+        "+r"(src_g),                      // %1
+        "+r"(src_b),                      // %2
+        "+r"(dst_rgb),                    // %3
+        "+r"(width)                       // %4
+      :                                   // Input registers
+      : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b, dst_a.
+void SplitARGBRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // next 8 ARGB
+      "subs        %5, %5, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%3]!                   \n"  // store B
+      "vst1.8      {q1}, [%2]!                   \n"  // store G
+      "vst1.8      {q2}, [%1]!                   \n"  // store R
+      "vst1.8      {q3}, [%4]!                   \n"  // store A
+      "bgt         1b                            \n"
+      : "+r"(src_argb),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(dst_a),                            // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Reads 16 planar R's, G's and B's and writes out 16 packed ARGB at a time
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q2}, [%0]!                   \n"  // load R
+      "vld1.8      {q1}, [%1]!                   \n"  // load G
+      "vld1.8      {q0}, [%2]!                   \n"  // load B
+      "vld1.8      {q3}, [%3]!                   \n"  // load A
+      "subs        %5, %5, #16                   \n"  // 16 processed per loop
+      "vst4.8      {d0, d2, d4, d6}, [%4]!       \n"  // store 8 ARGB
+      "vst4.8      {d1, d3, d5, d7}, [%4]!       \n"  // next 8 ARGB
+      "bgt         1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(src_a),                            // %3
+        "+r"(dst_argb),                         // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b.
+void SplitXRGBRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // next 8 ARGB
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%3]!                   \n"  // store B
+      "vst1.8      {q1}, [%2]!                   \n"  // store G
+      "vst1.8      {q2}, [%1]!                   \n"  // store R
+      "bgt         1b                            \n"
+      : "+r"(src_argb),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Reads 16 planar R's, G's, B's and A's and writes out 16 packed ARGB at a time
+void MergeXRGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "vmov.u8     q3, #255                      \n"  // load A(255)
+      "1:                                        \n"
+      "vld1.8      {q2}, [%0]!                   \n"  // load R
+      "vld1.8      {q1}, [%1]!                   \n"  // load G
+      "vld1.8      {q0}, [%2]!                   \n"  // load B
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst4.8      {d0, d2, d4, d6}, [%3]!       \n"  // store 8 ARGB
+      "vst4.8      {d1, d3, d5, d7}, [%3]!       \n"  // next 8 ARGB
+      "bgt         1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(dst_argb),                         // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void MergeXR30Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width) {
+  int shift = 10 - depth;
+  asm volatile(
+      "vmov.u32    q14, #1023                    \n"
+      "vdup.32     q15, %5                       \n"
+      "1:                                        \n"
+      "vld1.16     {d4}, [%2]!                   \n"  // B
+      "vld1.16     {d2}, [%1]!                   \n"  // G
+      "vld1.16     {d0}, [%0]!                   \n"  // R
+      "vmovl.u16   q2, d4                        \n"  // B
+      "vmovl.u16   q1, d2                        \n"  // G
+      "vmovl.u16   q0, d0                        \n"  // R
+      "vshl.u32    q2, q2, q15                   \n"  // 000B
+      "vshl.u32    q1, q1, q15                   \n"
+      "vshl.u32    q0, q0, q15                   \n"
+      "vmin.u32    q2, q2, q14                   \n"
+      "vmin.u32    q1, q1, q14                   \n"
+      "vmin.u32    q0, q0, q14                   \n"
+      "vsli.u32    q2, q1, #10                   \n"  // 00GB
+      "vsli.u32    q2, q0, #20                   \n"  // 0RGB
+      "vorr.u32    q2, #0xc0000000               \n"  // ARGB (AR30)
+      "subs        %4, %4, #4                    \n"
+      "vst1.8      {q2}, [%3]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "q0", "q1", "q2", "q14", "q15");
+}
+
+void MergeXR30Row_10_NEON(const uint16_t* src_r,
+                          const uint16_t* src_g,
+                          const uint16_t* src_b,
+                          uint8_t* dst_ar30,
+                          int /* depth */,
+                          int width) {
+  asm volatile(
+      "vmov.u32    q14, #1023                    \n"
+      "1:                                        \n"
+      "vld1.16     {d4}, [%2]!                   \n"  // B
+      "vld1.16     {d2}, [%1]!                   \n"  // G
+      "vld1.16     {d0}, [%0]!                   \n"  // R
+      "vmovl.u16   q2, d4                        \n"  // 000B
+      "vmovl.u16   q1, d2                        \n"  // G
+      "vmovl.u16   q0, d0                        \n"  // R
+      "vmin.u32    q2, q2, q14                   \n"
+      "vmin.u32    q1, q1, q14                   \n"
+      "vmin.u32    q0, q0, q14                   \n"
+      "vsli.u32    q2, q1, #10                   \n"  // 00GB
+      "vsli.u32    q2, q0, #20                   \n"  // 0RGB
+      "vorr.u32    q2, #0xc0000000               \n"  // ARGB (AR30)
+      "subs        %4, %4, #4                    \n"
+      "vst1.8      {q2}, [%3]!                   \n"
+      "bgt         1b                            \n"
+      "3:                                        \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q14");
+}
+
+void MergeAR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "vdup.u16    q15, %6                       \n"
+      "vdup.u16    q14, %7                       \n"
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vld1.16     {q3}, [%3]!                   \n"  // A
+      "vmin.u16    q2, q2, q14                   \n"
+      "vmin.u16    q1, q1, q14                   \n"
+      "vmin.u16    q0, q0, q14                   \n"
+      "vmin.u16    q3, q3, q14                   \n"
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "vshl.u16    q3, q3, q15                   \n"
+      "subs        %5, %5, #8                    \n"
+      "vst4.16     {d0, d2, d4, d6}, [%4]!       \n"
+      "vst4.16     {d1, d3, d5, d7}, [%4]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_ar64),  // %4
+        "+r"(width)      // %5
+      : "r"(shift),      // %6
+        "r"(mask)        // %7
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q15");
+}
+
+void MergeXR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "vmov.u8     q3, #0xff                     \n"  // A (0xffff)
+      "vdup.u16    q15, %5                       \n"
+      "vdup.u16    q14, %6                       \n"
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vmin.u16    q2, q2, q14                   \n"
+      "vmin.u16    q1, q1, q14                   \n"
+      "vmin.u16    q0, q0, q14                   \n"
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "subs        %4, %4, #8                    \n"
+      "vst4.16     {d0, d2, d4, d6}, [%3]!       \n"
+      "vst4.16     {d1, d3, d5, d7}, [%3]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar64),  // %3
+        "+r"(width)      // %4
+      : "r"(shift),      // %5
+        "r"(mask)        // %6
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q15");
+}
+
+void MergeARGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "vdup.16     q15, %6                       \n"
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vld1.16     {q3}, [%3]!                   \n"  // A
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "vshl.u16    q3, q3, q15                   \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vqmovn.u16  d1, q1                        \n"
+      "vqmovn.u16  d2, q2                        \n"
+      "vqmovn.u16  d3, q3                        \n"
+      "subs        %5, %5, #8                    \n"
+      "vst4.8      {d0, d1, d2, d3}, [%4]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      : "r"(shift)       // %6
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q15");
+}
+
+void MergeXRGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "vdup.16     q15, %5                       \n"
+      "vmov.u8     d6, #0xff                     \n"  // A (0xff)
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "vqmovn.u16  d5, q2                        \n"
+      "vqmovn.u16  d4, q1                        \n"
+      "vqmovn.u16  d3, q0                        \n"
+      "subs        %4, %4, #8                    \n"
+      "vst4.u8     {d3, d4, d5, d6}, [%3]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "q0", "q1", "q2", "d6", "q15");
 }
 
-#define ARGB1555TOARGB                                                         \
-    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
-    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
-    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
-    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
-    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
-    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
-    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
-    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
-    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
-    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
-    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
-    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 32
+      "subs        %2, %2, #32                   \n"  // 32 processed per loop
+      "vst1.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 32
+      "bgt         1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst),                  // %1
+        "+r"(width)                 // %2  // Output registers
+      :                             // Input registers
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SetRow writes 'width' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8_t* dst, uint8_t v8, int width) {
+  asm volatile(
+      "vdup.8      q0, %2                        \n"  // duplicate 16 bytes
+      "1:                                        \n"
+      "subs        %1, %1, #16                   \n"  // 16 bytes per loop
+      "vst1.8      {q0}, [%0]!                   \n"  // store
+      "bgt         1b                            \n"
+      : "+r"(dst),   // %0
+        "+r"(width)  // %1
+      : "r"(v8)      // %2
+      : "cc", "memory", "q0");
+}
+
+// ARGBSetRow writes 'width' pixels using an 32 bit value repeated.
+void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width) {
+  asm volatile(
+      "vdup.u32    q0, %2                        \n"  // duplicate 4 ints
+      "1:                                        \n"
+      "subs        %1, %1, #4                    \n"  // 4 pixels per loop
+      "vst1.8      {q0}, [%0]!                   \n"  // store
+      "bgt         1b                            \n"
+      : "+r"(dst),   // %0
+        "+r"(width)  // %1
+      : "r"(v32)     // %2
+      : "cc", "memory", "q0");
+}
+
+void MirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "add         %0, %0, %2                    \n"
+      "sub         %0, %0, #32                   \n"  // 32 bytes per loop
+
+      "1:                                        \n"
+      "vld1.8      {q1, q2}, [%0], %3            \n"  // src -= 32
+      "subs        %2, #32                       \n"  // 32 pixels per loop.
+      "vrev64.8    q0, q2                        \n"
+      "vrev64.8    q1, q1                        \n"
+      "vswp        d0, d1                        \n"
+      "vswp        d2, d3                        \n"
+      "vst1.8      {q0, q1}, [%1]!               \n"  // dst += 32
+      "bgt         1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      : "r"(-32)     // %3
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void MirrorUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "mov         r12, #-16                     \n"
+      "add         %0, %0, %2, lsl #1            \n"
+      "sub         %0, #16                       \n"
+
+      "1:                                        \n"
+      "vld2.8      {d0, d1}, [%0], r12           \n"  // src -= 16
+      "subs        %2, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    q0, q0                        \n"
+      "vst2.8      {d0, d1}, [%1]!               \n"  // dst += 16
+      "bgt         1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_uv),  // %1
+        "+r"(width)    // %2
+      :
+      : "cc", "memory", "r12", "q0");
+}
+
+void MirrorSplitUVRow_NEON(const uint8_t* src_uv,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
+  asm volatile(
+      // Start at end of source row.
+      "mov         r12, #-16                     \n"
+      "add         %0, %0, %3, lsl #1            \n"
+      "sub         %0, #16                       \n"
+
+      "1:                                        \n"
+      "vld2.8      {d0, d1}, [%0], r12           \n"  // src -= 16
+      "subs        %3, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    q0, q0                        \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // dst += 8
+      "vst1.8      {d1}, [%2]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      :
+      : "cc", "memory", "r12", "q0");
+}
+
+void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "add         %0, %0, %2, lsl #2            \n"
+      "sub         %0, #32                       \n"
+
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0], %3    \n"  // src -= 32
+      "subs        %2, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    d0, d0                        \n"
+      "vrev64.8    d1, d1                        \n"
+      "vrev64.8    d2, d2                        \n"
+      "vrev64.8    d3, d3                        \n"
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // dst += 32
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(-32)         // %3
+      : "cc", "memory", "d0", "d1", "d2", "d3");
+}
+
+void RGB24MirrorRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_rgb24,
+                         int width) {
+  src_rgb24 += width * 3 - 24;
+  asm volatile(
+      "1:                                        \n"
+      "vld3.8      {d0, d1, d2}, [%0], %3        \n"  // src -= 24
+      "subs        %2, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    d0, d0                        \n"
+      "vrev64.8    d1, d1                        \n"
+      "vrev64.8    d2, d2                        \n"
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"  // dst += 24
+      "bgt         1b                            \n"
+      : "+r"(src_rgb24),  // %0
+        "+r"(dst_rgb24),  // %1
+        "+r"(width)       // %2
+      : "r"(-24)          // %3
+      : "cc", "memory", "d0", "d1", "d2");
+}
+
+void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width) {
+  asm volatile(
+      "vmov.u8     d4, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RGB24.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst4.8      {d1, d2, d3, d4}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_rgb24),  // %0
+        "+r"(dst_argb),   // %1
+        "+r"(width)       // %2
+      :
+      : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "vmov.u8     d4, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RAW.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst4.8      {d1, d2, d3, d4}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_raw),   // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  asm volatile(
+      "vmov.u8     d0, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RAW.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of RGBA.
+      "bgt         1b                            \n"
+      : "+r"(src_raw),   // %0
+        "+r"(dst_rgba),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+void RAWToRGB24Row_NEON(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RAW.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst3.8      {d1, d2, d3}, [%1]!           \n"  // store 8 pixels of
+                                                      // RGB24.
+      "bgt         1b                            \n"
+      : "+r"(src_raw),    // %0
+        "+r"(dst_rgb24),  // %1
+        "+r"(width)       // %2
+      :
+      : "cc", "memory", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                        \
+  "vshrn.u16  d6, q0, #5                     \n" /* G xxGGGGGG           */ \
+  "vuzp.u8    d0, d1                         \n" /* d0 xxxBBBBB RRRRRxxx */ \
+  "vshl.u8    d6, d6, #2                     \n" /* G GGGGGG00 upper 6   */ \
+  "vshr.u8    d1, d1, #3                     \n" /* R 000RRRRR lower 5   */ \
+  "vshl.u8    q0, q0, #3                     \n" /* B,R BBBBB000 upper 5 */ \
+  "vshr.u8    q2, q0, #5                     \n" /* B,R 00000BBB lower 3 */ \
+  "vorr.u8    d0, d0, d4                     \n" /* B                    */ \
+  "vshr.u8    d4, d6, #6                     \n" /* G 000000GG lower 2   */ \
+  "vorr.u8    d2, d1, d5                     \n" /* R                    */ \
+  "vorr.u8    d1, d4, d6                     \n" /* G                    */
+
+void RGB565ToARGBRow_NEON(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      "vmov.u8     d3, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 RGB565 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      RGB565TOARGB
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_rgb565),  // %0
+        "+r"(dst_argb),    // %1
+        "+r"(width)        // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB1555TOARGB                                                      \
+  "vshrn.u16  d7, q0, #8                     \n" /* A Arrrrrxx           */ \
+  "vshr.u8    d6, d7, #2                     \n" /* R xxxRRRRR           */ \
+  "vshrn.u16  d5, q0, #5                     \n" /* G xxxGGGGG           */ \
+  "vmovn.u16  d4, q0                         \n" /* B xxxBBBBB           */ \
+  "vshr.u8    d7, d7, #7                     \n" /* A 0000000A           */ \
+  "vneg.s8    d7, d7                         \n" /* A AAAAAAAA upper 8   */ \
+  "vshl.u8    d6, d6, #3                     \n" /* R RRRRR000 upper 5   */ \
+  "vshr.u8    q1, q3, #5                     \n" /* R,A 00000RRR lower 3 */ \
+  "vshl.u8    q0, q2, #3                     \n" /* B,G BBBBB000 upper 5 */ \
+  "vshr.u8    q2, q0, #5                     \n" /* B,G 00000BBB lower 3 */ \
+  "vorr.u8    q1, q1, q3                     \n" /* R,A                  */ \
+  "vorr.u8    q0, q0, q2                     \n" /* B,G                  */
 
 // RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
-#define RGB555TOARGB                                                           \
-    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
-    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
-    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
-    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
-    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
-    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
-    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
-    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
-    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
-    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
-
-void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+#define RGB555TOARGB                                                        \
+  "vshrn.u16  d6, q0, #5                     \n" /* G xxxGGGGG           */ \
+  "vuzp.u8    d0, d1                         \n" /* d0 xxxBBBBB xRRRRRxx */ \
+  "vshl.u8    d6, d6, #3                     \n" /* G GGGGG000 upper 5   */ \
+  "vshr.u8    d1, d1, #2                     \n" /* R 00xRRRRR lower 5   */ \
+  "vshl.u8    q0, q0, #3                     \n" /* B,R BBBBB000 upper 5 */ \
+  "vshr.u8    q2, q0, #5                     \n" /* B,R 00000BBB lower 3 */ \
+  "vorr.u8    d0, d0, d4                     \n" /* B                    */ \
+  "vshr.u8    d4, d6, #5                     \n" /* G 00000GGG lower 3   */ \
+  "vorr.u8    d2, d1, d5                     \n" /* R                    */ \
+  "vorr.u8    d1, d4, d6                     \n" /* G                    */
+
+void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
                             int width) {
-  asm volatile (
-    "vmov.u8    d3, #255                       \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGB1555TOARGB
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_argb1555),  // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
-  );
-}
-
-#define ARGB4444TOARGB                                                         \
-    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
-    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
-    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
-    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
-    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
-    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
-    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
-    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
-
-void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+  asm volatile(
+      "vmov.u8     d3, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB1555 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGB1555TOARGB
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_argb1555),  // %0
+        "+r"(dst_argb),      // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB4444TOARGB                                                      \
+  "vuzp.u8    d0, d1                         \n" /* d0 BG, d1 RA         */ \
+  "vshl.u8    q2, q0, #4                     \n" /* B,R BBBB0000         */ \
+  "vshr.u8    q1, q0, #4                     \n" /* G,A 0000GGGG         */ \
+  "vshr.u8    q0, q2, #4                     \n" /* B,R 0000BBBB         */ \
+  "vorr.u8    q0, q0, q2                     \n" /* B,R BBBBBBBB         */ \
+  "vshl.u8    q2, q1, #4                     \n" /* G,A GGGG0000         */ \
+  "vorr.u8    q1, q1, q2                     \n" /* G,A GGGGGGGG         */ \
+  "vswp.u8    d1, d2                         \n" /* B,R,G,A -> B,G,R,A   */
+
+void ARGB4444ToARGBRow_NEON(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
                             int width) {
-  asm volatile (
-    "vmov.u8    d3, #255                       \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGB4444TOARGB
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_argb4444),  // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
-  );
-}
-
-void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    MEMACCESS(1)
-    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_rgb24),  // %1
-    "+r"(width)         // %2
-  :
-  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
-  );
-}
-
-void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vswp.u8    d1, d3                         \n"  // swap R, B
-    MEMACCESS(1)
-    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_raw),   // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
-  );
-}
-
-void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
-    "bgt        1b                             \n"
-  : "+r"(src_yuy2),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-
-void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
-    MEMACCESS(1)
-    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
-    "bgt        1b                             \n"
-  : "+r"(src_uyvy),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-
-void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+  asm volatile(
+      "vmov.u8     d3, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB4444 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGB4444TOARGB
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_argb4444),  // %0
+        "+r"(dst_argb),      // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void ARGBToRGB24Row_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_rgb24,
                          int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
-    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
-    MEMACCESS(2)
-    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
-    "bgt        1b                             \n"
-  : "+r"(src_yuy2),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
-  );
-}
-
-void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 16 pixels of ARGB.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vst3.8      {d0, d2, d4}, [%1]!           \n"  // store 16 RGB24 pixels.
+      "vst3.8      {d1, d3, d5}, [%1]!           \n"
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(dst_rgb24),  // %1
+        "+r"(width)       // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d1, d2, d3, d4}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst3.8      {d1, d2, d3}, [%1]!           \n"  // store 8 pixels of RAW.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_raw),   // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pixels of YUY2.
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vst1.8      {q0}, [%1]!                   \n"  // store 16 pixels of Y.
+      "bgt         1b                            \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void UYVYToYRow_NEON(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pixels of UYVY.
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vst1.8      {q1}, [%1]!                   \n"  // store 16 pixels of Y.
+      "bgt         1b                            \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void YUY2ToUV422Row_NEON(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
-    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
-    MEMACCESS(2)
-    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
-    "bgt        1b                             \n"
-  : "+r"(src_uyvy),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
-  );
-}
-
-void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "add        %1, %0, %1                     \n"  // stride + src_yuy2
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
-    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
-    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
-    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
-    MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
-    MEMACCESS(3)
-    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
-    "bgt        1b                             \n"
-  : "+r"(src_yuy2),     // %0
-    "+r"(stride_yuy2),  // %1
-    "+r"(dst_u),        // %2
-    "+r"(dst_v),        // %3
-    "+r"(width)           // %4
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
-  );
-}
-
-void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "add        %1, %0, %1                     \n"  // stride + src_uyvy
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
-    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
-    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
-    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
-    MEMACCESS(3)
-    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
-    "bgt        1b                             \n"
-  : "+r"(src_uyvy),     // %0
-    "+r"(stride_uyvy),  // %1
-    "+r"(dst_u),        // %2
-    "+r"(dst_v),        // %3
-    "+r"(width)           // %4
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of YUY2.
+      "subs        %3, %3, #16                   \n"  // 16 pixels = 8 UVs.
+      "vst1.8      {d1}, [%1]!                   \n"  // store 8 U.
+      "vst1.8      {d3}, [%2]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void UYVYToUV422Row_NEON(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of UYVY.
+      "subs        %3, %3, #16                   \n"  // 16 pixels = 8 UVs.
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 U.
+      "vst1.8      {d2}, [%2]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void YUY2ToUVRow_NEON(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // stride + src_yuy2
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of YUY2.
+      "subs        %4, %4, #16                   \n"  // 16 pixels = 8 UVs.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load next row YUY2.
+      "vrhadd.u8   d1, d1, d5                    \n"  // average rows of U
+      "vrhadd.u8   d3, d3, d7                    \n"  // average rows of V
+      "vst1.8      {d1}, [%2]!                   \n"  // store 8 U.
+      "vst1.8      {d3}, [%3]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
+      : "+r"(src_yuy2),     // %0
+        "+r"(stride_yuy2),  // %1
+        "+r"(dst_u),        // %2
+        "+r"(dst_v),        // %3
+        "+r"(width)         // %4
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6",
+        "d7"  // Clobber List
+  );
+}
+
+void UYVYToUVRow_NEON(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // stride + src_uyvy
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of UYVY.
+      "subs        %4, %4, #16                   \n"  // 16 pixels = 8 UVs.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load next row UYVY.
+      "vrhadd.u8   d0, d0, d4                    \n"  // average rows of U
+      "vrhadd.u8   d2, d2, d6                    \n"  // average rows of V
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 U.
+      "vst1.8      {d2}, [%3]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
+      : "+r"(src_uyvy),     // %0
+        "+r"(stride_uyvy),  // %1
+        "+r"(dst_u),        // %2
+        "+r"(dst_v),        // %3
+        "+r"(width)         // %4
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6",
+        "d7"  // Clobber List
+  );
+}
+
+void YUY2ToNVUVRow_NEON(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // stride + src_yuy2
+      "1:                                        \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pixels of YUY2.
+      "subs        %3, %3, #16                   \n"  // 16 pixels = 8 UVs.
+      "vld2.8      {q2, q3}, [%1]!               \n"  // load next row YUY2.
+      "vrhadd.u8   q4, q1, q3                    \n"  // average rows of UV
+      "vst1.8      {q4}, [%2]!                   \n"  // store 8 UV.
+      "bgt         1b                            \n"
+      : "+r"(src_yuy2),     // %0
+        "+r"(stride_yuy2),  // %1
+        "+r"(dst_uv),       // %2
+        "+r"(width)         // %3
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6",
+        "d7"  // Clobber List
   );
 }
 
 // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
-void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width) {
-  asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {q2}, [%3]                     \n"  // shuffler
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
-    "subs       %2, %2, #4                     \n"  // 4 processed per loop
-    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
-    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
-    MEMACCESS(1)
-    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  : "r"(shuffler)    // %3
-  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
-  );
-}
-
-void I422ToYUY2Row_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_yuy2, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
-    MEMACCESS(1)
-    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
-    MEMACCESS(2)
-    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
-    "subs       %4, %4, #16                    \n"  // 16 pixels
-    MEMACCESS(3)
-    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_y),     // %0
-    "+r"(src_u),     // %1
-    "+r"(src_v),     // %2
-    "+r"(dst_yuy2),  // %3
-    "+r"(width)      // %4
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3"
-  );
-}
-
-void I422ToUYVYRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_uyvy, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
-    MEMACCESS(1)
-    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
-    MEMACCESS(2)
-    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
-    "subs       %4, %4, #16                    \n"  // 16 pixels
-    MEMACCESS(3)
-    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_y),     // %0
-    "+r"(src_u),     // %1
-    "+r"(src_v),     // %2
-    "+r"(dst_uyvy),  // %3
-    "+r"(width)      // %4
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3"
-  );
-}
-
-void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGBTORGB565
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_rgb565),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
-  );
-}
-
-void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width) {
-  asm volatile (
-    "vdup.32    d2, %2                         \n"  // dither4
-  "1:                                          \n"
-    MEMACCESS(1)
-    "vld4.8     {d20, d21, d22, d23}, [%1]!    \n"  // load 8 pixels of ARGB.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vqadd.u8   d20, d20, d2                   \n"
-    "vqadd.u8   d21, d21, d2                   \n"
-    "vqadd.u8   d22, d22, d2                   \n"
-    ARGBTORGB565
-    MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"  // store 8 pixels RGB565.
-    "bgt        1b                             \n"
-  : "+r"(dst_rgb)    // %0
-  : "r"(src_argb),   // %1
-    "r"(dither4),    // %2
-    "r"(width)       // %3
-  : "cc", "memory", "q0", "q1", "q8", "q9", "q10", "q11"
-  );
-}
-
-void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+void ARGBShuffleRow_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width) {
+  asm volatile(
+      "vld1.8      {q2}, [%3]                    \n"  // shuffler
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 4 pixels.
+      "subs        %2, %2, #4                    \n"  // 4 processed per loop
+      "vtbl.8      d2, {d0, d1}, d4              \n"  // look up 2 first pixels
+      "vtbl.8      d3, {d0, d1}, d5              \n"  // look up 2 next pixels
+      "vst1.8      {q1}, [%1]!                   \n"  // store 4.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),                   // %0
+        "+r"(dst_argb),                   // %1
+        "+r"(width)                       // %2
+      : "r"(shuffler)                     // %3
+      : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void I422ToYUY2Row_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {d0, d2}, [%0]!               \n"  // load 16 Ys
+      "vld1.8      {d1}, [%1]!                   \n"  // load 8 Us
+      "vld1.8      {d3}, [%2]!                   \n"  // load 8 Vs
+      "subs        %4, %4, #16                   \n"  // 16 pixels
+      "vst4.8      {d0, d1, d2, d3}, [%3]!       \n"  // Store 8 YUY2/16 pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_yuy2),  // %3
+        "+r"(width)      // %4
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3");
+}
+
+void I422ToUYVYRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {d1, d3}, [%0]!               \n"  // load 16 Ys
+      "vld1.8      {d0}, [%1]!                   \n"  // load 8 Us
+      "vld1.8      {d2}, [%2]!                   \n"  // load 8 Vs
+      "subs        %4, %4, #16                   \n"  // 16 pixels
+      "vst4.8      {d0, d1, d2, d3}, [%3]!       \n"  // Store 8 UYVY/16 pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_uyvy),  // %3
+        "+r"(width)      // %4
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3");
+}
+
+void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
+                          uint8_t* dst_rgb565,
+                          int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGBTORGB565
+      "vst1.8      {q2}, [%1]!                   \n"  // store 8 pixels RGB565.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(dst_rgb565),  // %1
+        "+r"(width)        // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "d6");
+}
+
+void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width) {
+  asm volatile(
+      "vdup.32     d7, %2                        \n"  // dither4
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%1]!       \n"  // load 8 pixels of ARGB.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    d0, d0, d7                    \n"
+      "vqadd.u8    d2, d2, d7                    \n"
+      "vqadd.u8    d4, d4, d7                    \n"  // add for dither
+      ARGBTORGB565
+      "vst1.8      {q2}, [%0]!                   \n"  // store 8 RGB565.
+      "bgt         1b                            \n"
+      : "+r"(dst_rgb)   // %0
+      : "r"(src_argb),  // %1
+        "r"(dither4),   // %2
+        "r"(width)      // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
+                            uint8_t* dst_argb1555,
                             int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGBTOARGB1555
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb1555),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
-  );
-}
-
-void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGBTOARGB1555
+      "vst1.8      {q3}, [%1]!                   \n"  // store 8 ARGB1555.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),      // %0
+        "+r"(dst_argb1555),  // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
+                            uint8_t* dst_argb4444,
                             int width) {
-  asm volatile (
-    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGBTOARGB4444
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),      // %0
-    "+r"(dst_argb4444),  // %1
-    "+r"(width)            // %2
-  :
-  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
-  );
-}
-
-void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
-  );
-}
-
-void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {q3}, [%1]!                    \n"  // store 16 A's.
-    "bgt       1b                              \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_a),      // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
-  );
-}
-
-void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
-    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
-    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  asm volatile(
+      "vmov.u8     d7, #0x0f                     \n"  // bits to clear with
+                                                      // vbic.
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGBTOARGB4444
+      "vst1.8      {q0}, [%1]!                   \n"  // store 8 ARGB4444.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),      // %0
+        "+r"(dst_argb4444),  // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q3}, [%1]!                   \n"  // store 16 A's.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_a),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
   );
 }
 
 // 8x1 pixels.
-void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-                         int width) {
-  asm volatile (
-    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
-    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
-    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
-    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
-    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlsl.u8   q2, d1, d25                    \n"  // G
-    "vmlsl.u8   q2, d2, d26                    \n"  // R
-    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
-
-    "vmull.u8   q3, d2, d24                    \n"  // R
-    "vmlsl.u8   q3, d1, d28                    \n"  // G
-    "vmlsl.u8   q3, d0, d27                    \n"  // B
-    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
-
-    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
-
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
-    MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
-  );
-}
-
-// 32x1 pixels -> 8x1.  width is number of argb pixels. e.g. 32.
-void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+void ARGBToUV444Row_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width) {
-  asm volatile (
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(0)
-    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
-    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
-    "vpadd.u16  d1, d8, d9                     \n"  // B
-    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
-    "vpadd.u16  d3, d10, d11                   \n"  // G
-    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
-    "vpadd.u16  d5, d12, d13                   \n"  // R
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
-    "vmul.s16   q8, q0, q10                    \n"  // B
-    "vmls.s16   q8, q1, q11                    \n"  // G
-    "vmls.s16   q8, q2, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q2, q10                    \n"  // R
-    "vmls.s16   q9, q1, q14                    \n"  // G
-    "vmls.s16   q9, q0, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
-    MEMACCESS(2)
-    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
-    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
-  );
-}
-
+  asm volatile(
+      "vmov.u8     d24, #112                     \n"  // UB / VR 0.875
+                                                      // coefficient
+      "vmov.u8     d25, #74                      \n"  // UG -0.5781 coefficient
+      "vmov.u8     d26, #38                      \n"  // UR -0.2969 coefficient
+      "vmov.u8     d27, #18                      \n"  // VB -0.1406 coefficient
+      "vmov.u8     d28, #94                      \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlsl.u8    q2, d1, d25                   \n"  // G
+      "vmlsl.u8    q2, d2, d26                   \n"  // R
+
+      "vmull.u8    q3, d2, d24                   \n"  // R
+      "vmlsl.u8    q3, d1, d28                   \n"  // G
+      "vmlsl.u8    q3, d0, d27                   \n"  // B
+
+      "vaddhn.u16  d0, q2, q15                   \n"  // +128 -> unsigned
+      "vaddhn.u16  d1, q3, q15                   \n"  // +128 -> unsigned
+
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%2]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14",
+        "q15");
+}
+
+// clang-format off
 // 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-#define RGBTOUV(QB, QG, QR) \
-    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
-    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
-    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
-    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
-    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
-    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
-    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
-    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
-    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
-    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+#define RGBTOUV(QB, QG, QR)                                                 \
+  "vmul.s16   q8, " #QB ", q10               \n" /* B                    */ \
+  "vmls.s16   q8, " #QG ", q11               \n" /* G                    */ \
+  "vmls.s16   q8, " #QR ", q12               \n" /* R                    */ \
+  "vmul.s16   q9, " #QR ", q10               \n" /* R                    */ \
+  "vmls.s16   q9, " #QG ", q14               \n" /* G                    */ \
+  "vmls.s16   q9, " #QB ", q13               \n" /* B                    */ \
+  "vaddhn.u16 d0, q8, q15                    \n" /* +128 -> unsigned     */ \
+  "vaddhn.u16 d1, q9, q15                    \n" /* +128 -> unsigned     */
+// clang-format on
 
 // TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
-void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
+void ARGBToUVRow_NEON(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ARGB pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ARGB pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q0, q1, q2)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_stride_argb),  // %1
     "+r"(dst_u),     // %2
@@ -1506,44 +1937,41 @@ void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
   );
 }
 
-// TODO(fbarchard): Subsample match C code.
-void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
-    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
-    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
-    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
-    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+// TODO(fbarchard): Subsample match Intel code.
+void ARGBToUVJRow_NEON(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile (
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ARGB pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ARGB pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q0, q1, q2)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_stride_argb),  // %1
     "+r"(dst_u),     // %2
@@ -1555,47 +1983,44 @@ void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
   );
 }
 
-void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
-    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
-    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
-    "vrshr.u16  q2, q2, #1                     \n"
-    "vrshr.u16  q3, q3, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q3, q2, q1)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_bgra),  // %0
-    "+r"(src_stride_bgra),  // %1
-    "+r"(dst_u),     // %2
-    "+r"(dst_v),     // %3
+void ABGRToUVJRow_NEON(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_uj,
+                       uint8_t* dst_vj,
+                       int width) {
+  asm volatile (
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ABGR pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ABGR pixels.
+      "vpaddl.u8   q0, q0                        \n"  // R 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // B 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ABGR pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ABGR pixels.
+      "vpadal.u8   q0, q4                        \n"  // R 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // B 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q2, q1, q0)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_uj),     // %2
+    "+r"(dst_vj),     // %3
     "+r"(width)        // %4
   :
   : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
@@ -1603,45 +2028,43 @@ void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
   );
 }
 
-void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
-                      uint8* dst_u, uint8* dst_v, int width) {
+// TODO(fbarchard): Subsample match C code.
+void RGB24ToUVJRow_NEON(const uint8_t* src_rgb24,
+                        int src_stride_rgb24,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
-    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
-    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q2, q1, q0)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_abgr),  // %0
-    "+r"(src_stride_abgr),  // %1
+      "add         %1, %0, %1                    \n"  // src_stride + src_rgb24
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RGB24 pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RGB24 pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RGB24 pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RGB24 pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q0, q1, q2)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1651,45 +2074,43 @@ void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
   );
 }
 
-void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
-                      uint8* dst_u, uint8* dst_v, int width) {
+// TODO(fbarchard): Subsample match C code.
+void RAWToUVJRow_NEON(const uint8_t* src_raw,
+                      int src_stride_raw,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
-    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
-    MEMACCESS(1)
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
-    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q0, q1, q2)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_rgba),  // %0
-    "+r"(src_stride_rgba),  // %1
+      "add         %1, %0, %1                    \n"  // src_stride + src_raw
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RAW pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RAW pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RAW pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RAW pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q2, q1, q0)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1699,45 +2120,42 @@ void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
   );
 }
 
-void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
-                       uint8* dst_u, uint8* dst_v, int width) {
+void BGRAToUVRow_NEON(const uint8_t* src_bgra,
+                      int src_stride_bgra,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
-    MEMACCESS(0)
-    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
-    MEMACCESS(1)
-    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
-    RGBTOUV(q0, q1, q2)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_rgb24),  // %0
-    "+r"(src_stride_rgb24),  // %1
+      "add         %1, %0, %1                    \n"  // src_stride + src_bgra
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 BGRA pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 BGRA pixels.
+      "vpaddl.u8   q3, q3                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more BGRA pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 BGRA pixels.
+      "vpadal.u8   q3, q7                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q1, q1, #1                    \n"  // 2x average
+      "vrshr.u16   q2, q2, #1                    \n"
+      "vrshr.u16   q3, q3, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q3, q2, q1)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1747,45 +2165,42 @@ void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
   );
 }
 
-void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
-                     uint8* dst_u, uint8* dst_v, int width) {
+void ABGRToUVRow_NEON(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_raw
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
-    MEMACCESS(0)
-    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
-    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
-    MEMACCESS(1)
-    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
-    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_abgr
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ABGR pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ABGR pixels.
+      "vpaddl.u8   q2, q2                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q0, q0                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ABGR pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ABGR pixels.
+      "vpadal.u8   q2, q6                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q0, q4                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q2, q1, q0)
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_raw),  // %0
-    "+r"(src_stride_raw),  // %1
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1795,66 +2210,42 @@ void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
   );
 }
 
-// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
-                        uint8* dst_u, uint8* dst_v, int width) {
+void RGBAToUVRow_NEON(const uint8_t* src_rgba,
+                      int src_stride_rgba,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
-    RGB565TOARGB
-    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
-    RGB565TOARGB
-    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
-    RGB565TOARGB
-    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
-    RGB565TOARGB
-    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-    "vrshr.u16  q5, q5, #1                     \n"
-    "vrshr.u16  q6, q6, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q4, q10                    \n"  // B
-    "vmls.s16   q8, q5, q11                    \n"  // G
-    "vmls.s16   q8, q6, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q6, q10                    \n"  // R
-    "vmls.s16   q9, q5, q14                    \n"  // G
-    "vmls.s16   q9, q4, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_rgb565),  // %0
-    "+r"(src_stride_rgb565),  // %1
+      "add         %1, %0, %1                    \n"  // src_stride + src_rgba
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 RGBA pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 RGBA pixels.
+      "vpaddl.u8   q0, q1                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q2                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q3                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more RGBA pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 RGBA pixels.
+      "vpadal.u8   q0, q5                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q6                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q7                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q0, q1, q2)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1864,66 +2255,42 @@ void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
   );
 }
 
-// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
-                        uint8* dst_u, uint8* dst_v, int width) {
-  asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-    "vrshr.u16  q5, q5, #1                     \n"
-    "vrshr.u16  q6, q6, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q4, q10                    \n"  // B
-    "vmls.s16   q8, q5, q11                    \n"  // G
-    "vmls.s16   q8, q6, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q6, q10                    \n"  // R
-    "vmls.s16   q9, q5, q14                    \n"  // G
-    "vmls.s16   q9, q4, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_argb1555),  // %0
-    "+r"(src_stride_argb1555),  // %1
+void RGB24ToUVRow_NEON(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  asm volatile (
+      "add         %1, %0, %1                    \n"  // src_stride + src_rgb24
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RGB24 pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RGB24 pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RGB24 pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RGB24 pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q0, q1, q2)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1933,66 +2300,42 @@ void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
   );
 }
 
-// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
-                          uint8* dst_u, uint8* dst_v, int width) {
+void RAWToUVRow_NEON(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-    "vrshr.u16  q5, q5, #1                     \n"
-    "vrshr.u16  q6, q6, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-    "vmul.s16   q8, q4, q10                    \n"  // B
-    "vmls.s16   q8, q5, q11                    \n"  // G
-    "vmls.s16   q8, q6, q12                    \n"  // R
-    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-    "vmul.s16   q9, q6, q10                    \n"  // R
-    "vmls.s16   q9, q5, q14                    \n"  // G
-    "vmls.s16   q9, q4, q13                    \n"  // B
-    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
-  : "+r"(src_argb4444),  // %0
-    "+r"(src_stride_argb4444),  // %1
+      "add         %1, %0, %1                    \n"  // src_stride + src_raw
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RAW pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RAW pixels.
+      "vpaddl.u8   q2, q2                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q0, q0                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RAW pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RAW pixels.
+      "vpadal.u8   q2, q6                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q0, q4                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q2, q1, q0)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -2002,669 +2345,1028 @@ void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
   );
 }
 
-void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    RGB565TOARGB
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_rgb565),  // %0
-    "+r"(dst_y),       // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
-  );
-}
-
-void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGB1555TOARGB
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_argb1555),  // %0
-    "+r"(dst_y),         // %1
-    "+r"(width)            // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
-  );
-}
-
-void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    ARGB4444TOARGB
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_argb4444),  // %0
-    "+r"(dst_y),         // %1
-    "+r"(width)            // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
-  );
-}
-
-void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d1, d4                     \n"  // R
-    "vmlal.u8   q8, d2, d5                     \n"  // G
-    "vmlal.u8   q8, d3, d6                     \n"  // B
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_bgra),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
-  );
-}
-
-void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d0, d4                     \n"  // R
-    "vmlal.u8   q8, d1, d5                     \n"  // G
-    "vmlal.u8   q8, d2, d6                     \n"  // B
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_abgr),  // %0
-    "+r"(dst_y),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
-  );
+// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875
+                                                      // coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 RGB565 pixels.
+      RGB565TOARGB
+      "vpaddl.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%0]!                   \n"  // next 8 RGB565 pixels.
+      RGB565TOARGB
+      "vpaddl.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vld1.8      {q0}, [%1]!                   \n"  // load 8 RGB565 pixels.
+      RGB565TOARGB
+      "vpadal.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%1]!                   \n"  // next 8 RGB565 pixels.
+      RGB565TOARGB
+      "vpadal.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vrshr.u16   q4, q4, #1                    \n"  // 2x average
+      "vrshr.u16   q5, q5, #1                    \n"
+      "vrshr.u16   q6, q6, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+      "vmul.s16    q8, q4, q10                   \n"  // B
+      "vmls.s16    q8, q5, q11                   \n"  // G
+      "vmls.s16    q8, q6, q12                   \n"  // R
+      "vadd.u16    q8, q8, q15                   \n"  // +128 -> unsigned
+      "vmul.s16    q9, q6, q10                   \n"  // R
+      "vmls.s16    q9, q5, q14                   \n"  // G
+      "vmls.s16    q9, q4, q13                   \n"  // B
+      "vadd.u16    q9, q9, q15                   \n"  // +128 -> unsigned
+      "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+      "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+      : "+r"(src_rgb565),         // %0
+        "+r"(src_stride_rgb565),  // %1
+        "+r"(dst_u),              // %2
+        "+r"(dst_v),              // %3
+        "+r"(width)               // %4
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8",
+        "q9", "q10", "q11", "q12", "q13", "q14", "q15");
 }
 
-void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d1, d4                     \n"  // B
-    "vmlal.u8   q8, d2, d5                     \n"  // G
-    "vmlal.u8   q8, d3, d6                     \n"  // R
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_rgba),  // %0
-    "+r"(dst_y),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
-  );
+// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
+void ARGB1555ToUVRow_NEON(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875
+                                                      // coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "vpaddl.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%0]!                   \n"  // next 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "vpaddl.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vld1.8      {q0}, [%1]!                   \n"  // load 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "vpadal.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%1]!                   \n"  // next 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "vpadal.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vrshr.u16   q4, q4, #1                    \n"  // 2x average
+      "vrshr.u16   q5, q5, #1                    \n"
+      "vrshr.u16   q6, q6, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+      "vmul.s16    q8, q4, q10                   \n"  // B
+      "vmls.s16    q8, q5, q11                   \n"  // G
+      "vmls.s16    q8, q6, q12                   \n"  // R
+      "vadd.u16    q8, q8, q15                   \n"  // +128 -> unsigned
+      "vmul.s16    q9, q6, q10                   \n"  // R
+      "vmls.s16    q9, q5, q14                   \n"  // G
+      "vmls.s16    q9, q4, q13                   \n"  // B
+      "vadd.u16    q9, q9, q15                   \n"  // +128 -> unsigned
+      "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+      "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+      : "+r"(src_argb1555),         // %0
+        "+r"(src_stride_argb1555),  // %1
+        "+r"(dst_u),                // %2
+        "+r"(dst_v),                // %3
+        "+r"(width)                 // %4
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8",
+        "q9", "q10", "q11", "q12", "q13", "q14", "q15");
 }
 
-void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d0, d4                     \n"  // B
-    "vmlal.u8   q8, d1, d5                     \n"  // G
-    "vmlal.u8   q8, d2, d6                     \n"  // R
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_rgb24),  // %0
-    "+r"(dst_y),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
-  );
+// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
+void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
+                          int src_stride_argb4444,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875
+                                                      // coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "vpaddl.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%0]!                   \n"  // next 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "vpaddl.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vld1.8      {q0}, [%1]!                   \n"  // load 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "vpadal.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%1]!                   \n"  // next 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "vpadal.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vrshr.u16   q0, q4, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q5, #1                    \n"
+      "vrshr.u16   q2, q6, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+      RGBTOUV(q0, q1, q2)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+      : "+r"(src_argb4444),         // %0
+        "+r"(src_stride_argb4444),  // %1
+        "+r"(dst_u),                // %2
+        "+r"(dst_v),                // %3
+        "+r"(width)                 // %4
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8",
+        "q9", "q10", "q11", "q12", "q13", "q14", "q15");
+}
+
+void RGB565ToYRow_NEON(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vmov.u8     d24, #25                      \n"  // B * 0.1016 coefficient
+      "vmov.u8     d25, #129                     \n"  // G * 0.5078 coefficient
+      "vmov.u8     d26, #66                      \n"  // R * 0.2578 coefficient
+      "vmov.u8     d27, #16                      \n"  // Add 16 constant
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 RGB565 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      RGB565TOARGB
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit Y
+      "vqadd.u8    d0, d27                       \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_rgb565),  // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13");
+}
+
+void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width) {
+  asm volatile(
+      "vmov.u8     d24, #25                      \n"  // B * 0.1016 coefficient
+      "vmov.u8     d25, #129                     \n"  // G * 0.5078 coefficient
+      "vmov.u8     d26, #66                      \n"  // R * 0.2578 coefficient
+      "vmov.u8     d27, #16                      \n"  // Add 16 constant
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB1555 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGB1555TOARGB
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit Y
+      "vqadd.u8    d0, d27                       \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_argb1555),  // %0
+        "+r"(dst_y),         // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13");
+}
+
+void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
+                         uint8_t* dst_y,
+                         int width) {
+  asm volatile(
+      "vmov.u8     d24, #25                      \n"  // B * 0.1016 coefficient
+      "vmov.u8     d25, #129                     \n"  // G * 0.5078 coefficient
+      "vmov.u8     d26, #66                      \n"  // R * 0.2578 coefficient
+      "vmov.u8     d27, #16                      \n"  // Add 16 constant
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB4444 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      ARGB4444TOARGB
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit Y
+      "vqadd.u8    d0, d27                       \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_argb4444),  // %0
+        "+r"(dst_y),         // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13");
+}
+
+void ARGBToAR64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"
+      "vld1.8      {q2}, [%0]!                   \n"
+      "vmov.u8     q1, q0                        \n"
+      "vmov.u8     q3, q2                        \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst2.8      {q0, q1}, [%1]!               \n"  // store 4 pixels
+      "vst2.8      {q2, q3}, [%1]!               \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+static const uvec8 kShuffleARGBToABGR = {2,  1, 0, 3,  6,  5,  4,  7,
+                                         10, 9, 8, 11, 14, 13, 12, 15};
+
+void ARGBToAB64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
+  asm volatile(
+      "vld1.8      {q4}, [%3]                    \n"  // shuffler
+
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"
+      "vld1.8      {q2}, [%0]!                   \n"
+      "vtbl.8      d2, {d0, d1}, d8              \n"
+      "vtbl.8      d3, {d0, d1}, d9              \n"
+      "vtbl.8      d6, {d4, d5}, d8              \n"
+      "vtbl.8      d7, {d4, d5}, d9              \n"
+      "vmov.u8     q0, q1                        \n"
+      "vmov.u8     q2, q3                        \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst2.8      {q0, q1}, [%1]!               \n"  // store 4 pixels
+      "vst2.8      {q2, q3}, [%1]!               \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_argb),           // %0
+        "+r"(dst_ab64),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleARGBToABGR)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+void AR64ToARGBRow_NEON(const uint16_t* src_ar64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vld1.16     {q2}, [%0]!                   \n"
+      "vld1.16     {q3}, [%0]!                   \n"
+      "vshrn.u16   d0, q0, #8                    \n"
+      "vshrn.u16   d1, q1, #8                    \n"
+      "vshrn.u16   d4, q2, #8                    \n"
+      "vshrn.u16   d5, q3, #8                    \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst1.8      {q0}, [%1]!                   \n"  // store 4 pixels
+      "vst1.8      {q2}, [%1]!                   \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_ar64),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+static const uvec8 kShuffleAB64ToARGB = {5, 3, 1, 7, 13, 11, 9, 15};
+
+void AB64ToARGBRow_NEON(const uint16_t* src_ab64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "vld1.8      {d8}, [%3]                    \n"  // shuffler
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vld1.16     {q2}, [%0]!                   \n"
+      "vld1.16     {q3}, [%0]!                   \n"
+      "vtbl.8      d0, {d0, d1}, d8              \n"
+      "vtbl.8      d1, {d2, d3}, d8              \n"
+      "vtbl.8      d4, {d4, d5}, d8              \n"
+      "vtbl.8      d5, {d6, d7}, d8              \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst1.8      {q0}, [%1]!                   \n"  // store 4 pixels
+      "vst1.8      {q2}, [%1]!                   \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_ab64),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleAB64ToARGB)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+struct RgbConstants {
+  uint8_t kRGBToY[4];
+  uint16_t kAddY;
+  uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+                                                        128,
+                                                        0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+                                                        0x1080,
+                                                        0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+                                                      0x1080,
+                                                      0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+void ARGBToYMatrixRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vld1.8      {d0}, [%3]                    \n"  // load rgbconstants
+      "vdup.u8     d20, d0[0]                    \n"
+      "vdup.u8     d21, d0[1]                    \n"
+      "vdup.u8     d22, d0[2]                    \n"
+      "vdup.u16    q12, d0[2]                    \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 16 pixels of ARGB
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vmull.u8    q8, d0, d20                   \n"  // B
+      "vmull.u8    q9, d1, d20                   \n"
+      "vmlal.u8    q8, d2, d21                   \n"  // G
+      "vmlal.u8    q9, d3, d21                   \n"
+      "vmlal.u8    q8, d4, d22                   \n"  // R
+      "vmlal.u8    q9, d5, d22                   \n"
+      "vaddhn.u16  d0, q8, q12                   \n"  // 16 bit to 8 bit Y
+      "vaddhn.u16  d1, q9, q12                   \n"
+      "vst1.8      {d0, d1}, [%1]!               \n"  // store 16 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+        "q12");
+}
+
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+void ABGRToYJRow_NEON(const uint8_t* src_abgr, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_NEON(src_abgr, dst_yj, width, &kRawJPEGConstants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+void RGBAToYMatrixRow_NEON(const uint8_t* src_rgba,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vld1.8      {d0}, [%3]                    \n"  // load rgbconstants
+      "vdup.u8     d20, d0[0]                    \n"
+      "vdup.u8     d21, d0[1]                    \n"
+      "vdup.u8     d22, d0[2]                    \n"
+      "vdup.u16    q12, d0[2]                    \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 16 pixels of RGBA
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vmull.u8    q8, d2, d20                   \n"  // B
+      "vmull.u8    q9, d3, d20                   \n"
+      "vmlal.u8    q8, d4, d21                   \n"  // G
+      "vmlal.u8    q9, d5, d21                   \n"
+      "vmlal.u8    q8, d6, d22                   \n"  // R
+      "vmlal.u8    q9, d7, d22                   \n"
+      "vaddhn.u16  d0, q8, q12                   \n"  // 16 bit to 8 bit Y
+      "vaddhn.u16  d1, q9, q12                   \n"
+      "vst1.8      {d0, d1}, [%1]!               \n"  // store 16 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_rgba),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+        "q12");
+}
+
+void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+void RGBToYMatrixRow_NEON(const uint8_t* src_rgb,
+                          uint8_t* dst_y,
+                          int width,
+                          const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vld1.8      {d0}, [%3]                    \n"  // load rgbconstants
+      "vdup.u8     d20, d0[0]                    \n"
+      "vdup.u8     d21, d0[1]                    \n"
+      "vdup.u8     d22, d0[2]                    \n"
+      "vdup.u16    q12, d0[2]                    \n"
+      "1:                                        \n"
+      "vld3.8      {d2, d4, d6}, [%0]!           \n"  // load 16 pixels of
+                                                      // RGB24.
+      "vld3.8      {d3, d5, d7}, [%0]!           \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vmull.u8    q8, d2, d20                   \n"  // B
+      "vmull.u8    q9, d3, d20                   \n"
+      "vmlal.u8    q8, d4, d21                   \n"  // G
+      "vmlal.u8    q9, d5, d21                   \n"
+      "vmlal.u8    q8, d6, d22                   \n"  // R
+      "vmlal.u8    q9, d7, d22                   \n"
+      "vaddhn.u16  d0, q8, q12                   \n"  // 16 bit to 8 bit Y
+      "vaddhn.u16  d1, q9, q12                   \n"
+      "vst1.8      {d0, d1}, [%1]!               \n"  // store 16 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_rgb),     // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+        "q12");
+}
+
+void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
 }
 
-void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int width) {
-  asm volatile (
-    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-    "vmov.u8    d7, #16                        \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q8, d0, d4                     \n"  // B
-    "vmlal.u8   q8, d1, d5                     \n"  // G
-    "vmlal.u8   q8, d2, d6                     \n"  // R
-    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d7                         \n"
-    MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-    "bgt        1b                             \n"
-  : "+r"(src_raw),  // %0
-    "+r"(dst_y),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
-  );
+void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_NEON(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_NEON(src_raw, dst_y, width, &kRawI601Constants);
 }
 
 // Bilinear filter 16x2 -> 16x1
-void InterpolateRow_NEON(uint8* dst_ptr,
-                         const uint8* src_ptr, ptrdiff_t src_stride,
-                         int dst_width, int source_y_fraction) {
+void InterpolateRow_NEON(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         int dst_width,
+                         int source_y_fraction) {
   int y1_fraction = source_y_fraction;
-  asm volatile (
-    "cmp        %4, #0                         \n"
-    "beq        100f                           \n"
-    "add        %2, %1                         \n"
-    "cmp        %4, #128                       \n"
-    "beq        50f                            \n"
-
-    "vdup.8     d5, %4                         \n"
-    "rsb        %4, #256                       \n"
-    "vdup.8     d4, %4                         \n"
-    // General purpose row blend.
-  "1:                                          \n"
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
-    MEMACCESS(2)
-    "vld1.8     {q1}, [%2]!                    \n"
-    "subs       %3, %3, #16                    \n"
-    "vmull.u8   q13, d0, d4                    \n"
-    "vmull.u8   q14, d1, d4                    \n"
-    "vmlal.u8   q13, d2, d5                    \n"
-    "vmlal.u8   q14, d3, d5                    \n"
-    "vrshrn.u16 d0, q13, #8                    \n"
-    "vrshrn.u16 d1, q14, #8                    \n"
-    MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        1b                             \n"
-    "b          99f                            \n"
-
-    // Blend 50 / 50.
-  "50:                                         \n"
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
-    MEMACCESS(2)
-    "vld1.8     {q1}, [%2]!                    \n"
-    "subs       %3, %3, #16                    \n"
-    "vrhadd.u8  q0, q1                         \n"
-    MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        50b                            \n"
-    "b          99f                            \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-  "100:                                        \n"
-    MEMACCESS(1)
-    "vld1.8     {q0}, [%1]!                    \n"
-    "subs       %3, %3, #16                    \n"
-    MEMACCESS(0)
-    "vst1.8     {q0}, [%0]!                    \n"
-    "bgt        100b                           \n"
-
-  "99:                                         \n"
-  : "+r"(dst_ptr),          // %0
-    "+r"(src_ptr),          // %1
-    "+r"(src_stride),       // %2
-    "+r"(dst_width),        // %3
-    "+r"(y1_fraction)       // %4
-  :
-  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
-  );
+  asm volatile(
+      "cmp         %4, #0                        \n"
+      "beq         100f                          \n"
+      "add         %2, %1                        \n"
+      "cmp         %4, #128                      \n"
+      "beq         50f                           \n"
+
+      "vdup.8      d5, %4                        \n"
+      "rsb         %4, #256                      \n"
+      "vdup.8      d4, %4                        \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vmull.u8    q13, d0, d4                   \n"
+      "vmull.u8    q14, d1, d4                   \n"
+      "vmlal.u8    q13, d2, d5                   \n"
+      "vmlal.u8    q14, d3, d5                   \n"
+      "vrshrn.u16  d0, q13, #8                   \n"
+      "vrshrn.u16  d1, q14, #8                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_stride),  // %2
+        "+r"(dst_width),   // %3
+        "+r"(y1_fraction)  // %4
+      :
+      : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14");
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRow_16_NEON(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int dst_width,
+                            int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+
+  asm volatile(
+      "cmp         %4, #0                        \n"
+      "beq         100f                          \n"
+      "cmp         %4, #128                      \n"
+      "beq         50f                           \n"
+
+      "vdup.16     d17, %4                       \n"
+      "vdup.16     d16, %5                       \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "vld1.16     {q0}, [%1]!                   \n"
+      "vld1.16     {q1}, [%2]!                   \n"
+      "subs        %3, %3, #8                    \n"
+      "vmull.u16   q2, d0, d16                   \n"
+      "vmull.u16   q3, d1, d16                   \n"
+      "vmlal.u16   q2, d2, d17                   \n"
+      "vmlal.u16   q3, d3, d17                   \n"
+      "vrshrn.u32  d0, q2, #8                    \n"
+      "vrshrn.u32  d1, q3, #8                    \n"
+      "vst1.16     {q0}, [%0]!                   \n"
+      "bgt         1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "vld1.16     {q0}, [%1]!                   \n"
+      "vld1.16     {q1}, [%2]!                   \n"
+      "subs        %3, %3, #8                    \n"
+      "vrhadd.u16  q0, q1                        \n"
+      "vst1.16     {q0}, [%0]!                   \n"
+      "bgt         50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "vld1.16     {q0}, [%1]!                   \n"
+      "subs        %3, %3, #8                    \n"
+      "vst1.16     {q0}, [%0]!                   \n"
+      "bgt         100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_ptr1),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(y1_fraction),  // %4
+        "r"(y0_fraction)   // %5
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8");
 }
 
 // dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
-void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                       uint8* dst_argb, int width) {
-  asm volatile (
-    "subs       %3, #8                         \n"
-    "blt        89f                            \n"
-    // Blend 8 pixels.
-  "8:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
-    MEMACCESS(1)
-    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q10, d4, d3                    \n"  // db * a
-    "vmull.u8   q11, d5, d3                    \n"  // dg * a
-    "vmull.u8   q12, d6, d3                    \n"  // dr * a
-    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
-    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
-    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
-    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
-    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
-    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
-    "vqadd.u8   d2, d2, d6                     \n"  // + sr
-    "vmov.u8    d3, #255                       \n"  // a = 255
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
-    "bge        8b                             \n"
-
-  "89:                                         \n"
-    "adds       %3, #8-1                       \n"
-    "blt        99f                            \n"
-
-    // Blend 1 pixels.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
-    MEMACCESS(1)
-    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
-    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
-    "vmull.u8   q10, d4, d3                    \n"  // db * a
-    "vmull.u8   q11, d5, d3                    \n"  // dg * a
-    "vmull.u8   q12, d6, d3                    \n"  // dr * a
-    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
-    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
-    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
-    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
-    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
-    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
-    "vqadd.u8   d2, d2, d6                     \n"  // + sr
-    "vmov.u8    d3, #255                       \n"  // a = 255
-    MEMACCESS(2)
-    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
-    "bge        1b                             \n"
-
-  "99:                                         \n"
-
-  : "+r"(src_argb0),    // %0
-    "+r"(src_argb1),    // %1
-    "+r"(dst_argb),     // %2
-    "+r"(width)         // %3
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
-  );
+void ARGBBlendRow_NEON(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "subs        %3, #8                        \n"
+      "blt         89f                           \n"
+      // Blend 8 pixels.
+      "8:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 pixels of ARGB0.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load 8 pixels of ARGB1.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q10, d4, d3                   \n"  // db * a
+      "vmull.u8    q11, d5, d3                   \n"  // dg * a
+      "vmull.u8    q12, d6, d3                   \n"  // dr * a
+      "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+      "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+      "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+      "vqsub.u8    q2, q2, q10                   \n"  // dbg - dbg * a / 256
+      "vqsub.u8    d6, d6, d22                   \n"  // dr - dr * a / 256
+      "vqadd.u8    q0, q0, q2                    \n"  // + sbg
+      "vqadd.u8    d2, d2, d6                    \n"  // + sr
+      "vmov.u8     d3, #255                      \n"  // a = 255
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 pixels of ARGB.
+      "bge         8b                            \n"
+
+      "89:                                       \n"
+      "adds        %3, #8-1                      \n"
+      "blt         99f                           \n"
+
+      // Blend 1 pixels.
+      "1:                                        \n"
+      "vld4.8      {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+      "vld4.8      {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+      "subs        %3, %3, #1                    \n"  // 1 processed per loop.
+      "vmull.u8    q10, d4, d3                   \n"  // db * a
+      "vmull.u8    q11, d5, d3                   \n"  // dg * a
+      "vmull.u8    q12, d6, d3                   \n"  // dr * a
+      "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+      "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+      "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+      "vqsub.u8    q2, q2, q10                   \n"  // dbg - dbg * a / 256
+      "vqsub.u8    d6, d6, d22                   \n"  // dr - dr * a / 256
+      "vqadd.u8    q0, q0, q2                    \n"  // + sbg
+      "vqadd.u8    d2, d2, d6                    \n"  // + sr
+      "vmov.u8     d3, #255                      \n"  // a = 255
+      "vst4.8      {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+      "bge         1b                            \n"
+
+      "99:                                       \n"
+
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12");
 }
 
 // Attenuate 8 pixels at a time.
-void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    // Attenuate 8 pixels.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q10, d0, d3                    \n"  // b * a
-    "vmull.u8   q11, d1, d3                    \n"  // g * a
-    "vmull.u8   q12, d2, d3                    \n"  // r * a
-    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
-    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
-    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
-  );
+void ARGBAttenuateRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width) {
+  asm volatile(
+      // Attenuate 8 pixels.
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q10, d0, d3                   \n"  // b * a
+      "vmull.u8    q11, d1, d3                   \n"  // g * a
+      "vmull.u8    q12, d2, d3                   \n"  // r * a
+      "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+      "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+      "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1", "q10", "q11", "q12");
 }
 
 // Quantize 8 ARGB pixels (32 bytes).
 // dst = (dst * scale >> 16) * interval_size + interval_offset;
-void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
-                          int interval_offset, int width) {
-  asm volatile (
-    "vdup.u16   q8, %2                         \n"
-    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
-    "vdup.u16   q9, %3                         \n"  // interval multiply.
-    "vdup.u16   q10, %4                        \n"  // interval add
-
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
-    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
-    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
-    "vmovl.u8   q1, d2                         \n"
-    "vmovl.u8   q2, d4                         \n"
-    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
-    "vqdmulh.s16 q1, q1, q8                    \n"  // g
-    "vqdmulh.s16 q2, q2, q8                    \n"  // r
-    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
-    "vmul.u16   q1, q1, q9                     \n"  // g
-    "vmul.u16   q2, q2, q9                     \n"  // r
-    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
-    "vadd.u16   q1, q1, q10                    \n"  // g
-    "vadd.u16   q2, q2, q10                    \n"  // r
-    "vqmovn.u16 d0, q0                         \n"
-    "vqmovn.u16 d2, q1                         \n"
-    "vqmovn.u16 d4, q2                         \n"
-    MEMACCESS(0)
-    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(dst_argb),       // %0
-    "+r"(width)           // %1
-  : "r"(scale),           // %2
-    "r"(interval_size),   // %3
-    "r"(interval_offset)  // %4
-  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
-  );
+void ARGBQuantizeRow_NEON(uint8_t* dst_argb,
+                          int scale,
+                          int interval_size,
+                          int interval_offset,
+                          int width) {
+  asm volatile(
+      "vdup.u16    q8, %2                        \n"
+      "vshr.u16    q8, q8, #1                    \n"  // scale >>= 1
+      "vdup.u16    q9, %3                        \n"  // interval multiply.
+      "vdup.u16    q10, %4                       \n"  // interval add
+
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]        \n"  // load 8 pixels of ARGB.
+      "subs        %1, %1, #8                    \n"  // 8 processed per loop.
+      "vmovl.u8    q0, d0                        \n"  // b (0 .. 255)
+      "vmovl.u8    q1, d2                        \n"
+      "vmovl.u8    q2, d4                        \n"
+      "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+      "vqdmulh.s16 q1, q1, q8                    \n"  // g
+      "vqdmulh.s16 q2, q2, q8                    \n"  // r
+      "vmul.u16    q0, q0, q9                    \n"  // b * interval_size
+      "vmul.u16    q1, q1, q9                    \n"  // g
+      "vmul.u16    q2, q2, q9                    \n"  // r
+      "vadd.u16    q0, q0, q10                   \n"  // b + interval_offset
+      "vadd.u16    q1, q1, q10                   \n"  // g
+      "vadd.u16    q2, q2, q10                   \n"  // r
+      "vqmovn.u16  d0, q0                        \n"
+      "vqmovn.u16  d2, q1                        \n"
+      "vqmovn.u16  d4, q2                        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%0]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(dst_argb),       // %0
+        "+r"(width)           // %1
+      : "r"(scale),           // %2
+        "r"(interval_size),   // %3
+        "r"(interval_offset)  // %4
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10");
 }
 
 // Shade 8 pixels at a time by specified value.
 // NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
 // Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
-void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value) {
-  asm volatile (
-    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
-    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
-    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
-
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
-    "vmovl.u8   q11, d22                       \n"
-    "vmovl.u8   q12, d24                       \n"
-    "vmovl.u8   q13, d26                       \n"
-    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
-    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
-    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
-    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
-    "vqmovn.u16 d20, q10                       \n"
-    "vqmovn.u16 d22, q11                       \n"
-    "vqmovn.u16 d24, q12                       \n"
-    "vqmovn.u16 d26, q13                       \n"
-    MEMACCESS(1)
-    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),       // %0
-    "+r"(dst_argb),       // %1
-    "+r"(width)           // %2
-  : "r"(value)            // %3
-  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
-  );
+void ARGBShadeRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value) {
+  asm volatile(
+      "vdup.u32    q0, %3                        \n"  // duplicate scale value.
+      "vzip.u8     d0, d1                        \n"  // d0 aarrggbb.
+      "vshr.u16    q0, q0, #1                    \n"  // scale / 2.
+
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld4.8      {d20, d22, d24, d26}, [%0]!   \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmovl.u8    q10, d20                      \n"  // b (0 .. 255)
+      "vmovl.u8    q11, d22                      \n"
+      "vmovl.u8    q12, d24                      \n"
+      "vmovl.u8    q13, d26                      \n"
+      "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+      "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+      "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+      "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+      "vqmovn.u16  d20, q10                      \n"
+      "vqmovn.u16  d22, q11                      \n"
+      "vqmovn.u16  d24, q12                      \n"
+      "vqmovn.u16  d26, q13                      \n"
+      "vst4.8      {d20, d22, d24, d26}, [%1]!   \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(value)       // %3
+      : "cc", "memory", "q0", "q10", "q11", "q12", "q13");
 }
 
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 // Similar to ARGBToYJ but stores ARGB.
-// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
-void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
-    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
-    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
-    "vmov       d1, d0                         \n"  // G
-    "vmov       d2, d0                         \n"  // R
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
-  );
+// C code is (29 * b + 150 * g + 77 * r + 128) >> 8;
+void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "vmov.u8     d24, #29                      \n"  // B * 0.1140 coefficient
+      "vmov.u8     d25, #150                     \n"  // G * 0.5870 coefficient
+      "vmov.u8     d26, #77                      \n"  // R * 0.2990 coefficient
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit B
+      "vmov        d1, d0                        \n"  // G
+      "vmov        d2, d0                        \n"  // R
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q12", "q13");
 }
 
 // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
 //    b = (r * 35 + g * 68 + b * 17) >> 7
 //    g = (r * 45 + g * 88 + b * 22) >> 7
 //    r = (r * 50 + g * 98 + b * 24) >> 7
-void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d20, #17                       \n"  // BB coefficient
-    "vmov.u8    d21, #68                       \n"  // BG coefficient
-    "vmov.u8    d22, #35                       \n"  // BR coefficient
-    "vmov.u8    d24, #22                       \n"  // GB coefficient
-    "vmov.u8    d25, #88                       \n"  // GG coefficient
-    "vmov.u8    d26, #45                       \n"  // GR coefficient
-    "vmov.u8    d28, #24                       \n"  // BB coefficient
-    "vmov.u8    d29, #98                       \n"  // BG coefficient
-    "vmov.u8    d30, #50                       \n"  // BR coefficient
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
-    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
-    "vmlal.u8   q2, d1, d21                    \n"  // G
-    "vmlal.u8   q2, d2, d22                    \n"  // R
-    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
-    "vmlal.u8   q3, d1, d25                    \n"  // G
-    "vmlal.u8   q3, d2, d26                    \n"  // R
-    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
-    "vmlal.u8   q8, d1, d29                    \n"  // G
-    "vmlal.u8   q8, d2, d30                    \n"  // R
-    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
-    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
-    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
-    MEMACCESS(0)
-    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-  : "+r"(dst_argb),  // %0
-    "+r"(width)      // %1
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3",
-    "q10", "q11", "q12", "q13", "q14", "q15"
-  );
+void ARGBSepiaRow_NEON(uint8_t* dst_argb, int width) {
+  asm volatile(
+      "vmov.u8     d20, #17                      \n"  // BB coefficient
+      "vmov.u8     d21, #68                      \n"  // BG coefficient
+      "vmov.u8     d22, #35                      \n"  // BR coefficient
+      "vmov.u8     d24, #22                      \n"  // GB coefficient
+      "vmov.u8     d25, #88                      \n"  // GG coefficient
+      "vmov.u8     d26, #45                      \n"  // GR coefficient
+      "vmov.u8     d28, #24                      \n"  // BB coefficient
+      "vmov.u8     d29, #98                      \n"  // BG coefficient
+      "vmov.u8     d30, #50                      \n"  // BR coefficient
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]        \n"  // load 8 ARGB pixels.
+      "subs        %1, %1, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q2, d0, d20                   \n"  // B to Sepia B
+      "vmlal.u8    q2, d1, d21                   \n"  // G
+      "vmlal.u8    q2, d2, d22                   \n"  // R
+      "vmull.u8    q3, d0, d24                   \n"  // B to Sepia G
+      "vmlal.u8    q3, d1, d25                   \n"  // G
+      "vmlal.u8    q3, d2, d26                   \n"  // R
+      "vmull.u8    q8, d0, d28                   \n"  // B to Sepia R
+      "vmlal.u8    q8, d1, d29                   \n"  // G
+      "vmlal.u8    q8, d2, d30                   \n"  // R
+      "vqshrn.u16  d0, q2, #7                    \n"  // 16 bit to 8 bit B
+      "vqshrn.u16  d1, q3, #7                    \n"  // 16 bit to 8 bit G
+      "vqshrn.u16  d2, q8, #7                    \n"  // 16 bit to 8 bit R
+      "vst4.8      {d0, d1, d2, d3}, [%0]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(dst_argb),  // %0
+        "+r"(width)      // %1
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12", "q13",
+        "q14", "q15");
 }
 
 // Tranform 8 ARGB pixels (32 bytes) with color matrix.
 // TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
 // needs to saturate.  Consider doing a non-saturating version.
-void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
-                             const int8* matrix_argb, int width) {
-  asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
-    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
-    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
-
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
-    "vmovl.u8   q9, d18                        \n"  // g
-    "vmovl.u8   q10, d20                       \n"  // r
-    "vmovl.u8   q11, d22                       \n"  // a
-    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
-    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
-    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
-    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
-    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
-    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
-    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
-    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
-    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
-    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
-    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
-    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
-    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vmul.s16   q4, q11, d0[3]                 \n"  // B += A * Matrix B
-    "vmul.s16   q5, q11, d1[3]                 \n"  // G += A * Matrix G
-    "vmul.s16   q6, q11, d2[3]                 \n"  // R += A * Matrix R
-    "vmul.s16   q7, q11, d3[3]                 \n"  // A += A * Matrix A
-    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
-    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
-    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
-    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
-    MEMACCESS(1)
-    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)       // %2
-  : "r"(matrix_argb)  // %3
-  : "cc", "memory", "q0", "q1", "q2", "q4", "q5", "q6", "q7", "q8", "q9",
-    "q10", "q11", "q12", "q13", "q14", "q15"
-  );
+void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             const int8_t* matrix_argb,
+                             int width) {
+  asm volatile(
+      "vld1.8      {q2}, [%3]                    \n"  // load 3 ARGB vectors.
+      "vmovl.s8    q0, d4                        \n"  // B,G coefficients s16.
+      "vmovl.s8    q1, d5                        \n"  // R,A coefficients s16.
+
+      "1:                                        \n"
+      "vld4.8      {d16, d18, d20, d22}, [%0]!   \n"  // load 8 ARGB pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmovl.u8    q8, d16                       \n"  // b (0 .. 255) 16 bit
+      "vmovl.u8    q9, d18                       \n"  // g
+      "vmovl.u8    q10, d20                      \n"  // r
+      "vmovl.u8    q11, d22                      \n"  // a
+      "vmul.s16    q12, q8, d0[0]                \n"  // B = B * Matrix B
+      "vmul.s16    q13, q8, d1[0]                \n"  // G = B * Matrix G
+      "vmul.s16    q14, q8, d2[0]                \n"  // R = B * Matrix R
+      "vmul.s16    q15, q8, d3[0]                \n"  // A = B * Matrix A
+      "vmul.s16    q4, q9, d0[1]                 \n"  // B += G * Matrix B
+      "vmul.s16    q5, q9, d1[1]                 \n"  // G += G * Matrix G
+      "vmul.s16    q6, q9, d2[1]                 \n"  // R += G * Matrix R
+      "vmul.s16    q7, q9, d3[1]                 \n"  // A += G * Matrix A
+      "vqadd.s16   q12, q12, q4                  \n"  // Accumulate B
+      "vqadd.s16   q13, q13, q5                  \n"  // Accumulate G
+      "vqadd.s16   q14, q14, q6                  \n"  // Accumulate R
+      "vqadd.s16   q15, q15, q7                  \n"  // Accumulate A
+      "vmul.s16    q4, q10, d0[2]                \n"  // B += R * Matrix B
+      "vmul.s16    q5, q10, d1[2]                \n"  // G += R * Matrix G
+      "vmul.s16    q6, q10, d2[2]                \n"  // R += R * Matrix R
+      "vmul.s16    q7, q10, d3[2]                \n"  // A += R * Matrix A
+      "vqadd.s16   q12, q12, q4                  \n"  // Accumulate B
+      "vqadd.s16   q13, q13, q5                  \n"  // Accumulate G
+      "vqadd.s16   q14, q14, q6                  \n"  // Accumulate R
+      "vqadd.s16   q15, q15, q7                  \n"  // Accumulate A
+      "vmul.s16    q4, q11, d0[3]                \n"  // B += A * Matrix B
+      "vmul.s16    q5, q11, d1[3]                \n"  // G += A * Matrix G
+      "vmul.s16    q6, q11, d2[3]                \n"  // R += A * Matrix R
+      "vmul.s16    q7, q11, d3[3]                \n"  // A += A * Matrix A
+      "vqadd.s16   q12, q12, q4                  \n"  // Accumulate B
+      "vqadd.s16   q13, q13, q5                  \n"  // Accumulate G
+      "vqadd.s16   q14, q14, q6                  \n"  // Accumulate R
+      "vqadd.s16   q15, q15, q7                  \n"  // Accumulate A
+      "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+      "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+      "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+      "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+      "vst4.8      {d16, d18, d20, d22}, [%1]!   \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(dst_argb),   // %1
+        "+r"(width)       // %2
+      : "r"(matrix_argb)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q4", "q5", "q6", "q7", "q8", "q9",
+        "q10", "q11", "q12", "q13", "q14", "q15");
 }
 
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q0, d0, d1                     \n"  // multiply B
-    "vmull.u8   q1, d2, d3                     \n"  // multiply G
-    "vmull.u8   q2, d4, d5                     \n"  // multiply R
-    "vmull.u8   q3, d6, d7                     \n"  // multiply A
-    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
-    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
-    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
-    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3"
-  );
+void ARGBMultiplyRow_NEON(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%1]!       \n"  // load 8 more ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q0, d0, d1                    \n"  // multiply B
+      "vmull.u8    q1, d2, d3                    \n"  // multiply G
+      "vmull.u8    q2, d4, d5                    \n"  // multiply R
+      "vmull.u8    q3, d6, d7                    \n"  // multiply A
+      "vrshrn.u16  d0, q0, #8                    \n"  // 16 bit to 8 bit B
+      "vrshrn.u16  d1, q1, #8                    \n"  // 16 bit to 8 bit G
+      "vrshrn.u16  d2, q2, #8                    \n"  // 16 bit to 8 bit R
+      "vrshrn.u16  d3, q3, #8                    \n"  // 16 bit to 8 bit A
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
 // Add 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
-    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3"
-  );
+void ARGBAddRow_NEON(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load 8 more ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    q0, q0, q2                    \n"  // add B, G
+      "vqadd.u8    q1, q1, q3                    \n"  // add R, A
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
 // Subtract 2 rows of ARGB pixels, 8 pixels at a time.
-void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
-    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "cc", "memory", "q0", "q1", "q2", "q3"
-  );
+void ARGBSubtractRow_NEON(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load 8 more ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqsub.u8    q0, q0, q2                    \n"  // subtract B, G
+      "vqsub.u8    q1, q1, q3                    \n"  // subtract R, A
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
 // Adds Sobel X and Sobel Y and stores Sobel into ARGB.
@@ -2672,54 +3374,50 @@ void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
 // R = Sobel
 // G = Sobel
 // B = Sobel
-void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d3, #255                       \n"  // alpha
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
-    MEMACCESS(1)
-    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vqadd.u8   d0, d0, d1                     \n"  // add
-    "vmov.u8    d1, d0                         \n"
-    "vmov.u8    d2, d0                         \n"
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "q0", "q1"
-  );
+void SobelRow_NEON(const uint8_t* src_sobelx,
+                   const uint8_t* src_sobely,
+                   uint8_t* dst_argb,
+                   int width) {
+  asm volatile(
+      "vmov.u8     d3, #255                      \n"  // alpha
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0]!                   \n"  // load 8 sobelx.
+      "vld1.8      {d1}, [%1]!                   \n"  // load 8 sobely.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    d0, d0, d1                    \n"  // add
+      "vmov.u8     d1, d0                        \n"
+      "vmov.u8     d2, d0                        \n"
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "q0", "q1");
 }
 
 // Adds Sobel X and Sobel Y and stores Sobel into plane.
-void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_y, int width) {
-  asm volatile (
-    // 16 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
-    MEMACCESS(1)
-    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
-    "vqadd.u8   q0, q0, q1                     \n"  // add
-    MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_y),       // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "q0", "q1"
-  );
+void SobelToPlaneRow_NEON(const uint8_t* src_sobelx,
+                          const uint8_t* src_sobely,
+                          uint8_t* dst_y,
+                          int width) {
+  asm volatile(
+      // 16 pixel loop.
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 16 sobelx.
+      "vld1.8      {q1}, [%1]!                   \n"  // load 16 sobely.
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop.
+      "vqadd.u8    q0, q0, q1                    \n"  // add
+      "vst1.8      {q0}, [%2]!                   \n"  // store 16 pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_y),       // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "q0", "q1");
 }
 
 // Mixes Sobel X, Sobel Y and Sobel into ARGB.
@@ -2727,70 +3425,64 @@ void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
 // R = Sobel X
 // G = Sobel
 // B = Sobel Y
-void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    "vmov.u8    d3, #255                       \n"  // alpha
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
-    MEMACCESS(1)
-    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vqadd.u8   d1, d0, d2                     \n"  // add
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-    "bgt        1b                             \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "q0", "q1"
-  );
+void SobelXYRow_NEON(const uint8_t* src_sobelx,
+                     const uint8_t* src_sobely,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      "vmov.u8     d3, #255                      \n"  // alpha
+      // 8 pixel loop.
+      "1:                                        \n"
+      "vld1.8      {d2}, [%0]!                   \n"  // load 8 sobelx.
+      "vld1.8      {d0}, [%1]!                   \n"  // load 8 sobely.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    d1, d0, d2                    \n"  // add
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "q0", "q1");
 }
 
 // SobelX as a matrix is
 // -1  0  1
 // -2  0  2
 // -1  0  1
-void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobelx, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d0}, [%0],%5                  \n"  // top
-    MEMACCESS(0)
-    "vld1.8     {d1}, [%0],%6                  \n"
-    "vsubl.u8   q0, d0, d1                     \n"
-    MEMACCESS(1)
-    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
-    MEMACCESS(1)
-    "vld1.8     {d3}, [%1],%6                  \n"
-    "vsubl.u8   q1, d2, d3                     \n"
-    "vadd.s16   q0, q0, q1                     \n"
-    "vadd.s16   q0, q0, q1                     \n"
-    MEMACCESS(2)
-    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
-    MEMACCESS(2)
-    "vld1.8     {d3}, [%2],%6                  \n"
-    "subs       %4, %4, #8                     \n"  // 8 pixels
-    "vsubl.u8   q1, d2, d3                     \n"
-    "vadd.s16   q0, q0, q1                     \n"
-    "vabs.s16   q0, q0                         \n"
-    "vqmovn.u16 d0, q0                         \n"
-    MEMACCESS(3)
-    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
-    "bgt        1b                             \n"
-  : "+r"(src_y0),      // %0
-    "+r"(src_y1),      // %1
-    "+r"(src_y2),      // %2
-    "+r"(dst_sobelx),  // %3
-    "+r"(width)        // %4
-  : "r"(2),            // %5
-    "r"(6)             // %6
-  : "cc", "memory", "q0", "q1"  // Clobber List
+void SobelXRow_NEON(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    const uint8_t* src_y2,
+                    uint8_t* dst_sobelx,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0],%5                 \n"  // top
+      "vld1.8      {d1}, [%0],%6                 \n"
+      "vsubl.u8    q0, d0, d1                    \n"
+      "vld1.8      {d2}, [%1],%5                 \n"  // center * 2
+      "vld1.8      {d3}, [%1],%6                 \n"
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vld1.8      {d2}, [%2],%5                 \n"  // bottom
+      "vld1.8      {d3}, [%2],%6                 \n"
+      "subs        %4, %4, #8                    \n"  // 8 pixels
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vabs.s16    q0, q0                        \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vst1.8      {d0}, [%3]!                   \n"  // store 8 sobelx
+      "bgt         1b                            \n"
+      : "+r"(src_y0),               // %0
+        "+r"(src_y1),               // %1
+        "+r"(src_y2),               // %2
+        "+r"(dst_sobelx),           // %3
+        "+r"(width)                 // %4
+      : "r"(2),                     // %5
+        "r"(6)                      // %6
+      : "cc", "memory", "q0", "q1"  // Clobber List
   );
 }
 
@@ -2798,44 +3490,478 @@ void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
 // -1 -2 -1
 //  0  0  0
 //  1  2  1
-void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d0}, [%0],%4                  \n"  // left
-    MEMACCESS(1)
-    "vld1.8     {d1}, [%1],%4                  \n"
-    "vsubl.u8   q0, d0, d1                     \n"
-    MEMACCESS(0)
-    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
-    MEMACCESS(1)
-    "vld1.8     {d3}, [%1],%4                  \n"
-    "vsubl.u8   q1, d2, d3                     \n"
-    "vadd.s16   q0, q0, q1                     \n"
-    "vadd.s16   q0, q0, q1                     \n"
-    MEMACCESS(0)
-    "vld1.8     {d2}, [%0],%5                  \n"  // right
-    MEMACCESS(1)
-    "vld1.8     {d3}, [%1],%5                  \n"
-    "subs       %3, %3, #8                     \n"  // 8 pixels
-    "vsubl.u8   q1, d2, d3                     \n"
-    "vadd.s16   q0, q0, q1                     \n"
-    "vabs.s16   q0, q0                         \n"
-    "vqmovn.u16 d0, q0                         \n"
-    MEMACCESS(2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
-    "bgt        1b                             \n"
-  : "+r"(src_y0),      // %0
-    "+r"(src_y1),      // %1
-    "+r"(dst_sobely),  // %2
-    "+r"(width)        // %3
-  : "r"(1),            // %4
-    "r"(6)             // %5
-  : "cc", "memory", "q0", "q1"  // Clobber List
-  );
-}
-#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+void SobelYRow_NEON(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    uint8_t* dst_sobely,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0],%4                 \n"  // left
+      "vld1.8      {d1}, [%1],%4                 \n"
+      "vsubl.u8    q0, d0, d1                    \n"
+      "vld1.8      {d2}, [%0],%4                 \n"  // center * 2
+      "vld1.8      {d3}, [%1],%4                 \n"
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vld1.8      {d2}, [%0],%5                 \n"  // right
+      "vld1.8      {d3}, [%1],%5                 \n"
+      "subs        %3, %3, #8                    \n"  // 8 pixels
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vabs.s16    q0, q0                        \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 sobely
+      "bgt         1b                            \n"
+      : "+r"(src_y0),               // %0
+        "+r"(src_y1),               // %1
+        "+r"(dst_sobely),           // %2
+        "+r"(width)                 // %3
+      : "r"(1),                     // %4
+        "r"(6)                      // %5
+      : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// %y passes a float as a scalar vector for vector * scalar multiply.
+// the regoster must be d0 to d15 and indexed with [0] or [1] to access
+// the float in the first or second float of the d-reg
+
+void HalfFloat1Row_NEON(const uint16_t* src,
+                        uint16_t* dst,
+                        float /*unused*/,
+                        int width) {
+  asm volatile(
+
+      "1:                                        \n"
+      "vld1.8      {q1}, [%0]!                   \n"  // load 8 shorts
+      "subs        %2, %2, #8                    \n"  // 8 pixels per loop
+      "vmovl.u16   q2, d2                        \n"  // 8 int's
+      "vmovl.u16   q3, d3                        \n"
+      "vcvt.f32.u32 q2, q2                       \n"  // 8 floats
+      "vcvt.f32.u32 q3, q3                       \n"
+      "vmul.f32    q2, q2, %y3                   \n"  // adjust exponent
+      "vmul.f32    q3, q3, %y3                   \n"
+      "vqshrn.u32  d2, q2, #13                   \n"  // isolate halffloat
+      "vqshrn.u32  d3, q3, #13                   \n"
+      "vst1.8      {q1}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src),              // %0
+        "+r"(dst),              // %1
+        "+r"(width)             // %2
+      : "w"(1.9259299444e-34f)  // %3
+      : "cc", "memory", "q1", "q2", "q3");
+}
+
+void HalfFloatRow_NEON(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width) {
+  asm volatile(
+
+      "1:                                        \n"
+      "vld1.8      {q1}, [%0]!                   \n"  // load 8 shorts
+      "subs        %2, %2, #8                    \n"  // 8 pixels per loop
+      "vmovl.u16   q2, d2                        \n"  // 8 int's
+      "vmovl.u16   q3, d3                        \n"
+      "vcvt.f32.u32 q2, q2                       \n"  // 8 floats
+      "vcvt.f32.u32 q3, q3                       \n"
+      "vmul.f32    q2, q2, %y3                   \n"  // adjust exponent
+      "vmul.f32    q3, q3, %y3                   \n"
+      "vqshrn.u32  d2, q2, #13                   \n"  // isolate halffloat
+      "vqshrn.u32  d3, q3, #13                   \n"
+      "vst1.8      {q1}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src),                      // %0
+        "+r"(dst),                      // %1
+        "+r"(width)                     // %2
+      : "w"(scale * 1.9259299444e-34f)  // %3
+      : "cc", "memory", "q1", "q2", "q3");
+}
+
+void ByteToFloatRow_NEON(const uint8_t* src,
+                         float* dst,
+                         float scale,
+                         int width) {
+  asm volatile(
+
+      "1:                                        \n"
+      "vld1.8      {d2}, [%0]!                   \n"  // load 8 bytes
+      "subs        %2, %2, #8                    \n"  // 8 pixels per loop
+      "vmovl.u8    q1, d2                        \n"  // 8 shorts
+      "vmovl.u16   q2, d2                        \n"  // 8 ints
+      "vmovl.u16   q3, d3                        \n"
+      "vcvt.f32.u32 q2, q2                       \n"  // 8 floats
+      "vcvt.f32.u32 q3, q3                       \n"
+      "vmul.f32    q2, q2, %y3                   \n"  // scale
+      "vmul.f32    q3, q3, %y3                   \n"
+      "vst1.8      {q2, q3}, [%1]!               \n"  // store 8 floats
+      "bgt         1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      : "w"(scale)   // %3
+      : "cc", "memory", "q1", "q2", "q3");
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_NEON(const uint16_t* src0,
+                   const uint16_t* src1,
+                   const uint16_t* src2,
+                   const uint16_t* src3,
+                   const uint16_t* src4,
+                   uint32_t* dst,
+                   int width) {
+  asm volatile(
+      "vmov.u16    d6, #4                        \n"  // constant 4
+      "vmov.u16    d7, #6                        \n"  // constant 6
+
+      "1:                                        \n"
+      "vld1.16     {q1}, [%0]!                   \n"  // load 8 samples, 5 rows
+      "vld1.16     {q2}, [%4]!                   \n"
+      "vaddl.u16   q0, d2, d4                    \n"  // * 1
+      "vaddl.u16   q1, d3, d5                    \n"  // * 1
+      "vld1.16     {q2}, [%1]!                   \n"
+      "vmlal.u16   q0, d4, d6                    \n"  // * 4
+      "vmlal.u16   q1, d5, d6                    \n"  // * 4
+      "vld1.16     {q2}, [%2]!                   \n"
+      "vmlal.u16   q0, d4, d7                    \n"  // * 6
+      "vmlal.u16   q1, d5, d7                    \n"  // * 6
+      "vld1.16     {q2}, [%3]!                   \n"
+      "vmlal.u16   q0, d4, d6                    \n"  // * 4
+      "vmlal.u16   q1, d5, d6                    \n"  // * 4
+      "subs        %6, %6, #8                    \n"  // 8 processed per loop
+      "vst1.32     {q0, q1}, [%5]!               \n"  // store 8 samples
+      "bgt         1b                            \n"
+      : "+r"(src0),  // %0
+        "+r"(src1),  // %1
+        "+r"(src2),  // %2
+        "+r"(src3),  // %3
+        "+r"(src4),  // %4
+        "+r"(dst),   // %5
+        "+r"(width)  // %6
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width) {
+  const uint32_t* src1 = src + 1;
+  const uint32_t* src2 = src + 2;
+  const uint32_t* src3 = src + 3;
+  asm volatile(
+      "vmov.u32    q10, #4                       \n"  // constant 4
+      "vmov.u32    q11, #6                       \n"  // constant 6
+
+      "1:                                        \n"
+      "vld1.32     {q0, q1}, [%0]!               \n"  // load 12 source samples
+      "vld1.32     {q2}, [%0]                    \n"
+      "vadd.u32    q0, q0, q1                    \n"  // * 1
+      "vadd.u32    q1, q1, q2                    \n"  // * 1
+      "vld1.32     {q2, q3}, [%2]!               \n"
+      "vmla.u32    q0, q2, q11                   \n"  // * 6
+      "vmla.u32    q1, q3, q11                   \n"  // * 6
+      "vld1.32     {q2, q3}, [%1]!               \n"
+      "vld1.32     {q8, q9}, [%3]!               \n"
+      "vadd.u32    q2, q2, q8                    \n"  // add rows for * 4
+      "vadd.u32    q3, q3, q9                    \n"
+      "vmla.u32    q0, q2, q10                   \n"  // * 4
+      "vmla.u32    q1, q3, q10                   \n"  // * 4
+      "subs        %5, %5, #8                    \n"  // 8 processed per loop
+      "vqshrn.u32  d0, q0, #8                    \n"  // round and pack
+      "vqshrn.u32  d1, q1, #8                    \n"
+      "vst1.u16    {q0}, [%4]!                   \n"  // store 8 samples
+      "bgt         1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(src1),  // %1
+        "+r"(src2),  // %2
+        "+r"(src3),  // %3
+        "+r"(dst),   // %4
+        "+r"(width)  // %5
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+}
+
+// Convert biplanar NV21 to packed YUV24
+void NV21ToYUV24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q2}, [%0]!                   \n"  // load 16 Y values
+      "vld2.8      {d0, d2}, [%1]!               \n"  // load 8 VU values
+      "vmov        d1, d0                        \n"
+      "vzip.u8     d0, d1                        \n"  // VV
+      "vmov        d3, d2                        \n"
+      "vzip.u8     d2, d3                        \n"  // UU
+      "subs        %3, %3, #16                   \n"  // 16 pixels per loop
+      "vst3.8      {d0, d2, d4}, [%2]!           \n"  // store 16 YUV pixels
+      "vst3.8      {d1, d3, d5}, [%2]!           \n"
+      "bgt         1b                            \n"
+      : "+r"(src_y),      // %0
+        "+r"(src_vu),     // %1
+        "+r"(dst_yuv24),  // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
+                      int src_stride_ayuv,
+                      uint8_t* dst_uv,
+                      int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // src_stride + src_AYUV
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 AYUV pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 AYUV
+                                                      // pixels.
+      "vpaddl.u8   q0, q0                        \n"  // V 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // U 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more AYUV
+                                                      // pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 AYUV
+                                                      // pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vqrshrun.s16 d1, q0, #2                   \n"  // 2x2 average
+      "vqrshrun.s16 d0, q1, #2                   \n"
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop.
+      "vst2.8      {d0, d1}, [%2]!               \n"  // store 8 pixels UV.
+      "bgt         1b                            \n"
+      : "+r"(src_ayuv),         // %0
+        "+r"(src_stride_ayuv),  // %1
+        "+r"(dst_uv),           // %2
+        "+r"(width)             // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7");
+}
+
+void AYUVToVURow_NEON(const uint8_t* src_ayuv,
+                      int src_stride_ayuv,
+                      uint8_t* dst_vu,
+                      int width) {
+  asm volatile(
+      "add         %1, %0, %1                    \n"  // src_stride + src_AYUV
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 AYUV pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 AYUV
+                                                      // pixels.
+      "vpaddl.u8   q0, q0                        \n"  // V 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // U 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more AYUV
+                                                      // pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 AYUV
+                                                      // pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vqrshrun.s16 d0, q0, #2                   \n"  // 2x2 average
+      "vqrshrun.s16 d1, q1, #2                   \n"
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop.
+      "vst2.8      {d0, d1}, [%2]!               \n"  // store 8 pixels VU.
+      "bgt         1b                            \n"
+      : "+r"(src_ayuv),         // %0
+        "+r"(src_stride_ayuv),  // %1
+        "+r"(dst_vu),           // %2
+        "+r"(width)             // %3
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7");
+}
+
+// Copy row of AYUV Y's into Y.
+// Similar to ARGBExtractAlphaRow_NEON
+void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 AYUV pixels
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 AYUV pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q2}, [%1]!                   \n"  // store 16 Y's.
+      "bgt         1b                            \n"
+      : "+r"(src_ayuv),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {d0, d2}, [%0]!               \n"  // load 16 UV values
+      "vld2.8      {d1, d3}, [%0]!               \n"
+      "vorr.u8     q2, q0, q0                    \n"  // move U after V
+      "subs        %2, %2, #16                   \n"  // 16 pixels per loop
+      "vst2.8      {q1, q2}, [%1]!               \n"  // store 16 VU pixels
+      "bgt         1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_vu),  // %1
+        "+r"(width)    // %2
+      :
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void HalfMergeUVRow_NEON(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width) {
+  const uint8_t* src_u_1 = src_u + src_stride_u;
+  const uint8_t* src_v_1 = src_v + src_stride_v;
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 16 U values
+      "vld1.8      {q1}, [%2]!                   \n"  // load 16 V values
+      "vld1.8      {q2}, [%1]!                   \n"
+      "vld1.8      {q3}, [%3]!                   \n"
+      "vpaddl.u8   q0, q0                        \n"  // half size
+      "vpaddl.u8   q1, q1                        \n"
+      "vpadal.u8   q0, q2                        \n"
+      "vpadal.u8   q1, q3                        \n"
+      "vqrshrn.u16 d0, q0, #2                    \n"
+      "vqrshrn.u16 d1, q1, #2                    \n"
+      "subs        %5, %5, #16                   \n"  // 16 src pixels per loop
+      "vst2.8      {d0, d1}, [%4]!               \n"  // store 8 UV pixels
+      "bgt         1b                            \n"
+      : "+r"(src_u),    // %0
+        "+r"(src_u_1),  // %1
+        "+r"(src_v),    // %2
+        "+r"(src_v_1),  // %3
+        "+r"(dst_uv),   // %4
+        "+r"(width)     // %5
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+void SplitUVRow_16_NEON(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width) {
+  int shift = depth - 16;  // Negative for right shift.
+  asm volatile(
+      "vdup.16     q2, %4                        \n"
+      "1:                                        \n"
+      "vld2.16     {q0, q1}, [%0]!               \n"  // load 8 UV
+      "vshl.u16    q0, q0, q2                    \n"
+      "vshl.u16    q1, q1, q2                    \n"
+      "subs        %3, %3, #8                    \n"  // 8 src pixels per loop
+      "vst1.16     {q0}, [%1]!                   \n"  // store 8 U pixels
+      "vst1.16     {q1}, [%2]!                   \n"  // store 8 V pixels
+      "bgt         1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  int shift = 16 - depth;
+  asm volatile(
+      "vdup.16     q2, %4                        \n"
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // load 8 U
+      "vld1.16     {q1}, [%1]!                   \n"  // load 8 V
+      "vshl.u16    q0, q0, q2                    \n"
+      "vshl.u16    q1, q1, q2                    \n"
+      "subs        %3, %3, #8                    \n"  // 8 src pixels per loop
+      "vst2.16     {q0, q1}, [%2]!               \n"  // store 8 UV pixels
+      "bgt         1b                            \n"
+      : "+r"(src_u),   // %0
+        "+r"(src_v),   // %1
+        "+r"(dst_uv),  // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void MultiplyRow_16_NEON(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width) {
+  asm volatile(
+      "vdup.16     q2, %3                        \n"
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vmul.u16    q0, q0, q2                    \n"
+      "vmul.u16    q1, q1, q2                    \n"
+      "vst1.16     {q0}, [%1]!                   \n"
+      "vst1.16     {q1}, [%1]!                   \n"
+      "subs        %2, %2, #16                   \n"  // 16 src pixels per loop
+      "bgt         1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void DivideRow_16_NEON(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  asm volatile(
+      "vdup.16     d8, %3                        \n"
+      "1:                                        \n"
+      "vld1.16     {q2, q3}, [%0]!               \n"
+      "vmull.u16   q0, d4, d8                    \n"
+      "vmull.u16   q1, d5, d8                    \n"
+      "vmull.u16   q2, d6, d8                    \n"
+      "vmull.u16   q3, d7, d8                    \n"
+      "vshrn.u32   d0, q0, #16                   \n"
+      "vshrn.u32   d1, q1, #16                   \n"
+      "vshrn.u32   d2, q2, #16                   \n"
+      "vshrn.u32   d3, q3, #16                   \n"
+      "vst1.16     {q0, q1}, [%1]!               \n"  // store 16 pixels
+      "subs        %2, %2, #16                   \n"  // 16 src pixels per loop
+      "bgt         1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "d8");
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits = shr 1
+// 16384 = 10 bits = shr 2
+// 4096 = 12 bits = shr 4
+// 256 = 16 bits = shr 8
+void Convert16To8Row_NEON(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width) {
+  int shift = 15 - __builtin_clz((int32_t)scale);  // Negative shl is shr
+  asm volatile(
+      "vdup.16     q2, %3                        \n"
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vshl.u16    q0, q0, q2                    \n"  // shr = q2 is negative
+      "vshl.u16    q1, q1, q2                    \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vqmovn.u16  d1, q1                        \n"
+      "subs        %2, %2, #16                   \n"  // 16 src pixels per loop
+      "vst1.8      {q0}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(shift)    // %3
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)..
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/media/libyuv/source/row_neon64.cc b/media/libyuv/source/row_neon64.cc
index 6375d4f55f..7f04b60687 100644
--- a/media/libyuv/source/row_neon64.cc
+++ b/media/libyuv/source/row_neon64.cc
@@ -15,1482 +15,2340 @@ namespace libyuv {
 extern "C" {
 #endif
 
+// Enable LIBYUV_USE_ST2, LIBYUV_USE_ST3, LIBYUV_USE_ST4 for CPUs that prefer
+// STn over ZIP1+ST1
+// Exynos M1, M2, M3 are slow with ST2, ST3 and ST4 instructions.
+
 // This module is for GCC Neon armv8 64 bit.
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
+// v0.8h: Y
+// v1.16b: 8U, 8V
+
 // Read 8 Y, 4 U and 4 V from 422
-#define READYUV422                                                             \
-    MEMACCESS(0)                                                               \
-    "ld1        {v0.8b}, [%0], #8              \n"                             \
-    MEMACCESS(1)                                                               \
-    "ld1        {v1.s}[0], [%1], #4            \n"                             \
-    MEMACCESS(2)                                                               \
-    "ld1        {v1.s}[1], [%2], #4            \n"
-
-// Read 8 Y, 2 U and 2 V from 422
-#define READYUV411                                                             \
-    MEMACCESS(0)                                                               \
-    "ld1        {v0.8b}, [%0], #8              \n"                             \
-    MEMACCESS(1)                                                               \
-    "ld1        {v2.h}[0], [%1], #2            \n"                             \
-    MEMACCESS(2)                                                               \
-    "ld1        {v2.h}[1], [%2], #2            \n"                             \
-    "zip1       v1.8b, v2.8b, v2.8b            \n"
+#define READYUV422                               \
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "ld1        {v1.s}[0], [%[src_u]], #4      \n" \
+  "ld1        {v1.s}[1], [%[src_v]], #4      \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "zip1       v1.16b, v1.16b, v1.16b         \n" \
+  "prfm       pldl1keep, [%[src_u], 128]     \n" \
+  "prfm       pldl1keep, [%[src_v], 128]     \n"
 
 // Read 8 Y, 8 U and 8 V from 444
-#define READYUV444                                                             \
-    MEMACCESS(0)                                                               \
-    "ld1        {v0.8b}, [%0], #8              \n"                             \
-    MEMACCESS(1)                                                               \
-    "ld1        {v1.d}[0], [%1], #8            \n"                             \
-    MEMACCESS(2)                                                               \
-    "ld1        {v1.d}[1], [%2], #8            \n"                             \
-    "uaddlp     v1.8h, v1.16b                  \n"                             \
-    "rshrn      v1.8b, v1.8h, #1               \n"
+#define READYUV444                               \
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "ld1        {v1.d}[0], [%[src_u]], #8      \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "ld1        {v1.d}[1], [%[src_v]], #8      \n" \
+  "prfm       pldl1keep, [%[src_u], 448]     \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n" \
+  "prfm       pldl1keep, [%[src_v], 448]     \n"
 
 // Read 8 Y, and set 4 U and 4 V to 128
-#define READYUV400                                                             \
-    MEMACCESS(0)                                                               \
-    "ld1        {v0.8b}, [%0], #8              \n"                             \
-    "movi       v1.8b , #128                   \n"
-
-// Read 8 Y and 4 UV from NV12
-#define READNV12                                                               \
-    MEMACCESS(0)                                                               \
-    "ld1        {v0.8b}, [%0], #8              \n"                             \
-    MEMACCESS(1)                                                               \
-    "ld1        {v2.8b}, [%1], #8              \n"                             \
-    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
-    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
-    "ins        v1.s[1], v3.s[0]               \n"
-
-// Read 8 Y and 4 VU from NV21
-#define READNV21                                                               \
-    MEMACCESS(0)                                                               \
-    "ld1        {v0.8b}, [%0], #8              \n"                             \
-    MEMACCESS(1)                                                               \
-    "ld1        {v2.8b}, [%1], #8              \n"                             \
-    "uzp1       v3.8b, v2.8b, v2.8b            \n"                             \
-    "uzp2       v1.8b, v2.8b, v2.8b            \n"                             \
-    "ins        v1.s[1], v3.s[0]               \n"
+#define READYUV400                               \
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "movi       v1.16b, #128                   \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n"
+
+static const uvec8 kNV12Table = {0, 0, 2, 2, 4, 4, 6, 6,
+                                 1, 1, 3, 3, 5, 5, 7, 7};
+static const uvec8 kNV21Table = {1, 1, 3, 3, 5, 5, 7, 7,
+                                 0, 0, 2, 2, 4, 4, 6, 6};
+
+// Read 8 Y and 4 UV from NV12 or NV21
+#define READNV12                                 \
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "ldr        d1, [%[src_uv]], #8            \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "tbl        v1.16b, {v1.16b}, v2.16b       \n" \
+  "prfm       pldl1keep, [%[src_uv], 448]    \n"
 
 // Read 8 YUY2
-#define READYUY2                                                               \
-    MEMACCESS(0)                                                               \
-    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"                             \
-    "uzp2       v3.8b, v1.8b, v1.8b            \n"                             \
-    "uzp1       v1.8b, v1.8b, v1.8b            \n"                             \
-    "ins        v1.s[1], v3.s[0]               \n"
+#define READYUY2                                     \
+  "ld2        {v0.8b, v1.8b}, [%[src_yuy2]], #16 \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n"     \
+  "prfm       pldl1keep, [%[src_yuy2], 448]  \n"     \
+  "tbl        v1.16b, {v1.16b}, v2.16b       \n"
 
 // Read 8 UYVY
-#define READUYVY                                                               \
-    MEMACCESS(0)                                                               \
-    "ld2        {v2.8b, v3.8b}, [%0], #16      \n"                             \
-    "orr        v0.8b, v3.8b, v3.8b            \n"                             \
-    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
-    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
-    "ins        v1.s[1], v3.s[0]               \n"
-
-#define YUVTORGB_SETUP                                                         \
-    "ld1r       {v24.8h}, [%[kUVBiasBGR]], #2  \n"                             \
-    "ld1r       {v25.8h}, [%[kUVBiasBGR]], #2  \n"                             \
-    "ld1r       {v26.8h}, [%[kUVBiasBGR]]      \n"                             \
-    "ld1r       {v31.4s}, [%[kYToRgb]]         \n"                             \
-    "ld2        {v27.8h, v28.8h}, [%[kUVToRB]] \n"                             \
-    "ld2        {v29.8h, v30.8h}, [%[kUVToG]]  \n"
-
-#define YUVTORGB(vR, vG, vB)                                                   \
-    "uxtl       v0.8h, v0.8b                   \n" /* Extract Y    */          \
-    "shll       v2.8h, v1.8b, #8               \n" /* Replicate UV */          \
-    "ushll2     v3.4s, v0.8h, #0               \n" /* Y */                     \
-    "ushll      v0.4s, v0.4h, #0               \n"                             \
-    "mul        v3.4s, v3.4s, v31.4s           \n"                             \
-    "mul        v0.4s, v0.4s, v31.4s           \n"                             \
-    "sqshrun    v0.4h, v0.4s, #16              \n"                             \
-    "sqshrun2   v0.8h, v3.4s, #16              \n" /* Y */                     \
-    "uaddw      v1.8h, v2.8h, v1.8b            \n" /* Replicate UV */          \
-    "mov        v2.d[0], v1.d[1]               \n" /* Extract V */             \
-    "uxtl       v2.8h, v2.8b                   \n"                             \
-    "uxtl       v1.8h, v1.8b                   \n" /* Extract U */             \
-    "mul        v3.8h, v1.8h, v27.8h           \n"                             \
-    "mul        v5.8h, v1.8h, v29.8h           \n"                             \
-    "mul        v6.8h, v2.8h, v30.8h           \n"                             \
-    "mul        v7.8h, v2.8h, v28.8h           \n"                             \
-    "sqadd      v6.8h, v6.8h, v5.8h            \n"                             \
-    "sqadd      " #vB ".8h, v24.8h, v0.8h      \n" /* B */                     \
-    "sqadd      " #vG ".8h, v25.8h, v0.8h      \n" /* G */                     \
-    "sqadd      " #vR ".8h, v26.8h, v0.8h      \n" /* R */                     \
-    "sqadd      " #vB ".8h, " #vB ".8h, v3.8h  \n" /* B */                     \
-    "sqsub      " #vG ".8h, " #vG ".8h, v6.8h  \n" /* G */                     \
-    "sqadd      " #vR ".8h, " #vR ".8h, v7.8h  \n" /* R */                     \
-    "sqshrun    " #vB ".8b, " #vB ".8h, #6     \n" /* B */                     \
-    "sqshrun    " #vG ".8b, " #vG ".8h, #6     \n" /* G */                     \
-    "sqshrun    " #vR ".8b, " #vR ".8h, #6     \n" /* R */                     \
-
-void I444ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+#define READUYVY                                     \
+  "ld2        {v3.8b, v4.8b}, [%[src_uyvy]], #16 \n" \
+  "zip1       v0.16b, v4.16b, v4.16b         \n"     \
+  "prfm       pldl1keep, [%[src_uyvy], 448]  \n"     \
+  "tbl        v1.16b, {v3.16b}, v2.16b       \n"
+
+// UB VR UG VG
+// YG BB BG BR
+#define YUVTORGB_SETUP                                                \
+  "ld4r       {v28.16b, v29.16b, v30.16b, v31.16b}, [%[kUVCoeff]] \n" \
+  "ld4r       {v24.8h, v25.8h, v26.8h, v27.8h}, [%[kRGBCoeffBias]] \n"
+
+// v16.8h: B
+// v17.8h: G
+// v18.8h: R
+
+// Convert from YUV to 2.14 fixed point RGB
+#define YUVTORGB                                          \
+  "umull2     v3.4s, v0.8h, v24.8h           \n"          \
+  "umull      v6.8h, v1.8b, v30.8b           \n"          \
+  "umull      v0.4s, v0.4h, v24.4h           \n"          \
+  "umlal2     v6.8h, v1.16b, v31.16b         \n" /* DG */ \
+  "uqshrn     v0.4h, v0.4s, #16              \n"          \
+  "uqshrn2    v0.8h, v3.4s, #16              \n" /* Y */  \
+  "umull      v4.8h, v1.8b, v28.8b           \n" /* DB */ \
+  "umull2     v5.8h, v1.16b, v29.16b         \n" /* DR */ \
+  "add        v17.8h, v0.8h, v26.8h          \n" /* G */  \
+  "add        v16.8h, v0.8h, v4.8h           \n" /* B */  \
+  "add        v18.8h, v0.8h, v5.8h           \n" /* R */  \
+  "uqsub      v17.8h, v17.8h, v6.8h          \n" /* G */  \
+  "uqsub      v16.8h, v16.8h, v25.8h         \n" /* B */  \
+  "uqsub      v18.8h, v18.8h, v27.8h         \n" /* R */
+
+// Convert from 2.14 fixed point RGB To 8 bit RGB
+#define RGBTORGB8                                \
+  "uqshrn     v17.8b, v17.8h, #6             \n" \
+  "uqshrn     v16.8b, v16.8h, #6             \n" \
+  "uqshrn     v18.8b, v18.8h, #6             \n"
+
+#define YUVTORGB_REGS                                                          \
+  "v0", "v1", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v24", "v25", \
+      "v26", "v27", "v28", "v29", "v30", "v31"
+
+void I444ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV444
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    MEMACCESS(3)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n" /* A */
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
+}
+
+void I444ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
-void I422ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
+void I422ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    MEMACCESS(3)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n" /* A */
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
-void I422AlphaToARGBRow_NEON(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             const uint8* src_a,
-                             uint8* dst_argb,
+void I444AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    MEMACCESS(3)
-    "ld1        {v23.8b}, [%3], #8             \n"
-    "subs       %w5, %w5, #8                   \n"
-    MEMACCESS(4)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%4], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(src_a),     // %3
-      "+r"(dst_argb),  // %4
-      "+r"(width)      // %5
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n"
+      "ld1         {v19.8b}, [%[src_a]], #8      \n" READYUV444
+      "prfm        pldl1keep, [%[src_a], 448]    \n" YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
-void I411ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV411
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    MEMACCESS(3)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n"
+      "ld1         {v19.8b}, [%[src_a]], #8      \n" READYUV422
+      "prfm        pldl1keep, [%[src_a], 448]    \n" YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
-void I422ToRGBARow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_rgba,
+void I422ToRGBARow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_rgba,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v20.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v23, v22, v21)
-    "subs       %w4, %w4, #8                   \n"
-    MEMACCESS(3)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_rgba),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v15.8b, #255                  \n" /* A */
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v15.8b,v16.8b,v17.8b,v18.8b}, [%[dst_rgba]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgba] "+r"(dst_rgba),                         // %[dst_rgba]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v15");
 }
 
-void I422ToRGB24Row_NEON(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* dst_rgb24,
+void I422ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    MEMACCESS(3)
-    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_rgb24), // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
-#define ARGBTORGB565                                                           \
-    "shll       v0.8h,  v22.8b, #8             \n"  /* R                    */ \
-    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
-    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
-    "sri        v0.8h,  v21.8h, #5             \n"  /* RG                   */ \
-    "sri        v0.8h,  v20.8h, #11            \n"  /* RGB                  */
-
-void I422ToRGB565Row_NEON(const uint8* src_y,
-                          const uint8* src_u,
-                          const uint8* src_v,
-                          uint8* dst_rgb565,
+#define ARGBTORGB565                                                        \
+  "shll       v18.8h, v18.8b, #8             \n" /* R                    */ \
+  "shll       v17.8h, v17.8b, #8             \n" /* G                    */ \
+  "shll       v16.8h, v16.8b, #8             \n" /* B                    */ \
+  "sri        v18.8h, v17.8h, #5             \n" /* RG                   */ \
+  "sri        v18.8h, v16.8h, #11            \n" /* RGB                  */
+
+void I422ToRGB565Row_NEON(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    ARGBTORGB565
-    MEMACCESS(3)
-    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
-    "b.gt       1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_rgb565),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV422 YUVTORGB
+      RGBTORGB8 "subs        %w[width], %w[width], #8      \n" ARGBTORGB565
+      "st1         {v18.8h}, [%[dst_rgb565]], #16 \n"  // store 8 pixels RGB565.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb565] "+r"(dst_rgb565),                     // %[dst_rgb565]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
-#define ARGBTOARGB1555                                                         \
-    "shll       v0.8h,  v23.8b, #8             \n"  /* A                    */ \
-    "shll       v22.8h, v22.8b, #8             \n"  /* R                    */ \
-    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
-    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
-    "sri        v0.8h,  v22.8h, #1             \n"  /* AR                   */ \
-    "sri        v0.8h,  v21.8h, #6             \n"  /* ARG                  */ \
-    "sri        v0.8h,  v20.8h, #11            \n"  /* ARGB                 */
-
-void I422ToARGB1555Row_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb1555,
+#define ARGBTOARGB1555                                                      \
+  "shll       v0.8h,  v19.8b, #8             \n" /* A                    */ \
+  "shll       v18.8h, v18.8b, #8             \n" /* R                    */ \
+  "shll       v17.8h, v17.8b, #8             \n" /* G                    */ \
+  "shll       v16.8h, v16.8b, #8             \n" /* B                    */ \
+  "sri        v0.8h,  v18.8h, #1             \n" /* AR                   */ \
+  "sri        v0.8h,  v17.8h, #6             \n" /* ARG                  */ \
+  "sri        v0.8h,  v16.8h, #11            \n" /* ARGB                 */
+
+void I422ToARGB1555Row_NEON(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
                             const struct YuvConstants* yuvconstants,
                             int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    ARGBTOARGB1555
-    MEMACCESS(3)
-    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
-    "b.gt       1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_argb1555),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n" ARGBTOARGB1555
+      "st1         {v0.8h}, [%[dst_argb1555]], #16 \n"  // store 8 pixels
+                                                        // RGB565.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb1555] "+r"(dst_argb1555),                 // %[dst_argb1555]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
-#define ARGBTOARGB4444                                                         \
-    /* Input v20.8b<=B, v21.8b<=G, v22.8b<=R, v23.8b<=A, v4.8b<=0x0f        */ \
-    "ushr       v20.8b, v20.8b, #4             \n"  /* B                    */ \
-    "bic        v21.8b, v21.8b, v4.8b          \n"  /* G                    */ \
-    "ushr       v22.8b, v22.8b, #4             \n"  /* R                    */ \
-    "bic        v23.8b, v23.8b, v4.8b          \n"  /* A                    */ \
-    "orr        v0.8b,  v20.8b, v21.8b         \n"  /* BG                   */ \
-    "orr        v1.8b,  v22.8b, v23.8b         \n"  /* RA                   */ \
-    "zip1       v0.16b, v0.16b, v1.16b         \n"  /* BGRA                 */
-
-void I422ToARGB4444Row_NEON(const uint8* src_y,
-                            const uint8* src_u,
-                            const uint8* src_v,
-                            uint8* dst_argb4444,
+#define ARGBTOARGB4444                                                       \
+  /* Input v16.8b<=B, v17.8b<=G, v18.8b<=R, v19.8b<=A, v23.8b<=0x0f       */ \
+  "ushr       v16.8b, v16.8b, #4             \n" /* B                    */  \
+  "bic        v17.8b, v17.8b, v23.8b         \n" /* G                    */  \
+  "ushr       v18.8b, v18.8b, #4             \n" /* R                    */  \
+  "bic        v19.8b, v19.8b, v23.8b         \n" /* A                    */  \
+  "orr        v0.8b,  v16.8b, v17.8b         \n" /* BG                   */  \
+  "orr        v1.8b,  v18.8b, v19.8b         \n" /* RA                   */  \
+  "zip1       v0.16b, v0.16b, v1.16b         \n" /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
                             const struct YuvConstants* yuvconstants,
                             int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    "movi       v23.8b, #255                   \n"
-    ARGBTOARGB4444
-    MEMACCESS(3)
-    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels ARGB4444.
-    "b.gt       1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_argb4444),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v23.16b, #0x0f                \n"  // bits to clear with
+                                                      // vbic.
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "movi        v19.8b, #255                  \n" ARGBTOARGB4444
+      "st1         {v0.8h}, [%[dst_argb4444]], #16 \n"  // store 8
+                                                        // pixels
+                                                        // ARGB4444.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb4444] "+r"(dst_argb4444),                 // %[dst_argb4444]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19", "v23");
 }
 
-void I400ToARGBRow_NEON(const uint8* src_y,
-                        uint8* dst_argb,
+void I400ToARGBRow_NEON(const uint8_t* src_y,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READYUV400
-    YUVTORGB(v22, v21, v20)
-    "subs       %w2, %w2, #8                   \n"
-    MEMACCESS(1)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&kYuvI601Constants.kUVToRB),
-      [kUVToG]"r"(&kYuvI601Constants.kUVToG),
-      [kUVBiasBGR]"r"(&kYuvI601Constants.kUVBiasBGR),
-      [kYToRgb]"r"(&kYuvI601Constants.kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "1:                                        \n" READYUV400 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
-void J400ToARGBRow_NEON(const uint8* src_y,
-                        uint8* dst_argb,
-                        int width) {
-  asm volatile (
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v20.8b}, [%0], #8             \n"
-    "orr        v21.8b, v20.8b, v20.8b         \n"
-    "orr        v22.8b, v20.8b, v20.8b         \n"
-    "subs       %w2, %w2, #8                   \n"
-    MEMACCESS(1)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    :
-    : "cc", "memory", "v20", "v21", "v22", "v23"
-  );
+#if LIBYUV_USE_ST4
+void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movi        v23.8b, #255                  \n"
+      "1:                                        \n"
+      "ld1         {v20.8b}, [%0], #8            \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v21.8b, v20.8b, v20.8b        \n"
+      "orr         v22.8b, v20.8b, v20.8b        \n"
+      "subs        %w2, %w2, #8                  \n"
+      "st4         {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v20", "v21", "v22", "v23");
+}
+#else
+void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movi        v20.8b, #255                  \n"
+      "1:                                        \n"
+      "ldr         d16, [%0], #8                 \n"
+      "subs        %w2, %w2, #8                  \n"
+      "zip1        v18.16b, v16.16b, v16.16b     \n"  // YY
+      "zip1        v19.16b, v16.16b, v20.16b     \n"  // YA
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v16.16b, v18.16b, v19.16b     \n"  // YYYA
+      "zip2        v17.16b, v18.16b, v19.16b     \n"
+      "stp         q16, q17, [%1], #32           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20");
 }
+#endif  // LIBYUV_USE_ST4
 
-void NV12ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_uv,
-                        uint8* dst_argb,
+void NV12ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READNV12
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    MEMACCESS(2)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_uv),    // %1
-      "+r"(dst_argb),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_uv),                              // %[src_uv]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
-void NV21ToARGBRow_NEON(const uint8* src_y,
-                        const uint8* src_vu,
-                        uint8* dst_argb,
+void NV21ToARGBRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READNV21
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    MEMACCESS(2)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_vu),    // %1
-      "+r"(dst_argb),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_vu),                              // %[src_uv]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV21Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
-void NV12ToRGB565Row_NEON(const uint8* src_y,
-                          const uint8* src_uv,
-                          uint8* dst_rgb565,
+void NV12ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_uv),                              // %[src_uv]
+        [dst_rgb24] "+r"(dst_rgb24),                        // %[dst_rgb24]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2");
+}
+
+void NV21ToRGB24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_rgb24,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_vu),                              // %[src_uv]
+        [dst_rgb24] "+r"(dst_rgb24),                        // %[dst_rgb24]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV21Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2");
+}
+
+void NV12ToRGB565Row_NEON(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
                           const struct YuvConstants* yuvconstants,
                           int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READNV12
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    ARGBTORGB565
-    MEMACCESS(2)
-    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_uv),    // %1
-      "+r"(dst_rgb565),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n" ARGBTORGB565
+      "st1         {v18.8h}, [%[dst_rgb565]], #16 \n"  // store 8
+                                                       // pixels
+                                                       // RGB565.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_uv),                              // %[src_uv]
+        [dst_rgb565] "+r"(dst_rgb565),                      // %[dst_rgb565]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2");
 }
 
-void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
-                        uint8* dst_argb,
+void YUY2ToARGBRow_NEON(const uint8_t* src_yuy2,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READYUY2
-    YUVTORGB(v22, v21, v20)
-    "subs       %w2, %w2, #8                   \n"
-    MEMACCESS(1)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32      \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_yuy2),  // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READYUY2 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_yuy2] "+r"(src_yuy2),                          // %[src_yuy2]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
-void UYVYToARGBRow_NEON(const uint8* src_uyvy,
-                        uint8* dst_argb,
+void UYVYToARGBRow_NEON(const uint8_t* src_uyvy,
+                        uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READUYVY
-    YUVTORGB(v22, v21, v20)
-    "subs       %w2, %w2, #8                   \n"
-    MEMACCESS(1)
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], 32      \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_uyvy),  // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READUYVY YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_uyvy] "+r"(src_uyvy),                          // %[src_yuy2]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
 // Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
-void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+void SplitUVRow_NEON(const uint8_t* src_uv,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
                      int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pairs of UV
-    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"  // store U
-    MEMACCESS(2)
-    "st1        {v1.16b}, [%2], #16            \n"  // store V
-    "b.gt       1b                             \n"
-    : "+r"(src_uv),  // %0
-      "+r"(dst_u),   // %1
-      "+r"(dst_v),   // %2
-      "+r"(width)    // %3  // Output registers
-    :                       // Input registers
-    : "cc", "memory", "v0", "v1"  // Clobber List
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pairs of UV
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%1], #16           \n"  // store U
+      "st1         {v1.16b}, [%2], #16           \n"  // store V
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),               // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(width)                 // %3  // Output registers
+      :                             // Input registers
+      : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+
+// Reads 16 byte Y's from tile and writes out 16 Y's.
+// MM21 Y tiles are 16x32 so src_tile_stride = 512 bytes
+// MM21 UV tiles are 8x16 so src_tile_stride = 256 bytes
+// width measured in bytes so 8 UV = 16.
+void DetileRow_NEON(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], %3            \n"  // load 16 bytes
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 1792]         \n"  // 7 tiles of 256b ahead
+      "st1         {v0.16b}, [%1], #16           \n"  // store 16 bytes
+      "b.gt        1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "v0"  // Clobber List
+  );
+}
+
+// Reads 16 byte Y's of 16 bits from tile and writes out 16 Y's.
+void DetileRow_16_NEON(const uint16_t* src,
+                       ptrdiff_t src_tile_stride,
+                       uint16_t* dst,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.8h,v1.8h}, [%0], %3       \n"  // load 16 pixels
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 3584]         \n"  // 7 tiles of 512b ahead
+      "st1         {v0.8h,v1.8h}, [%1], #32      \n"  // store 16 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst),                  // %1
+        "+r"(width)                 // %2
+      : "r"(src_tile_stride * 2)    // %3
+      : "cc", "memory", "v0", "v1"  // Clobber List
   );
 }
 
+// Read 16 bytes of UV, detile, and write 8 bytes of U and 8 bytes of V.
+void DetileSplitUVRow_NEON(const uint8_t* src_uv,
+                           ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.8b,v1.8b}, [%0], %4       \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%0, 1792]         \n"
+      "st1         {v0.8b}, [%1], #8             \n"
+      "st1         {v1.8b}, [%2], #8             \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),               // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(width)                 // %3
+      : "r"(src_tile_stride)        // %4
+      : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+
+#if LIBYUV_USE_ST2
+// Read 16 Y, 8 UV, and write 8 YUY2
+void DetileToYUY2_NEON(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], %4            \n"  // load 16 Ys
+      "prfm        pldl1keep, [%0, 1792]         \n"
+      "ld1         {v1.16b}, [%1], %5            \n"  // load 8 UVs
+      "prfm        pldl1keep, [%1, 1792]         \n"
+      "subs        %w3, %w3, #16                 \n"  // store 8 YUY2
+      "st2         {v0.16b,v1.16b}, [%2], #32    \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_y),                // %0
+        "+r"(src_uv),               // %1
+        "+r"(dst_yuy2),             // %2
+        "+r"(width)                 // %3
+      : "r"(src_y_tile_stride),     // %4
+        "r"(src_uv_tile_stride)     // %5
+      : "cc", "memory", "v0", "v1"  // Clobber list
+  );
+}
+#else
+// Read 16 Y, 8 UV, and write 8 YUY2
+void DetileToYUY2_NEON(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], %4            \n"  // load 16 Ys
+      "ld1         {v1.16b}, [%1], %5            \n"  // load 8 UVs
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%0, 1792]         \n"
+      "zip1        v2.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 1792]         \n"
+      "zip2        v3.16b, v0.16b, v1.16b        \n"
+      "st1         {v2.16b,v3.16b}, [%2], #32    \n"  // store 8 YUY2
+      "b.gt        1b                            \n"
+      : "+r"(src_y),                            // %0
+        "+r"(src_uv),                           // %1
+        "+r"(dst_yuy2),                         // %2
+        "+r"(width)                             // %3
+      : "r"(src_y_tile_stride),                 // %4
+        "r"(src_uv_tile_stride)                 // %5
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber list
+  );
+}
+#endif
+
+// Unpack MT2T into tiled P010 64 pixels at a time. See
+// tinyurl.com/mtk-10bit-video-format for format documentation.
+void UnpackMT2T_NEON(const uint8_t* src, uint16_t* dst, size_t size) {
+  const uint8_t* src_lower_bits = src;
+  const uint8_t* src_upper_bits = src + 16;
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b, v1.8b, v2.8b, v3.8b}, [%1], #32 \n"
+      "ld1         {v7.8b}, [%0], #8             \n"
+      "shl         v6.8b, v7.8b, #2              \n"
+      "shl         v5.8b, v7.8b, #4              \n"
+      "shl         v4.8b, v7.8b, #6              \n"
+      "zip1        v0.16b, v4.16b, v0.16b        \n"
+      "zip1        v1.16b, v5.16b, v1.16b        \n"
+      "zip1        v2.16b, v6.16b, v2.16b        \n"
+      "zip1        v3.16b, v7.16b, v3.16b        \n"
+      "sri         v0.8h, v0.8h, #10             \n"
+      "sri         v1.8h, v1.8h, #10             \n"
+      "sri         v2.8h, v2.8h, #10             \n"
+      "sri         v3.8h, v3.8h, #10             \n"
+      "st4         {v0.8h, v1.8h, v2.8h, v3.8h}, [%2], #64 \n"
+      "ld4         {v0.8b, v1.8b, v2.8b, v3.8b}, [%1], #32 \n"
+      "ld1         {v7.8b}, [%0], #8             \n"
+      "shl         v6.8b, v7.8b, #2              \n"
+      "shl         v5.8b, v7.8b, #4              \n"
+      "shl         v4.8b, v7.8b, #6              \n"
+      "zip1        v0.16b, v4.16b, v0.16b        \n"
+      "zip1        v1.16b, v5.16b, v1.16b        \n"
+      "zip1        v2.16b, v6.16b, v2.16b        \n"
+      "zip1        v3.16b, v7.16b, v3.16b        \n"
+      "sri         v0.8h, v0.8h, #10             \n"
+      "sri         v1.8h, v1.8h, #10             \n"
+      "sri         v2.8h, v2.8h, #10             \n"
+      "sri         v3.8h, v3.8h, #10             \n"
+      "st4         {v0.8h, v1.8h, v2.8h, v3.8h}, [%2], #64 \n"
+      "mov         %0, %1                        \n"
+      "add         %1, %0, #16                   \n"
+      "subs        %3, %3, #80                   \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_lower_bits),  // %0
+        "+r"(src_upper_bits),  // %1
+        "+r"(dst),             // %2
+        "+r"(size)             // %3
+      :
+      : "cc", "memory", "w0", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+        "v8", "v9", "v10", "v11", "v12");
+}
+
+#if LIBYUV_USE_ST2
 // Reads 16 U's and V's and writes out 16 pairs of UV.
-void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+void MergeUVRow_NEON(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
                      int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load U
-    MEMACCESS(1)
-    "ld1        {v1.16b}, [%1], #16            \n"  // load V
-    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
-    MEMACCESS(2)
-    "st2        {v0.16b,v1.16b}, [%2], #32     \n"  // store 16 pairs of UV
-    "b.gt       1b                             \n"
-    :
-      "+r"(src_u),   // %0
-      "+r"(src_v),   // %1
-      "+r"(dst_uv),  // %2
-      "+r"(width)    // %3  // Output registers
-    :                       // Input registers
-    : "cc", "memory", "v0", "v1"  // Clobber List
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load U
+      "ld1         {v1.16b}, [%1], #16           \n"  // load V
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st2         {v0.16b,v1.16b}, [%2], #32    \n"  // store 16 pairs of UV
+      "b.gt        1b                            \n"
+      : "+r"(src_u),                // %0
+        "+r"(src_v),                // %1
+        "+r"(dst_uv),               // %2
+        "+r"(width)                 // %3  // Output registers
+      :                             // Input registers
+      : "cc", "memory", "v0", "v1"  // Clobber List
   );
 }
 
-// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
-void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32       \n"  // load 32
-    "subs       %w2, %w2, #32                  \n"  // 32 processed per loop
-    MEMACCESS(1)
-    "st1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32       \n"  // store 32
-    "b.gt       1b                             \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(count)  // %2  // Output registers
-  :                     // Input registers
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  int shift = 16 - depth;
+  asm volatile(
+      "dup         v2.8h, %w4                    \n"
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // load 8 U
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "ld1         {v1.8h}, [%1], #16            \n"  // load 8 V
+      "ushl        v0.8h, v0.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st2         {v0.8h, v1.8h}, [%2], #32     \n"  // store 8 UV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_u),   // %0
+        "+r"(src_v),   // %1
+        "+r"(dst_uv),  // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "v0", "v1", "v2");
+}
+#else
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
+                     int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load U
+      "ld1         {v1.16b}, [%1], #16           \n"  // load V
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "zip1        v2.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip2        v3.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v2.16b,v3.16b}, [%2], #32    \n"  // store 16 pairs of UV
+      "b.gt        1b                            \n"
+      : "+r"(src_u),                            // %0
+        "+r"(src_v),                            // %1
+        "+r"(dst_uv),                           // %2
+        "+r"(width)                             // %3  // Output registers
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-// SetRow writes 'count' bytes using an 8 bit value repeated.
-void SetRow_NEON(uint8* dst, uint8 v8, int count) {
-  asm volatile (
-    "dup        v0.16b, %w2                    \n"  // duplicate 16 bytes
-  "1:                                          \n"
-    "subs       %w1, %w1, #16                  \n"  // 16 bytes per loop
-    MEMACCESS(0)
-    "st1        {v0.16b}, [%0], #16            \n"  // store
-    "b.gt       1b                             \n"
-  : "+r"(dst),   // %0
-    "+r"(count)  // %1
-  : "r"(v8)      // %2
-  : "cc", "memory", "v0"
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  int shift = 16 - depth;
+  asm volatile(
+      "dup         v4.8h, %w4                    \n"
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // load 8 U
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "ld1         {v1.8h}, [%1], #16            \n"  // load 8 V
+      "ushl        v0.8h, v0.8h, v4.8h           \n"
+      "ushl        v1.8h, v1.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v2.8h, v0.8h, v1.8h           \n"
+      "zip2        v3.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v2.8h, v3.8h}, [%2], #32     \n"  // store 8 UV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_u),   // %0
+        "+r"(src_v),   // %1
+        "+r"(dst_uv),  // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "v0", "v1", "v2", "v1", "v2", "v3", "v4");
+}
+#endif  // LIBYUV_USE_ST2
+
+// Reads 16 packed RGB and write to planar dst_r, dst_g, dst_b.
+void SplitRGBRow_NEON(const uint8_t* src_rgb,
+                      uint8_t* dst_r,
+                      uint8_t* dst_g,
+                      uint8_t* dst_b,
+                      int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 RGB
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%1], #16           \n"  // store R
+      "st1         {v1.16b}, [%2], #16           \n"  // store G
+      "st1         {v2.16b}, [%3], #16           \n"  // store B
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb),                    // %0
+        "+r"(dst_r),                      // %1
+        "+r"(dst_g),                      // %2
+        "+r"(dst_b),                      // %3
+        "+r"(width)                       // %4
+      :                                   // Input registers
+      : "cc", "memory", "v0", "v1", "v2"  // Clobber List
   );
 }
 
-void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
-  asm volatile (
-    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
-  "1:                                          \n"
-    "subs       %w1, %w1, #4                   \n"  // 4 ints per loop
-    MEMACCESS(0)
-    "st1        {v0.16b}, [%0], #16            \n"  // store
-    "b.gt       1b                             \n"
-  : "+r"(dst),   // %0
-    "+r"(count)  // %1
-  : "r"(v32)     // %2
-  : "cc", "memory", "v0"
+// Reads 16 planar R's, G's and B's and writes out 16 packed RGB at a time
+void MergeRGBRow_NEON(const uint8_t* src_r,
+                      const uint8_t* src_g,
+                      const uint8_t* src_b,
+                      uint8_t* dst_rgb,
+                      int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load R
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v2.16b}, [%2], #16           \n"  // load B
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st3         {v0.16b,v1.16b,v2.16b}, [%3], #48 \n"  // store 16 RGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),                      // %0
+        "+r"(src_g),                      // %1
+        "+r"(src_b),                      // %2
+        "+r"(dst_rgb),                    // %3
+        "+r"(width)                       // %4
+      :                                   // Input registers
+      : "cc", "memory", "v0", "v1", "v2"  // Clobber List
   );
 }
 
-void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-    // Start at end of source row.
-    "add        %0, %0, %w2, sxtw              \n"
-    "sub        %0, %0, #16                    \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
-    "subs       %w2, %w2, #16                  \n"  // 16 pixels per loop.
-    "rev64      v0.16b, v0.16b                 \n"
-    MEMACCESS(1)
-    "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
-    MEMACCESS(1)
-    "st1        {v0.D}[0], [%1], #8            \n"
-    "b.gt       1b                             \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  : "r"((ptrdiff_t)-16)    // %3
-  : "cc", "memory", "v0"
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b, dst_a.
+void SplitARGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w5, %w5, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%3], #16           \n"  // store B
+      "st1         {v1.16b}, [%2], #16           \n"  // store G
+      "st1         {v2.16b}, [%1], #16           \n"  // store R
+      "st1         {v3.16b}, [%4], #16           \n"  // store A
+      "b.gt        1b                            \n"
+      : "+r"(src_rgba),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(dst_a),                            // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                      int width) {
-  asm volatile (
-    // Start at end of source row.
-    "add        %0, %0, %w3, sxtw #1           \n"
-    "sub        %0, %0, #16                    \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld2        {v0.8b, v1.8b}, [%0], %4       \n"  // src -= 16
-    "subs       %w3, %w3, #8                   \n"  // 8 pixels per loop.
-    "rev64      v0.8b, v0.8b                   \n"
-    "rev64      v1.8b, v1.8b                   \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // dst += 8
-    MEMACCESS(2)
-    "st1        {v1.8b}, [%2], #8              \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_uv),  // %0
-    "+r"(dst_u),   // %1
-    "+r"(dst_v),   // %2
-    "+r"(width)    // %3
-  : "r"((ptrdiff_t)-16)      // %4
-  : "cc", "memory", "v0", "v1"
+#if LIBYUV_USE_ST4
+// Reads 16 planar R's, G's, B's and A's and writes out 16 packed ARGB at a time
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%2], #16           \n"  // load B
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v2.16b}, [%0], #16           \n"  // load R
+      "ld1         {v3.16b}, [%3], #16           \n"  // load A
+      "subs        %w5, %w5, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "st4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%4], #64 \n"  // store 16ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(src_a),                            // %3
+        "+r"(dst_argb),                         // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#else
+// Reads 16 planar R's, G's, B's and A's and writes out 16 packed ARGB at a time
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%2], #16           \n"  // load B
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v2.16b}, [%0], #16           \n"  // load R
+      "ld1         {v3.16b}, [%3], #16           \n"  // load A
+      "subs        %w5, %w5, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "zip1        v4.16b, v0.16b, v1.16b        \n"  // BG
+      "zip1        v5.16b, v2.16b, v3.16b        \n"  // RA
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "zip2        v6.16b, v0.16b, v1.16b        \n"  // BG
+      "zip2        v7.16b, v2.16b, v3.16b        \n"  // RA
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v0.8h, v4.8h, v5.8h           \n"  // BGRA
+      "zip2        v1.8h, v4.8h, v5.8h           \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "zip1        v2.8h, v6.8h, v7.8h           \n"
+      "zip2        v3.8h, v6.8h, v7.8h           \n"
+      "st1         {v0.16b,v1.16b,v2.16b,v3.16b}, [%4], #64 \n"  // store 16ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      :                  // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+        "v7"  // Clobber List
+  );
+}
+#endif  // LIBYUV_USE_ST4
+
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b.
+void SplitXRGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%3], #16           \n"  // store B
+      "st1         {v1.16b}, [%2], #16           \n"  // store G
+      "st1         {v2.16b}, [%1], #16           \n"  // store R
+      "b.gt        1b                            \n"
+      : "+r"(src_rgba),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
-  asm volatile (
-  // Start at end of source row.
-    "add        %0, %0, %w2, sxtw #2           \n"
-    "sub        %0, %0, #16                    \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
-    "subs       %w2, %w2, #4                   \n"  // 4 pixels per loop.
-    "rev64      v0.4s, v0.4s                   \n"
-    MEMACCESS(1)
-    "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
-    MEMACCESS(1)
-    "st1        {v0.D}[0], [%1], #8            \n"
-    "b.gt       1b                             \n"
-  : "+r"(src),   // %0
-    "+r"(dst),   // %1
-    "+r"(width)  // %2
-  : "r"((ptrdiff_t)-16)    // %3
-  : "cc", "memory", "v0"
+// Reads 16 planar R's, G's and B's and writes out 16 packed ARGB at a time
+void MergeXRGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "movi        v3.16b, #255                  \n"  // load A(255)
+      "1:                                        \n"
+      "ld1         {v2.16b}, [%0], #16           \n"  // load R
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v0.16b}, [%2], #16           \n"  // load B
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%3], #64 \n"  // store 16ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(dst_argb),                         // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v4.8b, #255                    \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of RGB24.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    MEMACCESS(1)
-    "st4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_rgb24),  // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+void MergeXR30Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width) {
+  int shift = 10 - depth;
+  asm volatile(
+      "movi        v30.16b, #255                 \n"
+      "ushr        v30.4s, v30.4s, #22           \n"  // 1023
+      "dup         v31.4s, %w5                   \n"
+      "1:                                        \n"
+      "ldr         d2, [%2], #8                  \n"  // B
+      "ldr         d1, [%1], #8                  \n"  // G
+      "ldr         d0, [%0], #8                  \n"  // R
+      "ushll       v2.4s, v2.4h, #0              \n"  // B
+      "ushll       v1.4s, v1.4h, #0              \n"  // G
+      "ushll       v0.4s, v0.4h, #0              \n"  // R
+      "ushl        v2.4s, v2.4s, v31.4s          \n"  // 000B
+      "ushl        v1.4s, v1.4s, v31.4s          \n"  // G
+      "ushl        v0.4s, v0.4s, v31.4s          \n"  // R
+      "umin        v2.4s, v2.4s, v30.4s          \n"
+      "umin        v1.4s, v1.4s, v30.4s          \n"
+      "umin        v0.4s, v0.4s, v30.4s          \n"
+      "sli         v2.4s, v1.4s, #10             \n"  // 00GB
+      "sli         v2.4s, v0.4s, #20             \n"  // 0RGB
+      "orr         v2.4s, #0xc0, lsl #24         \n"  // ARGB (AR30)
+      "subs        %w4, %w4, #4                  \n"
+      "str         q2, [%3], #16                 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "v0", "v1", "v2", "v30", "v31");
+}
+
+void MergeXR30Row_10_NEON(const uint16_t* src_r,
+                          const uint16_t* src_g,
+                          const uint16_t* src_b,
+                          uint8_t* dst_ar30,
+                          int /* depth */,
+                          int width) {
+  asm volatile(
+      "movi        v30.16b, #255                 \n"
+      "ushr        v30.4s, v30.4s, #22           \n"  // 1023
+      "1:                                        \n"
+      "ldr         d2, [%2], #8                  \n"  // B
+      "ldr         d1, [%1], #8                  \n"  // G
+      "ldr         d0, [%0], #8                  \n"  // R
+      "ushll       v2.4s, v2.4h, #0              \n"  // 000B
+      "ushll       v1.4s, v1.4h, #0              \n"  // G
+      "ushll       v0.4s, v0.4h, #0              \n"  // R
+      "umin        v2.4s, v2.4s, v30.4s          \n"
+      "umin        v1.4s, v1.4s, v30.4s          \n"
+      "umin        v0.4s, v0.4s, v30.4s          \n"
+      "sli         v2.4s, v1.4s, #10             \n"  // 00GB
+      "sli         v2.4s, v0.4s, #20             \n"  // 0RGB
+      "orr         v2.4s, #0xc0, lsl #24         \n"  // ARGB (AR30)
+      "subs        %w4, %w4, #4                  \n"
+      "str         q2, [%3], #16                 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v30");
+}
+
+void MergeAR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "dup         v30.8h, %w7                   \n"
+      "dup         v31.8h, %w6                   \n"
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "ldr         q3, [%3], #16                 \n"  // A
+      "umin        v2.8h, v2.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umin        v1.8h, v1.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umin        v0.8h, v0.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umin        v3.8h, v3.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "ushl        v3.8h, v3.8h, v31.8h          \n"
+      "subs        %w5, %w5, #8                  \n"
+      "st4         {v0.8h, v1.8h, v2.8h, v3.8h}, [%4], #64 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_ar64),  // %4
+        "+r"(width)      // %5
+      : "r"(shift),      // %6
+        "r"(mask)        // %7
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+void MergeXR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "movi        v3.16b, #0xff                 \n"  // A (0xffff)
+      "dup         v30.8h, %w6                   \n"
+      "dup         v31.8h, %w5                   \n"
+
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "umin        v2.8h, v2.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umin        v1.8h, v1.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umin        v0.8h, v0.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "subs        %w4, %w4, #8                  \n"
+      "st4         {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar64),  // %3
+        "+r"(width)      // %4
+      : "r"(shift),      // %5
+        "r"(mask)        // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+void MergeARGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "dup         v31.8h, %w6                   \n"
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "ldr         q3, [%3], #16                 \n"  // A
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "ushl        v3.8h, v3.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "uqxtn       v2.8b, v2.8h                  \n"
+      "uqxtn       v1.8b, v1.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "uqxtn       v3.8b, v3.8h                  \n"
+      "subs        %w5, %w5, #8                  \n"
+      "st4         {v0.8b, v1.8b, v2.8b, v3.8b}, [%4], #32 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      : "r"(shift)       // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+void MergeXRGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "dup         v31.8h, %w5                   \n"
+      "movi        v3.8b, #0xff                  \n"  // A (0xff)
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "uqxtn       v2.8b, v2.8h                  \n"
+      "uqxtn       v1.8b, v1.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "subs        %w4, %w4, #8                  \n"
+      "st4         {v0.8b, v1.8b, v2.8b, v3.8b}, [%3], #32 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+// Copy multiple of 32.
+void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #32                 \n"  // 32 processed per loop
+      "stp         q0, q1, [%1], #32             \n"
+      "b.gt        1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst),                  // %1
+        "+r"(width)                 // %2  // Output registers
+      :                             // Input registers
+      : "cc", "memory", "v0", "v1"  // Clobber List
   );
 }
 
-void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v5.8b, #255                    \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
-    "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
-    MEMACCESS(1)
-    "st4        {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
-    "b.gt       1b                             \n"
-  : "+r"(src_raw),   // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+// SetRow writes 'width' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8_t* dst, uint8_t v8, int width) {
+  asm volatile(
+      "dup         v0.16b, %w2                   \n"  // duplicate 16 bytes
+      "1:                                        \n"
+      "subs        %w1, %w1, #16                 \n"  // 16 bytes per loop
+      "st1         {v0.16b}, [%0], #16           \n"  // store
+      "b.gt        1b                            \n"
+      : "+r"(dst),   // %0
+        "+r"(width)  // %1
+      : "r"(v8)      // %2
+      : "cc", "memory", "v0");
+}
+
+void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width) {
+  asm volatile(
+      "dup         v0.4s, %w2                    \n"  // duplicate 4 ints
+      "1:                                        \n"
+      "subs        %w1, %w1, #4                  \n"  // 4 ints per loop
+      "st1         {v0.16b}, [%0], #16           \n"  // store
+      "b.gt        1b                            \n"
+      : "+r"(dst),   // %0
+        "+r"(width)  // %1
+      : "r"(v32)     // %2
+      : "cc", "memory", "v0");
+}
+
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u,
+                                     7u,  6u,  5u,  4u,  3u,  2u,  1u, 0u};
+
+void MirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "ld1         {v3.16b}, [%3]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw             \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q2, [%0, 16]                  \n"
+      "ldr         q1, [%0], -32                 \n"  // src -= 32
+      "subs        %w2, %w2, #32                 \n"  // 32 pixels per loop.
+      "tbl         v0.16b, {v2.16b}, v3.16b      \n"
+      "tbl         v1.16b, {v1.16b}, v3.16b      \n"
+      "st1         {v0.16b, v1.16b}, [%1], #32   \n"  // store 32 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(&kShuffleMirror)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3");
+}
+
+// Shuffle table for reversing the UV.
+static const uvec8 kShuffleMirrorUV = {14u, 15u, 12u, 13u, 10u, 11u, 8u, 9u,
+                                       6u,  7u,  4u,  5u,  2u,  3u,  0u, 1u};
+
+void MirrorUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "ld1         {v4.16b}, [%3]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw #1          \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q1, [%0, 16]                  \n"
+      "ldr         q0, [%0], -32                 \n"  // src -= 32
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop.
+      "tbl         v2.16b, {v1.16b}, v4.16b      \n"
+      "tbl         v3.16b, {v0.16b}, v4.16b      \n"
+      "st1         {v2.16b, v3.16b}, [%1], #32   \n"  // dst += 32
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),           // %0
+        "+r"(dst_uv),           // %1
+        "+r"(width)             // %2
+      : "r"(&kShuffleMirrorUV)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+void MirrorSplitUVRow_NEON(const uint8_t* src_uv,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
+  asm volatile(
+      // Start at end of source row.
+      "ld1         {v4.16b}, [%4]                \n"  // shuffler
+      "add         %0, %0, %w3, sxtw #1          \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q1, [%0, 16]                  \n"
+      "ldr         q0, [%0], -32                 \n"  // src -= 32
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels per loop.
+      "tbl         v2.16b, {v1.16b}, v4.16b      \n"
+      "tbl         v3.16b, {v0.16b}, v4.16b      \n"
+      "uzp1        v0.16b, v2.16b, v3.16b        \n"  // U
+      "uzp2        v1.16b, v2.16b, v3.16b        \n"  // V
+      "st1         {v0.16b}, [%1], #16           \n"  // dst += 16
+      "st1         {v1.16b}, [%2], #16           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),           // %0
+        "+r"(dst_u),            // %1
+        "+r"(dst_v),            // %2
+        "+r"(width)             // %3
+      : "r"(&kShuffleMirrorUV)  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+// Shuffle table for reversing the ARGB.
+static const uvec8 kShuffleMirrorARGB = {12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u,
+                                         4u,  5u,  6u,  7u,  0u, 1u, 2u,  3u};
+
+void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "ld1         {v4.16b}, [%3]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw #2          \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q1, [%0, 16]                  \n"
+      "ldr         q0, [%0], -32                 \n"  // src -= 32
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop.
+      "tbl         v2.16b, {v1.16b}, v4.16b      \n"
+      "tbl         v3.16b, {v0.16b}, v4.16b      \n"
+      "st1         {v2.16b, v3.16b}, [%1], #32   \n"  // dst += 32
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleMirrorARGB)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+void RGB24MirrorRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_rgb24,
+                         int width) {
+  asm volatile(
+      "ld1         {v3.16b}, [%4]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw #1          \n"  // Start at end of row.
+      "add         %0, %0, %w2, sxtw             \n"
+      "sub         %0, %0, #48                   \n"
+
+      "1:                                        \n"
+      "ld3         {v0.16b, v1.16b, v2.16b}, [%0], %3 \n"  // src -= 48
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop.
+      "tbl         v0.16b, {v0.16b}, v3.16b      \n"
+      "tbl         v1.16b, {v1.16b}, v3.16b      \n"
+      "tbl         v2.16b, {v2.16b}, v3.16b      \n"
+      "st3         {v0.16b, v1.16b, v2.16b}, [%1], #48 \n"  // dst += 48
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb24),      // %0
+        "+r"(dst_rgb24),      // %1
+        "+r"(width)           // %2
+      : "r"((ptrdiff_t)-48),  // %3
+        "r"(&kShuffleMirror)  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3");
+}
+
+void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width) {
+  asm volatile(
+      "movi        v4.8b, #255                   \n"  // Alpha
+      "1:                                        \n"
+      "ld3         {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of
+                                                       // RGB24.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "st4         {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb24),  // %0
+        "+r"(dst_argb),   // %1
+        "+r"(width)       // %2
+      :
+      : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
   );
 }
 
-void RAWToRGB24Row_NEON(const uint8* src_raw, uint8* dst_rgb24, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
-    "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
-    MEMACCESS(1)
-    "st3        {v2.8b,v3.8b,v4.8b}, [%1], #24 \n"  // store b g r
-    "b.gt       1b                             \n"
-  : "+r"(src_raw),    // %0
-    "+r"(dst_rgb24),  // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
+void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movi        v5.8b, #255                   \n"  // Alpha
+      "1:                                        \n"
+      "ld3         {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
+      "subs        %w2, %w2, #8                  \n"   // 8 processed per loop.
+      "orr         v3.8b, v1.8b, v1.8b           \n"   // move g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v4.8b, v0.8b, v0.8b           \n"         // move r
+      "st4         {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
+      "b.gt        1b                            \n"
+      : "+r"(src_raw),   // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5"  // Clobber List
   );
 }
 
-#define RGB565TOARGB                                                           \
-    "shrn       v6.8b, v0.8h, #5               \n"  /* G xxGGGGGG           */ \
-    "shl        v6.8b, v6.8b, #2               \n"  /* G GGGGGG00 upper 6   */ \
-    "ushr       v4.8b, v6.8b, #6               \n"  /* G 000000GG lower 2   */ \
-    "orr        v1.8b, v4.8b, v6.8b            \n"  /* G                    */ \
-    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
-    "ushr       v0.8h, v0.8h, #11              \n"  /* R 000RRRRR           */ \
-    "xtn2       v2.16b,v0.8h                   \n"  /* R in upper part      */ \
-    "shl        v2.16b, v2.16b, #3             \n"  /* R,B BBBBB000 upper 5 */ \
-    "ushr       v0.16b, v2.16b, #5             \n"  /* R,B 00000BBB lower 3 */ \
-    "orr        v0.16b, v0.16b, v2.16b         \n"  /* R,B                  */ \
-    "dup        v2.2D, v0.D[1]                 \n"  /* R                    */
-
-void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v3.8b, #255                    \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    RGB565TOARGB
-    MEMACCESS(1)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_rgb565),  // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6"  // Clobber List
+void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  asm volatile(
+      "movi        v0.8b, #255                   \n"  // Alpha
+      "1:                                        \n"
+      "ld3         {v3.8b,v4.8b,v5.8b}, [%0], #24 \n"  // read r g b
+      "subs        %w2, %w2, #8                  \n"   // 8 processed per loop.
+      "orr         v2.8b, v4.8b, v4.8b           \n"   // move g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v1.8b, v5.8b, v5.8b           \n"         // move r
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store a b g r
+      "b.gt        1b                            \n"
+      : "+r"(src_raw),   // %0
+        "+r"(dst_rgba),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+  );
+}
+
+void RAWToRGB24Row_NEON(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld3         {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
+      "subs        %w2, %w2, #8                  \n"   // 8 processed per loop.
+      "orr         v3.8b, v1.8b, v1.8b           \n"   // move g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v4.8b, v0.8b, v0.8b           \n"   // move r
+      "st3         {v2.8b,v3.8b,v4.8b}, [%1], #24 \n"  // store b g r
+      "b.gt        1b                            \n"
+      : "+r"(src_raw),    // %0
+        "+r"(dst_rgb24),  // %1
+        "+r"(width)       // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                        \
+  "shrn       v6.8b, v0.8h, #5               \n" /* G xxGGGGGG           */ \
+  "shl        v6.8b, v6.8b, #2               \n" /* G GGGGGG00 upper 6   */ \
+  "ushr       v4.8b, v6.8b, #6               \n" /* G 000000GG lower 2   */ \
+  "orr        v1.8b, v4.8b, v6.8b            \n" /* G                    */ \
+  "xtn        v2.8b, v0.8h                   \n" /* B xxxBBBBB           */ \
+  "ushr       v0.8h, v0.8h, #11              \n" /* R 000RRRRR           */ \
+  "xtn2       v2.16b,v0.8h                   \n" /* R in upper part      */ \
+  "shl        v2.16b, v2.16b, #3             \n" /* R,B BBBBB000 upper 5 */ \
+  "ushr       v0.16b, v2.16b, #5             \n" /* R,B 00000BBB lower 3 */ \
+  "orr        v0.16b, v0.16b, v2.16b         \n" /* R,B                  */ \
+  "dup        v2.2D, v0.D[1]                 \n" /* R                    */
+
+void RGB565ToARGBRow_NEON(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      "movi        v3.8b, #255                   \n"  // Alpha
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 RGB565 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" RGB565TOARGB
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb565),  // %0
+        "+r"(dst_argb),    // %1
+        "+r"(width)        // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6"  // Clobber List
   );
 }
 
-#define ARGB1555TOARGB                                                         \
-    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
-    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
-    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000 AAAAAAAA    */ \
-                                                                               \
-    "sshr       v2.8h, v0.8h, #15              \n"  /* A AAAAAAAA           */ \
-    "xtn2       v3.16b, v2.8h                  \n"                             \
-                                                                               \
-    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
-    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
-                                                                               \
-    "ushr       v1.16b, v3.16b, #5             \n"  /* R,A 00000RRR lower 3 */ \
-    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
-    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
-                                                                               \
-    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
-    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R,A                  */ \
-    "dup        v1.2D, v0.D[1]                 \n"                             \
-    "dup        v3.2D, v2.D[1]                 \n"
+#define ARGB1555TOARGB                                                      \
+  "ushr       v2.8h, v0.8h, #10              \n" /* R xxxRRRRR           */ \
+  "shl        v2.8h, v2.8h, #3               \n" /* R RRRRR000 upper 5   */ \
+  "xtn        v3.8b, v2.8h                   \n" /* RRRRR000 AAAAAAAA    */ \
+                                                                            \
+  "sshr       v2.8h, v0.8h, #15              \n" /* A AAAAAAAA           */ \
+  "xtn2       v3.16b, v2.8h                  \n"                            \
+                                                                            \
+  "xtn        v2.8b, v0.8h                   \n" /* B xxxBBBBB           */ \
+  "shrn2      v2.16b,v0.8h, #5               \n" /* G xxxGGGGG           */ \
+                                                                            \
+  "ushr       v1.16b, v3.16b, #5             \n" /* R,A 00000RRR lower 3 */ \
+  "shl        v0.16b, v2.16b, #3             \n" /* B,G BBBBB000 upper 5 */ \
+  "ushr       v2.16b, v0.16b, #5             \n" /* B,G 00000BBB lower 3 */ \
+                                                                            \
+  "orr        v0.16b, v0.16b, v2.16b         \n" /* B,G                  */ \
+  "orr        v2.16b, v1.16b, v3.16b         \n" /* R,A                  */ \
+  "dup        v1.2D, v0.D[1]                 \n"                            \
+  "dup        v3.2D, v2.D[1]                 \n"
 
 // RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
-#define RGB555TOARGB                                                           \
-    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
-    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
-    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000             */ \
-                                                                               \
-    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
-    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
-                                                                               \
-    "ushr       v1.16b, v3.16b, #5             \n"  /* R   00000RRR lower 3 */ \
-    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
-    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
-                                                                               \
-    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
-    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R                    */ \
-    "dup        v1.2D, v0.D[1]                 \n"  /* G */                    \
-
-void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+#define RGB555TOARGB                                                        \
+  "ushr       v2.8h, v0.8h, #10              \n" /* R xxxRRRRR           */ \
+  "shl        v2.8h, v2.8h, #3               \n" /* R RRRRR000 upper 5   */ \
+  "xtn        v3.8b, v2.8h                   \n" /* RRRRR000             */ \
+                                                                            \
+  "xtn        v2.8b, v0.8h                   \n" /* B xxxBBBBB           */ \
+  "shrn2      v2.16b,v0.8h, #5               \n" /* G xxxGGGGG           */ \
+                                                                            \
+  "ushr       v1.16b, v3.16b, #5             \n" /* R   00000RRR lower 3 */ \
+  "shl        v0.16b, v2.16b, #3             \n" /* B,G BBBBB000 upper 5 */ \
+  "ushr       v2.16b, v0.16b, #5             \n" /* B,G 00000BBB lower 3 */ \
+                                                                            \
+  "orr        v0.16b, v0.16b, v2.16b         \n" /* B,G                  */ \
+  "orr        v2.16b, v1.16b, v3.16b         \n" /* R                    */ \
+  "dup        v1.2D, v0.D[1]                 \n" /* G */
+
+void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
                             int width) {
-  asm volatile (
-    "movi       v3.8b, #255                    \n"  // Alpha
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGB1555TOARGB
-    MEMACCESS(1)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_argb1555),  // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  asm volatile(
+      "movi        v3.8b, #255                   \n"  // Alpha
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB1555 pixels.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      ARGB1555TOARGB
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb1555),  // %0
+        "+r"(dst_argb),      // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-#define ARGB4444TOARGB                                                         \
-    "shrn       v1.8b,  v0.8h, #8              \n"  /* v1(l) AR             */ \
-    "xtn2       v1.16b, v0.8h                  \n"  /* v1(h) GB             */ \
-    "shl        v2.16b, v1.16b, #4             \n"  /* B,R BBBB0000         */ \
-    "ushr       v3.16b, v1.16b, #4             \n"  /* G,A 0000GGGG         */ \
-    "ushr       v0.16b, v2.16b, #4             \n"  /* B,R 0000BBBB         */ \
-    "shl        v1.16b, v3.16b, #4             \n"  /* G,A GGGG0000         */ \
-    "orr        v2.16b, v0.16b, v2.16b         \n"  /* B,R BBBBBBBB         */ \
-    "orr        v3.16b, v1.16b, v3.16b         \n"  /* G,A GGGGGGGG         */ \
-    "dup        v0.2D, v2.D[1]                 \n"                             \
-    "dup        v1.2D, v3.D[1]                 \n"
-
-void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+// Convert v0.8h to b = v0.8b g = v1.8b r = v2.8b
+// clobbers v3
+#define ARGB4444TOARGB                                                      \
+  "shrn       v1.8b,  v0.8h, #8              \n" /* v1(l) AR             */ \
+  "xtn2       v1.16b, v0.8h                  \n" /* v1(h) GB             */ \
+  "shl        v2.16b, v1.16b, #4             \n" /* B,R BBBB0000         */ \
+  "ushr       v3.16b, v1.16b, #4             \n" /* G,A 0000GGGG         */ \
+  "ushr       v0.16b, v2.16b, #4             \n" /* B,R 0000BBBB         */ \
+  "shl        v1.16b, v3.16b, #4             \n" /* G,A GGGG0000         */ \
+  "orr        v2.16b, v0.16b, v2.16b         \n" /* B,R BBBBBBBB         */ \
+  "orr        v3.16b, v1.16b, v3.16b         \n" /* G,A GGGGGGGG         */ \
+  "dup        v0.2D, v2.D[1]                 \n"                            \
+  "dup        v1.2D, v3.D[1]                 \n"
+
+void ARGB4444ToARGBRow_NEON(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
                             int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGB4444TOARGB
-    MEMACCESS(1)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_argb4444),  // %0
-    "+r"(dst_argb),    // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB4444 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGB4444TOARGB
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb4444),  // %0
+        "+r"(dst_argb),      // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
   );
 }
 
-void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load 8 ARGB pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    MEMACCESS(1)
-    "st3        {v1.8b,v2.8b,v3.8b}, [%1], #24 \n"  // store 8 pixels of RGB24.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_rgb24),  // %1
-    "+r"(width)         // %2
-  :
-  : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+void ARGBToRGB24Row_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_rgb24,
+                         int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st3         {v0.16b,v1.16b,v2.16b}, [%1], #48 \n"  // store 8 RGB24
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(dst_rgb24),  // %1
+        "+r"(width)       // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "orr        v4.8b, v2.8b, v2.8b            \n"  // mov g
-    "orr        v5.8b, v1.8b, v1.8b            \n"  // mov b
-    MEMACCESS(1)
-    "st3        {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_raw),   // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "orr         v4.8b, v2.8b, v2.8b           \n"  // mov g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v5.8b, v1.8b, v1.8b           \n"   // mov b
+      "st3         {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_raw),   // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v1", "v2", "v3", "v4", "v5"  // Clobber List
   );
 }
 
-void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of YUY2.
-    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"  // store 16 pixels of Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_yuy2),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1"  // Clobber List
+void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pixels of YUY2.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%1], #16           \n"  // store 16 pixels of Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1"  // Clobber List
   );
 }
 
-void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of UYVY.
-    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
-    MEMACCESS(1)
-    "st1        {v1.16b}, [%1], #16            \n"  // store 16 pixels of Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_uyvy),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1"  // Clobber List
+void UYVYToYRow_NEON(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pixels of UYVY.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v1.16b}, [%1], #16           \n"  // store 16 pixels of Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1"  // Clobber List
   );
 }
 
-void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+void YUY2ToUV422Row_NEON(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2 pixels
-    "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "st1        {v1.8b}, [%1], #8              \n"  // store 8 U.
-    MEMACCESS(2)
-    "st1        {v3.8b}, [%2], #8              \n"  // store 8 V.
-    "b.gt       1b                             \n"
-  : "+r"(src_yuy2),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels = 8 UVs.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v1.8b}, [%1], #8             \n"  // store 8 U.
+      "st1         {v3.8b}, [%2], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
+      : "+r"(src_yuy2),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+void UYVYToUV422Row_NEON(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY pixels
-    "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 U.
-    MEMACCESS(2)
-    "st1        {v2.8b}, [%2], #8              \n"  // store 8 V.
-    "b.gt       1b                             \n"
-  : "+r"(src_uyvy),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels = 8 UVs.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 U.
+      "st1         {v2.8b}, [%2], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
+      : "+r"(src_uyvy),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_yuy2b = src_yuy2 + stride_yuy2;
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
-    "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
-    "urhadd     v1.8b, v1.8b, v5.8b            \n"  // average rows of U
-    "urhadd     v3.8b, v3.8b, v7.8b            \n"  // average rows of V
-    MEMACCESS(2)
-    "st1        {v1.8b}, [%2], #8              \n"  // store 8 U.
-    MEMACCESS(3)
-    "st1        {v3.8b}, [%3], #8              \n"  // store 8 V.
-    "b.gt       1b                             \n"
-  : "+r"(src_yuy2),     // %0
-    "+r"(src_yuy2b),    // %1
-    "+r"(dst_u),        // %2
-    "+r"(dst_v),        // %3
-    "+r"(width)           // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
-    "v5", "v6", "v7"  // Clobber List
+void YUY2ToUVRow_NEON(const uint8_t* src_yuy2,
+                      int stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_yuy2b = src_yuy2 + stride_yuy2;
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+      "subs        %w4, %w4, #16                 \n"  // 16 pixels = 8 UVs.
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+      "urhadd      v1.8b, v1.8b, v5.8b           \n"  // average rows of U
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "urhadd      v3.8b, v3.8b, v7.8b           \n"  // average rows of V
+      "st1         {v1.8b}, [%2], #8             \n"  // store 8 U.
+      "st1         {v3.8b}, [%3], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
+      : "+r"(src_yuy2),   // %0
+        "+r"(src_yuy2b),  // %1
+        "+r"(dst_u),      // %2
+        "+r"(dst_v),      // %3
+        "+r"(width)       // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+        "v7"  // Clobber List
   );
 }
 
-void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_uyvyb = src_uyvy + stride_uyvy;
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
-    "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
-    MEMACCESS(1)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
-    "urhadd     v0.8b, v0.8b, v4.8b            \n"  // average rows of U
-    "urhadd     v2.8b, v2.8b, v6.8b            \n"  // average rows of V
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 U.
-    MEMACCESS(3)
-    "st1        {v2.8b}, [%3], #8              \n"  // store 8 V.
-    "b.gt       1b                             \n"
-  : "+r"(src_uyvy),     // %0
-    "+r"(src_uyvyb),    // %1
-    "+r"(dst_u),        // %2
-    "+r"(dst_v),        // %3
-    "+r"(width)           // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
-    "v5", "v6", "v7"  // Clobber List
+void UYVYToUVRow_NEON(const uint8_t* src_uyvy,
+                      int stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_uyvyb = src_uyvy + stride_uyvy;
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+      "subs        %w4, %w4, #16                 \n"  // 16 pixels = 8 UVs.
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+      "urhadd      v0.8b, v0.8b, v4.8b           \n"  // average rows of U
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "urhadd      v2.8b, v2.8b, v6.8b           \n"  // average rows of V
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 U.
+      "st1         {v2.8b}, [%3], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
+      : "+r"(src_uyvy),   // %0
+        "+r"(src_uyvyb),  // %1
+        "+r"(dst_u),      // %2
+        "+r"(dst_v),      // %3
+        "+r"(width)       // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+        "v7"  // Clobber List
   );
 }
 
-// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
-void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width) {
-  asm volatile (
-    MEMACCESS(3)
-    "ld1        {v2.16b}, [%3]                 \n"  // shuffler
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 4 pixels.
-    "subs       %w2, %w2, #4                   \n"  // 4 processed per loop
-    "tbl        v1.16b, {v0.16b}, v2.16b       \n"  // look up 4 pixels
-    MEMACCESS(1)
-    "st1        {v1.16b}, [%1], #16            \n"  // store 4.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)        // %2
-  : "r"(shuffler)    // %3
-  : "cc", "memory", "v0", "v1", "v2"  // Clobber List
+void YUY2ToNVUVRow_NEON(const uint8_t* src_yuy2,
+                        int stride_yuy2,
+                        uint8_t* dst_uv,
+                        int width) {
+  const uint8_t* src_yuy2b = src_yuy2 + stride_yuy2;
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pixels
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels = 8 UVs.
+      "ld2         {v2.16b,v3.16b}, [%1], #32    \n"  // load next row
+      "urhadd      v4.16b, v1.16b, v3.16b        \n"  // average rows of UV
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v4.16b}, [%2], #16           \n"  // store 8 UV.
+      "b.gt        1b                            \n"
+      : "+r"(src_yuy2),   // %0
+        "+r"(src_yuy2b),  // %1
+        "+r"(dst_uv),     // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
   );
 }
 
-void I422ToYUY2Row_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_yuy2, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"  // load 16 Ys
-    "orr        v2.8b, v1.8b, v1.8b            \n"
-    MEMACCESS(1)
-    "ld1        {v1.8b}, [%1], #8              \n"  // load 8 Us
-    MEMACCESS(2)
-    "ld1        {v3.8b}, [%2], #8              \n"  // load 8 Vs
-    "subs       %w4, %w4, #16                  \n"  // 16 pixels
-    MEMACCESS(3)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
-    "b.gt       1b                             \n"
-  : "+r"(src_y),     // %0
-    "+r"(src_u),     // %1
-    "+r"(src_v),     // %2
-    "+r"(dst_yuy2),  // %3
-    "+r"(width)      // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width) {
+  asm volatile(
+      "ld1         {v2.16b}, [%3]                \n"  // shuffler
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 4 pixels.
+      "subs        %w2, %w2, #4                  \n"  // 4 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v1.16b, {v0.16b}, v2.16b      \n"  // look up 4 pixels
+      "st1         {v1.16b}, [%1], #16           \n"  // store 4.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),                   // %0
+        "+r"(dst_argb),                   // %1
+        "+r"(width)                       // %2
+      : "r"(shuffler)                     // %3
+      : "cc", "memory", "v0", "v1", "v2"  // Clobber List
   );
 }
 
-void I422ToUYVYRow_NEON(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_uyvy, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld2        {v1.8b,v2.8b}, [%0], #16       \n"  // load 16 Ys
-    "orr        v3.8b, v2.8b, v2.8b            \n"
-    MEMACCESS(1)
-    "ld1        {v0.8b}, [%1], #8              \n"  // load 8 Us
-    MEMACCESS(2)
-    "ld1        {v2.8b}, [%2], #8              \n"  // load 8 Vs
-    "subs       %w4, %w4, #16                  \n"  // 16 pixels
-    MEMACCESS(3)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
-    "b.gt       1b                             \n"
-  : "+r"(src_y),     // %0
-    "+r"(src_u),     // %1
-    "+r"(src_v),     // %2
-    "+r"(dst_uyvy),  // %3
-    "+r"(width)      // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"
-  );
+void I422ToYUY2Row_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.8b, v1.8b}, [%0], #16     \n"  // load 16 Ys
+      "subs        %w4, %w4, #16                 \n"  // 16 pixels
+      "orr         v2.8b, v1.8b, v1.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ld1         {v1.8b}, [%1], #8             \n"         // load 8 Us
+      "ld1         {v3.8b}, [%2], #8             \n"         // load 8 Vs
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+      "b.gt        1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_yuy2),  // %3
+        "+r"(width)      // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
-void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGBTORGB565
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels RGB565.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_rgb565),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
-  );
+void I422ToUYVYRow_NEON(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v1.8b,v2.8b}, [%0], #16      \n"  // load 16 Ys
+      "orr         v3.8b, v2.8b, v2.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ld1         {v0.8b}, [%1], #8             \n"         // load 8 Us
+      "ld1         {v2.8b}, [%2], #8             \n"         // load 8 Vs
+      "subs        %w4, %w4, #16                 \n"         // 16 pixels
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+      "b.gt        1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(src_u),     // %1
+        "+r"(src_v),     // %2
+        "+r"(dst_uyvy),  // %3
+        "+r"(width)      // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
-void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width) {
-  asm volatile (
-    "dup        v1.4s, %w2                     \n"  // dither4
-  "1:                                          \n"
-    MEMACCESS(1)
-    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n"  // load 8 pixels
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "uqadd      v20.8b, v20.8b, v1.8b          \n"
-    "uqadd      v21.8b, v21.8b, v1.8b          \n"
-    "uqadd      v22.8b, v22.8b, v1.8b          \n"
-    ARGBTORGB565
-    MEMACCESS(0)
-    "st1        {v0.16b}, [%0], #16            \n"  // store 8 pixels RGB565.
-    "b.gt       1b                             \n"
-  : "+r"(dst_rgb)    // %0
-  : "r"(src_argb),   // %1
-    "r"(dither4),    // %2
-    "r"(width)       // %3
-  : "cc", "memory", "v0", "v1", "v20", "v21", "v22", "v23"
-  );
+void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
+                          uint8_t* dst_rgb565,
+                          int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGBTORGB565
+      "st1         {v18.16b}, [%1], #16          \n"  // store 8 pixels RGB565.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(dst_rgb565),  // %1
+        "+r"(width)        // %2
+      :
+      : "cc", "memory", "v16", "v17", "v18", "v19");
 }
 
-void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width) {
+  asm volatile(
+      "dup         v1.4s, %w2                    \n"  // dither4
+      "1:                                        \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqadd       v16.8b, v16.8b, v1.8b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v17.8b, v17.8b, v1.8b         \n"
+      "uqadd       v18.8b, v18.8b, v1.8b         \n" ARGBTORGB565
+      "st1         {v18.16b}, [%0], #16          \n"  // store 8 pixels RGB565.
+      "b.gt        1b                            \n"
+      : "+r"(dst_rgb)   // %0
+      : "r"(src_argb),  // %1
+        "r"(dither4),   // %2
+        "r"(width)      // %3
+      : "cc", "memory", "v1", "v16", "v17", "v18", "v19");
+}
+
+void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
+                            uint8_t* dst_argb1555,
                             int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGBTOARGB1555
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB1555.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb1555),  // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
-  );
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGBTOARGB1555
+      "st1         {v0.16b}, [%1], #16           \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),      // %0
+        "+r"(dst_argb1555),  // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "v0", "v16", "v17", "v18", "v19");
 }
 
-void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
+                            uint8_t* dst_argb4444,
                             int width) {
-  asm volatile (
-    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGBTOARGB4444
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB4444.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),      // %0
-    "+r"(dst_argb4444),  // %1
-    "+r"(width)            // %2
-  :
-  : "cc", "memory", "v0", "v1", "v4", "v20", "v21", "v22", "v23"
-  );
+  asm volatile(
+      "movi        v23.16b, #0x0f                \n"  // bits to clear with
+                                                      // vbic.
+      "1:                                        \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGBTOARGB4444
+      "st1         {v0.16b}, [%1], #16           \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),      // %0
+        "+r"(dst_argb4444),  // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v23");
 }
 
-void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
-    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
-    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
-    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-  );
+#if LIBYUV_USE_ST2
+void ARGBToAR64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ldp         q0, q2, [%0], #32             \n"  // load 8 pixels
+      "mov         v1.16b, v0.16b                \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mov         v3.16b, v2.16b                \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "st2         {v0.16b, v1.16b}, [%1], #32   \n"  // store 4 pixels
+      "st2         {v2.16b, v3.16b}, [%1], #32   \n"  // store 4 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
-void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load row 16 pixels
-    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "st1        {v3.16b}, [%1], #16            \n"  // store 16 A's.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_a),      // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
-  );
+static const uvec8 kShuffleARGBToABGR = {2,  1, 0, 3,  6,  5,  4,  7,
+                                         10, 9, 8, 11, 14, 13, 12, 15};
+
+void ARGBToAB64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
+  asm volatile(
+      "ldr         q4, [%3]                      \n"  // shuffler
+      "1:                                        \n"
+      "ldp         q0, q2, [%0], #32             \n"  // load 8 pixels
+      "tbl         v0.16b, {v0.16b}, v4.16b      \n"
+      "tbl         v2.16b, {v2.16b}, v4.16b      \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mov         v1.16b, v0.16b                \n"
+      "mov         v3.16b, v2.16b                \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "st2         {v0.16b, v1.16b}, [%1], #32   \n"  // store 4 pixels
+      "st2         {v2.16b, v3.16b}, [%1], #32   \n"  // store 4 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),           // %0
+        "+r"(dst_ab64),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleARGBToABGR)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+#else
+void ARGBToAR64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 8 ARGB pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "zip1        v2.16b, v0.16b, v0.16b        \n"
+      "zip2        v3.16b, v0.16b, v0.16b        \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v4.16b, v1.16b, v1.16b        \n"
+      "zip2        v5.16b, v1.16b, v1.16b        \n"
+      "st1         {v2.8h, v3.8h, v4.8h, v5.8h}, [%1], #64 \n"  // 8 AR64
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5");
 }
 
-void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #15                     \n"  // B * 0.11400 coefficient
-    "movi       v5.8b, #75                     \n"  // G * 0.58700 coefficient
-    "movi       v6.8b, #38                     \n"  // R * 0.29900 coefficient
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
-    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
-    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
-    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 15 bit to 8 bit Y
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+static const uvec8 kShuffleARGBToAB64[2] = {
+    {2, 2, 1, 1, 0, 0, 3, 3, 6, 6, 5, 5, 4, 4, 7, 7},
+    {10, 10, 9, 9, 8, 8, 11, 11, 14, 14, 13, 13, 12, 12, 15, 15}};
+
+void ARGBToAB64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
+  asm volatile(
+      "ldp         q6, q7, [%3]                  \n"  // 2 shufflers
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 8 pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "tbl         v2.16b, {v0.16b}, v6.16b      \n"  // ARGB to AB64
+      "tbl         v3.16b, {v0.16b}, v7.16b      \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v4.16b, {v1.16b}, v6.16b      \n"
+      "tbl         v5.16b, {v1.16b}, v7.16b      \n"
+      "st1         {v2.8h, v3.8h, v4.8h, v5.8h}, [%1], #64 \n"  // 8 AR64
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),              // %0
+        "+r"(dst_ab64),              // %1
+        "+r"(width)                  // %2
+      : "r"(&kShuffleARGBToAB64[0])  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+#endif  // LIBYUV_USE_ST2
+
+static const uvec8 kShuffleAR64ToARGB = {1,  3,  5,  7,  9,  11, 13, 15,
+                                         17, 19, 21, 23, 25, 27, 29, 31};
+
+void AR64ToARGBRow_NEON(const uint16_t* src_ar64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "ldr         q4, [%3]                      \n"  // shuffler
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 4 pixels
+      "ldp         q2, q3, [%0], #32             \n"  // load 4 pixels
+      "tbl         v0.16b, {v0.16b, v1.16b}, v4.16b \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v2.16b, {v2.16b, v3.16b}, v4.16b \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "stp         q0, q2, [%1], #32             \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ar64),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleAR64ToARGB)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+static const uvec8 kShuffleAB64ToARGB = {5,  3,  1,  7,  13, 11, 9,  15,
+                                         21, 19, 17, 23, 29, 27, 25, 31};
+
+void AB64ToARGBRow_NEON(const uint16_t* src_ab64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "ldr         q4, [%3]                      \n"  // shuffler
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 4 pixels
+      "ldp         q2, q3, [%0], #32             \n"  // load 4 pixels
+      "tbl         v0.16b, {v0.16b, v1.16b}, v4.16b \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v2.16b, {v2.16b, v3.16b}, v4.16b \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "stp         q0, q2, [%1], #32             \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ab64),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleAB64ToARGB)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "st1         {v3.16b}, [%1], #16           \n"  // store 16 A's.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_a),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
 // 8x1 pixels.
-void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+void ARGBToUV444Row_NEON(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
                          int width) {
-  asm volatile (
-    "movi       v24.8b, #112                   \n"  // UB / VR 0.875 coefficient
-    "movi       v25.8b, #74                    \n"  // UG -0.5781 coefficient
-    "movi       v26.8b, #38                    \n"  // UR -0.2969 coefficient
-    "movi       v27.8b, #18                    \n"  // VB -0.1406 coefficient
-    "movi       v28.8b, #94                    \n"  // VG -0.7344 coefficient
-    "movi       v29.16b,#0x80                  \n"  // 128.5
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
-    "umlsl      v4.8h, v1.8b, v25.8b           \n"  // G
-    "umlsl      v4.8h, v2.8b, v26.8b           \n"  // R
-    "add        v4.8h, v4.8h, v29.8h           \n"  // +128 -> unsigned
-
-    "umull      v3.8h, v2.8b, v24.8b           \n"  // R
-    "umlsl      v3.8h, v1.8b, v28.8b           \n"  // G
-    "umlsl      v3.8h, v0.8b, v27.8b           \n"  // B
-    "add        v3.8h, v3.8h, v29.8h           \n"  // +128 -> unsigned
-
-    "uqshrn     v0.8b, v4.8h, #8               \n"  // 16 bit to 8 bit U
-    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
-
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
-    MEMACCESS(2)
-    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
-    "v24", "v25", "v26", "v27", "v28", "v29"
-  );
+  asm volatile(
+      "movi        v24.8b, #112                  \n"  // UB / VR 0.875
+                                                      // coefficient
+      "movi        v25.8b, #74                   \n"  // UG -0.5781 coefficient
+      "movi        v26.8b, #38                   \n"  // UR -0.2969 coefficient
+      "movi        v27.8b, #18                   \n"  // VB -0.1406 coefficient
+      "movi        v28.8b, #94                   \n"  // VG -0.7344 coefficient
+      "movi        v29.16b,#0x80                 \n"  // 128.5
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v24.8b          \n"  // B
+      "umlsl       v4.8h, v1.8b, v25.8b          \n"  // G
+      "umlsl       v4.8h, v2.8b, v26.8b          \n"  // R
+      "prfm        pldl1keep, [%0, 448]          \n"
+
+      "umull       v3.8h, v2.8b, v24.8b          \n"  // R
+      "umlsl       v3.8h, v1.8b, v28.8b          \n"  // G
+      "umlsl       v3.8h, v0.8b, v27.8b          \n"  // B
+
+      "addhn       v0.8b, v4.8h, v29.8h          \n"  // +128 -> unsigned
+      "addhn       v1.8b, v3.8h, v29.8h          \n"  // +128 -> unsigned
+
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%2], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_u),     // %1
+        "+r"(dst_v),     // %2
+        "+r"(width)      // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26",
+        "v27", "v28", "v29");
 }
 
-#define RGBTOUV_SETUP_REG                                                      \
-    "movi       v20.8h, #56, lsl #0  \n"  /* UB/VR coefficient (0.875) / 2 */  \
-    "movi       v21.8h, #37, lsl #0  \n"  /* UG coefficient (-0.5781) / 2  */  \
-    "movi       v22.8h, #19, lsl #0  \n"  /* UR coefficient (-0.2969) / 2  */  \
-    "movi       v23.8h, #9,  lsl #0  \n"  /* VB coefficient (-0.1406) / 2  */  \
-    "movi       v24.8h, #47, lsl #0  \n"  /* VG coefficient (-0.7344) / 2  */  \
-    "movi       v25.16b, #0x80       \n"  /* 128.5 (0x8080 in 16-bit)      */
+#define RGBTOUV_SETUP_REG                                                  \
+  "movi       v20.8h, #56, lsl #0  \n" /* UB/VR coefficient (0.875) / 2 */ \
+  "movi       v21.8h, #37, lsl #0  \n" /* UG coefficient (-0.5781) / 2  */ \
+  "movi       v22.8h, #19, lsl #0  \n" /* UR coefficient (-0.2969) / 2  */ \
+  "movi       v23.8h, #9,  lsl #0  \n" /* VB coefficient (-0.1406) / 2  */ \
+  "movi       v24.8h, #47, lsl #0  \n" /* VG coefficient (-0.7344) / 2  */ \
+  "movi       v25.16b, #0x80       \n" /* 128.5 (0x8080 in 16-bit)      */
 
-// 32x1 pixels -> 8x1.  width is number of argb pixels. e.g. 32.
-void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
-                         int width) {
+// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
+// clang-format off
+#define RGBTOUV(QB, QG, QR)                                                 \
+  "mul        v3.8h, " #QB ",v20.8h          \n" /* B                    */ \
+  "mul        v4.8h, " #QR ",v20.8h          \n" /* R                    */ \
+  "mls        v3.8h, " #QG ",v21.8h          \n" /* G                    */ \
+  "mls        v4.8h, " #QG ",v24.8h          \n" /* G                    */ \
+  "mls        v3.8h, " #QR ",v22.8h          \n" /* R                    */ \
+  "mls        v4.8h, " #QB ",v23.8h          \n" /* B                    */ \
+  "addhn      v0.8b, v3.8h, v25.8h           \n" /* +128 -> unsigned     */ \
+  "addhn      v1.8b, v4.8h, v25.8h           \n" /* +128 -> unsigned     */
+// clang-format on
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+// TODO(fbarchard): consider ptrdiff_t for all strides.
+
+void ARGBToUVRow_NEON(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_argb_1 = src_argb + src_stride_argb;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(0)
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%0], #64 \n"  // load next 16.
-    "uaddlp     v4.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v5.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v6.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "addp       v0.8h, v0.8h, v4.8h            \n"  // B 16 shorts -> 8 shorts.
-    "addp       v1.8h, v1.8h, v5.8h            \n"  // G 16 shorts -> 8 shorts.
-    "addp       v2.8h, v2.8h, v6.8h            \n"  // R 16 shorts -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w3, %w3, #32                  \n"  // 32 processed per loop.
-    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
-    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
-    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
-    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
-    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
-    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
-    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
-    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
-    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
-    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
-    MEMACCESS(2)
-    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_argb),  // %0
-    "+r"(dst_u),     // %1
-    "+r"(dst_v),     // %2
-    "+r"(width)        // %3
+    "+r"(src_argb_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(width)        // %4
   :
   : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
     "v20", "v21", "v22", "v23", "v24", "v25"
   );
 }
 
-// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-#define RGBTOUV(QB, QG, QR) \
-    "mul        v3.8h, " #QB ",v20.8h          \n"  /* B                    */ \
-    "mul        v4.8h, " #QR ",v20.8h          \n"  /* R                    */ \
-    "mls        v3.8h, " #QG ",v21.8h          \n"  /* G                    */ \
-    "mls        v4.8h, " #QG ",v24.8h          \n"  /* G                    */ \
-    "mls        v3.8h, " #QR ",v22.8h          \n"  /* R                    */ \
-    "mls        v4.8h, " #QB ",v23.8h          \n"  /* B                    */ \
-    "add        v3.8h, v3.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
-    "add        v4.8h, v4.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
-    "uqshrn     v0.8b, v3.8h, #8               \n"  /* 16 bit to 8 bit U    */ \
-    "uqshrn     v1.8b, v4.8h, #8               \n"  /* 16 bit to 8 bit V    */
-
-// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
-// TODO(fbarchard): consider ptrdiff_t for all strides.
-
-void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_argb_1 = src_argb + src_stride_argb;
+// TODO(fbarchard): Subsample match Intel code.
+void ARGBToUVJRow_NEON(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  const uint8_t* src_argb_1 = src_argb + src_stride_argb;
   asm volatile (
-    RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    MEMACCESS(1)
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
-    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_argb_1),  // %1
     "+r"(dst_u),     // %2
@@ -1502,42 +2360,132 @@ void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
   );
 }
 
-// TODO(fbarchard): Subsample match C code.
-void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_argb_1 = src_argb + src_stride_argb;
+void ABGRToUVJRow_NEON(const uint8_t* src_abgr,
+                       int src_stride_abgr,
+                       uint8_t* dst_uj,
+                       uint8_t* dst_vj,
+                       int width) {
+  const uint8_t* src_abgr_1 = src_abgr + src_stride_abgr;
   asm volatile (
-    "movi       v20.8h, #63, lsl #0            \n"  // UB/VR coeff (0.500) / 2
-    "movi       v21.8h, #42, lsl #0            \n"  // UG coeff (-0.33126) / 2
-    "movi       v22.8h, #21, lsl #0            \n"  // UR coeff (-0.16874) / 2
-    "movi       v23.8h, #10, lsl #0            \n"  // VB coeff (-0.08131) / 2
-    "movi       v24.8h, #53, lsl #0            \n"  // VG coeff (-0.41869) / 2
-    "movi       v25.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64  \n"  // load next 16
-    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // B 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+    RGBTOUV(v2.8h, v1.8h, v0.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_abgr_1),  // %1
+    "+r"(dst_uj),     // %2
+    "+r"(dst_vj),     // %3
+    "+r"(width)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+
+void RGB24ToUVJRow_NEON(const uint8_t* src_rgb24,
+                        int src_stride_rgb24,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  const uint8_t* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
+  asm volatile (
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(src_argb_1),  // %1
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_rgb24_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(width)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+
+void RAWToUVJRow_NEON(const uint8_t* src_raw,
+                      int src_stride_raw,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_raw_1 = src_raw + src_stride_raw;
+  asm volatile (
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+    RGBTOUV(v2.8h, v1.8h, v0.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_raw_1),  // %1
     "+r"(dst_u),     // %2
     "+r"(dst_v),     // %3
     "+r"(width)        // %4
@@ -1547,34 +2495,35 @@ void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
   );
 }
 
-void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_bgra_1 = src_bgra + src_stride_bgra;
+void BGRAToUVRow_NEON(const uint8_t* src_bgra,
+                      int src_stride_bgra,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_bgra_1 = src_bgra + src_stride_bgra;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v3.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v3.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v1.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
-    "uadalp     v0.8h, v7.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v3.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v5.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v3.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v3.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v3.8h, v2.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v1.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
+      "uadalp      v0.8h, v7.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v3.8h, v6.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v5.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v3.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_bgra),  // %0
     "+r"(src_bgra_1),  // %1
     "+r"(dst_u),     // %2
@@ -1586,34 +2535,35 @@ void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
   );
 }
 
-void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_abgr_1 = src_abgr + src_stride_abgr;
+void ABGRToUVRow_NEON(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_abgr_1 = src_abgr + src_stride_abgr;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v3.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
-    "uadalp     v3.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v3.8h, #1               \n"  // 2x average
-    "urshr      v2.8h, v2.8h, #1               \n"
-    "urshr      v1.8h, v1.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v3.8h, v2.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v2.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v1.8h, v0.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+      "uadalp      v3.8h, v6.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v2.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v4.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v3.8h, #1              \n"  // 2x average
+      "urshr       v2.8h, v2.8h, #1              \n"
+      "urshr       v1.8h, v1.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v2.8h, v1.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_abgr),  // %0
     "+r"(src_abgr_1),  // %1
     "+r"(dst_u),     // %2
@@ -1625,34 +2575,35 @@ void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
   );
 }
 
-void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
-                      uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_rgba_1 = src_rgba + src_stride_rgba;
+void RGBAToUVRow_NEON(const uint8_t* src_rgba,
+                      int src_stride_rgba,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_rgba_1 = src_rgba + src_stride_rgba;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v1.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v3.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
-    "uadalp     v0.8h, v5.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v7.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v1.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v2.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v3.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+      "uadalp      v0.8h, v5.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v6.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v7.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_rgba),  // %0
     "+r"(src_rgba_1),  // %1
     "+r"(dst_u),     // %2
@@ -1664,34 +2615,35 @@ void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
   );
 }
 
-void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
-                       uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
+void RGB24ToUVRow_NEON(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  const uint8_t* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
-    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_rgb24),  // %0
     "+r"(src_rgb24_1),  // %1
     "+r"(dst_u),     // %2
@@ -1703,34 +2655,35 @@ void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
   );
 }
 
-void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
-                     uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_raw_1 = src_raw + src_stride_raw;
+void RAWToUVRow_NEON(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  const uint8_t* src_raw_1 = src_raw + src_stride_raw;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 8 RAW pixels.
-    "uaddlp     v2.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v0.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
-    "uadalp     v2.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v0.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v2.8h, v2.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v0.8h, v0.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 RAW pixels.
+      "uaddlp      v2.8h, v2.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v0.8h, v0.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
+      "uadalp      v2.8h, v6.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v0.8h, v4.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v2.8h, v2.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v0.8h, v0.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v2.8h, v1.8h, v0.8h)
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_raw),  // %0
     "+r"(src_raw_1),  // %1
     "+r"(dst_u),     // %2
@@ -1742,700 +2695,850 @@ void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
   );
 }
 
-// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
-                        uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_rgb565_1 = src_rgb565 + src_stride_rgb565;
-  asm volatile (
-    "movi       v22.8h, #56, lsl #0            \n"  // UB / VR coeff (0.875) / 2
-    "movi       v23.8h, #37, lsl #0            \n"  // UG coeff (-0.5781) / 2
-    "movi       v24.8h, #19, lsl #0            \n"  // UR coeff (-0.2969) / 2
-    "movi       v25.8h, #9 , lsl #0            \n"  // VB coeff (-0.1406) / 2
-    "movi       v26.8h, #47, lsl #0            \n"  // VG coeff (-0.7344) / 2
-    "movi       v27.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
-    RGB565TOARGB
-    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uaddlp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uaddlp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 RGB565 pixels.
-    RGB565TOARGB
-    "uaddlp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uaddlp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uaddlp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 RGB565 pixels.
-    RGB565TOARGB
-    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uadalp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uadalp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 RGB565 pixels.
-    RGB565TOARGB
-    "uadalp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uadalp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uadalp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-    "ins        v16.D[1], v17.D[0]             \n"
-    "ins        v18.D[1], v19.D[0]             \n"
-    "ins        v20.D[1], v21.D[0]             \n"
-
-    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
-    "urshr      v5.8h, v18.8h, #1              \n"
-    "urshr      v6.8h, v20.8h, #1              \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
-    "mul        v16.8h, v4.8h, v22.8h          \n"  // B
-    "mls        v16.8h, v5.8h, v23.8h          \n"  // G
-    "mls        v16.8h, v6.8h, v24.8h          \n"  // R
-    "add        v16.8h, v16.8h, v27.8h         \n"  // +128 -> unsigned
-    "mul        v17.8h, v6.8h, v22.8h          \n"  // R
-    "mls        v17.8h, v5.8h, v26.8h          \n"  // G
-    "mls        v17.8h, v4.8h, v25.8h          \n"  // B
-    "add        v17.8h, v17.8h, v27.8h         \n"  // +128 -> unsigned
-    "uqshrn     v0.8b, v16.8h, #8              \n"  // 16 bit to 8 bit U
-    "uqshrn     v1.8b, v17.8h, #8              \n"  // 16 bit to 8 bit V
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
-  : "+r"(src_rgb565),  // %0
-    "+r"(src_rgb565_1),  // %1
-    "+r"(dst_u),     // %2
-    "+r"(dst_v),     // %3
-    "+r"(width)        // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
-    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24",
-    "v25", "v26", "v27"
-  );
+// 16x2 pixels -> 8x1.  width is number of rgb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  const uint8_t* src_rgb565_1 = src_rgb565 + src_stride_rgb565;
+  asm volatile(
+      RGBTOUV_SETUP_REG
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 RGB565 pixels.
+      RGB565TOARGB
+      "uaddlp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%0], #16           \n"  // next 8 RGB565 pixels.
+      RGB565TOARGB
+      "uaddlp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uaddlp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ld1         {v0.16b}, [%1], #16           \n"  // load 8 RGB565 pixels.
+      RGB565TOARGB
+      "uadalp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%1], #16           \n"  // next 8 RGB565 pixels.
+      RGB565TOARGB
+      "uadalp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uadalp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ins         v16.D[1], v26.D[0]            \n"
+      "ins         v17.D[1], v27.D[0]            \n"
+      "ins         v18.D[1], v28.D[0]            \n"
+
+      "urshr       v0.8h, v16.8h, #1             \n"  // 2x average
+      "urshr       v1.8h, v17.8h, #1             \n"
+      "urshr       v2.8h, v18.8h, #1             \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+      RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb565),    // %0
+        "+r"(src_rgb565_1),  // %1
+        "+r"(dst_u),           // %2
+        "+r"(dst_v),           // %3
+        "+r"(width)            // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17",
+        "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27",
+        "v28");
 }
 
 // 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
-                        uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_argb1555_1 = src_argb1555 + src_stride_argb1555;
-  asm volatile (
-    RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB1555 pixels.
-    RGB555TOARGB
-    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-    "ins        v16.D[1], v26.D[0]             \n"
-    "ins        v17.D[1], v27.D[0]             \n"
-    "ins        v18.D[1], v28.D[0]             \n"
-
-    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
-    "urshr      v5.8h, v17.8h, #1              \n"
-    "urshr      v6.8h, v18.8h, #1              \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
-    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
-    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
-    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
-    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
-    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
-    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
-    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
-    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
-    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
-    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb1555),  // %0
-    "+r"(src_argb1555_1),  // %1
-    "+r"(dst_u),     // %2
-    "+r"(dst_v),     // %3
-    "+r"(width)        // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
-    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
-    "v26", "v27", "v28"
-  );
+void ARGB1555ToUVRow_NEON(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  const uint8_t* src_argb1555_1 = src_argb1555 + src_stride_argb1555;
+  asm volatile(
+      RGBTOUV_SETUP_REG
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "uaddlp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%0], #16           \n"  // next 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "uaddlp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uaddlp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ld1         {v0.16b}, [%1], #16           \n"  // load 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "uadalp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%1], #16           \n"  // next 8 ARGB1555 pixels.
+      RGB555TOARGB
+      "uadalp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uadalp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ins         v16.D[1], v26.D[0]            \n"
+      "ins         v17.D[1], v27.D[0]            \n"
+      "ins         v18.D[1], v28.D[0]            \n"
+
+      "urshr       v0.8h, v16.8h, #1             \n"  // 2x average
+      "urshr       v1.8h, v17.8h, #1             \n"
+      "urshr       v2.8h, v18.8h, #1             \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+      RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb1555),    // %0
+        "+r"(src_argb1555_1),  // %1
+        "+r"(dst_u),           // %2
+        "+r"(dst_v),           // %3
+        "+r"(width)            // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17",
+        "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27",
+        "v28");
 }
 
 // 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
-void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
-                          uint8* dst_u, uint8* dst_v, int width) {
-  const uint8* src_argb4444_1 = src_argb4444 + src_stride_argb4444;
-  asm volatile (
-    RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB4444 pixels.
-    ARGB4444TOARGB
-    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-    "ins        v16.D[1], v26.D[0]             \n"
-    "ins        v17.D[1], v27.D[0]             \n"
-    "ins        v18.D[1], v28.D[0]             \n"
-
-    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
-    "urshr      v5.8h, v17.8h, #1              \n"
-    "urshr      v6.8h, v18.8h, #1              \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
-    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
-    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
-    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
-    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
-    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
-    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
-    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
-    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
-    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
-    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    MEMACCESS(3)
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb4444),  // %0
-    "+r"(src_argb4444_1),  // %1
-    "+r"(dst_u),     // %2
-    "+r"(dst_v),     // %3
-    "+r"(width)        // %4
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
-    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
-    "v26", "v27", "v28"
+void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
+                          int src_stride_argb4444,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  const uint8_t* src_argb4444_1 = src_argb4444 + src_stride_argb4444;
+  asm volatile(
+      RGBTOUV_SETUP_REG  // sets v20-v25
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "uaddlp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%0], #16           \n"  // next 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "uaddlp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uaddlp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ld1         {v0.16b}, [%1], #16           \n"  // load 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "uadalp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%1], #16           \n"  // next 8 ARGB4444 pixels.
+      ARGB4444TOARGB
+      "uadalp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uadalp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ins         v16.D[1], v26.D[0]            \n"
+      "ins         v17.D[1], v27.D[0]            \n"
+      "ins         v18.D[1], v28.D[0]            \n"
+
+      "urshr       v0.8h, v16.8h, #1             \n"  // 2x average
+      "urshr       v1.8h, v17.8h, #1             \n"
+      "urshr       v2.8h, v18.8h, #1             \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+      RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb4444),    // %0
+        "+r"(src_argb4444_1),  // %1
+        "+r"(dst_u),           // %2
+        "+r"(dst_v),           // %3
+        "+r"(width)            // %4
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17",
+        "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27",
+        "v28"
 
   );
 }
 
-void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
-    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
-    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
-    "movi       v27.8b, #16                    \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    RGB565TOARGB
-    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
-    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
-    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
-    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v27.8b           \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_rgb565),  // %0
-    "+r"(dst_y),       // %1
-    "+r"(width)          // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6",
-    "v24", "v25", "v26", "v27"
-  );
+void RGB565ToYRow_NEON(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movi        v24.8b, #25                   \n"  // B * 0.1016 coefficient
+      "movi        v25.8b, #129                  \n"  // G * 0.5078 coefficient
+      "movi        v26.8b, #66                   \n"  // R * 0.2578 coefficient
+      "movi        v27.8b, #16                   \n"  // Add 16 constant
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 RGB565 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      RGB565TOARGB
+      "umull       v3.8h, v0.8b, v24.8b          \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v3.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v3.8h, v2.8b, v26.8b          \n"  // R
+      "uqrshrn     v0.8b, v3.8h, #8              \n"  // 16 bit to 8 bit Y
+      "uqadd       v0.8b, v0.8b, v27.8b          \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb565),  // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6", "v24", "v25", "v26",
+        "v27");
 }
 
-void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGB1555TOARGB
-    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
-    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
-    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
-    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb1555),  // %0
-    "+r"(dst_y),         // %1
-    "+r"(width)            // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-  );
+void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width) {
+  asm volatile(
+      "movi        v4.8b, #25                    \n"  // B * 0.1016 coefficient
+      "movi        v5.8b, #129                   \n"  // G * 0.5078 coefficient
+      "movi        v6.8b, #66                    \n"  // R * 0.2578 coefficient
+      "movi        v7.8b, #16                    \n"  // Add 16 constant
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB1555 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      ARGB1555TOARGB
+      "umull       v3.8h, v0.8b, v4.8b           \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v3.8h, v1.8b, v5.8b           \n"  // G
+      "umlal       v3.8h, v2.8b, v6.8b           \n"  // R
+      "uqrshrn     v0.8b, v3.8h, #8              \n"  // 16 bit to 8 bit Y
+      "uqadd       v0.8b, v0.8b, v7.8b           \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb1555),  // %0
+        "+r"(dst_y),         // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
 }
 
-void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
-    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
-    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
-    "movi       v27.8b, #16                    \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    ARGB4444TOARGB
-    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
-    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
-    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
-    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v27.8b           \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb4444),  // %0
-    "+r"(dst_y),         // %1
-    "+r"(width)            // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27"
-  );
+void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
+                         uint8_t* dst_y,
+                         int width) {
+  asm volatile(
+      "movi        v24.8b, #25                   \n"  // B * 0.1016 coefficient
+      "movi        v25.8b, #129                  \n"  // G * 0.5078 coefficient
+      "movi        v26.8b, #66                   \n"  // R * 0.2578 coefficient
+      "movi        v27.8b, #16                   \n"  // Add 16 constant
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB4444 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      ARGB4444TOARGB
+      "umull       v3.8h, v0.8b, v24.8b          \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v3.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v3.8h, v2.8b, v26.8b          \n"  // R
+      "uqrshrn     v0.8b, v3.8h, #8              \n"  // 16 bit to 8 bit Y
+      "uqadd       v0.8b, v0.8b, v27.8b          \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb4444),  // %0
+        "+r"(dst_y),         // %1
+        "+r"(width)          // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27");
 }
 
-void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v16.8h, v1.8b, v4.8b           \n"  // R
-    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
-    "umlal      v16.8h, v3.8b, v6.8b           \n"  // B
-    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_bgra),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
-  );
+struct RgbConstants {
+  uint8_t kRGBToY[4];
+  uint16_t kAddY;
+  uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+                                                        128,
+                                                        0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+                                                        0x1080,
+                                                        0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+                                                      0x1080,
+                                                      0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+void ARGBToYMatrixRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "ldr         d0, [%3]                      \n"  // load rgbconstants
+      "dup         v6.16b, v0.b[0]               \n"
+      "dup         v7.16b, v0.b[1]               \n"
+      "dup         v16.16b, v0.b[2]              \n"
+      "dup         v17.8h,  v0.h[2]              \n"
+      "1:                                        \n"
+      "ld4         {v2.16b,v3.16b,v4.16b,v5.16b}, [%0], #64 \n"  // load 16
+                                                                 // pixels.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "umull       v0.8h, v2.8b, v6.8b           \n"  // B
+      "umull2      v1.8h, v2.16b, v6.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v0.8h, v3.8b, v7.8b           \n"  // G
+      "umlal2      v1.8h, v3.16b, v7.16b         \n"
+      "umlal       v0.8h, v4.8b, v16.8b          \n"  // R
+      "umlal2      v1.8h, v4.16b, v16.16b        \n"
+      "addhn       v0.8b, v0.8h, v17.8h          \n"  // 16 bit to 8 bit Y
+      "addhn       v1.8b, v1.8h, v17.8h          \n"
+      "st1         {v0.8b, v1.8b}, [%1], #16     \n"  // store 16 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17");
 }
 
-void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v16.8h, v0.8b, v4.8b           \n"  // R
-    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
-    "umlal      v16.8h, v2.8b, v6.8b           \n"  // B
-    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_abgr),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
-  );
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_y, width, &kRgb24I601Constants);
 }
 
-void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v16.8h, v1.8b, v4.8b           \n"  // B
-    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
-    "umlal      v16.8h, v3.8b, v6.8b           \n"  // R
-    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_rgba),  // %0
-    "+r"(dst_y),     // %1
-    "+r"(width)        // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
-  );
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_yj, width, &kRgb24JPEGConstants);
 }
 
-void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
-    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
-    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
-    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_rgb24),  // %0
-    "+r"(dst_y),      // %1
-    "+r"(width)         // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
-  );
+void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_abgr, dst_y, width, &kRawI601Constants);
 }
 
-void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int width) {
-  asm volatile (
-    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
-    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
-    "movi       v7.8b, #16                     \n"  // Add 16 constant
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
-    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
-    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
-    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-    "uqadd      v0.8b, v0.8b, v7.8b            \n"
-    MEMACCESS(1)
-    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-    "b.gt       1b                             \n"
-  : "+r"(src_raw),  // %0
-    "+r"(dst_y),    // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
-  );
+void ABGRToYJRow_NEON(const uint8_t* src_abgr, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_NEON(src_abgr, dst_yj, width, &kRawJPEGConstants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+void RGBAToYMatrixRow_NEON(const uint8_t* src_rgba,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "ldr         d0, [%3]                      \n"  // load rgbconstants
+      "dup         v6.16b, v0.b[0]               \n"
+      "dup         v7.16b, v0.b[1]               \n"
+      "dup         v16.16b, v0.b[2]              \n"
+      "dup         v17.8h,  v0.h[2]              \n"
+      "1:                                        \n"
+      "ld4         {v1.16b,v2.16b,v3.16b,v4.16b}, [%0], #64 \n"  // load 16
+                                                                 // pixels.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "umull       v0.8h, v2.8b, v6.8b           \n"  // B
+      "umull2      v1.8h, v2.16b, v6.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v0.8h, v3.8b, v7.8b           \n"  // G
+      "umlal2      v1.8h, v3.16b, v7.16b         \n"
+      "umlal       v0.8h, v4.8b, v16.8b          \n"  // R
+      "umlal2      v1.8h, v4.16b, v16.16b        \n"
+      "addhn       v0.8b, v0.8h, v17.8h          \n"  // 16 bit to 8 bit Y
+      "addhn       v1.8b, v1.8h, v17.8h          \n"
+      "st1         {v0.8b, v1.8b}, [%1], #16     \n"  // store 16 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_rgba),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17");
+}
+
+void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+void RGBToYMatrixRow_NEON(const uint8_t* src_rgb,
+                          uint8_t* dst_y,
+                          int width,
+                          const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "ldr         d0, [%3]                      \n"  // load rgbconstants
+      "dup         v5.16b, v0.b[0]               \n"
+      "dup         v6.16b, v0.b[1]               \n"
+      "dup         v7.16b, v0.b[2]               \n"
+      "dup         v16.8h,  v0.h[2]              \n"
+      "1:                                        \n"
+      "ld3         {v2.16b,v3.16b,v4.16b}, [%0], #48 \n"  // load 16 pixels.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "umull       v0.8h, v2.8b, v5.8b           \n"  // B
+      "umull2      v1.8h, v2.16b, v5.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v0.8h, v3.8b, v6.8b           \n"  // G
+      "umlal2      v1.8h, v3.16b, v6.16b         \n"
+      "umlal       v0.8h, v4.8b, v7.8b           \n"  // R
+      "umlal2      v1.8h, v4.16b, v7.16b         \n"
+      "addhn       v0.8b, v0.8h, v16.8h          \n"  // 16 bit to 8 bit Y
+      "addhn       v1.8b, v1.8h, v16.8h          \n"
+      "st1         {v0.8b, v1.8b}, [%1], #16     \n"  // store 16 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb),     // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+}
+
+void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_NEON(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+  RGBToYMatrixRow_NEON(src_raw, dst_y, width, &kRawI601Constants);
 }
 
 // Bilinear filter 16x2 -> 16x1
-void InterpolateRow_NEON(uint8* dst_ptr,
-                         const uint8* src_ptr, ptrdiff_t src_stride,
-                         int dst_width, int source_y_fraction) {
+void InterpolateRow_NEON(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         int dst_width,
+                         int source_y_fraction) {
   int y1_fraction = source_y_fraction;
   int y0_fraction = 256 - y1_fraction;
-  const uint8* src_ptr1 = src_ptr + src_stride;
-  asm volatile (
-    "cmp        %w4, #0                        \n"
-    "b.eq       100f                           \n"
-    "cmp        %w4, #128                      \n"
-    "b.eq       50f                            \n"
-
-    "dup        v5.16b, %w4                    \n"
-    "dup        v4.16b, %w5                    \n"
-    // General purpose row blend.
-  "1:                                          \n"
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"
-    MEMACCESS(2)
-    "ld1        {v1.16b}, [%2], #16            \n"
-    "subs       %w3, %w3, #16                  \n"
-    "umull      v2.8h, v0.8b,  v4.8b           \n"
-    "umull2     v3.8h, v0.16b, v4.16b          \n"
-    "umlal      v2.8h, v1.8b,  v5.8b           \n"
-    "umlal2     v3.8h, v1.16b, v5.16b          \n"
-    "rshrn      v0.8b,  v2.8h, #8              \n"
-    "rshrn2     v0.16b, v3.8h, #8              \n"
-    MEMACCESS(0)
-    "st1        {v0.16b}, [%0], #16            \n"
-    "b.gt       1b                             \n"
-    "b          99f                            \n"
-
-    // Blend 50 / 50.
-  "50:                                         \n"
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"
-    MEMACCESS(2)
-    "ld1        {v1.16b}, [%2], #16            \n"
-    "subs       %w3, %w3, #16                  \n"
-    "urhadd     v0.16b, v0.16b, v1.16b         \n"
-    MEMACCESS(0)
-    "st1        {v0.16b}, [%0], #16            \n"
-    "b.gt       50b                            \n"
-    "b          99f                            \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-  "100:                                        \n"
-    MEMACCESS(1)
-    "ld1        {v0.16b}, [%1], #16            \n"
-    "subs       %w3, %w3, #16                  \n"
-    MEMACCESS(0)
-    "st1        {v0.16b}, [%0], #16            \n"
-    "b.gt       100b                           \n"
-
-  "99:                                         \n"
-  : "+r"(dst_ptr),          // %0
-    "+r"(src_ptr),          // %1
-    "+r"(src_ptr1),         // %2
-    "+r"(dst_width),        // %3
-    "+r"(y1_fraction),      // %4
-    "+r"(y0_fraction)       // %5
-  :
-  : "cc", "memory", "v0", "v1", "v3", "v4", "v5"
-  );
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  asm volatile(
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+
+      "dup         v5.16b, %w4                   \n"
+      "dup         v4.16b, %w5                   \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "umull       v2.8h, v0.8b,  v4.8b          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull2      v3.8h, v0.16b, v4.16b         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal       v2.8h, v1.8b,  v5.8b          \n"
+      "umlal2      v3.8h, v1.16b, v5.16b         \n"
+      "rshrn       v0.8b,  v2.8h, #8             \n"
+      "rshrn2      v0.16b, v3.8h, #8             \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),      // %0
+        "+r"(src_ptr),      // %1
+        "+r"(src_ptr1),     // %2
+        "+r"(dst_width),    // %3
+        "+r"(y1_fraction),  // %4
+        "+r"(y0_fraction)   // %5
+      :
+      : "cc", "memory", "v0", "v1", "v3", "v4", "v5");
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRow_16_NEON(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int dst_width,
+                            int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+
+  asm volatile(
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+
+      "dup         v5.8h, %w4                    \n"
+      "dup         v4.8h, %w5                    \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "umull       v2.4s, v0.4h, v4.4h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull2      v3.4s, v0.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal       v2.4s, v1.4h, v5.4h           \n"
+      "umlal2      v3.4s, v1.8h, v5.8h           \n"
+      "rshrn       v0.4h, v2.4s, #8              \n"
+      "rshrn2      v0.8h, v3.4s, #8              \n"
+      "st1         {v0.8h}, [%0], #16            \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "urhadd      v0.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.8h}, [%0], #16            \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v0.8h}, [%0], #16            \n"
+      "b.gt        100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_ptr1),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(y1_fraction),  // %4
+        "r"(y0_fraction)   // %5
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5");
+}
+
+// Bilinear filter 8x2 -> 8x1
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+void InterpolateRow_16To8_NEON(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int dst_width,
+                               int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  int shift = 15 - __builtin_clz((int32_t)scale);  // Negative shl is shr
+
+  asm volatile(
+      "dup         v6.8h, %w6                    \n"
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+
+      "dup         v5.8h, %w4                    \n"
+      "dup         v4.8h, %w5                    \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "umull       v2.4s, v0.4h, v4.4h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull2      v3.4s, v0.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal       v2.4s, v1.4h, v5.4h           \n"
+      "umlal2      v3.4s, v1.8h, v5.8h           \n"
+      "rshrn       v0.4h, v2.4s, #8              \n"
+      "rshrn2      v0.8h, v3.4s, #8              \n"
+      "ushl        v0.8h, v0.8h, v6.8h           \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%0], #8             \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "urhadd      v0.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "ushl        v0.8h, v0.8h, v6.8h           \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%0], #8             \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "ldr         q0, [%1], #16                 \n"
+      "ushl        v0.8h, v0.8h, v2.8h           \n"  // shr = v2 is negative
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "str         d0, [%0], #8                  \n"  // store 8 pixels
+      "b.gt        100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_ptr1),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(y1_fraction),  // %4
+        "r"(y0_fraction),  // %5
+        "r"(shift)         // %6
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
 }
 
 // dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
-void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                       uint8* dst_argb, int width) {
-  asm volatile (
-    "subs       %w3, %w3, #8                   \n"
-    "b.lt       89f                            \n"
-    // Blend 8 pixels.
-  "8:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0 pixels
-    MEMACCESS(1)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1 pixels
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
-    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
-    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
-    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
-    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
-    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
-    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
-    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
-    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
-    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
-    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
-    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
-    "movi       v3.8b, #255                    \n"  // a = 255
-    MEMACCESS(2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
-    "b.ge       8b                             \n"
-
-  "89:                                         \n"
-    "adds       %w3, %w3, #8-1                 \n"
-    "b.lt       99f                            \n"
-
-    // Blend 1 pixels.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel ARGB0.
-    MEMACCESS(1)
-    "ld4        {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel ARGB1.
-    "subs       %w3, %w3, #1                   \n"  // 1 processed per loop.
-    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
-    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
-    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
-    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
-    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
-    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
-    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
-    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
-    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
-    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
-    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
-    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
-    "movi       v3.8b, #255                    \n"  // a = 255
-    MEMACCESS(2)
-    "st4        {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
-    "b.ge       1b                             \n"
-
-  "99:                                         \n"
-
-  : "+r"(src_argb0),    // %0
-    "+r"(src_argb1),    // %1
-    "+r"(dst_argb),     // %2
-    "+r"(width)         // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
-    "v16", "v17", "v18"
-  );
+void ARGBBlendRow_NEON(const uint8_t* src_argb,
+                       const uint8_t* src_argb1,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "subs        %w3, %w3, #8                  \n"
+      "b.lt        89f                           \n"
+      // Blend 8 pixels.
+      "8:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "umull       v16.8h, v4.8b, v3.8b          \n"  // db * a
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v17.8h, v5.8b, v3.8b          \n"  // dg * a
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull       v18.8h, v6.8b, v3.8b          \n"  // dr * a
+      "uqrshrn     v16.8b, v16.8h, #8            \n"  // db >>= 8
+      "uqrshrn     v17.8b, v17.8h, #8            \n"  // dg >>= 8
+      "uqrshrn     v18.8b, v18.8h, #8            \n"  // dr >>= 8
+      "uqsub       v4.8b, v4.8b, v16.8b          \n"  // db - (db * a / 256)
+      "uqsub       v5.8b, v5.8b, v17.8b          \n"  // dg - (dg * a / 256)
+      "uqsub       v6.8b, v6.8b, v18.8b          \n"  // dr - (dr * a / 256)
+      "uqadd       v0.8b, v0.8b, v4.8b           \n"  // + sb
+      "uqadd       v1.8b, v1.8b, v5.8b           \n"  // + sg
+      "uqadd       v2.8b, v2.8b, v6.8b           \n"  // + sr
+      "movi        v3.8b, #255                   \n"  // a = 255
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+                                                             // pixels
+      "b.ge        8b                            \n"
+
+      "89:                                       \n"
+      "adds        %w3, %w3, #8-1                \n"
+      "b.lt        99f                           \n"
+
+      // Blend 1 pixels.
+      "1:                                        \n"
+      "ld4         {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel
+                                                           // ARGB0.
+      "ld4         {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel
+                                                           // ARGB1.
+      "subs        %w3, %w3, #1                  \n"  // 1 processed per loop.
+      "umull       v16.8h, v4.8b, v3.8b          \n"  // db * a
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v17.8h, v5.8b, v3.8b          \n"  // dg * a
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull       v18.8h, v6.8b, v3.8b          \n"  // dr * a
+      "uqrshrn     v16.8b, v16.8h, #8            \n"  // db >>= 8
+      "uqrshrn     v17.8b, v17.8h, #8            \n"  // dg >>= 8
+      "uqrshrn     v18.8b, v18.8h, #8            \n"  // dr >>= 8
+      "uqsub       v4.8b, v4.8b, v16.8b          \n"  // db - (db * a / 256)
+      "uqsub       v5.8b, v5.8b, v17.8b          \n"  // dg - (dg * a / 256)
+      "uqsub       v6.8b, v6.8b, v18.8b          \n"  // dr - (dr * a / 256)
+      "uqadd       v0.8b, v0.8b, v4.8b           \n"  // + sb
+      "uqadd       v1.8b, v1.8b, v5.8b           \n"  // + sg
+      "uqadd       v2.8b, v2.8b, v6.8b           \n"  // + sr
+      "movi        v3.8b, #255                   \n"  // a = 255
+      "st4         {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
+      "b.ge        1b                            \n"
+
+      "99:                                       \n"
+
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17", "v18");
 }
 
 // Attenuate 8 pixels at a time.
-void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    // Attenuate 8 pixels.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v4.8h, v0.8b, v3.8b            \n"  // b * a
-    "umull      v5.8h, v1.8b, v3.8b            \n"  // g * a
-    "umull      v6.8h, v2.8b, v3.8b            \n"  // r * a
-    "uqrshrn    v0.8b, v4.8h, #8               \n"  // b >>= 8
-    "uqrshrn    v1.8b, v5.8h, #8               \n"  // g >>= 8
-    "uqrshrn    v2.8b, v6.8h, #8               \n"  // r >>= 8
-    MEMACCESS(1)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)       // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
-  );
+void ARGBAttenuateRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width) {
+  asm volatile(
+      // Attenuate 8 pixels.
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v3.8b           \n"  // b * a
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v5.8h, v1.8b, v3.8b           \n"         // g * a
+      "umull       v6.8h, v2.8b, v3.8b           \n"         // r * a
+      "uqrshrn     v0.8b, v4.8h, #8              \n"         // b >>= 8
+      "uqrshrn     v1.8b, v5.8h, #8              \n"         // g >>= 8
+      "uqrshrn     v2.8b, v6.8h, #8              \n"         // r >>= 8
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
 }
 
 // Quantize 8 ARGB pixels (32 bytes).
 // dst = (dst * scale >> 16) * interval_size + interval_offset;
-void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
-                          int interval_offset, int width) {
-  asm volatile (
-    "dup        v4.8h, %w2                     \n"
-    "ushr       v4.8h, v4.8h, #1               \n"  // scale >>= 1
-    "dup        v5.8h, %w3                     \n"  // interval multiply.
-    "dup        v6.8h, %w4                     \n"  // interval add
-
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0]  \n"  // load 8 pixels of ARGB.
-    "subs       %w1, %w1, #8                   \n"  // 8 processed per loop.
-    "uxtl       v0.8h, v0.8b                   \n"  // b (0 .. 255)
-    "uxtl       v1.8h, v1.8b                   \n"
-    "uxtl       v2.8h, v2.8b                   \n"
-    "sqdmulh    v0.8h, v0.8h, v4.8h            \n"  // b * scale
-    "sqdmulh    v1.8h, v1.8h, v4.8h            \n"  // g
-    "sqdmulh    v2.8h, v2.8h, v4.8h            \n"  // r
-    "mul        v0.8h, v0.8h, v5.8h            \n"  // b * interval_size
-    "mul        v1.8h, v1.8h, v5.8h            \n"  // g
-    "mul        v2.8h, v2.8h, v5.8h            \n"  // r
-    "add        v0.8h, v0.8h, v6.8h            \n"  // b + interval_offset
-    "add        v1.8h, v1.8h, v6.8h            \n"  // g
-    "add        v2.8h, v2.8h, v6.8h            \n"  // r
-    "uqxtn      v0.8b, v0.8h                   \n"
-    "uqxtn      v1.8b, v1.8h                   \n"
-    "uqxtn      v2.8b, v2.8h                   \n"
-    MEMACCESS(0)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(dst_argb),       // %0
-    "+r"(width)           // %1
-  : "r"(scale),           // %2
-    "r"(interval_size),   // %3
-    "r"(interval_offset)  // %4
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
-  );
+void ARGBQuantizeRow_NEON(uint8_t* dst_argb,
+                          int scale,
+                          int interval_size,
+                          int interval_offset,
+                          int width) {
+  asm volatile(
+      "dup         v4.8h, %w2                    \n"
+      "ushr        v4.8h, v4.8h, #1              \n"  // scale >>= 1
+      "dup         v5.8h, %w3                    \n"  // interval multiply.
+      "dup         v6.8h, %w4                    \n"  // interval add
+
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8  ARGB.
+      "subs        %w1, %w1, #8                  \n"    // 8 processed per loop.
+      "uxtl        v0.8h, v0.8b                  \n"    // b (0 .. 255)
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v1.8h, v1.8b                  \n"
+      "uxtl        v2.8h, v2.8b                  \n"
+      "sqdmulh     v0.8h, v0.8h, v4.8h           \n"  // b * scale
+      "sqdmulh     v1.8h, v1.8h, v4.8h           \n"  // g
+      "sqdmulh     v2.8h, v2.8h, v4.8h           \n"  // r
+      "mul         v0.8h, v0.8h, v5.8h           \n"  // b * interval_size
+      "mul         v1.8h, v1.8h, v5.8h           \n"  // g
+      "mul         v2.8h, v2.8h, v5.8h           \n"  // r
+      "add         v0.8h, v0.8h, v6.8h           \n"  // b + interval_offset
+      "add         v1.8h, v1.8h, v6.8h           \n"  // g
+      "add         v2.8h, v2.8h, v6.8h           \n"  // r
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "uqxtn       v1.8b, v1.8h                  \n"
+      "uqxtn       v2.8b, v2.8h                  \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(dst_argb),       // %0
+        "+r"(width)           // %1
+      : "r"(scale),           // %2
+        "r"(interval_size),   // %3
+        "r"(interval_offset)  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
 }
 
 // Shade 8 pixels at a time by specified value.
 // NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
 // Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
-void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value) {
-  asm volatile (
-    "dup        v0.4s, %w3                     \n"  // duplicate scale value.
-    "zip1       v0.8b, v0.8b, v0.8b            \n"  // v0.8b aarrggbb.
-    "ushr       v0.8h, v0.8h, #1               \n"  // scale / 2.
-
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "uxtl       v4.8h, v4.8b                   \n"  // b (0 .. 255)
-    "uxtl       v5.8h, v5.8b                   \n"
-    "uxtl       v6.8h, v6.8b                   \n"
-    "uxtl       v7.8h, v7.8b                   \n"
-    "sqrdmulh   v4.8h, v4.8h, v0.h[0]          \n"  // b * scale * 2
-    "sqrdmulh   v5.8h, v5.8h, v0.h[1]          \n"  // g
-    "sqrdmulh   v6.8h, v6.8h, v0.h[2]          \n"  // r
-    "sqrdmulh   v7.8h, v7.8h, v0.h[3]          \n"  // a
-    "uqxtn      v4.8b, v4.8h                   \n"
-    "uqxtn      v5.8b, v5.8h                   \n"
-    "uqxtn      v6.8b, v6.8h                   \n"
-    "uqxtn      v7.8b, v7.8h                   \n"
-    MEMACCESS(1)
-    "st4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),       // %0
-    "+r"(dst_argb),       // %1
-    "+r"(width)           // %2
-  : "r"(value)            // %3
-  : "cc", "memory", "v0", "v4", "v5", "v6", "v7"
-  );
+void ARGBShadeRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value) {
+  asm volatile(
+      "dup         v0.4s, %w3                    \n"  // duplicate scale value.
+      "zip1        v0.8b, v0.8b, v0.8b           \n"  // v0.8b aarrggbb.
+      "ushr        v0.8h, v0.8h, #1              \n"  // scale / 2.
+
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "uxtl        v4.8h, v4.8b                  \n"  // b (0 .. 255)
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v5.8h, v5.8b                  \n"
+      "uxtl        v6.8h, v6.8b                  \n"
+      "uxtl        v7.8h, v7.8b                  \n"
+      "sqrdmulh    v4.8h, v4.8h, v0.h[0]         \n"  // b * scale * 2
+      "sqrdmulh    v5.8h, v5.8h, v0.h[1]         \n"  // g
+      "sqrdmulh    v6.8h, v6.8h, v0.h[2]         \n"  // r
+      "sqrdmulh    v7.8h, v7.8h, v0.h[3]         \n"  // a
+      "uqxtn       v4.8b, v4.8h                  \n"
+      "uqxtn       v5.8b, v5.8h                  \n"
+      "uqxtn       v6.8b, v6.8h                  \n"
+      "uqxtn       v7.8b, v7.8h                  \n"
+      "st4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(value)       // %3
+      : "cc", "memory", "v0", "v4", "v5", "v6", "v7");
 }
 
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 // Similar to ARGBToYJ but stores ARGB.
-// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
-void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v24.8b, #15                    \n"  // B * 0.11400 coefficient
-    "movi       v25.8b, #75                    \n"  // G * 0.58700 coefficient
-    "movi       v26.8b, #38                    \n"  // R * 0.29900 coefficient
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
-    "umlal      v4.8h, v1.8b, v25.8b           \n"  // G
-    "umlal      v4.8h, v2.8b, v26.8b           \n"  // R
-    "sqrshrun   v0.8b, v4.8h, #7               \n"  // 15 bit to 8 bit B
-    "orr        v1.8b, v0.8b, v0.8b            \n"  // G
-    "orr        v2.8b, v0.8b, v0.8b            \n"  // R
-    MEMACCESS(1)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(width)      // %2
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26"
-  );
+// C code is (29 * b + 150 * g + 77 * r + 128) >> 8;
+void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movi        v24.8b, #29                   \n"  // B * 0.1140 coefficient
+      "movi        v25.8b, #150                  \n"  // G * 0.5870 coefficient
+      "movi        v26.8b, #77                   \n"  // R * 0.2990 coefficient
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v24.8b          \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v4.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v4.8h, v2.8b, v26.8b          \n"  // R
+      "uqrshrn     v0.8b, v4.8h, #8              \n"  // 16 bit to 8 bit B
+      "orr         v1.8b, v0.8b, v0.8b           \n"  // G
+      "orr         v2.8b, v0.8b, v0.8b           \n"  // R
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26");
 }
 
 // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
@@ -2443,194 +3546,188 @@ void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
 //    g = (r * 45 + g * 88 + b * 22) >> 7
 //    r = (r * 50 + g * 98 + b * 24) >> 7
 
-void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v20.8b, #17                    \n"  // BB coefficient
-    "movi       v21.8b, #68                    \n"  // BG coefficient
-    "movi       v22.8b, #35                    \n"  // BR coefficient
-    "movi       v24.8b, #22                    \n"  // GB coefficient
-    "movi       v25.8b, #88                    \n"  // GG coefficient
-    "movi       v26.8b, #45                    \n"  // GR coefficient
-    "movi       v28.8b, #24                    \n"  // BB coefficient
-    "movi       v29.8b, #98                    \n"  // BG coefficient
-    "movi       v30.8b, #50                    \n"  // BR coefficient
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
-    "subs       %w1, %w1, #8                   \n"  // 8 processed per loop.
-    "umull      v4.8h, v0.8b, v20.8b           \n"  // B to Sepia B
-    "umlal      v4.8h, v1.8b, v21.8b           \n"  // G
-    "umlal      v4.8h, v2.8b, v22.8b           \n"  // R
-    "umull      v5.8h, v0.8b, v24.8b           \n"  // B to Sepia G
-    "umlal      v5.8h, v1.8b, v25.8b           \n"  // G
-    "umlal      v5.8h, v2.8b, v26.8b           \n"  // R
-    "umull      v6.8h, v0.8b, v28.8b           \n"  // B to Sepia R
-    "umlal      v6.8h, v1.8b, v29.8b           \n"  // G
-    "umlal      v6.8h, v2.8b, v30.8b           \n"  // R
-    "uqshrn     v0.8b, v4.8h, #7               \n"  // 16 bit to 8 bit B
-    "uqshrn     v1.8b, v5.8h, #7               \n"  // 16 bit to 8 bit G
-    "uqshrn     v2.8b, v6.8h, #7               \n"  // 16 bit to 8 bit R
-    MEMACCESS(0)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
-    "b.gt       1b                             \n"
-  : "+r"(dst_argb),  // %0
-    "+r"(width)      // %1
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
-    "v20", "v21", "v22", "v24", "v25", "v26", "v28", "v29", "v30"
-  );
+void ARGBSepiaRow_NEON(uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movi        v20.8b, #17                   \n"  // BB coefficient
+      "movi        v21.8b, #68                   \n"  // BG coefficient
+      "movi        v22.8b, #35                   \n"  // BR coefficient
+      "movi        v24.8b, #22                   \n"  // GB coefficient
+      "movi        v25.8b, #88                   \n"  // GG coefficient
+      "movi        v26.8b, #45                   \n"  // GR coefficient
+      "movi        v28.8b, #24                   \n"  // BB coefficient
+      "movi        v29.8b, #98                   \n"  // BG coefficient
+      "movi        v30.8b, #50                   \n"  // BR coefficient
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
+      "subs        %w1, %w1, #8                  \n"    // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v20.8b          \n"    // B to Sepia B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v4.8h, v1.8b, v21.8b          \n"  // G
+      "umlal       v4.8h, v2.8b, v22.8b          \n"  // R
+      "umull       v5.8h, v0.8b, v24.8b          \n"  // B to Sepia G
+      "umlal       v5.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v5.8h, v2.8b, v26.8b          \n"  // R
+      "umull       v6.8h, v0.8b, v28.8b          \n"  // B to Sepia R
+      "umlal       v6.8h, v1.8b, v29.8b          \n"  // G
+      "umlal       v6.8h, v2.8b, v30.8b          \n"  // R
+      "uqshrn      v0.8b, v4.8h, #7              \n"  // 16 bit to 8 bit B
+      "uqshrn      v1.8b, v5.8h, #7              \n"  // 16 bit to 8 bit G
+      "uqshrn      v2.8b, v6.8h, #7              \n"  // 16 bit to 8 bit R
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
+      "b.gt        1b                            \n"
+      : "+r"(dst_argb),  // %0
+        "+r"(width)      // %1
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+        "v21", "v22", "v24", "v25", "v26", "v28", "v29", "v30");
 }
 
 // Tranform 8 ARGB pixels (32 bytes) with color matrix.
 // TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
 // needs to saturate.  Consider doing a non-saturating version.
-void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
-                             const int8* matrix_argb, int width) {
-  asm volatile (
-    MEMACCESS(3)
-    "ld1        {v2.16b}, [%3]                 \n"  // load 3 ARGB vectors.
-    "sxtl       v0.8h, v2.8b                   \n"  // B,G coefficients s16.
-    "sxtl2      v1.8h, v2.16b                  \n"  // R,A coefficients s16.
-
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 pixels.
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "uxtl       v16.8h, v16.8b                 \n"  // b (0 .. 255) 16 bit
-    "uxtl       v17.8h, v17.8b                 \n"  // g
-    "uxtl       v18.8h, v18.8b                 \n"  // r
-    "uxtl       v19.8h, v19.8b                 \n"  // a
-    "mul        v22.8h, v16.8h, v0.h[0]        \n"  // B = B * Matrix B
-    "mul        v23.8h, v16.8h, v0.h[4]        \n"  // G = B * Matrix G
-    "mul        v24.8h, v16.8h, v1.h[0]        \n"  // R = B * Matrix R
-    "mul        v25.8h, v16.8h, v1.h[4]        \n"  // A = B * Matrix A
-    "mul        v4.8h, v17.8h, v0.h[1]         \n"  // B += G * Matrix B
-    "mul        v5.8h, v17.8h, v0.h[5]         \n"  // G += G * Matrix G
-    "mul        v6.8h, v17.8h, v1.h[1]         \n"  // R += G * Matrix R
-    "mul        v7.8h, v17.8h, v1.h[5]         \n"  // A += G * Matrix A
-    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
-    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
-    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
-    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
-    "mul        v4.8h, v18.8h, v0.h[2]         \n"  // B += R * Matrix B
-    "mul        v5.8h, v18.8h, v0.h[6]         \n"  // G += R * Matrix G
-    "mul        v6.8h, v18.8h, v1.h[2]         \n"  // R += R * Matrix R
-    "mul        v7.8h, v18.8h, v1.h[6]         \n"  // A += R * Matrix A
-    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
-    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
-    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
-    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
-    "mul        v4.8h, v19.8h, v0.h[3]         \n"  // B += A * Matrix B
-    "mul        v5.8h, v19.8h, v0.h[7]         \n"  // G += A * Matrix G
-    "mul        v6.8h, v19.8h, v1.h[3]         \n"  // R += A * Matrix R
-    "mul        v7.8h, v19.8h, v1.h[7]         \n"  // A += A * Matrix A
-    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
-    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
-    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
-    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
-    "sqshrun    v16.8b, v22.8h, #6             \n"  // 16 bit to 8 bit B
-    "sqshrun    v17.8b, v23.8h, #6             \n"  // 16 bit to 8 bit G
-    "sqshrun    v18.8b, v24.8h, #6             \n"  // 16 bit to 8 bit R
-    "sqshrun    v19.8b, v25.8h, #6             \n"  // 16 bit to 8 bit A
-    MEMACCESS(1)
-    "st4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 pixels.
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_argb),   // %1
-    "+r"(width)       // %2
-  : "r"(matrix_argb)  // %3
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
-    "v18", "v19", "v22", "v23", "v24", "v25"
-  );
+void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
+                             uint8_t* dst_argb,
+                             const int8_t* matrix_argb,
+                             int width) {
+  asm volatile(
+      "ld1         {v2.16b}, [%3]                \n"  // load 3 ARGB vectors.
+      "sxtl        v0.8h, v2.8b                  \n"  // B,G coefficients s16.
+      "sxtl2       v1.8h, v2.16b                 \n"  // R,A coefficients s16.
+
+      "1:                                        \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "uxtl        v16.8h, v16.8b                \n"  // b (0 .. 255) 16 bit
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v17.8h, v17.8b                \n"  // g
+      "uxtl        v18.8h, v18.8b                \n"  // r
+      "uxtl        v19.8h, v19.8b                \n"  // a
+      "mul         v22.8h, v16.8h, v0.h[0]       \n"  // B = B * Matrix B
+      "mul         v23.8h, v16.8h, v0.h[4]       \n"  // G = B * Matrix G
+      "mul         v24.8h, v16.8h, v1.h[0]       \n"  // R = B * Matrix R
+      "mul         v25.8h, v16.8h, v1.h[4]       \n"  // A = B * Matrix A
+      "mul         v4.8h, v17.8h, v0.h[1]        \n"  // B += G * Matrix B
+      "mul         v5.8h, v17.8h, v0.h[5]        \n"  // G += G * Matrix G
+      "mul         v6.8h, v17.8h, v1.h[1]        \n"  // R += G * Matrix R
+      "mul         v7.8h, v17.8h, v1.h[5]        \n"  // A += G * Matrix A
+      "sqadd       v22.8h, v22.8h, v4.8h         \n"  // Accumulate B
+      "sqadd       v23.8h, v23.8h, v5.8h         \n"  // Accumulate G
+      "sqadd       v24.8h, v24.8h, v6.8h         \n"  // Accumulate R
+      "sqadd       v25.8h, v25.8h, v7.8h         \n"  // Accumulate A
+      "mul         v4.8h, v18.8h, v0.h[2]        \n"  // B += R * Matrix B
+      "mul         v5.8h, v18.8h, v0.h[6]        \n"  // G += R * Matrix G
+      "mul         v6.8h, v18.8h, v1.h[2]        \n"  // R += R * Matrix R
+      "mul         v7.8h, v18.8h, v1.h[6]        \n"  // A += R * Matrix A
+      "sqadd       v22.8h, v22.8h, v4.8h         \n"  // Accumulate B
+      "sqadd       v23.8h, v23.8h, v5.8h         \n"  // Accumulate G
+      "sqadd       v24.8h, v24.8h, v6.8h         \n"  // Accumulate R
+      "sqadd       v25.8h, v25.8h, v7.8h         \n"  // Accumulate A
+      "mul         v4.8h, v19.8h, v0.h[3]        \n"  // B += A * Matrix B
+      "mul         v5.8h, v19.8h, v0.h[7]        \n"  // G += A * Matrix G
+      "mul         v6.8h, v19.8h, v1.h[3]        \n"  // R += A * Matrix R
+      "mul         v7.8h, v19.8h, v1.h[7]        \n"  // A += A * Matrix A
+      "sqadd       v22.8h, v22.8h, v4.8h         \n"  // Accumulate B
+      "sqadd       v23.8h, v23.8h, v5.8h         \n"  // Accumulate G
+      "sqadd       v24.8h, v24.8h, v6.8h         \n"  // Accumulate R
+      "sqadd       v25.8h, v25.8h, v7.8h         \n"  // Accumulate A
+      "sqshrun     v16.8b, v22.8h, #6            \n"  // 16 bit to 8 bit B
+      "sqshrun     v17.8b, v23.8h, #6            \n"  // 16 bit to 8 bit G
+      "sqshrun     v18.8b, v24.8h, #6            \n"  // 16 bit to 8 bit R
+      "sqshrun     v19.8b, v25.8h, #6            \n"  // 16 bit to 8 bit A
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(dst_argb),   // %1
+        "+r"(width)       // %2
+      : "r"(matrix_argb)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17", "v18", "v19", "v22", "v23", "v24", "v25");
 }
 
 // TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    MEMACCESS(1)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "umull      v0.8h, v0.8b, v4.8b            \n"  // multiply B
-    "umull      v1.8h, v1.8b, v5.8b            \n"  // multiply G
-    "umull      v2.8h, v2.8b, v6.8b            \n"  // multiply R
-    "umull      v3.8h, v3.8b, v7.8b            \n"  // multiply A
-    "rshrn      v0.8b, v0.8h, #8               \n"  // 16 bit to 8 bit B
-    "rshrn      v1.8b, v1.8h, #8               \n"  // 16 bit to 8 bit G
-    "rshrn      v2.8b, v2.8h, #8               \n"  // 16 bit to 8 bit R
-    "rshrn      v3.8b, v3.8h, #8               \n"  // 16 bit to 8 bit A
-    MEMACCESS(2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-  );
+void ARGBMultiplyRow_NEON(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "umull       v0.8h, v0.8b, v4.8b           \n"  // multiply B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v1.8h, v1.8b, v5.8b           \n"  // multiply G
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull       v2.8h, v2.8b, v6.8b           \n"  // multiply R
+      "umull       v3.8h, v3.8b, v7.8b           \n"  // multiply A
+      "rshrn       v0.8b, v0.8h, #8              \n"  // 16 bit to 8 bit B
+      "rshrn       v1.8b, v1.8h, #8              \n"  // 16 bit to 8 bit G
+      "rshrn       v2.8b, v2.8h, #8              \n"  // 16 bit to 8 bit R
+      "rshrn       v3.8b, v3.8h, #8              \n"  // 16 bit to 8 bit A
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
 }
 
 // Add 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    MEMACCESS(1)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "uqadd      v0.8b, v0.8b, v4.8b            \n"
-    "uqadd      v1.8b, v1.8b, v5.8b            \n"
-    "uqadd      v2.8b, v2.8b, v6.8b            \n"
-    "uqadd      v3.8b, v3.8b, v7.8b            \n"
-    MEMACCESS(2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-  );
+void ARGBAddRow_NEON(const uint8_t* src_argb,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqadd       v0.8b, v0.8b, v4.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v1.8b, v1.8b, v5.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqadd       v2.8b, v2.8b, v6.8b           \n"
+      "uqadd       v3.8b, v3.8b, v7.8b           \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
 }
 
 // Subtract 2 rows of ARGB pixels, 8 pixels at a time.
-void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
-  asm volatile (
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
-    MEMACCESS(1)
-    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "uqsub      v0.8b, v0.8b, v4.8b            \n"
-    "uqsub      v1.8b, v1.8b, v5.8b            \n"
-    "uqsub      v2.8b, v2.8b, v6.8b            \n"
-    "uqsub      v3.8b, v3.8b, v7.8b            \n"
-    MEMACCESS(2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-
-  : "+r"(src_argb0),  // %0
-    "+r"(src_argb1),  // %1
-    "+r"(dst_argb),   // %2
-    "+r"(width)       // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-  );
+void ARGBSubtractRow_NEON(const uint8_t* src_argb,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  asm volatile(
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqsub       v0.8b, v0.8b, v4.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqsub       v1.8b, v1.8b, v5.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqsub       v2.8b, v2.8b, v6.8b           \n"
+      "uqsub       v3.8b, v3.8b, v7.8b           \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
+        "+r"(src_argb1),  // %1
+        "+r"(dst_argb),   // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
 }
 
 // Adds Sobel X and Sobel Y and stores Sobel into ARGB.
@@ -2638,54 +3735,54 @@ void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
 // R = Sobel
 // G = Sobel
 // B = Sobel
-void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v3.8b, #255                    \n"  // alpha
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.8b}, [%0], #8              \n"  // load 8 sobelx.
-    MEMACCESS(1)
-    "ld1        {v1.8b}, [%1], #8              \n"  // load 8 sobely.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "uqadd      v0.8b, v0.8b, v1.8b            \n"  // add
-    "orr        v1.8b, v0.8b, v0.8b            \n"
-    "orr        v2.8b, v0.8b, v0.8b            \n"
-    MEMACCESS(2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"
-  );
+void SobelRow_NEON(const uint8_t* src_sobelx,
+                   const uint8_t* src_sobely,
+                   uint8_t* dst_argb,
+                   int width) {
+  asm volatile(
+      "movi        v3.8b, #255                   \n"  // alpha
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld1         {v0.8b}, [%0], #8             \n"  // load 8 sobelx.
+      "ld1         {v1.8b}, [%1], #8             \n"  // load 8 sobely.
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqadd       v0.8b, v0.8b, v1.8b           \n"  // add
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v1.8b, v0.8b, v0.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "orr         v2.8b, v0.8b, v0.8b           \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
 // Adds Sobel X and Sobel Y and stores Sobel into plane.
-void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_y, int width) {
-  asm volatile (
-    // 16 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b}, [%0], #16            \n"  // load 16 sobelx.
-    MEMACCESS(1)
-    "ld1        {v1.16b}, [%1], #16            \n"  // load 16 sobely.
-    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
-    "uqadd      v0.16b, v0.16b, v1.16b         \n"  // add
-    MEMACCESS(2)
-    "st1        {v0.16b}, [%2], #16            \n"  // store 16 pixels.
-    "b.gt       1b                             \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_y),       // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "v0", "v1"
-  );
+void SobelToPlaneRow_NEON(const uint8_t* src_sobelx,
+                          const uint8_t* src_sobely,
+                          uint8_t* dst_y,
+                          int width) {
+  asm volatile(
+      // 16 pixel loop.
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 16 sobelx.
+      "ld1         {v1.16b}, [%1], #16           \n"  // load 16 sobely.
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v0.16b, v0.16b, v1.16b        \n"  // add
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v0.16b}, [%2], #16           \n"  // store 16 pixels.
+      "b.gt        1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_y),       // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "v0", "v1");
 }
 
 // Mixes Sobel X, Sobel Y and Sobel into ARGB.
@@ -2693,70 +3790,69 @@ void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
 // R = Sobel X
 // G = Sobel
 // B = Sobel Y
-void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) {
-  asm volatile (
-    "movi       v3.8b, #255                    \n"  // alpha
-    // 8 pixel loop.
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v2.8b}, [%0], #8              \n"  // load 8 sobelx.
-    MEMACCESS(1)
-    "ld1        {v0.8b}, [%1], #8              \n"  // load 8 sobely.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "uqadd      v1.8b, v0.8b, v2.8b            \n"  // add
-    MEMACCESS(2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_sobelx),  // %0
-    "+r"(src_sobely),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(width)        // %3
-  :
-  : "cc", "memory", "v0", "v1", "v2", "v3"
-  );
+void SobelXYRow_NEON(const uint8_t* src_sobelx,
+                     const uint8_t* src_sobely,
+                     uint8_t* dst_argb,
+                     int width) {
+  asm volatile(
+      "movi        v3.8b, #255                   \n"  // alpha
+      // 8 pixel loop.
+      "1:                                        \n"
+      "ld1         {v2.8b}, [%0], #8             \n"  // load 8 sobelx.
+      "ld1         {v0.8b}, [%1], #8             \n"  // load 8 sobely.
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v1.8b, v0.8b, v2.8b           \n"  // add
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_sobelx),  // %0
+        "+r"(src_sobely),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
 // SobelX as a matrix is
 // -1  0  1
 // -2  0  2
 // -1  0  1
-void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobelx, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.8b}, [%0],%5               \n"  // top
-    MEMACCESS(0)
-    "ld1        {v1.8b}, [%0],%6               \n"
-    "usubl      v0.8h, v0.8b, v1.8b            \n"
-    MEMACCESS(1)
-    "ld1        {v2.8b}, [%1],%5               \n"  // center * 2
-    MEMACCESS(1)
-    "ld1        {v3.8b}, [%1],%6               \n"
-    "usubl      v1.8h, v2.8b, v3.8b            \n"
-    "add        v0.8h, v0.8h, v1.8h            \n"
-    "add        v0.8h, v0.8h, v1.8h            \n"
-    MEMACCESS(2)
-    "ld1        {v2.8b}, [%2],%5               \n"  // bottom
-    MEMACCESS(2)
-    "ld1        {v3.8b}, [%2],%6               \n"
-    "subs       %w4, %w4, #8                   \n"  // 8 pixels
-    "usubl      v1.8h, v2.8b, v3.8b            \n"
-    "add        v0.8h, v0.8h, v1.8h            \n"
-    "abs        v0.8h, v0.8h                   \n"
-    "uqxtn      v0.8b, v0.8h                   \n"
-    MEMACCESS(3)
-    "st1        {v0.8b}, [%3], #8              \n"  // store 8 sobelx
-    "b.gt       1b                             \n"
-  : "+r"(src_y0),      // %0
-    "+r"(src_y1),      // %1
-    "+r"(src_y2),      // %2
-    "+r"(dst_sobelx),  // %3
-    "+r"(width)        // %4
-  : "r"(2LL),          // %5
-    "r"(6LL)           // %6
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+void SobelXRow_NEON(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    const uint8_t* src_y2,
+                    uint8_t* dst_sobelx,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.8b}, [%0],%5              \n"  // top
+      "ld1         {v1.8b}, [%0],%6              \n"
+      "usubl       v0.8h, v0.8b, v1.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ld1         {v2.8b}, [%1],%5              \n"  // center * 2
+      "ld1         {v3.8b}, [%1],%6              \n"
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "ld1         {v2.8b}, [%2],%5              \n"  // bottom
+      "ld1         {v3.8b}, [%2],%6              \n"
+      "subs        %w4, %w4, #8                  \n"  // 8 pixels
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "abs         v0.8h, v0.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%3], #8             \n"  // store 8 sobelx
+      "b.gt        1b                            \n"
+      : "+r"(src_y0),                           // %0
+        "+r"(src_y1),                           // %1
+        "+r"(src_y2),                           // %2
+        "+r"(dst_sobelx),                       // %3
+        "+r"(width)                             // %4
+      : "r"(2LL),                               // %5
+        "r"(6LL)                                // %6
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
@@ -2764,43 +3860,658 @@ void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
 // -1 -2 -1
 //  0  0  0
 //  1  2  1
-void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.8b}, [%0],%4               \n"  // left
-    MEMACCESS(1)
-    "ld1        {v1.8b}, [%1],%4               \n"
-    "usubl      v0.8h, v0.8b, v1.8b            \n"
-    MEMACCESS(0)
-    "ld1        {v2.8b}, [%0],%4               \n"  // center * 2
-    MEMACCESS(1)
-    "ld1        {v3.8b}, [%1],%4               \n"
-    "usubl      v1.8h, v2.8b, v3.8b            \n"
-    "add        v0.8h, v0.8h, v1.8h            \n"
-    "add        v0.8h, v0.8h, v1.8h            \n"
-    MEMACCESS(0)
-    "ld1        {v2.8b}, [%0],%5               \n"  // right
-    MEMACCESS(1)
-    "ld1        {v3.8b}, [%1],%5               \n"
-    "subs       %w3, %w3, #8                   \n"  // 8 pixels
-    "usubl      v1.8h, v2.8b, v3.8b            \n"
-    "add        v0.8h, v0.8h, v1.8h            \n"
-    "abs        v0.8h, v0.8h                   \n"
-    "uqxtn      v0.8b, v0.8h                   \n"
-    MEMACCESS(2)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 sobely
-    "b.gt       1b                             \n"
-  : "+r"(src_y0),      // %0
-    "+r"(src_y1),      // %1
-    "+r"(dst_sobely),  // %2
-    "+r"(width)        // %3
-  : "r"(1LL),          // %4
-    "r"(6LL)           // %5
-  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+void SobelYRow_NEON(const uint8_t* src_y0,
+                    const uint8_t* src_y1,
+                    uint8_t* dst_sobely,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.8b}, [%0],%4              \n"  // left
+      "ld1         {v1.8b}, [%1],%4              \n"
+      "usubl       v0.8h, v0.8b, v1.8b           \n"
+      "ld1         {v2.8b}, [%0],%4              \n"  // center * 2
+      "ld1         {v3.8b}, [%1],%4              \n"
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "ld1         {v2.8b}, [%0],%5              \n"  // right
+      "ld1         {v3.8b}, [%1],%5              \n"
+      "subs        %w3, %w3, #8                  \n"  // 8 pixels
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "abs         v0.8h, v0.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 sobely
+      "b.gt        1b                            \n"
+      : "+r"(src_y0),                           // %0
+        "+r"(src_y1),                           // %1
+        "+r"(dst_sobely),                       // %2
+        "+r"(width)                             // %3
+      : "r"(1LL),                               // %4
+        "r"(6LL)                                // %5
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
+
+// Caveat - rounds float to half float whereas scaling version truncates.
+void HalfFloat1Row_NEON(const uint16_t* src,
+                        uint16_t* dst,
+                        float /*unused*/,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v1.16b}, [%0], #16           \n"  // load 8 shorts
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop
+      "uxtl        v2.4s, v1.4h                  \n"  // 8 int's
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl2       v3.4s, v1.8h                  \n"
+      "scvtf       v2.4s, v2.4s                  \n"  // 8 floats
+      "scvtf       v3.4s, v3.4s                  \n"
+      "fcvtn       v1.4h, v2.4s                  \n"  // 8 half floats
+      "fcvtn2      v1.8h, v3.4s                  \n"
+      "st1         {v1.16b}, [%1], #16           \n"  // store 8 shorts
+      "b.gt        1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      :
+      : "cc", "memory", "v1", "v2", "v3");
+}
+
+void HalfFloatRow_NEON(const uint16_t* src,
+                       uint16_t* dst,
+                       float scale,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v1.16b}, [%0], #16           \n"  // load 8 shorts
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop
+      "uxtl        v2.4s, v1.4h                  \n"  // 8 int's
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl2       v3.4s, v1.8h                  \n"
+      "scvtf       v2.4s, v2.4s                  \n"  // 8 floats
+      "scvtf       v3.4s, v3.4s                  \n"
+      "fmul        v2.4s, v2.4s, %3.s[0]         \n"  // adjust exponent
+      "fmul        v3.4s, v3.4s, %3.s[0]         \n"
+      "uqshrn      v1.4h, v2.4s, #13             \n"  // isolate halffloat
+      "uqshrn2     v1.8h, v3.4s, #13             \n"
+      "st1         {v1.16b}, [%1], #16           \n"  // store 8 shorts
+      "b.gt        1b                            \n"
+      : "+r"(src),                      // %0
+        "+r"(dst),                      // %1
+        "+r"(width)                     // %2
+      : "w"(scale * 1.9259299444e-34f)  // %3
+      : "cc", "memory", "v1", "v2", "v3");
+}
+
+void ByteToFloatRow_NEON(const uint8_t* src,
+                         float* dst,
+                         float scale,
+                         int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v1.8b}, [%0], #8             \n"  // load 8 bytes
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop
+      "uxtl        v1.8h, v1.8b                  \n"  // 8 shorts
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v2.4s, v1.4h                  \n"  // 8 ints
+      "uxtl2       v3.4s, v1.8h                  \n"
+      "scvtf       v2.4s, v2.4s                  \n"  // 8 floats
+      "scvtf       v3.4s, v3.4s                  \n"
+      "fmul        v2.4s, v2.4s, %3.s[0]         \n"  // scale
+      "fmul        v3.4s, v3.4s, %3.s[0]         \n"
+      "st1         {v2.16b, v3.16b}, [%1], #32   \n"  // store 8 floats
+      "b.gt        1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      : "w"(scale)   // %3
+      : "cc", "memory", "v1", "v2", "v3");
+}
+
+float ScaleMaxSamples_NEON(const float* src,
+                           float* dst,
+                           float scale,
+                           int width) {
+  float fmax;
+  asm volatile(
+      "movi        v5.4s, #0                     \n"  // max
+      "movi        v6.4s, #0                     \n"
+
+      "1:                                        \n"
+      "ld1         {v1.4s, v2.4s}, [%0], #32     \n"  // load 8 samples
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "fmul        v3.4s, v1.4s, %4.s[0]         \n"  // scale
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmul        v4.4s, v2.4s, %4.s[0]         \n"  // scale
+      "fmax        v5.4s, v5.4s, v1.4s           \n"  // max
+      "fmax        v6.4s, v6.4s, v2.4s           \n"
+      "st1         {v3.4s, v4.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      "fmax        v5.4s, v5.4s, v6.4s           \n"  // max
+      "fmaxv       %s3, v5.4s                    \n"  // signed max acculator
+      : "+r"(src),                                    // %0
+        "+r"(dst),                                    // %1
+        "+r"(width),                                  // %2
+        "=w"(fmax)                                    // %3
+      : "w"(scale)                                    // %4
+      : "cc", "memory", "v1", "v2", "v3", "v4", "v5", "v6");
+  return fmax;
+}
+
+float ScaleSumSamples_NEON(const float* src,
+                           float* dst,
+                           float scale,
+                           int width) {
+  float fsum;
+  asm volatile(
+      "movi        v5.4s, #0                     \n"  // max
+      "movi        v6.4s, #0                     \n"  // max
+
+      "1:                                        \n"
+      "ld1         {v1.4s, v2.4s}, [%0], #32     \n"  // load 8 samples
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "fmul        v3.4s, v1.4s, %4.s[0]         \n"  // scale
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmul        v4.4s, v2.4s, %4.s[0]         \n"
+      "fmla        v5.4s, v1.4s, v1.4s           \n"  // sum of squares
+      "fmla        v6.4s, v2.4s, v2.4s           \n"
+      "st1         {v3.4s, v4.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      "faddp       v5.4s, v5.4s, v6.4s           \n"
+      "faddp       v5.4s, v5.4s, v5.4s           \n"
+      "faddp       %3.4s, v5.4s, v5.4s           \n"  // sum
+      : "+r"(src),                                    // %0
+        "+r"(dst),                                    // %1
+        "+r"(width),                                  // %2
+        "=w"(fsum)                                    // %3
+      : "w"(scale)                                    // %4
+      : "cc", "memory", "v1", "v2", "v3", "v4", "v5", "v6");
+  return fsum;
+}
+
+void ScaleSamples_NEON(const float* src, float* dst, float scale, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v1.4s, v2.4s}, [%0], #32     \n"  // load 8 samples
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "fmul        v1.4s, v1.4s, %3.s[0]         \n"  // scale
+      "fmul        v2.4s, v2.4s, %3.s[0]         \n"  // scale
+      "st1         {v1.4s, v2.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(dst),   // %1
+        "+r"(width)  // %2
+      : "w"(scale)   // %3
+      : "cc", "memory", "v1", "v2");
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_NEON(const uint16_t* src0,
+                   const uint16_t* src1,
+                   const uint16_t* src2,
+                   const uint16_t* src3,
+                   const uint16_t* src4,
+                   uint32_t* dst,
+                   int width) {
+  asm volatile(
+      "movi        v6.8h, #4                     \n"  // constant 4
+      "movi        v7.8h, #6                     \n"  // constant 6
+
+      "1:                                        \n"
+      "ld1         {v1.8h}, [%0], #16            \n"  // load 8 samples, 5 rows
+      "ld1         {v2.8h}, [%4], #16            \n"
+      "uaddl       v0.4s, v1.4h, v2.4h           \n"  // * 1
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddl2      v1.4s, v1.8h, v2.8h           \n"  // * 1
+      "ld1         {v2.8h}, [%1], #16            \n"
+      "umlal       v0.4s, v2.4h, v6.4h           \n"  // * 4
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umlal2      v1.4s, v2.8h, v6.8h           \n"  // * 4
+      "ld1         {v2.8h}, [%2], #16            \n"
+      "umlal       v0.4s, v2.4h, v7.4h           \n"  // * 6
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal2      v1.4s, v2.8h, v7.8h           \n"  // * 6
+      "ld1         {v2.8h}, [%3], #16            \n"
+      "umlal       v0.4s, v2.4h, v6.4h           \n"  // * 4
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "umlal2      v1.4s, v2.8h, v6.8h           \n"  // * 4
+      "subs        %w6, %w6, #8                  \n"  // 8 processed per loop
+      "st1         {v0.4s,v1.4s}, [%5], #32      \n"  // store 8 samples
+      "prfm        pldl1keep, [%4, 448]          \n"
+      "b.gt        1b                            \n"
+      : "+r"(src0),  // %0
+        "+r"(src1),  // %1
+        "+r"(src2),  // %2
+        "+r"(src3),  // %3
+        "+r"(src4),  // %4
+        "+r"(dst),   // %5
+        "+r"(width)  // %6
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v6", "v7");
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width) {
+  const uint32_t* src1 = src + 1;
+  const uint32_t* src2 = src + 2;
+  const uint32_t* src3 = src + 3;
+  asm volatile(
+      "movi        v6.4s, #4                     \n"  // constant 4
+      "movi        v7.4s, #6                     \n"  // constant 6
+
+      "1:                                        \n"
+      "ld1         {v0.4s,v1.4s,v2.4s}, [%0], %6 \n"  // load 12 source samples
+      "add         v0.4s, v0.4s, v1.4s           \n"  // * 1
+      "add         v1.4s, v1.4s, v2.4s           \n"  // * 1
+      "ld1         {v2.4s,v3.4s}, [%2], #32      \n"
+      "mla         v0.4s, v2.4s, v7.4s           \n"  // * 6
+      "mla         v1.4s, v3.4s, v7.4s           \n"  // * 6
+      "ld1         {v2.4s,v3.4s}, [%1], #32      \n"
+      "ld1         {v4.4s,v5.4s}, [%3], #32      \n"
+      "add         v2.4s, v2.4s, v4.4s           \n"  // add rows for * 4
+      "add         v3.4s, v3.4s, v5.4s           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mla         v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "mla         v1.4s, v3.4s, v6.4s           \n"  // * 4
+      "subs        %w5, %w5, #8                  \n"  // 8 processed per loop
+      "uqrshrn     v0.4h, v0.4s, #8              \n"  // round and pack
+      "uqrshrn2    v0.8h, v1.4s, #8              \n"
+      "st1         {v0.8h}, [%4], #16            \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      : "+r"(src),   // %0
+        "+r"(src1),  // %1
+        "+r"(src2),  // %2
+        "+r"(src3),  // %3
+        "+r"(dst),   // %4
+        "+r"(width)  // %5
+      : "r"(32LL)    // %6
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+
+static const vecf32 kGaussCoefficients = {4.0f, 6.0f, 1.0f / 256.0f, 0.0f};
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_F32_NEON(const float* src0,
+                       const float* src1,
+                       const float* src2,
+                       const float* src3,
+                       const float* src4,
+                       float* dst,
+                       int width) {
+  asm volatile(
+      "ld2r        {v6.4s, v7.4s}, [%7]          \n"  // constants 4 and 6
+
+      "1:                                        \n"
+      "ld1         {v0.4s, v1.4s}, [%0], #32     \n"  // load 8 samples, 5 rows
+      "ld1         {v2.4s, v3.4s}, [%1], #32     \n"
+      "fmla        v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "ld1         {v4.4s, v5.4s}, [%2], #32     \n"
+      "fmla        v1.4s, v3.4s, v6.4s           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmla        v0.4s, v4.4s, v7.4s           \n"  // * 6
+      "ld1         {v2.4s, v3.4s}, [%3], #32     \n"
+      "fmla        v1.4s, v5.4s, v7.4s           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "fmla        v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "ld1         {v4.4s, v5.4s}, [%4], #32     \n"
+      "fmla        v1.4s, v3.4s, v6.4s           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "fadd        v0.4s, v0.4s, v4.4s           \n"  // * 1
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "fadd        v1.4s, v1.4s, v5.4s           \n"
+      "prfm        pldl1keep, [%4, 448]          \n"
+      "subs        %w6, %w6, #8                  \n"  // 8 processed per loop
+      "st1         {v0.4s, v1.4s}, [%5], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      : "+r"(src0),               // %0
+        "+r"(src1),               // %1
+        "+r"(src2),               // %2
+        "+r"(src3),               // %3
+        "+r"(src4),               // %4
+        "+r"(dst),                // %5
+        "+r"(width)               // %6
+      : "r"(&kGaussCoefficients)  // %7
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussRow_F32_NEON(const float* src, float* dst, int width) {
+  asm volatile(
+      "ld3r        {v6.4s, v7.4s, v8.4s}, [%3]   \n"  // constants 4, 6, 1/256
+
+      "1:                                        \n"
+      "ld1         {v0.4s, v1.4s, v2.4s}, [%0], %4 \n"  // load 12 samples, 5
+                                                        // rows
+      "fadd        v0.4s, v0.4s, v1.4s           \n"    // * 1
+      "ld1         {v4.4s, v5.4s}, [%0], %5      \n"
+      "fadd        v1.4s, v1.4s, v2.4s           \n"
+      "fmla        v0.4s, v4.4s, v7.4s           \n"  // * 6
+      "ld1         {v2.4s, v3.4s}, [%0], %4      \n"
+      "fmla        v1.4s, v5.4s, v7.4s           \n"
+      "ld1         {v4.4s, v5.4s}, [%0], %6      \n"
+      "fadd        v2.4s, v2.4s, v4.4s           \n"
+      "fadd        v3.4s, v3.4s, v5.4s           \n"
+      "fmla        v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "fmla        v1.4s, v3.4s, v6.4s           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmul        v0.4s, v0.4s, v8.4s           \n"  // / 256
+      "fmul        v1.4s, v1.4s, v8.4s           \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "st1         {v0.4s, v1.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      : "+r"(src),                 // %0
+        "+r"(dst),                 // %1
+        "+r"(width)                // %2
+      : "r"(&kGaussCoefficients),  // %3
+        "r"(8LL),                  // %4
+        "r"(-4LL),                 // %5
+        "r"(20LL)                  // %6
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8");
+}
+
+#if LIBYUV_USE_ST3
+// Convert biplanar NV21 to packed YUV24
+void NV21ToYUV24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v2.16b}, [%0], #16           \n"  // load 16 Y values
+      "ld2         {v0.8b, v1.8b}, [%1], #16     \n"  // load 8 VU values
+      "zip1        v0.16b, v0.16b, v0.16b        \n"  // replicate V values
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v1.16b, v1.16b, v1.16b        \n"  // replicate U values
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "subs        %w3, %w3, #16                 \n"      // 16 pixels per loop
+      "st3         {v0.16b,v1.16b,v2.16b}, [%2], #48 \n"  // store 16 YUV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_y),      // %0
+        "+r"(src_vu),     // %1
+        "+r"(dst_yuv24),  // %2
+        "+r"(width)       // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2");
+}
+#else
+static const uvec8 kYUV24Shuffle[3] = {
+    {16, 17, 0, 16, 17, 1, 18, 19, 2, 18, 19, 3, 20, 21, 4, 20},
+    {21, 5, 22, 23, 6, 22, 23, 7, 24, 25, 8, 24, 25, 9, 26, 27},
+    {10, 26, 27, 11, 28, 29, 12, 28, 29, 13, 30, 31, 14, 30, 31, 15}};
+
+// Convert biplanar NV21 to packed YUV24
+// NV21 has VU in memory for chroma.
+// YUV24 is VUY in memory
+void NV21ToYUV24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width) {
+  asm volatile(
+      "ld1         {v5.16b,v6.16b,v7.16b}, [%4]  \n"  // 3 shuffler constants
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"    // load 16 Y values
+      "ld1         {v1.16b}, [%1], #16           \n"    // load 8 VU values
+      "tbl         v2.16b, {v0.16b,v1.16b}, v5.16b \n"  // weave into YUV24
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v3.16b, {v0.16b,v1.16b}, v6.16b \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "tbl         v4.16b, {v0.16b,v1.16b}, v7.16b \n"
+      "subs        %w3, %w3, #16                 \n"      // 16 pixels per loop
+      "st1         {v2.16b,v3.16b,v4.16b}, [%2], #48 \n"  // store 16 YUV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_y),            // %0
+        "+r"(src_vu),           // %1
+        "+r"(dst_yuv24),        // %2
+        "+r"(width)             // %3
+      : "r"(&kYUV24Shuffle[0])  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+#endif  // LIBYUV_USE_ST3
+
+// Note ST2 8b version is faster than zip+ST1
+
+// AYUV is VUYA in memory.  UV for NV12 is UV order in memory.
+void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
+                      int src_stride_ayuv,
+                      uint8_t* dst_uv,
+                      int width) {
+  const uint8_t* src_ayuv_1 = src_ayuv + src_stride_ayuv;
+  asm volatile(
+
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ayuv
+      "uaddlp      v0.8h, v0.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v5.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqrshrn     v3.8b, v0.8h, #2              \n"  // 2x2 average
+      "uqrshrn     v2.8b, v1.8h, #2              \n"
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop.
+      "st2         {v2.8b,v3.8b}, [%2], #16      \n"  // store 8 pixels UV.
+      "b.gt        1b                            \n"
+      : "+r"(src_ayuv),    // %0
+        "+r"(src_ayuv_1),  // %1
+        "+r"(dst_uv),      // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+
+void AYUVToVURow_NEON(const uint8_t* src_ayuv,
+                      int src_stride_ayuv,
+                      uint8_t* dst_vu,
+                      int width) {
+  const uint8_t* src_ayuv_1 = src_ayuv + src_stride_ayuv;
+  asm volatile(
+
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ayuv
+      "uaddlp      v0.8h, v0.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v5.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqrshrn     v0.8b, v0.8h, #2              \n"  // 2x2 average
+      "uqrshrn     v1.8b, v1.8h, #2              \n"
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop.
+      "st2         {v0.8b,v1.8b}, [%2], #16      \n"  // store 8 pixels VU.
+      "b.gt        1b                            \n"
+      : "+r"(src_ayuv),    // %0
+        "+r"(src_ayuv_1),  // %1
+        "+r"(dst_vu),      // %2
+        "+r"(width)        // %3
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+
+// Copy row of AYUV Y's into Y
+void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v2.16b}, [%1], #16           \n"  // store 16 Y pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ayuv),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
+}
+
+// Shuffle table for swapping UV bytes.
+static const uvec8 kShuffleSwapUV = {1u, 0u, 3u,  2u,  5u,  4u,  7u,  6u,
+                                     9u, 8u, 11u, 10u, 13u, 12u, 15u, 14u};
+
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+      "ld1         {v2.16b}, [%3]                \n"  // shuffler
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], 16            \n"  // load 16 UV values
+      "ld1         {v1.16b}, [%0], 16            \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop
+      "tbl         v0.16b, {v0.16b}, v2.16b      \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v1.16b, {v1.16b}, v2.16b      \n"
+      "stp         q0, q1, [%1], 32              \n"  // store 16 VU pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),         // %0
+        "+r"(dst_vu),         // %1
+        "+r"(width)           // %2
+      : "r"(&kShuffleSwapUV)  // %3
+      : "cc", "memory", "v0", "v1", "v2");
+}
+
+void HalfMergeUVRow_NEON(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width) {
+  const uint8_t* src_u_1 = src_u + src_stride_u;
+  const uint8_t* src_v_1 = src_v + src_stride_v;
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 16 U values
+      "ld1         {v1.16b}, [%2], #16           \n"  // load 16 V values
+      "ld1         {v2.16b}, [%1], #16           \n"
+      "ld1         {v3.16b}, [%3], #16           \n"
+      "uaddlp      v0.8h, v0.16b                 \n"  // half size
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "uadalp      v0.8h, v2.16b                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v3.16b                 \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "uqrshrn     v0.8b, v0.8h, #2              \n"
+      "uqrshrn     v1.8b, v1.8h, #2              \n"
+      "subs        %w5, %w5, #16                 \n"  // 16 src pixels per loop
+      "st2         {v0.8b, v1.8b}, [%4], #16     \n"  // store 8 UV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_u),    // %0
+        "+r"(src_u_1),  // %1
+        "+r"(src_v),    // %2
+        "+r"(src_v_1),  // %3
+        "+r"(dst_uv),   // %4
+        "+r"(width)     // %5
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
+}
+
+void SplitUVRow_16_NEON(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width) {
+  int shift = depth - 16;  // Negative for right shift.
+  asm volatile(
+      "dup         v2.8h, %w4                    \n"
+      "1:                                        \n"
+      "ld2         {v0.8h, v1.8h}, [%0], #32     \n"  // load 8 UV
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "ushl        v0.8h, v0.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v2.8h           \n"
+      "st1         {v0.8h}, [%1], #16            \n"  // store 8 U pixels
+      "st1         {v1.8h}, [%2], #16            \n"  // store 8 V pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "v0", "v1", "v2");
+}
+
+void MultiplyRow_16_NEON(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width) {
+  asm volatile(
+      "dup         v2.8h, %w3                    \n"
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"
+      "mul         v0.8h, v0.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mul         v1.8h, v1.8h, v2.8h           \n"
+      "stp         q0, q1, [%1], #32             \n"  // store 16 pixels
+      "subs        %w2, %w2, #16                 \n"  // 16 src pixels per loop
+      "b.gt        1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "v0", "v1", "v2");
+}
+
+void DivideRow_16_NEON(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  asm volatile(
+      "dup         v4.8h, %w3                    \n"
+      "1:                                        \n"
+      "ldp         q2, q3, [%0], #32             \n"
+      "umull       v0.4s, v2.4h, v4.4h           \n"
+      "umull2      v1.4s, v2.8h, v4.8h           \n"
+      "umull       v2.4s, v3.4h, v4.4h           \n"
+      "umull2      v3.4s, v3.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "shrn        v0.4h, v0.4s, #16             \n"
+      "shrn2       v0.8h, v1.4s, #16             \n"
+      "shrn        v1.4h, v2.4s, #16             \n"
+      "shrn2       v1.8h, v3.4s, #16             \n"
+      "stp         q0, q1, [%1], #32             \n"  // store 16 pixels
+      "subs        %w2, %w2, #16                 \n"  // 16 src pixels per loop
+      "b.gt        1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits = shr 1
+// 16384 = 10 bits = shr 2
+// 4096 = 12 bits = shr 4
+// 256 = 16 bits = shr 8
+void Convert16To8Row_NEON(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width) {
+  int shift = 15 - __builtin_clz((int32_t)scale);  // Negative shl is shr
+  asm volatile(
+      "dup         v2.8h, %w3                    \n"
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"
+      "ushl        v0.8h, v0.8h, v2.8h           \n"  // shr = v2 is negative
+      "ushl        v1.8h, v1.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "uqxtn2      v0.16b, v1.8h                 \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 src pixels per loop
+      "str         q0, [%1], #16                 \n"  // store 16 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(shift)    // %3
+      : "cc", "memory", "v0", "v1", "v2");
+}
+
 #endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/row_win.cc b/media/libyuv/source/row_win.cc
index 2a3da8969f..c5a14f86fb 100644
--- a/media/libyuv/source/row_win.cc
+++ b/media/libyuv/source/row_win.cc
@@ -10,9 +10,9 @@
 
 #include "libyuv/row.h"
 
-// This module is for Visual C 32/64 bit and clangcl 32 bit
+// This module is for Visual C 32/64 bit
 #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
-    (defined(_M_IX86) || (defined(_M_X64) && !defined(__clang__)))
+    !defined(__clang__) && (defined(_M_IX86) || defined(_M_X64))
 
 #if defined(_M_X64)
 #include <emmintrin.h>
@@ -27,73 +27,91 @@ extern "C" {
 // 64 bit
 #if defined(_M_X64)
 
+// Read 8 UV from 444
+#define READYUV444                                    \
+  xmm3 = _mm_loadl_epi64((__m128i*)u_buf);            \
+  xmm1 = _mm_loadl_epi64((__m128i*)(u_buf + offset)); \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);               \
+  u_buf += 8;                                         \
+  xmm4 = _mm_loadl_epi64((__m128i*)y_buf);            \
+  xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);               \
+  y_buf += 8;
+
+// Read 8 UV from 444, With 8 Alpha.
+#define READYUVA444                                   \
+  xmm3 = _mm_loadl_epi64((__m128i*)u_buf);            \
+  xmm1 = _mm_loadl_epi64((__m128i*)(u_buf + offset)); \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);               \
+  u_buf += 8;                                         \
+  xmm4 = _mm_loadl_epi64((__m128i*)y_buf);            \
+  xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);               \
+  y_buf += 8;                                         \
+  xmm5 = _mm_loadl_epi64((__m128i*)a_buf);            \
+  a_buf += 8;
+
 // Read 4 UV from 422, upsample to 8 UV.
-#define READYUV422                                                             \
-    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);                                 \
-    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));                      \
-    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);                                      \
-    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);                                     \
-    u_buf += 4;                                                                \
-    xmm4 = _mm_loadl_epi64((__m128i*)y_buf);                                   \
-    xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);                                      \
-    y_buf += 8;
+#define READYUV422                                        \
+  xmm3 = _mm_cvtsi32_si128(*(uint32_t*)u_buf);            \
+  xmm1 = _mm_cvtsi32_si128(*(uint32_t*)(u_buf + offset)); \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);                   \
+  xmm3 = _mm_unpacklo_epi16(xmm3, xmm3);                  \
+  u_buf += 4;                                             \
+  xmm4 = _mm_loadl_epi64((__m128i*)y_buf);                \
+  xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);                   \
+  y_buf += 8;
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
-#define READYUVA422                                                            \
-    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);                                 \
-    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));                      \
-    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);                                      \
-    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);                                     \
-    u_buf += 4;                                                                \
-    xmm4 = _mm_loadl_epi64((__m128i*)y_buf);                                   \
-    xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);                                      \
-    y_buf += 8;                                                                \
-    xmm5 = _mm_loadl_epi64((__m128i*)a_buf);                                   \
-    a_buf += 8;
+#define READYUVA422                                       \
+  xmm3 = _mm_cvtsi32_si128(*(uint32_t*)u_buf);            \
+  xmm1 = _mm_cvtsi32_si128(*(uint32_t*)(u_buf + offset)); \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);                   \
+  xmm3 = _mm_unpacklo_epi16(xmm3, xmm3);                  \
+  u_buf += 4;                                             \
+  xmm4 = _mm_loadl_epi64((__m128i*)y_buf);                \
+  xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);                   \
+  y_buf += 8;                                             \
+  xmm5 = _mm_loadl_epi64((__m128i*)a_buf);                \
+  a_buf += 8;
 
 // Convert 8 pixels: 8 UV and 8 Y.
-#define YUVTORGB(yuvconstants)                                                 \
-    xmm1 = _mm_loadu_si128(&xmm0);                                             \
-    xmm2 = _mm_loadu_si128(&xmm0);                                             \
-    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)yuvconstants->kUVToB);           \
-    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)yuvconstants->kUVToG);           \
-    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)yuvconstants->kUVToR);           \
-    xmm0 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasB, xmm0);             \
-    xmm1 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasG, xmm1);             \
-    xmm2 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasR, xmm2);             \
-    xmm4 = _mm_mulhi_epu16(xmm4, *(__m128i*)yuvconstants->kYToRgb);            \
-    xmm0 = _mm_adds_epi16(xmm0, xmm4);                                         \
-    xmm1 = _mm_adds_epi16(xmm1, xmm4);                                         \
-    xmm2 = _mm_adds_epi16(xmm2, xmm4);                                         \
-    xmm0 = _mm_srai_epi16(xmm0, 6);                                            \
-    xmm1 = _mm_srai_epi16(xmm1, 6);                                            \
-    xmm2 = _mm_srai_epi16(xmm2, 6);                                            \
-    xmm0 = _mm_packus_epi16(xmm0, xmm0);                                       \
-    xmm1 = _mm_packus_epi16(xmm1, xmm1);                                       \
-    xmm2 = _mm_packus_epi16(xmm2, xmm2);
+#define YUVTORGB(yuvconstants)                                      \
+  xmm3 = _mm_sub_epi8(xmm3, _mm_set1_epi8((char)0x80));             \
+  xmm4 = _mm_mulhi_epu16(xmm4, *(__m128i*)yuvconstants->kYToRgb);   \
+  xmm4 = _mm_add_epi16(xmm4, *(__m128i*)yuvconstants->kYBiasToRgb); \
+  xmm0 = _mm_maddubs_epi16(*(__m128i*)yuvconstants->kUVToB, xmm3);  \
+  xmm1 = _mm_maddubs_epi16(*(__m128i*)yuvconstants->kUVToG, xmm3);  \
+  xmm2 = _mm_maddubs_epi16(*(__m128i*)yuvconstants->kUVToR, xmm3);  \
+  xmm0 = _mm_adds_epi16(xmm4, xmm0);                                \
+  xmm1 = _mm_subs_epi16(xmm4, xmm1);                                \
+  xmm2 = _mm_adds_epi16(xmm4, xmm2);                                \
+  xmm0 = _mm_srai_epi16(xmm0, 6);                                   \
+  xmm1 = _mm_srai_epi16(xmm1, 6);                                   \
+  xmm2 = _mm_srai_epi16(xmm2, 6);                                   \
+  xmm0 = _mm_packus_epi16(xmm0, xmm0);                              \
+  xmm1 = _mm_packus_epi16(xmm1, xmm1);                              \
+  xmm2 = _mm_packus_epi16(xmm2, xmm2);
 
 // Store 8 ARGB values.
-#define STOREARGB                                                              \
-    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);                                      \
-    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);                                      \
-    xmm1 = _mm_loadu_si128(&xmm0);                                             \
-    xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);                                     \
-    xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);                                     \
-    _mm_storeu_si128((__m128i *)dst_argb, xmm0);                               \
-    _mm_storeu_si128((__m128i *)(dst_argb + 16), xmm1);                        \
-    dst_argb += 32;
-
+#define STOREARGB                                    \
+  xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);              \
+  xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);              \
+  xmm1 = _mm_loadu_si128(&xmm0);                     \
+  xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);             \
+  xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);             \
+  _mm_storeu_si128((__m128i*)dst_argb, xmm0);        \
+  _mm_storeu_si128((__m128i*)(dst_argb + 16), xmm1); \
+  dst_argb += 32;
 
 #if defined(HAS_I422TOARGBROW_SSSE3)
-void I422ToARGBRow_SSSE3(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* dst_argb,
+void I422ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  __m128i xmm0, xmm1, xmm2, xmm4;
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
   const __m128i xmm5 = _mm_set1_epi8(-1);
-  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
+  const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
   while (width > 0) {
     READYUV422
     YUVTORGB(yuvconstants)
@@ -104,15 +122,15 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
 #endif
 
 #if defined(HAS_I422ALPHATOARGBROW_SSSE3)
-void I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
-                              const uint8* u_buf,
-                              const uint8* v_buf,
-                              const uint8* a_buf,
-                              uint8* dst_argb,
+void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              const uint8_t* a_buf,
+                              uint8_t* dst_argb,
                               const struct YuvConstants* yuvconstants,
                               int width) {
-  __m128i xmm0, xmm1, xmm2, xmm4, xmm5;
-  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5;
+  const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
   while (width > 0) {
     READYUVA422
     YUVTORGB(yuvconstants)
@@ -122,180 +140,186 @@ void I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
 }
 #endif
 
+#if defined(HAS_I444TOARGBROW_SSSE3)
+void I444ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
+  const __m128i xmm5 = _mm_set1_epi8(-1);
+  const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
+  while (width > 0) {
+    READYUV444
+    YUVTORGB(yuvconstants)
+    STOREARGB
+    width -= 8;
+  }
+}
+#endif
+
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+void I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              const uint8_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width) {
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5;
+  const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
+  while (width > 0) {
+    READYUVA444
+    YUVTORGB(yuvconstants)
+    STOREARGB
+    width -= 8;
+  }
+}
+#endif
+
 // 32 bit
 #else  // defined(_M_X64)
 #ifdef HAS_ARGBTOYROW_SSSE3
 
 // Constants for ARGB.
-static const vec8 kARGBToY = {
-  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
-};
+static const vec8 kARGBToY = {13, 65, 33, 0, 13, 65, 33, 0,
+                              13, 65, 33, 0, 13, 65, 33, 0};
 
 // JPeg full range.
-static const vec8 kARGBToYJ = {
-  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
-};
+static const vec8 kARGBToYJ = {15, 75, 38, 0, 15, 75, 38, 0,
+                               15, 75, 38, 0, 15, 75, 38, 0};
 
-static const vec8 kARGBToU = {
-  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
-};
+static const vec8 kARGBToU = {112, -74, -38, 0, 112, -74, -38, 0,
+                              112, -74, -38, 0, 112, -74, -38, 0};
 
-static const vec8 kARGBToUJ = {
-  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
-};
+static const vec8 kARGBToUJ = {127, -84, -43, 0, 127, -84, -43, 0,
+                               127, -84, -43, 0, 127, -84, -43, 0};
 
 static const vec8 kARGBToV = {
-  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+    -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
 };
 
-static const vec8 kARGBToVJ = {
-  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
-};
+static const vec8 kARGBToVJ = {-20, -107, 127, 0, -20, -107, 127, 0,
+                               -20, -107, 127, 0, -20, -107, 127, 0};
 
 // vpshufb for vphaddw + vpackuswb packed to shorts.
 static const lvec8 kShufARGBToUV_AVX = {
-  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
-  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
-};
+    0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+    0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
 
 // Constants for BGRA.
-static const vec8 kBGRAToY = {
-  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
-};
+static const vec8 kBGRAToY = {0, 33, 65, 13, 0, 33, 65, 13,
+                              0, 33, 65, 13, 0, 33, 65, 13};
 
-static const vec8 kBGRAToU = {
-  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
-};
+static const vec8 kBGRAToU = {0, -38, -74, 112, 0, -38, -74, 112,
+                              0, -38, -74, 112, 0, -38, -74, 112};
 
-static const vec8 kBGRAToV = {
-  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
-};
+static const vec8 kBGRAToV = {0, 112, -94, -18, 0, 112, -94, -18,
+                              0, 112, -94, -18, 0, 112, -94, -18};
 
 // Constants for ABGR.
-static const vec8 kABGRToY = {
-  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
-};
+static const vec8 kABGRToY = {33, 65, 13, 0, 33, 65, 13, 0,
+                              33, 65, 13, 0, 33, 65, 13, 0};
 
-static const vec8 kABGRToU = {
-  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
-};
+static const vec8 kABGRToU = {-38, -74, 112, 0, -38, -74, 112, 0,
+                              -38, -74, 112, 0, -38, -74, 112, 0};
 
-static const vec8 kABGRToV = {
-  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
-};
+static const vec8 kABGRToV = {112, -94, -18, 0, 112, -94, -18, 0,
+                              112, -94, -18, 0, 112, -94, -18, 0};
 
 // Constants for RGBA.
-static const vec8 kRGBAToY = {
-  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
-};
+static const vec8 kRGBAToY = {0, 13, 65, 33, 0, 13, 65, 33,
+                              0, 13, 65, 33, 0, 13, 65, 33};
 
-static const vec8 kRGBAToU = {
-  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
-};
+static const vec8 kRGBAToU = {0, 112, -74, -38, 0, 112, -74, -38,
+                              0, 112, -74, -38, 0, 112, -74, -38};
 
-static const vec8 kRGBAToV = {
-  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
-};
+static const vec8 kRGBAToV = {0, -18, -94, 112, 0, -18, -94, 112,
+                              0, -18, -94, 112, 0, -18, -94, 112};
 
-static const uvec8 kAddY16 = {
-  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
-};
+static const uvec8 kAddY16 = {16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
+                              16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u};
 
 // 7 bit fixed point 0.5.
-static const vec16 kAddYJ64 = {
-  64, 64, 64, 64, 64, 64, 64, 64
-};
+static const vec16 kAddYJ64 = {64, 64, 64, 64, 64, 64, 64, 64};
 
-static const uvec8 kAddUV128 = {
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
-
-static const uvec16 kAddUVJ128 = {
-  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
-};
+// 8 bit fixed point 0.5, for bias of UV.
+static const ulvec8 kBiasUV128 = {
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
 
 // Shuffle table for converting RGB24 to ARGB.
 static const uvec8 kShuffleMaskRGB24ToARGB = {
-  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
-};
+    0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u};
 
 // Shuffle table for converting RAW to ARGB.
-static const uvec8 kShuffleMaskRAWToARGB = {
-  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
-};
+static const uvec8 kShuffleMaskRAWToARGB = {2u, 1u, 0u, 12u, 5u,  4u,  3u, 13u,
+                                            8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u};
 
 // Shuffle table for converting RAW to RGB24.  First 8.
 static const uvec8 kShuffleMaskRAWToRGB24_0 = {
-  2u, 1u, 0u, 5u, 4u, 3u, 8u, 7u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+    2u,   1u,   0u,   5u,   4u,   3u,   8u,   7u,
+    128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting RAW to RGB24.  Middle 8.
 static const uvec8 kShuffleMaskRAWToRGB24_1 = {
-  2u, 7u, 6u, 5u, 10u, 9u, 8u, 13u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+    2u,   7u,   6u,   5u,   10u,  9u,   8u,   13u,
+    128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting RAW to RGB24.  Last 8.
 static const uvec8 kShuffleMaskRAWToRGB24_2 = {
-  8u, 7u, 12u, 11u, 10u, 15u, 14u, 13u,
-  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
-};
+    8u,   7u,   12u,  11u,  10u,  15u,  14u,  13u,
+    128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting ARGB to RGB24.
 static const uvec8 kShuffleMaskARGBToRGB24 = {
-  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
-};
+    0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting ARGB to RAW.
 static const uvec8 kShuffleMaskARGBToRAW = {
-  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
-};
+    2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u};
 
 // Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
 static const uvec8 kShuffleMaskARGBToRGB24_0 = {
-  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
-};
+    0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u};
 
 // YUY2 shuf 16 Y to 32 Y.
-static const lvec8 kShuffleYUY2Y = {
-  0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12, 14, 14,
-  0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12, 14, 14
-};
+static const lvec8 kShuffleYUY2Y = {0,  0,  2,  2,  4,  4,  6,  6,  8,  8, 10,
+                                    10, 12, 12, 14, 14, 0,  0,  2,  2,  4, 4,
+                                    6,  6,  8,  8,  10, 10, 12, 12, 14, 14};
 
 // YUY2 shuf 8 UV to 16 UV.
-static const lvec8 kShuffleYUY2UV = {
-  1, 3, 1, 3, 5, 7, 5, 7, 9, 11, 9, 11, 13, 15, 13, 15,
-  1, 3, 1, 3, 5, 7, 5, 7, 9, 11, 9, 11, 13, 15, 13, 15
-};
+static const lvec8 kShuffleYUY2UV = {1,  3,  1,  3,  5,  7,  5,  7,  9,  11, 9,
+                                     11, 13, 15, 13, 15, 1,  3,  1,  3,  5,  7,
+                                     5,  7,  9,  11, 9,  11, 13, 15, 13, 15};
 
 // UYVY shuf 16 Y to 32 Y.
-static const lvec8 kShuffleUYVYY = {
-  1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15,
-  1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15
-};
+static const lvec8 kShuffleUYVYY = {1,  1,  3,  3,  5,  5,  7,  7,  9,  9, 11,
+                                    11, 13, 13, 15, 15, 1,  1,  3,  3,  5, 5,
+                                    7,  7,  9,  9,  11, 11, 13, 13, 15, 15};
 
 // UYVY shuf 8 UV to 16 UV.
-static const lvec8 kShuffleUYVYUV = {
-  0, 2, 0, 2, 4, 6, 4, 6, 8, 10, 8, 10, 12, 14, 12, 14,
-  0, 2, 0, 2, 4, 6, 4, 6, 8, 10, 8, 10, 12, 14, 12, 14
-};
+static const lvec8 kShuffleUYVYUV = {0,  2,  0,  2,  4,  6,  4,  6,  8,  10, 8,
+                                     10, 12, 14, 12, 14, 0,  2,  0,  2,  4,  6,
+                                     4,  6,  8,  10, 8,  10, 12, 14, 12, 14};
 
 // NV21 shuf 8 VU to 16 UV.
 static const lvec8 kShuffleNV21 = {
-  1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
-  1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
+    1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
+    1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
 };
 
 // Duplicates gray value 3 times and fills in alpha opaque.
-__declspec(naked)
-void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width) {
+__declspec(naked) void J400ToARGBRow_SSE2(const uint8_t* src_y,
+                                          uint8_t* dst_argb,
+                                          int width) {
   __asm {
-    mov        eax, [esp + 4]        // src_y
-    mov        edx, [esp + 8]        // dst_argb
-    mov        ecx, [esp + 12]       // width
-    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    mov        eax, [esp + 4]  // src_y
+    mov        edx, [esp + 8]  // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0xff000000
     pslld      xmm5, 24
 
   convertloop:
@@ -318,13 +342,14 @@ void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width) {
 
 #ifdef HAS_J400TOARGBROW_AVX2
 // Duplicates gray value 3 times and fills in alpha opaque.
-__declspec(naked)
-void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width) {
-  __asm {
-    mov         eax, [esp + 4]        // src_y
-    mov         edx, [esp + 8]        // dst_argb
-    mov         ecx, [esp + 12]       // width
-    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0xff000000
+__declspec(naked) void J400ToARGBRow_AVX2(const uint8_t* src_y,
+                                          uint8_t* dst_argb,
+                                          int width) {
+  __asm {
+    mov         eax, [esp + 4]  // src_y
+    mov         edx, [esp + 8]  // dst_argb
+    mov         ecx, [esp + 12]  // width
+    vpcmpeqb    ymm5, ymm5, ymm5  // generate mask 0xff000000
     vpslld      ymm5, ymm5, 24
 
   convertloop:
@@ -348,13 +373,14 @@ void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width) {
 }
 #endif  // HAS_J400TOARGBROW_AVX2
 
-__declspec(naked)
-void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int width) {
+__declspec(naked) void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_rgb24
-    mov       edx, [esp + 8]   // dst_argb
+    mov       eax, [esp + 4]  // src_rgb24
+    mov       edx, [esp + 8]  // dst_argb
     mov       ecx, [esp + 12]  // width
-    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pcmpeqb   xmm5, xmm5  // generate mask 0xff000000
     pslld     xmm5, 24
     movdqa    xmm4, xmmword ptr kShuffleMaskRGB24ToARGB
 
@@ -364,17 +390,17 @@ void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int width) {
     movdqu    xmm3, [eax + 32]
     lea       eax, [eax + 48]
     movdqa    xmm2, xmm3
-    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    palignr   xmm2, xmm1, 8  // xmm2 = { xmm3[0:3] xmm1[8:15]}
     pshufb    xmm2, xmm4
     por       xmm2, xmm5
-    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    palignr   xmm1, xmm0, 12  // xmm1 = { xmm3[0:7] xmm0[12:15]}
     pshufb    xmm0, xmm4
     movdqu    [edx + 32], xmm2
     por       xmm0, xmm5
     pshufb    xmm1, xmm4
     movdqu    [edx], xmm0
     por       xmm1, xmm5
-    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    palignr   xmm3, xmm3, 4  // xmm3 = { xmm3[4:15]}
     pshufb    xmm3, xmm4
     movdqu    [edx + 16], xmm1
     por       xmm3, xmm5
@@ -386,14 +412,14 @@ void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int width) {
   }
 }
 
-__declspec(naked)
-void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
-                        int width) {
+__declspec(naked) void RAWToARGBRow_SSSE3(const uint8_t* src_raw,
+                                          uint8_t* dst_argb,
+                                          int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_raw
-    mov       edx, [esp + 8]   // dst_argb
+    mov       eax, [esp + 4]  // src_raw
+    mov       edx, [esp + 8]  // dst_argb
     mov       ecx, [esp + 12]  // width
-    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pcmpeqb   xmm5, xmm5  // generate mask 0xff000000
     pslld     xmm5, 24
     movdqa    xmm4, xmmword ptr kShuffleMaskRAWToARGB
 
@@ -403,17 +429,17 @@ void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
     movdqu    xmm3, [eax + 32]
     lea       eax, [eax + 48]
     movdqa    xmm2, xmm3
-    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    palignr   xmm2, xmm1, 8  // xmm2 = { xmm3[0:3] xmm1[8:15]}
     pshufb    xmm2, xmm4
     por       xmm2, xmm5
-    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    palignr   xmm1, xmm0, 12  // xmm1 = { xmm3[0:7] xmm0[12:15]}
     pshufb    xmm0, xmm4
     movdqu    [edx + 32], xmm2
     por       xmm0, xmm5
     pshufb    xmm1, xmm4
     movdqu    [edx], xmm0
     por       xmm1, xmm5
-    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    palignr   xmm3, xmm3, 4  // xmm3 = { xmm3[4:15]}
     pshufb    xmm3, xmm4
     movdqu    [edx + 16], xmm1
     por       xmm3, xmm5
@@ -425,11 +451,12 @@ void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
   }
 }
 
-__declspec(naked)
-void RAWToRGB24Row_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width) {
+__declspec(naked) void RAWToRGB24Row_SSSE3(const uint8_t* src_raw,
+                                           uint8_t* dst_rgb24,
+                                           int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_raw
-    mov       edx, [esp + 8]   // dst_rgb24
+    mov       eax, [esp + 4]  // src_raw
+    mov       edx, [esp + 8]  // dst_rgb24
     mov       ecx, [esp + 12]  // width
     movdqa    xmm3, xmmword ptr kShuffleMaskRAWToRGB24_0
     movdqa    xmm4, xmmword ptr kShuffleMaskRAWToRGB24_1
@@ -460,9 +487,9 @@ void RAWToRGB24Row_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width) {
 // v * (256 + 8)
 // G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
 // 20 instructions.
-__declspec(naked)
-void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
-                          int width) {
+__declspec(naked) void RGB565ToARGBRow_SSE2(const uint8_t* src_rgb565,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
     mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
     movd      xmm5, eax
@@ -470,33 +497,33 @@ void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
     mov       eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
     movd      xmm6, eax
     pshufd    xmm6, xmm6, 0
-    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    pcmpeqb   xmm3, xmm3  // generate mask 0xf800f800 for Red
     psllw     xmm3, 11
-    pcmpeqb   xmm4, xmm4       // generate mask 0x07e007e0 for Green
+    pcmpeqb   xmm4, xmm4  // generate mask 0x07e007e0 for Green
     psllw     xmm4, 10
     psrlw     xmm4, 5
-    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    pcmpeqb   xmm7, xmm7  // generate mask 0xff00ff00 for Alpha
     psllw     xmm7, 8
 
-    mov       eax, [esp + 4]   // src_rgb565
-    mov       edx, [esp + 8]   // dst_argb
+    mov       eax, [esp + 4]  // src_rgb565
+    mov       edx, [esp + 8]  // dst_argb
     mov       ecx, [esp + 12]  // width
     sub       edx, eax
     sub       edx, eax
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
+    movdqu    xmm0, [eax]  // fetch 8 pixels of bgr565
     movdqa    xmm1, xmm0
     movdqa    xmm2, xmm0
-    pand      xmm1, xmm3    // R in upper 5 bits
-    psllw     xmm2, 11      // B in upper 5 bits
-    pmulhuw   xmm1, xmm5    // * (256 + 8)
-    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    pand      xmm1, xmm3  // R in upper 5 bits
+    psllw     xmm2, 11  // B in upper 5 bits
+    pmulhuw   xmm1, xmm5  // * (256 + 8)
+    pmulhuw   xmm2, xmm5  // * (256 + 8)
     psllw     xmm1, 8
-    por       xmm1, xmm2    // RB
-    pand      xmm0, xmm4    // G in middle 6 bits
-    pmulhuw   xmm0, xmm6    // << 5 * (256 + 4)
-    por       xmm0, xmm7    // AG
+    por       xmm1, xmm2  // RB
+    pand      xmm0, xmm4  // G in middle 6 bits
+    pmulhuw   xmm0, xmm6  // << 5 * (256 + 4)
+    por       xmm0, xmm7  // AG
     movdqa    xmm2, xmm1
     punpcklbw xmm1, xmm0
     punpckhbw xmm2, xmm0
@@ -516,9 +543,9 @@ void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
 // v * 256 + v * 8
 // v * (256 + 8)
 // G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
-__declspec(naked)
-void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb,
-                          int width) {
+__declspec(naked) void RGB565ToARGBRow_AVX2(const uint8_t* src_rgb565,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
     mov        eax, 0x01080108  // generate multiplier to repeat 5 bits
     vmovd      xmm5, eax
@@ -526,32 +553,32 @@ void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb,
     mov        eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
     vmovd      xmm6, eax
     vbroadcastss ymm6, xmm6
-    vpcmpeqb   ymm3, ymm3, ymm3       // generate mask 0xf800f800 for Red
+    vpcmpeqb   ymm3, ymm3, ymm3  // generate mask 0xf800f800 for Red
     vpsllw     ymm3, ymm3, 11
-    vpcmpeqb   ymm4, ymm4, ymm4       // generate mask 0x07e007e0 for Green
+    vpcmpeqb   ymm4, ymm4, ymm4  // generate mask 0x07e007e0 for Green
     vpsllw     ymm4, ymm4, 10
     vpsrlw     ymm4, ymm4, 5
-    vpcmpeqb   ymm7, ymm7, ymm7       // generate mask 0xff00ff00 for Alpha
+    vpcmpeqb   ymm7, ymm7, ymm7  // generate mask 0xff00ff00 for Alpha
     vpsllw     ymm7, ymm7, 8
 
-    mov        eax, [esp + 4]   // src_rgb565
-    mov        edx, [esp + 8]   // dst_argb
+    mov        eax, [esp + 4]  // src_rgb565
+    mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
     sub        edx, eax
 
  convertloop:
-    vmovdqu    ymm0, [eax]   // fetch 16 pixels of bgr565
-    vpand      ymm1, ymm0, ymm3    // R in upper 5 bits
-    vpsllw     ymm2, ymm0, 11      // B in upper 5 bits
-    vpmulhuw   ymm1, ymm1, ymm5    // * (256 + 8)
-    vpmulhuw   ymm2, ymm2, ymm5    // * (256 + 8)
+    vmovdqu    ymm0, [eax]  // fetch 16 pixels of bgr565
+    vpand      ymm1, ymm0, ymm3  // R in upper 5 bits
+    vpsllw     ymm2, ymm0, 11  // B in upper 5 bits
+    vpmulhuw   ymm1, ymm1, ymm5  // * (256 + 8)
+    vpmulhuw   ymm2, ymm2, ymm5  // * (256 + 8)
     vpsllw     ymm1, ymm1, 8
-    vpor       ymm1, ymm1, ymm2    // RB
-    vpand      ymm0, ymm0, ymm4    // G in middle 6 bits
-    vpmulhuw   ymm0, ymm0, ymm6    // << 5 * (256 + 4)
-    vpor       ymm0, ymm0, ymm7    // AG
-    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpor       ymm1, ymm1, ymm2  // RB
+    vpand      ymm0, ymm0, ymm4  // G in middle 6 bits
+    vpmulhuw   ymm0, ymm0, ymm6  // << 5 * (256 + 4)
+    vpor       ymm0, ymm0, ymm7  // AG
+    vpermq     ymm0, ymm0, 0xd8  // mutate for unpack
     vpermq     ymm1, ymm1, 0xd8
     vpunpckhbw ymm2, ymm1, ymm0
     vpunpcklbw ymm1, ymm1, ymm0
@@ -567,9 +594,9 @@ void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb,
 #endif  // HAS_RGB565TOARGBROW_AVX2
 
 #ifdef HAS_ARGB1555TOARGBROW_AVX2
-__declspec(naked)
-void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
-                            int width) {
+__declspec(naked) void ARGB1555ToARGBRow_AVX2(const uint8_t* src_argb1555,
+                                              uint8_t* dst_argb,
+                                              int width) {
   __asm {
     mov        eax, 0x01080108  // generate multiplier to repeat 5 bits
     vmovd      xmm5, eax
@@ -577,33 +604,33 @@ void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
     mov        eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
     vmovd      xmm6, eax
     vbroadcastss ymm6, xmm6
-    vpcmpeqb   ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red
+    vpcmpeqb   ymm3, ymm3, ymm3  // generate mask 0xf800f800 for Red
     vpsllw     ymm3, ymm3, 11
-    vpsrlw     ymm4, ymm3, 6    // generate mask 0x03e003e0 for Green
-    vpcmpeqb   ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha
+    vpsrlw     ymm4, ymm3, 6  // generate mask 0x03e003e0 for Green
+    vpcmpeqb   ymm7, ymm7, ymm7  // generate mask 0xff00ff00 for Alpha
     vpsllw     ymm7, ymm7, 8
 
-    mov        eax,  [esp + 4]   // src_argb1555
-    mov        edx,  [esp + 8]   // dst_argb
+    mov        eax,  [esp + 4]  // src_argb1555
+    mov        edx,  [esp + 8]  // dst_argb
     mov        ecx,  [esp + 12]  // width
     sub        edx,  eax
     sub        edx,  eax
 
  convertloop:
-    vmovdqu    ymm0, [eax]         // fetch 16 pixels of 1555
-    vpsllw     ymm1, ymm0, 1       // R in upper 5 bits
-    vpsllw     ymm2, ymm0, 11      // B in upper 5 bits
+    vmovdqu    ymm0, [eax]  // fetch 16 pixels of 1555
+    vpsllw     ymm1, ymm0, 1  // R in upper 5 bits
+    vpsllw     ymm2, ymm0, 11  // B in upper 5 bits
     vpand      ymm1, ymm1, ymm3
-    vpmulhuw   ymm2, ymm2, ymm5    // * (256 + 8)
-    vpmulhuw   ymm1, ymm1, ymm5    // * (256 + 8)
+    vpmulhuw   ymm2, ymm2, ymm5  // * (256 + 8)
+    vpmulhuw   ymm1, ymm1, ymm5  // * (256 + 8)
     vpsllw     ymm1, ymm1, 8
-    vpor       ymm1, ymm1, ymm2    // RB
-    vpsraw     ymm2, ymm0, 8       // A
-    vpand      ymm0, ymm0, ymm4    // G in middle 5 bits
-    vpmulhuw   ymm0, ymm0, ymm6    // << 6 * (256 + 8)
+    vpor       ymm1, ymm1, ymm2  // RB
+    vpsraw     ymm2, ymm0, 8  // A
+    vpand      ymm0, ymm0, ymm4  // G in middle 5 bits
+    vpmulhuw   ymm0, ymm0, ymm6  // << 6 * (256 + 8)
     vpand      ymm2, ymm2, ymm7
-    vpor       ymm0, ymm0, ymm2    // AG
-    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpor       ymm0, ymm0, ymm2  // AG
+    vpermq     ymm0, ymm0, 0xd8  // mutate for unpack
     vpermq     ymm1, ymm1, 0xd8
     vpunpckhbw ymm2, ymm1, ymm0
     vpunpcklbw ymm1, ymm1, ymm0
@@ -619,29 +646,29 @@ void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
 #endif  // HAS_ARGB1555TOARGBROW_AVX2
 
 #ifdef HAS_ARGB4444TOARGBROW_AVX2
-__declspec(naked)
-void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
-                            int width) {
+__declspec(naked) void ARGB4444ToARGBRow_AVX2(const uint8_t* src_argb4444,
+                                              uint8_t* dst_argb,
+                                              int width) {
   __asm {
     mov       eax,  0x0f0f0f0f  // generate mask 0x0f0f0f0f
     vmovd     xmm4, eax
     vbroadcastss ymm4, xmm4
-    vpslld    ymm5, ymm4, 4     // 0xf0f0f0f0 for high nibbles
-    mov       eax,  [esp + 4]   // src_argb4444
-    mov       edx,  [esp + 8]   // dst_argb
+    vpslld    ymm5, ymm4, 4  // 0xf0f0f0f0 for high nibbles
+    mov       eax,  [esp + 4]  // src_argb4444
+    mov       edx,  [esp + 8]  // dst_argb
     mov       ecx,  [esp + 12]  // width
     sub       edx,  eax
     sub       edx,  eax
 
  convertloop:
-    vmovdqu    ymm0, [eax]         // fetch 16 pixels of bgra4444
-    vpand      ymm2, ymm0, ymm5    // mask high nibbles
-    vpand      ymm0, ymm0, ymm4    // mask low nibbles
+    vmovdqu    ymm0, [eax]  // fetch 16 pixels of bgra4444
+    vpand      ymm2, ymm0, ymm5  // mask high nibbles
+    vpand      ymm0, ymm0, ymm4  // mask low nibbles
     vpsrlw     ymm3, ymm2, 4
     vpsllw     ymm1, ymm0, 4
     vpor       ymm2, ymm2, ymm3
     vpor       ymm0, ymm0, ymm1
-    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm0, ymm0, 0xd8  // mutate for unpack
     vpermq     ymm2, ymm2, 0xd8
     vpunpckhbw ymm1, ymm0, ymm2
     vpunpcklbw ymm0, ymm0, ymm2
@@ -657,9 +684,9 @@ void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
 #endif  // HAS_ARGB4444TOARGBROW_AVX2
 
 // 24 instructions
-__declspec(naked)
-void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
-                            int width) {
+__declspec(naked) void ARGB1555ToARGBRow_SSE2(const uint8_t* src_argb1555,
+                                              uint8_t* dst_argb,
+                                              int width) {
   __asm {
     mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
     movd      xmm5, eax
@@ -667,36 +694,36 @@ void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
     mov       eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
     movd      xmm6, eax
     pshufd    xmm6, xmm6, 0
-    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    pcmpeqb   xmm3, xmm3  // generate mask 0xf800f800 for Red
     psllw     xmm3, 11
-    movdqa    xmm4, xmm3       // generate mask 0x03e003e0 for Green
+    movdqa    xmm4, xmm3  // generate mask 0x03e003e0 for Green
     psrlw     xmm4, 6
-    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    pcmpeqb   xmm7, xmm7  // generate mask 0xff00ff00 for Alpha
     psllw     xmm7, 8
 
-    mov       eax, [esp + 4]   // src_argb1555
-    mov       edx, [esp + 8]   // dst_argb
+    mov       eax, [esp + 4]  // src_argb1555
+    mov       edx, [esp + 8]  // dst_argb
     mov       ecx, [esp + 12]  // width
     sub       edx, eax
     sub       edx, eax
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
+    movdqu    xmm0, [eax]  // fetch 8 pixels of 1555
     movdqa    xmm1, xmm0
     movdqa    xmm2, xmm0
-    psllw     xmm1, 1       // R in upper 5 bits
-    psllw     xmm2, 11      // B in upper 5 bits
+    psllw     xmm1, 1  // R in upper 5 bits
+    psllw     xmm2, 11  // B in upper 5 bits
     pand      xmm1, xmm3
-    pmulhuw   xmm2, xmm5    // * (256 + 8)
-    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    pmulhuw   xmm2, xmm5  // * (256 + 8)
+    pmulhuw   xmm1, xmm5  // * (256 + 8)
     psllw     xmm1, 8
-    por       xmm1, xmm2    // RB
+    por       xmm1, xmm2  // RB
     movdqa    xmm2, xmm0
-    pand      xmm0, xmm4    // G in middle 5 bits
-    psraw     xmm2, 8       // A
-    pmulhuw   xmm0, xmm6    // << 6 * (256 + 8)
+    pand      xmm0, xmm4  // G in middle 5 bits
+    psraw     xmm2, 8  // A
+    pmulhuw   xmm0, xmm6  // << 6 * (256 + 8)
     pand      xmm2, xmm7
-    por       xmm0, xmm2    // AG
+    por       xmm0, xmm2  // AG
     movdqa    xmm2, xmm1
     punpcklbw xmm1, xmm0
     punpckhbw xmm2, xmm0
@@ -710,26 +737,26 @@ void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
 }
 
 // 18 instructions.
-__declspec(naked)
-void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
-                            int width) {
+__declspec(naked) void ARGB4444ToARGBRow_SSE2(const uint8_t* src_argb4444,
+                                              uint8_t* dst_argb,
+                                              int width) {
   __asm {
     mov       eax, 0x0f0f0f0f  // generate mask 0x0f0f0f0f
     movd      xmm4, eax
     pshufd    xmm4, xmm4, 0
-    movdqa    xmm5, xmm4       // 0xf0f0f0f0 for high nibbles
+    movdqa    xmm5, xmm4  // 0xf0f0f0f0 for high nibbles
     pslld     xmm5, 4
-    mov       eax, [esp + 4]   // src_argb4444
-    mov       edx, [esp + 8]   // dst_argb
+    mov       eax, [esp + 4]  // src_argb4444
+    mov       edx, [esp + 8]  // dst_argb
     mov       ecx, [esp + 12]  // width
     sub       edx, eax
     sub       edx, eax
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
+    movdqu    xmm0, [eax]  // fetch 8 pixels of bgra4444
     movdqa    xmm2, xmm0
-    pand      xmm0, xmm4    // mask low nibbles
-    pand      xmm2, xmm5    // mask high nibbles
+    pand      xmm0, xmm4  // mask low nibbles
+    pand      xmm2, xmm5  // mask high nibbles
     movdqa    xmm1, xmm0
     movdqa    xmm3, xmm2
     psllw     xmm1, 4
@@ -748,37 +775,38 @@ void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
   }
 }
 
-__declspec(naked)
-void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToRGB24Row_SSSE3(const uint8_t* src_argb,
+                                            uint8_t* dst_rgb,
+                                            int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_argb
-    mov       edx, [esp + 8]   // dst_rgb
+    mov       eax, [esp + 4]  // src_argb
+    mov       edx, [esp + 8]  // dst_rgb
     mov       ecx, [esp + 12]  // width
     movdqa    xmm6, xmmword ptr kShuffleMaskARGBToRGB24
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm0, [eax]  // fetch 16 pixels of argb
     movdqu    xmm1, [eax + 16]
     movdqu    xmm2, [eax + 32]
     movdqu    xmm3, [eax + 48]
     lea       eax, [eax + 64]
-    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm0, xmm6  // pack 16 bytes of ARGB to 12 bytes of RGB
     pshufb    xmm1, xmm6
     pshufb    xmm2, xmm6
     pshufb    xmm3, xmm6
-    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
-    psrldq    xmm1, 4      // 8 bytes from 1
-    pslldq    xmm4, 12     // 4 bytes from 1 for 0
-    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
-    por       xmm0, xmm4   // 4 bytes from 1 for 0
-    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqa    xmm4, xmm1  // 4 bytes from 1 for 0
+    psrldq    xmm1, 4  // 8 bytes from 1
+    pslldq    xmm4, 12  // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2  // 8 bytes from 2 for 1
+    por       xmm0, xmm4  // 4 bytes from 1 for 0
+    pslldq    xmm5, 8  // 8 bytes from 2 for 1
     movdqu    [edx], xmm0  // store 0
-    por       xmm1, xmm5   // 8 bytes from 2 for 1
-    psrldq    xmm2, 8      // 4 bytes from 2
-    pslldq    xmm3, 4      // 12 bytes from 3 for 2
-    por       xmm2, xmm3   // 12 bytes from 3 for 2
-    movdqu    [edx + 16], xmm1   // store 1
-    movdqu    [edx + 32], xmm2   // store 2
+    por       xmm1, xmm5  // 8 bytes from 2 for 1
+    psrldq    xmm2, 8  // 4 bytes from 2
+    pslldq    xmm3, 4  // 12 bytes from 3 for 2
+    por       xmm2, xmm3  // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1  // store 1
+    movdqu    [edx + 32], xmm2  // store 2
     lea       edx, [edx + 48]
     sub       ecx, 16
     jg        convertloop
@@ -786,37 +814,38 @@ void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
-__declspec(naked)
-void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToRAWRow_SSSE3(const uint8_t* src_argb,
+                                          uint8_t* dst_rgb,
+                                          int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_argb
-    mov       edx, [esp + 8]   // dst_rgb
+    mov       eax, [esp + 4]  // src_argb
+    mov       edx, [esp + 8]  // dst_rgb
     mov       ecx, [esp + 12]  // width
     movdqa    xmm6, xmmword ptr kShuffleMaskARGBToRAW
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm0, [eax]  // fetch 16 pixels of argb
     movdqu    xmm1, [eax + 16]
     movdqu    xmm2, [eax + 32]
     movdqu    xmm3, [eax + 48]
     lea       eax, [eax + 64]
-    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm0, xmm6  // pack 16 bytes of ARGB to 12 bytes of RGB
     pshufb    xmm1, xmm6
     pshufb    xmm2, xmm6
     pshufb    xmm3, xmm6
-    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
-    psrldq    xmm1, 4      // 8 bytes from 1
-    pslldq    xmm4, 12     // 4 bytes from 1 for 0
-    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
-    por       xmm0, xmm4   // 4 bytes from 1 for 0
-    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqa    xmm4, xmm1  // 4 bytes from 1 for 0
+    psrldq    xmm1, 4  // 8 bytes from 1
+    pslldq    xmm4, 12  // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2  // 8 bytes from 2 for 1
+    por       xmm0, xmm4  // 4 bytes from 1 for 0
+    pslldq    xmm5, 8  // 8 bytes from 2 for 1
     movdqu    [edx], xmm0  // store 0
-    por       xmm1, xmm5   // 8 bytes from 2 for 1
-    psrldq    xmm2, 8      // 4 bytes from 2
-    pslldq    xmm3, 4      // 12 bytes from 3 for 2
-    por       xmm2, xmm3   // 12 bytes from 3 for 2
-    movdqu    [edx + 16], xmm1   // store 1
-    movdqu    [edx + 32], xmm2   // store 2
+    por       xmm1, xmm5  // 8 bytes from 2 for 1
+    psrldq    xmm2, 8  // 4 bytes from 2
+    pslldq    xmm3, 4  // 12 bytes from 3 for 2
+    por       xmm2, xmm3  // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1  // store 1
+    movdqu    [edx + 32], xmm2  // store 2
     lea       edx, [edx + 48]
     sub       ecx, 16
     jg        convertloop
@@ -824,33 +853,34 @@ void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
-__declspec(naked)
-void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToRGB565Row_SSE2(const uint8_t* src_argb,
+                                            uint8_t* dst_rgb,
+                                            int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_argb
-    mov       edx, [esp + 8]   // dst_rgb
+    mov       eax, [esp + 4]  // src_argb
+    mov       edx, [esp + 8]  // dst_rgb
     mov       ecx, [esp + 12]  // width
-    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    pcmpeqb   xmm3, xmm3  // generate mask 0x0000001f
     psrld     xmm3, 27
-    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    pcmpeqb   xmm4, xmm4  // generate mask 0x000007e0
     psrld     xmm4, 26
     pslld     xmm4, 5
-    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pcmpeqb   xmm5, xmm5  // generate mask 0xfffff800
     pslld     xmm5, 11
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
-    movdqa    xmm1, xmm0    // B
-    movdqa    xmm2, xmm0    // G
-    pslld     xmm0, 8       // R
-    psrld     xmm1, 3       // B
-    psrld     xmm2, 5       // G
-    psrad     xmm0, 16      // R
-    pand      xmm1, xmm3    // B
-    pand      xmm2, xmm4    // G
-    pand      xmm0, xmm5    // R
-    por       xmm1, xmm2    // BG
-    por       xmm0, xmm1    // BGR
+    movdqu    xmm0, [eax]  // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0  // B
+    movdqa    xmm2, xmm0  // G
+    pslld     xmm0, 8  // R
+    psrld     xmm1, 3  // B
+    psrld     xmm2, 5  // G
+    psrad     xmm0, 16  // R
+    pand      xmm1, xmm3  // B
+    pand      xmm2, xmm4  // G
+    pand      xmm0, xmm5  // R
+    por       xmm1, xmm2  // BG
+    por       xmm0, xmm1  // BGR
     packssdw  xmm0, xmm0
     lea       eax, [eax + 16]
     movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
@@ -861,41 +891,42 @@ void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
-__declspec(naked)
-void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width) {
+__declspec(naked) void ARGBToRGB565DitherRow_SSE2(const uint8_t* src_argb,
+                                                  uint8_t* dst_rgb,
+                                                  const uint32_t dither4,
+                                                  int width) {
   __asm {
 
-    mov       eax, [esp + 4]   // src_argb
-    mov       edx, [esp + 8]   // dst_rgb
-    movd      xmm6, [esp + 12] // dither4
+    mov       eax, [esp + 4]  // src_argb
+    mov       edx, [esp + 8]  // dst_rgb
+    movd      xmm6, [esp + 12]  // dither4
     mov       ecx, [esp + 16]  // width
-    punpcklbw xmm6, xmm6       // make dither 16 bytes
+    punpcklbw xmm6, xmm6  // make dither 16 bytes
     movdqa    xmm7, xmm6
     punpcklwd xmm6, xmm6
     punpckhwd xmm7, xmm7
-    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    pcmpeqb   xmm3, xmm3  // generate mask 0x0000001f
     psrld     xmm3, 27
-    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    pcmpeqb   xmm4, xmm4  // generate mask 0x000007e0
     psrld     xmm4, 26
     pslld     xmm4, 5
-    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pcmpeqb   xmm5, xmm5  // generate mask 0xfffff800
     pslld     xmm5, 11
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
-    paddusb   xmm0, xmm6    // add dither
-    movdqa    xmm1, xmm0    // B
-    movdqa    xmm2, xmm0    // G
-    pslld     xmm0, 8       // R
-    psrld     xmm1, 3       // B
-    psrld     xmm2, 5       // G
-    psrad     xmm0, 16      // R
-    pand      xmm1, xmm3    // B
-    pand      xmm2, xmm4    // G
-    pand      xmm0, xmm5    // R
-    por       xmm1, xmm2    // BG
-    por       xmm0, xmm1    // BGR
+    movdqu    xmm0, [eax]  // fetch 4 pixels of argb
+    paddusb   xmm0, xmm6  // add dither
+    movdqa    xmm1, xmm0  // B
+    movdqa    xmm2, xmm0  // G
+    pslld     xmm0, 8  // R
+    psrld     xmm1, 3  // B
+    psrld     xmm2, 5  // G
+    psrad     xmm0, 16  // R
+    pand      xmm1, xmm3  // B
+    pand      xmm2, xmm4  // G
+    pand      xmm0, xmm5  // R
+    por       xmm1, xmm2  // BG
+    por       xmm0, xmm1  // BGR
     packssdw  xmm0, xmm0
     lea       eax, [eax + 16]
     movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
@@ -907,39 +938,40 @@ void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
 }
 
 #ifdef HAS_ARGBTORGB565DITHERROW_AVX2
-__declspec(naked)
-void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
-                                const uint32 dither4, int width) {
+__declspec(naked) void ARGBToRGB565DitherRow_AVX2(const uint8_t* src_argb,
+                                                  uint8_t* dst_rgb,
+                                                  const uint32_t dither4,
+                                                  int width) {
   __asm {
-    mov        eax, [esp + 4]      // src_argb
-    mov        edx, [esp + 8]      // dst_rgb
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_rgb
     vbroadcastss xmm6, [esp + 12]  // dither4
-    mov        ecx, [esp + 16]     // width
-    vpunpcklbw xmm6, xmm6, xmm6    // make dither 32 bytes
+    mov        ecx, [esp + 16]  // width
+    vpunpcklbw xmm6, xmm6, xmm6  // make dither 32 bytes
     vpermq     ymm6, ymm6, 0xd8
     vpunpcklwd ymm6, ymm6, ymm6
-    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    vpcmpeqb   ymm3, ymm3, ymm3  // generate mask 0x0000001f
     vpsrld     ymm3, ymm3, 27
-    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpcmpeqb   ymm4, ymm4, ymm4  // generate mask 0x000007e0
     vpsrld     ymm4, ymm4, 26
     vpslld     ymm4, ymm4, 5
-    vpslld     ymm5, ymm3, 11      // generate mask 0x0000f800
+    vpslld     ymm5, ymm3, 11  // generate mask 0x0000f800
 
  convertloop:
-    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
-    vpaddusb   ymm0, ymm0, ymm6    // add dither
-    vpsrld     ymm2, ymm0, 5       // G
-    vpsrld     ymm1, ymm0, 3       // B
-    vpsrld     ymm0, ymm0, 8       // R
-    vpand      ymm2, ymm2, ymm4    // G
-    vpand      ymm1, ymm1, ymm3    // B
-    vpand      ymm0, ymm0, ymm5    // R
-    vpor       ymm1, ymm1, ymm2    // BG
-    vpor       ymm0, ymm0, ymm1    // BGR
+    vmovdqu    ymm0, [eax]  // fetch 8 pixels of argb
+    vpaddusb   ymm0, ymm0, ymm6  // add dither
+    vpsrld     ymm2, ymm0, 5  // G
+    vpsrld     ymm1, ymm0, 3  // B
+    vpsrld     ymm0, ymm0, 8  // R
+    vpand      ymm2, ymm2, ymm4  // G
+    vpand      ymm1, ymm1, ymm3  // B
+    vpand      ymm0, ymm0, ymm5  // R
+    vpor       ymm1, ymm1, ymm2  // BG
+    vpor       ymm0, ymm0, ymm1  // BGR
     vpackusdw  ymm0, ymm0, ymm0
     vpermq     ymm0, ymm0, 0xd8
     lea        eax, [eax + 32]
-    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    vmovdqu    [edx], xmm0  // store 8 pixels of RGB565
     lea        edx, [edx + 16]
     sub        ecx, 8
     jg         convertloop
@@ -950,37 +982,38 @@ void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
 #endif  // HAS_ARGBTORGB565DITHERROW_AVX2
 
 // TODO(fbarchard): Improve sign extension/packing.
-__declspec(naked)
-void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToARGB1555Row_SSE2(const uint8_t* src_argb,
+                                              uint8_t* dst_rgb,
+                                              int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_argb
-    mov       edx, [esp + 8]   // dst_rgb
+    mov       eax, [esp + 4]  // src_argb
+    mov       edx, [esp + 8]  // dst_rgb
     mov       ecx, [esp + 12]  // width
-    pcmpeqb   xmm4, xmm4       // generate mask 0x0000001f
+    pcmpeqb   xmm4, xmm4  // generate mask 0x0000001f
     psrld     xmm4, 27
-    movdqa    xmm5, xmm4       // generate mask 0x000003e0
+    movdqa    xmm5, xmm4  // generate mask 0x000003e0
     pslld     xmm5, 5
-    movdqa    xmm6, xmm4       // generate mask 0x00007c00
+    movdqa    xmm6, xmm4  // generate mask 0x00007c00
     pslld     xmm6, 10
-    pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
+    pcmpeqb   xmm7, xmm7  // generate mask 0xffff8000
     pslld     xmm7, 15
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
-    movdqa    xmm1, xmm0    // B
-    movdqa    xmm2, xmm0    // G
-    movdqa    xmm3, xmm0    // R
-    psrad     xmm0, 16      // A
-    psrld     xmm1, 3       // B
-    psrld     xmm2, 6       // G
-    psrld     xmm3, 9       // R
-    pand      xmm0, xmm7    // A
-    pand      xmm1, xmm4    // B
-    pand      xmm2, xmm5    // G
-    pand      xmm3, xmm6    // R
-    por       xmm0, xmm1    // BA
-    por       xmm2, xmm3    // GR
-    por       xmm0, xmm2    // BGRA
+    movdqu    xmm0, [eax]  // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0  // B
+    movdqa    xmm2, xmm0  // G
+    movdqa    xmm3, xmm0  // R
+    psrad     xmm0, 16  // A
+    psrld     xmm1, 3  // B
+    psrld     xmm2, 6  // G
+    psrld     xmm3, 9  // R
+    pand      xmm0, xmm7  // A
+    pand      xmm1, xmm4  // B
+    pand      xmm2, xmm5  // G
+    pand      xmm3, xmm6  // R
+    por       xmm0, xmm1  // BA
+    por       xmm2, xmm3  // GR
+    por       xmm0, xmm2  // BGRA
     packssdw  xmm0, xmm0
     lea       eax, [eax + 16]
     movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB1555
@@ -991,22 +1024,23 @@ void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
-__declspec(naked)
-void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToARGB4444Row_SSE2(const uint8_t* src_argb,
+                                              uint8_t* dst_rgb,
+                                              int width) {
   __asm {
-    mov       eax, [esp + 4]   // src_argb
-    mov       edx, [esp + 8]   // dst_rgb
+    mov       eax, [esp + 4]  // src_argb
+    mov       edx, [esp + 8]  // dst_rgb
     mov       ecx, [esp + 12]  // width
-    pcmpeqb   xmm4, xmm4       // generate mask 0xf000f000
+    pcmpeqb   xmm4, xmm4  // generate mask 0xf000f000
     psllw     xmm4, 12
-    movdqa    xmm3, xmm4       // generate mask 0x00f000f0
+    movdqa    xmm3, xmm4  // generate mask 0x00f000f0
     psrlw     xmm3, 8
 
  convertloop:
-    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqu    xmm0, [eax]  // fetch 4 pixels of argb
     movdqa    xmm1, xmm0
-    pand      xmm0, xmm3    // low nibble
-    pand      xmm1, xmm4    // high nibble
+    pand      xmm0, xmm3  // low nibble
+    pand      xmm1, xmm4  // high nibble
     psrld     xmm0, 4
     psrld     xmm1, 8
     por       xmm0, xmm1
@@ -1021,33 +1055,34 @@ void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width) {
 }
 
 #ifdef HAS_ARGBTORGB565ROW_AVX2
-__declspec(naked)
-void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToRGB565Row_AVX2(const uint8_t* src_argb,
+                                            uint8_t* dst_rgb,
+                                            int width) {
   __asm {
-    mov        eax, [esp + 4]      // src_argb
-    mov        edx, [esp + 8]      // dst_rgb
-    mov        ecx, [esp + 12]     // width
-    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_rgb
+    mov        ecx, [esp + 12]  // width
+    vpcmpeqb   ymm3, ymm3, ymm3  // generate mask 0x0000001f
     vpsrld     ymm3, ymm3, 27
-    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpcmpeqb   ymm4, ymm4, ymm4  // generate mask 0x000007e0
     vpsrld     ymm4, ymm4, 26
     vpslld     ymm4, ymm4, 5
-    vpslld     ymm5, ymm3, 11      // generate mask 0x0000f800
+    vpslld     ymm5, ymm3, 11  // generate mask 0x0000f800
 
  convertloop:
-    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
-    vpsrld     ymm2, ymm0, 5       // G
-    vpsrld     ymm1, ymm0, 3       // B
-    vpsrld     ymm0, ymm0, 8       // R
-    vpand      ymm2, ymm2, ymm4    // G
-    vpand      ymm1, ymm1, ymm3    // B
-    vpand      ymm0, ymm0, ymm5    // R
-    vpor       ymm1, ymm1, ymm2    // BG
-    vpor       ymm0, ymm0, ymm1    // BGR
+    vmovdqu    ymm0, [eax]  // fetch 8 pixels of argb
+    vpsrld     ymm2, ymm0, 5  // G
+    vpsrld     ymm1, ymm0, 3  // B
+    vpsrld     ymm0, ymm0, 8  // R
+    vpand      ymm2, ymm2, ymm4  // G
+    vpand      ymm1, ymm1, ymm3  // B
+    vpand      ymm0, ymm0, ymm5  // R
+    vpor       ymm1, ymm1, ymm2  // BG
+    vpor       ymm0, ymm0, ymm1  // BGR
     vpackusdw  ymm0, ymm0, ymm0
     vpermq     ymm0, ymm0, 0xd8
     lea        eax, [eax + 32]
-    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    vmovdqu    [edx], xmm0  // store 8 pixels of RGB565
     lea        edx, [edx + 16]
     sub        ecx, 8
     jg         convertloop
@@ -1058,36 +1093,37 @@ void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width) {
 #endif  // HAS_ARGBTORGB565ROW_AVX2
 
 #ifdef HAS_ARGBTOARGB1555ROW_AVX2
-__declspec(naked)
-void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToARGB1555Row_AVX2(const uint8_t* src_argb,
+                                              uint8_t* dst_rgb,
+                                              int width) {
   __asm {
-    mov        eax, [esp + 4]      // src_argb
-    mov        edx, [esp + 8]      // dst_rgb
-    mov        ecx, [esp + 12]     // width
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_rgb
+    mov        ecx, [esp + 12]  // width
     vpcmpeqb   ymm4, ymm4, ymm4
-    vpsrld     ymm4, ymm4, 27      // generate mask 0x0000001f
-    vpslld     ymm5, ymm4, 5       // generate mask 0x000003e0
-    vpslld     ymm6, ymm4, 10      // generate mask 0x00007c00
-    vpcmpeqb   ymm7, ymm7, ymm7    // generate mask 0xffff8000
+    vpsrld     ymm4, ymm4, 27  // generate mask 0x0000001f
+    vpslld     ymm5, ymm4, 5  // generate mask 0x000003e0
+    vpslld     ymm6, ymm4, 10  // generate mask 0x00007c00
+    vpcmpeqb   ymm7, ymm7, ymm7  // generate mask 0xffff8000
     vpslld     ymm7, ymm7, 15
 
  convertloop:
-    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
-    vpsrld     ymm3, ymm0, 9       // R
-    vpsrld     ymm2, ymm0, 6       // G
-    vpsrld     ymm1, ymm0, 3       // B
-    vpsrad     ymm0, ymm0, 16      // A
-    vpand      ymm3, ymm3, ymm6    // R
-    vpand      ymm2, ymm2, ymm5    // G
-    vpand      ymm1, ymm1, ymm4    // B
-    vpand      ymm0, ymm0, ymm7    // A
-    vpor       ymm0, ymm0, ymm1    // BA
-    vpor       ymm2, ymm2, ymm3    // GR
-    vpor       ymm0, ymm0, ymm2    // BGRA
+    vmovdqu    ymm0, [eax]  // fetch 8 pixels of argb
+    vpsrld     ymm3, ymm0, 9  // R
+    vpsrld     ymm2, ymm0, 6  // G
+    vpsrld     ymm1, ymm0, 3  // B
+    vpsrad     ymm0, ymm0, 16  // A
+    vpand      ymm3, ymm3, ymm6  // R
+    vpand      ymm2, ymm2, ymm5  // G
+    vpand      ymm1, ymm1, ymm4  // B
+    vpand      ymm0, ymm0, ymm7  // A
+    vpor       ymm0, ymm0, ymm1  // BA
+    vpor       ymm2, ymm2, ymm3  // GR
+    vpor       ymm0, ymm0, ymm2  // BGRA
     vpackssdw  ymm0, ymm0, ymm0
     vpermq     ymm0, ymm0, 0xd8
     lea        eax, [eax + 32]
-    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB1555
+    vmovdqu    [edx], xmm0  // store 8 pixels of ARGB1555
     lea        edx, [edx + 16]
     sub        ecx, 8
     jg         convertloop
@@ -1098,27 +1134,28 @@ void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width) {
 #endif  // HAS_ARGBTOARGB1555ROW_AVX2
 
 #ifdef HAS_ARGBTOARGB4444ROW_AVX2
-__declspec(naked)
-void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width) {
+__declspec(naked) void ARGBToARGB4444Row_AVX2(const uint8_t* src_argb,
+                                              uint8_t* dst_rgb,
+                                              int width) {
   __asm {
-    mov        eax, [esp + 4]   // src_argb
-    mov        edx, [esp + 8]   // dst_rgb
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_rgb
     mov        ecx, [esp + 12]  // width
-    vpcmpeqb   ymm4, ymm4, ymm4   // generate mask 0xf000f000
+    vpcmpeqb   ymm4, ymm4, ymm4  // generate mask 0xf000f000
     vpsllw     ymm4, ymm4, 12
-    vpsrlw     ymm3, ymm4, 8      // generate mask 0x00f000f0
+    vpsrlw     ymm3, ymm4, 8  // generate mask 0x00f000f0
 
  convertloop:
-    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
-    vpand      ymm1, ymm0, ymm4    // high nibble
-    vpand      ymm0, ymm0, ymm3    // low nibble
+    vmovdqu    ymm0, [eax]  // fetch 8 pixels of argb
+    vpand      ymm1, ymm0, ymm4  // high nibble
+    vpand      ymm0, ymm0, ymm3  // low nibble
     vpsrld     ymm1, ymm1, 8
     vpsrld     ymm0, ymm0, 4
     vpor       ymm0, ymm0, ymm1
     vpackuswb  ymm0, ymm0, ymm0
     vpermq     ymm0, ymm0, 0xd8
     lea        eax, [eax + 32]
-    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB4444
+    vmovdqu    [edx], xmm0  // store 8 pixels of ARGB4444
     lea        edx, [edx + 16]
     sub        ecx, 8
     jg         convertloop
@@ -1129,12 +1166,13 @@ void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width) {
 #endif  // HAS_ARGBTOARGB4444ROW_AVX2
 
 // Convert 16 ARGB pixels (64 bytes) to 16 Y values.
-__declspec(naked)
-void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void ARGBToYRow_SSSE3(const uint8_t* src_argb,
+                                        uint8_t* dst_y,
+                                        int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     movdqa     xmm4, xmmword ptr kARGBToY
     movdqa     xmm5, xmmword ptr kAddY16
 
@@ -1164,12 +1202,13 @@ void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
 
 // Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
 // Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
-__declspec(naked)
-void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void ARGBToYJRow_SSSE3(const uint8_t* src_argb,
+                                         uint8_t* dst_y,
+                                         int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     movdqa     xmm4, xmmword ptr kARGBToYJ
     movdqa     xmm5, xmmword ptr kAddYJ64
 
@@ -1200,17 +1239,16 @@ void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
 
 #ifdef HAS_ARGBTOYROW_AVX2
 // vpermd for vphaddw + vpackuswb vpermd.
-static const lvec32 kPermdARGBToY_AVX = {
-  0, 4, 1, 5, 2, 6, 3, 7
-};
+static const lvec32 kPermdARGBToY_AVX = {0, 4, 1, 5, 2, 6, 3, 7};
 
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
-__declspec(naked)
-void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void ARGBToYRow_AVX2(const uint8_t* src_argb,
+                                       uint8_t* dst_y,
+                                       int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     vbroadcastf128 ymm4, xmmword ptr kARGBToY
     vbroadcastf128 ymm5, xmmword ptr kAddY16
     vmovdqu    ymm6, ymmword ptr kPermdARGBToY_AVX
@@ -1244,12 +1282,13 @@ void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int width) {
 
 #ifdef HAS_ARGBTOYJROW_AVX2
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
-__declspec(naked)
-void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void ARGBToYJRow_AVX2(const uint8_t* src_argb,
+                                        uint8_t* dst_y,
+                                        int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     vbroadcastf128 ymm4, xmmword ptr kARGBToYJ
     vbroadcastf128 ymm5, xmmword ptr kAddYJ64
     vmovdqu    ymm6, ymmword ptr kPermdARGBToY_AVX
@@ -1283,12 +1322,13 @@ void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int width) {
 }
 #endif  //  HAS_ARGBTOYJROW_AVX2
 
-__declspec(naked)
-void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void BGRAToYRow_SSSE3(const uint8_t* src_argb,
+                                        uint8_t* dst_y,
+                                        int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     movdqa     xmm4, xmmword ptr kBGRAToY
     movdqa     xmm5, xmmword ptr kAddY16
 
@@ -1316,12 +1356,13 @@ void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
   }
 }
 
-__declspec(naked)
-void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void ABGRToYRow_SSSE3(const uint8_t* src_argb,
+                                        uint8_t* dst_y,
+                                        int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     movdqa     xmm4, xmmword ptr kABGRToY
     movdqa     xmm5, xmmword ptr kAddY16
 
@@ -1349,12 +1390,13 @@ void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
   }
 }
 
-__declspec(naked)
-void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
+__declspec(naked) void RGBAToYRow_SSSE3(const uint8_t* src_argb,
+                                        uint8_t* dst_y,
+                                        int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_y */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_y */
+    mov        ecx, [esp + 12] /* width */
     movdqa     xmm4, xmmword ptr kRGBAToY
     movdqa     xmm5, xmmword ptr kAddY16
 
@@ -1382,24 +1424,26 @@ void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width) {
   }
 }
 
-__declspec(naked)
-void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
+                                         int src_stride_argb,
+                                         uint8_t* dst_u,
+                                         uint8_t* dst_v,
+                                         int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kARGBToV
     movdqa     xmm7, xmmword ptr kARGBToU
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx  // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+         /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
     movdqu     xmm4, [eax + esi]
     pavgb      xmm0, xmm4
@@ -1423,9 +1467,9 @@ void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     shufps     xmm4, xmm3, 0xdd
     pavgb      xmm2, xmm4
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 16 different pixels, its 8 pixels of U and 8 of V
     movdqa     xmm1, xmm0
     movdqa     xmm3, xmm2
     pmaddubsw  xmm0, xmm7  // U
@@ -1437,11 +1481,11 @@ void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     psraw      xmm0, 8
     psraw      xmm1, 8
     packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    paddb      xmm0, xmm5  // -> unsigned
 
-    // step 3 - store 8 U and 8 V values
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
+        // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0  // U
+    movhps     qword ptr [edx + edi], xmm0  // V
     lea        edx, [edx + 8]
     sub        ecx, 16
     jg         convertloop
@@ -1452,24 +1496,26 @@ void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
   }
 }
 
-__declspec(naked)
-void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                        uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
+                                          int src_stride_argb,
+                                          uint8_t* dst_u,
+                                          uint8_t* dst_v,
+                                          int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUVJ128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kARGBToVJ
     movdqa     xmm7, xmmword ptr kARGBToUJ
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx  // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+         /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
     movdqu     xmm4, [eax + esi]
     pavgb      xmm0, xmm4
@@ -1493,9 +1539,9 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     shufps     xmm4, xmm3, 0xdd
     pavgb      xmm2, xmm4
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 16 different pixels, its 8 pixels of U and 8 of V
     movdqa     xmm1, xmm0
     movdqa     xmm3, xmm2
     pmaddubsw  xmm0, xmm7  // U
@@ -1510,9 +1556,9 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     psraw      xmm1, 8
     packsswb   xmm0, xmm1
 
-    // step 3 - store 8 U and 8 V values
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
+        // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0  // U
+    movhps     qword ptr [edx + edi], xmm0  // V
     lea        edx, [edx + 8]
     sub        ecx, 16
     jg         convertloop
@@ -1524,24 +1570,26 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
 }
 
 #ifdef HAS_ARGBTOUVROW_AVX2
-__declspec(naked)
-void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb,
+                                        int src_stride_argb,
+                                        uint8_t* dst_u,
+                                        uint8_t* dst_v,
+                                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    vbroadcastf128 ymm5, xmmword ptr kAddUV128
+    vbroadcastf128 ymm5, xmmword ptr kBiasUV128
     vbroadcastf128 ymm6, xmmword ptr kARGBToV
     vbroadcastf128 ymm7, xmmword ptr kARGBToU
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx   // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+        /* step 1 - subsample 32x2 argb pixels to 16x1 */
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     vmovdqu    ymm2, [eax + 64]
@@ -1558,9 +1606,9 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     vshufps    ymm2, ymm2, ymm3, 0xdd
     vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 32 different pixels, its 16 pixels of U and 16 of V
     vpmaddubsw ymm1, ymm0, ymm7  // U
     vpmaddubsw ymm3, ymm2, ymm7
     vpmaddubsw ymm0, ymm0, ymm6  // V
@@ -1574,9 +1622,9 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     vpshufb    ymm0, ymm0, ymmword ptr kShufARGBToUV_AVX  // for vshufps/vphaddw
     vpaddb     ymm0, ymm0, ymm5  // -> unsigned
 
-    // step 3 - store 16 U and 16 V values
-    vextractf128 [edx], ymm0, 0 // U
-    vextractf128 [edx + edi], ymm0, 1 // V
+        // step 3 - store 16 U and 16 V values
+    vextractf128 [edx], ymm0, 0  // U
+    vextractf128 [edx + edi], ymm0, 1  // V
     lea        edx, [edx + 16]
     sub        ecx, 32
     jg         convertloop
@@ -1590,24 +1638,26 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
 #endif  // HAS_ARGBTOUVROW_AVX2
 
 #ifdef HAS_ARGBTOUVJROW_AVX2
-__declspec(naked)
-void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
+                                         int src_stride_argb,
+                                         uint8_t* dst_u,
+                                         uint8_t* dst_v,
+                                         int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    vbroadcastf128 ymm5, xmmword ptr kAddUV128
-    vbroadcastf128 ymm6, xmmword ptr kARGBToV
-    vbroadcastf128 ymm7, xmmword ptr kARGBToU
-    sub        edi, edx             // stride from u to v
+    vbroadcastf128 ymm5, xmmword ptr kBiasUV128
+    vbroadcastf128 ymm6, xmmword ptr kARGBToVJ
+    vbroadcastf128 ymm7, xmmword ptr kARGBToUJ
+    sub        edi, edx   // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+        /* step 1 - subsample 32x2 argb pixels to 16x1 */
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     vmovdqu    ymm2, [eax + 64]
@@ -1624,9 +1674,9 @@ void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     vshufps    ymm2, ymm2, ymm3, 0xdd
     vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 32 different pixels, its 16 pixels of U and 16 of V
     vpmaddubsw ymm1, ymm0, ymm7  // U
     vpmaddubsw ymm3, ymm2, ymm7
     vpmaddubsw ymm0, ymm0, ymm6  // V
@@ -1641,9 +1691,9 @@ void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
     vpshufb    ymm0, ymm0, ymmword ptr kShufARGBToUV_AVX  // for vshufps/vphaddw
 
-    // step 3 - store 16 U and 16 V values
-    vextractf128 [edx], ymm0, 0 // U
-    vextractf128 [edx + edi], ymm0, 1 // V
+        // step 3 - store 16 U and 16 V values
+    vextractf128 [edx], ymm0, 0  // U
+    vextractf128 [edx + edi], ymm0, 1  // V
     lea        edx, [edx + 16]
     sub        ecx, 32
     jg         convertloop
@@ -1656,23 +1706,24 @@ void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb,
 }
 #endif  // HAS_ARGBTOUVJROW_AVX2
 
-__declspec(naked)
-void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
-                          uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
+                                            uint8_t* dst_u,
+                                            uint8_t* dst_v,
+                                            int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]   // src_argb
-    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        edx, [esp + 4 + 8]  // dst_u
     mov        edi, [esp + 4 + 12]  // dst_v
     mov        ecx, [esp + 4 + 16]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kARGBToV
     movdqa     xmm7, xmmword ptr kARGBToU
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx    // stride from u to v
 
  convertloop:
-    /* convert to U and V */
-    movdqu     xmm0, [eax]          // U
+        /* convert to U and V */
+    movdqu     xmm0, [eax]  // U
     movdqu     xmm1, [eax + 16]
     movdqu     xmm2, [eax + 32]
     movdqu     xmm3, [eax + 48]
@@ -1688,7 +1739,7 @@ void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
     paddb      xmm0, xmm5
     movdqu     [edx], xmm0
 
-    movdqu     xmm0, [eax]          // V
+    movdqu     xmm0, [eax]  // V
     movdqu     xmm1, [eax + 16]
     movdqu     xmm2, [eax + 32]
     movdqu     xmm3, [eax + 48]
@@ -1713,24 +1764,26 @@ void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
   }
 }
 
-__declspec(naked)
-void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void BGRAToUVRow_SSSE3(const uint8_t* src_argb,
+                                         int src_stride_argb,
+                                         uint8_t* dst_u,
+                                         uint8_t* dst_v,
+                                         int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kBGRAToV
     movdqa     xmm7, xmmword ptr kBGRAToU
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx  // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+         /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
     movdqu     xmm4, [eax + esi]
     pavgb      xmm0, xmm4
@@ -1754,9 +1807,9 @@ void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     shufps     xmm4, xmm3, 0xdd
     pavgb      xmm2, xmm4
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 16 different pixels, its 8 pixels of U and 8 of V
     movdqa     xmm1, xmm0
     movdqa     xmm3, xmm2
     pmaddubsw  xmm0, xmm7  // U
@@ -1768,11 +1821,11 @@ void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     psraw      xmm0, 8
     psraw      xmm1, 8
     packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    paddb      xmm0, xmm5  // -> unsigned
 
-    // step 3 - store 8 U and 8 V values
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
+        // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0  // U
+    movhps     qword ptr [edx + edi], xmm0  // V
     lea        edx, [edx + 8]
     sub        ecx, 16
     jg         convertloop
@@ -1783,24 +1836,26 @@ void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
   }
 }
 
-__declspec(naked)
-void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void ABGRToUVRow_SSSE3(const uint8_t* src_argb,
+                                         int src_stride_argb,
+                                         uint8_t* dst_u,
+                                         uint8_t* dst_v,
+                                         int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kABGRToV
     movdqa     xmm7, xmmword ptr kABGRToU
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx  // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+         /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
     movdqu     xmm4, [eax + esi]
     pavgb      xmm0, xmm4
@@ -1824,9 +1879,9 @@ void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     shufps     xmm4, xmm3, 0xdd
     pavgb      xmm2, xmm4
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 16 different pixels, its 8 pixels of U and 8 of V
     movdqa     xmm1, xmm0
     movdqa     xmm3, xmm2
     pmaddubsw  xmm0, xmm7  // U
@@ -1838,11 +1893,11 @@ void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     psraw      xmm0, 8
     psraw      xmm1, 8
     packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    paddb      xmm0, xmm5  // -> unsigned
 
-    // step 3 - store 8 U and 8 V values
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
+        // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0  // U
+    movhps     qword ptr [edx + edi], xmm0  // V
     lea        edx, [edx + 8]
     sub        ecx, 16
     jg         convertloop
@@ -1853,24 +1908,26 @@ void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
   }
 }
 
-__declspec(naked)
-void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void RGBAToUVRow_SSSE3(const uint8_t* src_argb,
+                                         int src_stride_argb,
+                                         uint8_t* dst_u,
+                                         uint8_t* dst_v,
+                                         int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_argb
-    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        eax, [esp + 8 + 4]  // src_argb
+    mov        esi, [esp + 8 + 8]  // src_stride_argb
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kRGBAToV
     movdqa     xmm7, xmmword ptr kRGBAToU
-    sub        edi, edx             // stride from u to v
+    sub        edi, edx  // stride from u to v
 
  convertloop:
-    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+         /* step 1 - subsample 16x2 argb pixels to 8x1 */
     movdqu     xmm0, [eax]
     movdqu     xmm4, [eax + esi]
     pavgb      xmm0, xmm4
@@ -1894,9 +1951,9 @@ void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     shufps     xmm4, xmm3, 0xdd
     pavgb      xmm2, xmm4
 
-    // step 2 - convert to U and V
-    // from here down is very similar to Y code except
-    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+        // step 2 - convert to U and V
+        // from here down is very similar to Y code except
+        // instead of 16 different pixels, its 8 pixels of U and 8 of V
     movdqa     xmm1, xmm0
     movdqa     xmm3, xmm2
     pmaddubsw  xmm0, xmm7  // U
@@ -1908,11 +1965,11 @@ void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     psraw      xmm0, 8
     psraw      xmm1, 8
     packsswb   xmm0, xmm1
-    paddb      xmm0, xmm5            // -> unsigned
+    paddb      xmm0, xmm5  // -> unsigned
 
-    // step 3 - store 8 U and 8 V values
-    movlps     qword ptr [edx], xmm0 // U
-    movhps     qword ptr [edx + edi], xmm0 // V
+        // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0  // U
+    movhps     qword ptr [edx + edi], xmm0  // V
     lea        edx, [edx + 8]
     sub        ecx, 16
     jg         convertloop
@@ -1925,179 +1982,180 @@ void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
 #endif  // HAS_ARGBTOYROW_SSSE3
 
 // Read 16 UV from 444
-#define READYUV444_AVX2 __asm {                                                \
-    __asm vmovdqu    xmm0, [esi]                  /* U */                      \
-    __asm vmovdqu    xmm1, [esi + edi]            /* V */                      \
+#define READYUV444_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* U */                                       \
+    __asm vmovdqu    xmm1, [esi + edi] /* V */                                 \
     __asm lea        esi,  [esi + 16]                                          \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpermq     ymm1, ymm1, 0xd8                                          \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
+    __asm vpermq     ymm4, ymm4, 0xd8                                          \
+    __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
+    __asm lea        eax, [eax + 16]}
+
+// Read 16 UV from 444.  With 16 Alpha.
+#define READYUVA444_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* U */                                       \
+    __asm vmovdqu    xmm1, [esi + edi] /* V */                                 \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
     __asm vpermq     ymm1, ymm1, 0xd8                                          \
-    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
-    __asm vmovdqu    xmm4, [eax]                  /* Y */                      \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm vmovdqu    xmm5, [ebp] /* A */                                       \
+    __asm vpermq     ymm5, ymm5, 0xd8                                          \
+    __asm lea        ebp, [ebp + 16]}
 
 // Read 8 UV from 422, upsample to 16 UV.
-#define READYUV422_AVX2 __asm {                                                \
-    __asm vmovq      xmm0, qword ptr [esi]        /* U */                      \
-    __asm vmovq      xmm1, qword ptr [esi + edi]  /* V */                      \
+#define READYUV422_AVX2 \
+  __asm {                                                                      \
+    __asm vmovq      xmm3, qword ptr [esi] /* U */                             \
+    __asm vmovq      xmm1, qword ptr [esi + edi] /* V */                       \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
-    __asm vmovdqu    xmm4, [eax]                  /* Y */                      \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklwd ymm3, ymm3, ymm3 /* UVUV (upsample) */                    \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
-    __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm lea        eax, [eax + 16]}
 
 // Read 8 UV from 422, upsample to 16 UV.  With 16 Alpha.
-#define READYUVA422_AVX2 __asm {                                               \
-    __asm vmovq      xmm0, qword ptr [esi]        /* U */                      \
-    __asm vmovq      xmm1, qword ptr [esi + edi]  /* V */                      \
+#define READYUVA422_AVX2 \
+  __asm {                                                                      \
+    __asm vmovq      xmm3, qword ptr [esi] /* U */                             \
+    __asm vmovq      xmm1, qword ptr [esi + edi] /* V */                       \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
-    __asm vmovdqu    xmm4, [eax]                  /* Y */                      \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklwd ymm3, ymm3, ymm3 /* UVUV (upsample) */                    \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]                                           \
-    __asm vmovdqu    xmm5, [ebp]                  /* A */                      \
+    __asm vmovdqu    xmm5, [ebp] /* A */                                       \
     __asm vpermq     ymm5, ymm5, 0xd8                                          \
-    __asm lea        ebp, [ebp + 16]                                           \
-  }
-
-// Read 4 UV from 411, upsample to 16 UV.
-#define READYUV411_AVX2 __asm {                                                \
-    __asm vmovd      xmm0, dword ptr [esi]        /* U */                      \
-    __asm vmovd      xmm1, dword ptr [esi + edi]  /* V */                      \
-    __asm lea        esi,  [esi + 4]                                           \
-    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
-    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpckldq ymm0, ymm0, ymm0             /* UVUVUVUV (upsample) */    \
-    __asm vmovdqu    xmm4, [eax]                  /* Y */                      \
-    __asm vpermq     ymm4, ymm4, 0xd8                                          \
-    __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
-    __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm lea        ebp, [ebp + 16]}
 
 // Read 8 UV from NV12, upsample to 16 UV.
-#define READNV12_AVX2 __asm {                                                  \
-    __asm vmovdqu    xmm0, [esi]                  /* UV */                     \
+#define READNV12_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* UV */                                      \
     __asm lea        esi,  [esi + 16]                                          \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
-    __asm vmovdqu    xmm4, [eax]                  /* Y */                      \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklwd ymm3, ymm3, ymm3 /* UVUV (upsample) */                    \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
-    __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm lea        eax, [eax + 16]}
 
 // Read 8 UV from NV21, upsample to 16 UV.
-#define READNV21_AVX2 __asm {                                                  \
-    __asm vmovdqu    xmm0, [esi]                  /* UV */                     \
+#define READNV21_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* UV */                                      \
     __asm lea        esi,  [esi + 16]                                          \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpshufb    ymm0, ymm0, ymmword ptr kShuffleNV21                      \
-    __asm vmovdqu    xmm4, [eax]                  /* Y */                      \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpshufb    ymm3, ymm3, ymmword ptr kShuffleNV21                      \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
-    __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm lea        eax, [eax + 16]}
 
 // Read 8 YUY2 with 16 Y and upsample 8 UV to 16 UV.
-#define READYUY2_AVX2 __asm {                                                  \
-    __asm vmovdqu    ymm4, [eax]          /* YUY2 */                           \
+#define READYUY2_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    ymm4, [eax] /* YUY2 */                                    \
     __asm vpshufb    ymm4, ymm4, ymmword ptr kShuffleYUY2Y                     \
-    __asm vmovdqu    ymm0, [eax]          /* UV */                             \
-    __asm vpshufb    ymm0, ymm0, ymmword ptr kShuffleYUY2UV                    \
-    __asm lea        eax, [eax + 32]                                           \
-  }
+    __asm vmovdqu    ymm3, [eax] /* UV */                                      \
+    __asm vpshufb    ymm3, ymm3, ymmword ptr kShuffleYUY2UV                    \
+    __asm lea        eax, [eax + 32]}
 
 // Read 8 UYVY with 16 Y and upsample 8 UV to 16 UV.
-#define READUYVY_AVX2 __asm {                                                  \
-    __asm vmovdqu    ymm4, [eax]          /* UYVY */                           \
+#define READUYVY_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    ymm4, [eax] /* UYVY */                                    \
     __asm vpshufb    ymm4, ymm4, ymmword ptr kShuffleUYVYY                     \
-    __asm vmovdqu    ymm0, [eax]          /* UV */                             \
-    __asm vpshufb    ymm0, ymm0, ymmword ptr kShuffleUYVYUV                    \
-    __asm lea        eax, [eax + 32]                                           \
-  }
+    __asm vmovdqu    ymm3, [eax] /* UV */                                      \
+    __asm vpshufb    ymm3, ymm3, ymmword ptr kShuffleUYVYUV                    \
+    __asm lea        eax, [eax + 32]}
 
 // Convert 16 pixels: 16 UV and 16 Y.
-#define YUVTORGB_AVX2(YuvConstants) __asm {                                    \
-    __asm vpmaddubsw ymm2, ymm0, ymmword ptr [YuvConstants + KUVTOR] /* R UV */\
-    __asm vpmaddubsw ymm1, ymm0, ymmword ptr [YuvConstants + KUVTOG] /* G UV */\
-    __asm vpmaddubsw ymm0, ymm0, ymmword ptr [YuvConstants + KUVTOB] /* B UV */\
-    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KUVBIASR]               \
-    __asm vpsubw     ymm2, ymm3, ymm2                                          \
-    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KUVBIASG]               \
-    __asm vpsubw     ymm1, ymm3, ymm1                                          \
-    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KUVBIASB]               \
-    __asm vpsubw     ymm0, ymm3, ymm0                                          \
-    /* Step 2: Find Y contribution to 16 R,G,B values */                       \
+#define YUVTORGB_AVX2(YuvConstants) \
+  __asm {                                                                      \
+    __asm vpsubb     ymm3, ymm3, ymmword ptr kBiasUV128                        \
     __asm vpmulhuw   ymm4, ymm4, ymmword ptr [YuvConstants + KYTORGB]          \
-    __asm vpaddsw    ymm0, ymm0, ymm4           /* B += Y */                   \
-    __asm vpaddsw    ymm1, ymm1, ymm4           /* G += Y */                   \
-    __asm vpaddsw    ymm2, ymm2, ymm4           /* R += Y */                   \
+    __asm vmovdqa    ymm0, ymmword ptr [YuvConstants + KUVTOB]                 \
+    __asm vmovdqa    ymm1, ymmword ptr [YuvConstants + KUVTOG]                 \
+    __asm vmovdqa    ymm2, ymmword ptr [YuvConstants + KUVTOR]                 \
+    __asm vpmaddubsw ymm0, ymm0, ymm3 /* B UV */                               \
+    __asm vpmaddubsw ymm1, ymm1, ymm3 /* G UV */                               \
+    __asm vpmaddubsw ymm2, ymm2, ymm3 /* B UV */                               \
+    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KYBIASTORGB]            \
+    __asm vpaddw     ymm4, ymm3, ymm4                                          \
+    __asm vpaddsw    ymm0, ymm0, ymm4                                          \
+    __asm vpsubsw    ymm1, ymm4, ymm1                                          \
+    __asm vpaddsw    ymm2, ymm2, ymm4                                          \
     __asm vpsraw     ymm0, ymm0, 6                                             \
     __asm vpsraw     ymm1, ymm1, 6                                             \
     __asm vpsraw     ymm2, ymm2, 6                                             \
-    __asm vpackuswb  ymm0, ymm0, ymm0           /* B */                        \
-    __asm vpackuswb  ymm1, ymm1, ymm1           /* G */                        \
-    __asm vpackuswb  ymm2, ymm2, ymm2           /* R */                        \
-  }
+    __asm vpackuswb  ymm0, ymm0, ymm0                                          \
+    __asm vpackuswb  ymm1, ymm1, ymm1                                          \
+    __asm vpackuswb  ymm2, ymm2, ymm2}
 
 // Store 16 ARGB values.
-#define STOREARGB_AVX2 __asm {                                                 \
-    __asm vpunpcklbw ymm0, ymm0, ymm1           /* BG */                       \
+#define STOREARGB_AVX2 \
+  __asm {                                                                      \
+    __asm vpunpcklbw ymm0, ymm0, ymm1 /* BG */                                 \
     __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklbw ymm2, ymm2, ymm5           /* RA */                       \
+    __asm vpunpcklbw ymm2, ymm2, ymm5 /* RA */                                 \
     __asm vpermq     ymm2, ymm2, 0xd8                                          \
-    __asm vpunpcklwd ymm1, ymm0, ymm2           /* BGRA first 8 pixels */      \
-    __asm vpunpckhwd ymm0, ymm0, ymm2           /* BGRA next 8 pixels */       \
+    __asm vpunpcklwd ymm1, ymm0, ymm2 /* BGRA first 8 pixels */                \
+    __asm vpunpckhwd ymm0, ymm0, ymm2 /* BGRA next 8 pixels */                 \
     __asm vmovdqu    0[edx], ymm1                                              \
     __asm vmovdqu    32[edx], ymm0                                             \
-    __asm lea        edx,  [edx + 64]                                          \
-  }
+    __asm lea        edx,  [edx + 64]}
 
 // Store 16 RGBA values.
-#define STORERGBA_AVX2 __asm {                                                 \
-    __asm vpunpcklbw ymm1, ymm1, ymm2           /* GR */                       \
+#define STORERGBA_AVX2 \
+  __asm {                                                                      \
+    __asm vpunpcklbw ymm1, ymm1, ymm2 /* GR */                                 \
     __asm vpermq     ymm1, ymm1, 0xd8                                          \
-    __asm vpunpcklbw ymm2, ymm5, ymm0           /* AB */                       \
+    __asm vpunpcklbw ymm2, ymm5, ymm0 /* AB */                                 \
     __asm vpermq     ymm2, ymm2, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm2, ymm1           /* ABGR first 8 pixels */      \
-    __asm vpunpckhwd ymm1, ymm2, ymm1           /* ABGR next 8 pixels */       \
+    __asm vpunpcklwd ymm0, ymm2, ymm1 /* ABGR first 8 pixels */                \
+    __asm vpunpckhwd ymm1, ymm2, ymm1 /* ABGR next 8 pixels */                 \
     __asm vmovdqu    [edx], ymm0                                               \
     __asm vmovdqu    [edx + 32], ymm1                                          \
-    __asm lea        edx,  [edx + 64]                                          \
-  }
+    __asm lea        edx,  [edx + 64]}
 
 #ifdef HAS_I422TOARGBROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-__declspec(naked)
-void I422ToARGBRow_AVX2(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void I422ToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
     mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
 
  convertloop:
     READYUV422_AVX2
@@ -2119,21 +2177,21 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
 #ifdef HAS_I422ALPHATOARGBROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB.
-__declspec(naked)
-void I422AlphaToARGBRow_AVX2(const uint8* y_buf,
-                             const uint8* u_buf,
-                             const uint8* v_buf,
-                             const uint8* a_buf,
-                             uint8* dst_argb,
-                             const struct YuvConstants* yuvconstants,
-                             int width) {
+__declspec(naked) void I422AlphaToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    const uint8_t* a_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
     push       ebp
-    mov        eax, [esp + 16 + 4]   // Y
-    mov        esi, [esp + 16 + 8]   // U
+    mov        eax, [esp + 16 + 4]  // Y
+    mov        esi, [esp + 16 + 8]  // U
     mov        edi, [esp + 16 + 12]  // V
     mov        ebp, [esp + 16 + 16]  // A
     mov        edx, [esp + 16 + 20]  // argb
@@ -2162,25 +2220,25 @@ void I422AlphaToARGBRow_AVX2(const uint8* y_buf,
 #ifdef HAS_I444TOARGBROW_AVX2
 // 16 pixels
 // 16 UV values with 16 Y producing 16 ARGB (64 bytes).
-__declspec(naked)
-void I444ToARGBRow_AVX2(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void I444ToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
     mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
  convertloop:
     READYUV444_AVX2
     YUVTORGB_AVX2(ebx)
@@ -2198,64 +2256,66 @@ void I444ToARGBRow_AVX2(const uint8* y_buf,
 }
 #endif  // HAS_I444TOARGBROW_AVX2
 
-#ifdef HAS_I411TOARGBROW_AVX2
+#ifdef HAS_I444ALPHATOARGBROW_AVX2
 // 16 pixels
-// 4 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-__declspec(naked)
-void I411ToARGBRow_AVX2(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
-  __asm {
-    push       esi
-    push       edi
-    push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
-    mov        edi, [esp + 12 + 12]  // V
-    mov        edx, [esp + 12 + 16]  // abgr
-    mov        ebx, [esp + 12 + 20]  // yuvconstants
-    mov        ecx, [esp + 12 + 24]  // width
-    sub        edi, esi
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) void I444AlphaToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    const uint8_t* a_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
+  __asm {
+  push       esi
+  push       edi
+  push       ebx
+  push       ebp
+  mov        eax, [esp + 16 + 4]  // Y
+  mov        esi, [esp + 16 + 8]  // U
+  mov        edi, [esp + 16 + 12]  // V
+  mov        ebp, [esp + 16 + 16]  // A
+  mov        edx, [esp + 16 + 20]  // argb
+  mov        ebx, [esp + 16 + 24]  // yuvconstants
+  mov        ecx, [esp + 16 + 28]  // width
+  sub        edi, esi
+  convertloop:
+  READYUVA444_AVX2
+  YUVTORGB_AVX2(ebx)
+  STOREARGB_AVX2
 
- convertloop:
-    READYUV411_AVX2
-    YUVTORGB_AVX2(ebx)
-    STOREARGB_AVX2
+  sub        ecx, 16
+  jg         convertloop
 
-    sub        ecx, 16
-    jg         convertloop
-
-    pop        ebx
-    pop        edi
-    pop        esi
-    vzeroupper
-    ret
+  pop        ebp
+  pop        ebx
+  pop        edi
+  pop        esi
+  vzeroupper
+  ret
   }
 }
-#endif  // HAS_I411TOARGBROW_AVX2
+#endif  // HAS_I444AlphaTOARGBROW_AVX2
 
 #ifdef HAS_NV12TOARGBROW_AVX2
 // 16 pixels.
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-__declspec(naked)
-void NV12ToARGBRow_AVX2(const uint8* y_buf,
-                        const uint8* uv_buf,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void NV12ToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* uv_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       ebx
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // UV
+    mov        eax, [esp + 8 + 4]  // Y
+    mov        esi, [esp + 8 + 8]  // UV
     mov        edx, [esp + 8 + 12]  // argb
     mov        ebx, [esp + 8 + 16]  // yuvconstants
     mov        ecx, [esp + 8 + 20]  // width
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
 
  convertloop:
     READNV12_AVX2
@@ -2276,21 +2336,21 @@ void NV12ToARGBRow_AVX2(const uint8* y_buf,
 #ifdef HAS_NV21TOARGBROW_AVX2
 // 16 pixels.
 // 8 VU values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
-__declspec(naked)
-void NV21ToARGBRow_AVX2(const uint8* y_buf,
-                        const uint8* vu_buf,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void NV21ToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* vu_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       ebx
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // VU
+    mov        eax, [esp + 8 + 4]  // Y
+    mov        esi, [esp + 8 + 8]  // VU
     mov        edx, [esp + 8 + 12]  // argb
     mov        ebx, [esp + 8 + 16]  // yuvconstants
     mov        ecx, [esp + 8 + 20]  // width
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
 
  convertloop:
     READNV21_AVX2
@@ -2311,18 +2371,18 @@ void NV21ToARGBRow_AVX2(const uint8* y_buf,
 #ifdef HAS_YUY2TOARGBROW_AVX2
 // 16 pixels.
 // 8 YUY2 values with 16 Y and 8 UV producing 16 ARGB (64 bytes).
-__declspec(naked)
-void YUY2ToARGBRow_AVX2(const uint8* src_yuy2,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void YUY2ToARGBRow_AVX2(
+    const uint8_t* src_yuy2,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       ebx
-    mov        eax, [esp + 4 + 4]   // yuy2
-    mov        edx, [esp + 4 + 8]   // argb
+    mov        eax, [esp + 4 + 4]  // yuy2
+    mov        edx, [esp + 4 + 8]  // argb
     mov        ebx, [esp + 4 + 12]  // yuvconstants
     mov        ecx, [esp + 4 + 16]  // width
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
 
  convertloop:
     READYUY2_AVX2
@@ -2342,18 +2402,18 @@ void YUY2ToARGBRow_AVX2(const uint8* src_yuy2,
 #ifdef HAS_UYVYTOARGBROW_AVX2
 // 16 pixels.
 // 8 UYVY values with 16 Y and 8 UV producing 16 ARGB (64 bytes).
-__declspec(naked)
-void UYVYToARGBRow_AVX2(const uint8* src_uyvy,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void UYVYToARGBRow_AVX2(
+    const uint8_t* src_uyvy,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       ebx
-    mov        eax, [esp + 4 + 4]   // uyvy
-    mov        edx, [esp + 4 + 8]   // argb
+    mov        eax, [esp + 4 + 4]  // uyvy
+    mov        edx, [esp + 4 + 8]  // argb
     mov        ebx, [esp + 4 + 12]  // yuvconstants
     mov        ecx, [esp + 4 + 16]  // width
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
 
  convertloop:
     READUYVY_AVX2
@@ -2373,25 +2433,25 @@ void UYVYToARGBRow_AVX2(const uint8* src_uyvy,
 #ifdef HAS_I422TORGBAROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
-__declspec(naked)
-void I422ToRGBARow_AVX2(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width) {
+__declspec(naked) void I422ToRGBARow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // abgr
     mov        ebx, [esp + 12 + 20]  // yuvconstants
     mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
-    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate 0xffffffffffffffff for alpha
 
  convertloop:
     READYUV422_AVX2
@@ -2415,236 +2475,226 @@ void I422ToRGBARow_AVX2(const uint8* y_buf,
 // Allows a conversion with half size scaling.
 
 // Read 8 UV from 444.
-#define READYUV444 __asm {                                                     \
-    __asm movq       xmm0, qword ptr [esi] /* U */                             \
+#define READYUV444 \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* U */                             \
     __asm movq       xmm1, qword ptr [esi + edi] /* V */                       \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
+    __asm movq       xmm4, qword ptr [eax]                                     \
+    __asm punpcklbw  xmm4, xmm4                                                \
+    __asm lea        eax, [eax + 8]}
+
+// Read 4 UV from 444.  With 8 Alpha.
+#define READYUVA444 \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* U */                             \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */                       \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]                                            \
-  }
+    __asm movq       xmm5, qword ptr [ebp] /* A */                             \
+    __asm lea        ebp, [ebp + 8]}
 
 // Read 4 UV from 422, upsample to 8 UV.
-#define READYUV422 __asm {                                                     \
-    __asm movd       xmm0, [esi]          /* U */                              \
-    __asm movd       xmm1, [esi + edi]    /* V */                              \
+#define READYUV422 \
+  __asm {                                                                      \
+    __asm movd       xmm3, [esi] /* U */                                       \
+    __asm movd       xmm1, [esi + edi] /* V */                                 \
     __asm lea        esi,  [esi + 4]                                           \
-    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
-    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
+    __asm punpcklwd  xmm3, xmm3 /* UVUV (upsample) */                          \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
-    __asm lea        eax, [eax + 8]                                            \
-  }
+    __asm lea        eax, [eax + 8]}
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
-#define READYUVA422 __asm {                                                    \
-    __asm movd       xmm0, [esi]          /* U */                              \
-    __asm movd       xmm1, [esi + edi]    /* V */                              \
+#define READYUVA422 \
+  __asm {                                                                      \
+    __asm movd       xmm3, [esi] /* U */                                       \
+    __asm movd       xmm1, [esi + edi] /* V */                                 \
     __asm lea        esi,  [esi + 4]                                           \
-    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
-    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
-    __asm movq       xmm4, qword ptr [eax]   /* Y */                           \
-    __asm punpcklbw  xmm4, xmm4                                                \
-    __asm lea        eax, [eax + 8]                                            \
-    __asm movq       xmm5, qword ptr [ebp]   /* A */                           \
-    __asm lea        ebp, [ebp + 8]                                            \
-  }
-
-// Read 2 UV from 411, upsample to 8 UV.
-// drmemory fails with memory fault if pinsrw used. libyuv bug: 525
-//  __asm pinsrw     xmm0, [esi], 0        /* U */
-//  __asm pinsrw     xmm1, [esi + edi], 0  /* V */
-#define READYUV411_EBX __asm {                                                 \
-    __asm movzx      ebx, word ptr [esi]        /* U */                        \
-    __asm movd       xmm0, ebx                                                 \
-    __asm movzx      ebx, word ptr [esi + edi]  /* V */                        \
-    __asm movd       xmm1, ebx                                                 \
-    __asm lea        esi,  [esi + 2]                                           \
-    __asm punpcklbw  xmm0, xmm1            /* UV */                            \
-    __asm punpcklwd  xmm0, xmm0            /* UVUV (upsample) */               \
-    __asm punpckldq  xmm0, xmm0            /* UVUVUVUV (upsample) */           \
-    __asm movq       xmm4, qword ptr [eax]                                     \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
+    __asm punpcklwd  xmm3, xmm3 /* UVUV (upsample) */                          \
+    __asm movq       xmm4, qword ptr [eax] /* Y */                             \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]                                            \
-  }
+    __asm movq       xmm5, qword ptr [ebp] /* A */                             \
+    __asm lea        ebp, [ebp + 8]}
 
 // Read 4 UV from NV12, upsample to 8 UV.
-#define READNV12 __asm {                                                       \
-    __asm movq       xmm0, qword ptr [esi] /* UV */                            \
+#define READNV12 \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* UV */                            \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpcklwd  xmm3, xmm3 /* UVUV (upsample) */                          \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
-    __asm lea        eax, [eax + 8]                                            \
-  }
+    __asm lea        eax, [eax + 8]}
 
 // Read 4 VU from NV21, upsample to 8 UV.
-#define READNV21 __asm {                                                       \
-    __asm movq       xmm0, qword ptr [esi] /* UV */                            \
+#define READNV21 \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* UV */                            \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm pshufb     xmm0, xmmword ptr kShuffleNV21                            \
+    __asm pshufb     xmm3, xmmword ptr kShuffleNV21                            \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
-    __asm lea        eax, [eax + 8]                                            \
-  }
+    __asm lea        eax, [eax + 8]}
 
 // Read 4 YUY2 with 8 Y and upsample 4 UV to 8 UV.
-#define READYUY2 __asm {                                                       \
-    __asm movdqu     xmm4, [eax]          /* YUY2 */                           \
+#define READYUY2 \
+  __asm {                                                                      \
+    __asm movdqu     xmm4, [eax] /* YUY2 */                                    \
     __asm pshufb     xmm4, xmmword ptr kShuffleYUY2Y                           \
-    __asm movdqu     xmm0, [eax]          /* UV */                             \
-    __asm pshufb     xmm0, xmmword ptr kShuffleYUY2UV                          \
-    __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm movdqu     xmm3, [eax] /* UV */                                      \
+    __asm pshufb     xmm3, xmmword ptr kShuffleYUY2UV                          \
+    __asm lea        eax, [eax + 16]}
 
 // Read 4 UYVY with 8 Y and upsample 4 UV to 8 UV.
-#define READUYVY __asm {                                                       \
-    __asm movdqu     xmm4, [eax]          /* UYVY */                           \
+#define READUYVY \
+  __asm {                                                                      \
+    __asm movdqu     xmm4, [eax] /* UYVY */                                    \
     __asm pshufb     xmm4, xmmword ptr kShuffleUYVYY                           \
-    __asm movdqu     xmm0, [eax]          /* UV */                             \
-    __asm pshufb     xmm0, xmmword ptr kShuffleUYVYUV                          \
-    __asm lea        eax, [eax + 16]                                           \
-  }
+    __asm movdqu     xmm3, [eax] /* UV */                                      \
+    __asm pshufb     xmm3, xmmword ptr kShuffleUYVYUV                          \
+    __asm lea        eax, [eax + 16]}
 
 // Convert 8 pixels: 8 UV and 8 Y.
-#define YUVTORGB(YuvConstants) __asm {                                         \
-    __asm movdqa     xmm1, xmm0                                                \
-    __asm movdqa     xmm2, xmm0                                                \
-    __asm movdqa     xmm3, xmm0                                                \
-    __asm movdqa     xmm0, xmmword ptr [YuvConstants + KUVBIASB]               \
-    __asm pmaddubsw  xmm1, xmmword ptr [YuvConstants + KUVTOB]                 \
-    __asm psubw      xmm0, xmm1                                                \
-    __asm movdqa     xmm1, xmmword ptr [YuvConstants + KUVBIASG]               \
-    __asm pmaddubsw  xmm2, xmmword ptr [YuvConstants + KUVTOG]                 \
-    __asm psubw      xmm1, xmm2                                                \
-    __asm movdqa     xmm2, xmmword ptr [YuvConstants + KUVBIASR]               \
-    __asm pmaddubsw  xmm3, xmmword ptr [YuvConstants + KUVTOR]                 \
-    __asm psubw      xmm2, xmm3                                                \
+#define YUVTORGB(YuvConstants) \
+  __asm {                                                                      \
+    __asm psubb      xmm3, xmmword ptr kBiasUV128                              \
     __asm pmulhuw    xmm4, xmmword ptr [YuvConstants + KYTORGB]                \
-    __asm paddsw     xmm0, xmm4           /* B += Y */                         \
-    __asm paddsw     xmm1, xmm4           /* G += Y */                         \
-    __asm paddsw     xmm2, xmm4           /* R += Y */                         \
+    __asm movdqa     xmm0, xmmword ptr [YuvConstants + KUVTOB]                 \
+    __asm movdqa     xmm1, xmmword ptr [YuvConstants + KUVTOG]                 \
+    __asm movdqa     xmm2, xmmword ptr [YuvConstants + KUVTOR]                 \
+    __asm pmaddubsw  xmm0, xmm3                                                \
+    __asm pmaddubsw  xmm1, xmm3                                                \
+    __asm pmaddubsw  xmm2, xmm3                                                \
+    __asm movdqa     xmm3, xmmword ptr [YuvConstants + KYBIASTORGB]            \
+    __asm paddw      xmm4, xmm3                                                \
+    __asm paddsw     xmm0, xmm4                                                \
+    __asm paddsw     xmm2, xmm4                                                \
+    __asm psubsw     xmm4, xmm1                                                \
+    __asm movdqa     xmm1, xmm4                                                \
     __asm psraw      xmm0, 6                                                   \
     __asm psraw      xmm1, 6                                                   \
     __asm psraw      xmm2, 6                                                   \
-    __asm packuswb   xmm0, xmm0           /* B */                              \
-    __asm packuswb   xmm1, xmm1           /* G */                              \
-    __asm packuswb   xmm2, xmm2           /* R */                              \
+    __asm packuswb   xmm0, xmm0 /* B */                                        \
+    __asm packuswb   xmm1, xmm1 /* G */                                        \
+    __asm packuswb   xmm2, xmm2 /* R */             \
   }
 
 // Store 8 ARGB values.
-#define STOREARGB __asm {                                                      \
-    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
-    __asm punpcklbw  xmm2, xmm5           /* RA */                             \
+#define STOREARGB \
+  __asm {                                                                      \
+    __asm punpcklbw  xmm0, xmm1 /* BG */                                       \
+    __asm punpcklbw  xmm2, xmm5 /* RA */                                       \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm punpcklwd  xmm0, xmm2           /* BGRA first 4 pixels */            \
-    __asm punpckhwd  xmm1, xmm2           /* BGRA next 4 pixels */             \
+    __asm punpcklwd  xmm0, xmm2 /* BGRA first 4 pixels */                      \
+    __asm punpckhwd  xmm1, xmm2 /* BGRA next 4 pixels */                       \
     __asm movdqu     0[edx], xmm0                                              \
     __asm movdqu     16[edx], xmm1                                             \
-    __asm lea        edx,  [edx + 32]                                          \
-  }
+    __asm lea        edx,  [edx + 32]}
 
 // Store 8 BGRA values.
-#define STOREBGRA __asm {                                                      \
-    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
-    __asm punpcklbw  xmm1, xmm0           /* GB */                             \
-    __asm punpcklbw  xmm5, xmm2           /* AR */                             \
+#define STOREBGRA \
+  __asm {                                                                      \
+    __asm pcmpeqb    xmm5, xmm5 /* generate 0xffffffff for alpha */            \
+    __asm punpcklbw  xmm1, xmm0 /* GB */                                       \
+    __asm punpcklbw  xmm5, xmm2 /* AR */                                       \
     __asm movdqa     xmm0, xmm5                                                \
-    __asm punpcklwd  xmm5, xmm1           /* BGRA first 4 pixels */            \
-    __asm punpckhwd  xmm0, xmm1           /* BGRA next 4 pixels */             \
+    __asm punpcklwd  xmm5, xmm1 /* BGRA first 4 pixels */                      \
+    __asm punpckhwd  xmm0, xmm1 /* BGRA next 4 pixels */                       \
     __asm movdqu     0[edx], xmm5                                              \
     __asm movdqu     16[edx], xmm0                                             \
-    __asm lea        edx,  [edx + 32]                                          \
-  }
+    __asm lea        edx,  [edx + 32]}
 
 // Store 8 RGBA values.
-#define STORERGBA __asm {                                                      \
-    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
-    __asm punpcklbw  xmm1, xmm2           /* GR */                             \
-    __asm punpcklbw  xmm5, xmm0           /* AB */                             \
+#define STORERGBA \
+  __asm {                                                                      \
+    __asm pcmpeqb    xmm5, xmm5 /* generate 0xffffffff for alpha */            \
+    __asm punpcklbw  xmm1, xmm2 /* GR */                                       \
+    __asm punpcklbw  xmm5, xmm0 /* AB */                                       \
     __asm movdqa     xmm0, xmm5                                                \
-    __asm punpcklwd  xmm5, xmm1           /* RGBA first 4 pixels */            \
-    __asm punpckhwd  xmm0, xmm1           /* RGBA next 4 pixels */             \
+    __asm punpcklwd  xmm5, xmm1 /* RGBA first 4 pixels */                      \
+    __asm punpckhwd  xmm0, xmm1 /* RGBA next 4 pixels */                       \
     __asm movdqu     0[edx], xmm5                                              \
     __asm movdqu     16[edx], xmm0                                             \
-    __asm lea        edx,  [edx + 32]                                          \
-  }
+    __asm lea        edx,  [edx + 32]}
 
 // Store 8 RGB24 values.
-#define STORERGB24 __asm {                                                     \
-    /* Weave into RRGB */                                                      \
-    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
-    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+#define STORERGB24 \
+  __asm {/* Weave into RRGB */                                                 \
+    __asm punpcklbw  xmm0, xmm1 /* BG */                                       \
+    __asm punpcklbw  xmm2, xmm2 /* RR */                                       \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
-    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
-    /* RRGB -> RGB24 */                                                        \
-    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
-    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
-    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
-    __asm movq       qword ptr 0[edx], xmm0  /* First 8 bytes */               \
-    __asm movdqu     8[edx], xmm1         /* Last 16 bytes */                  \
-    __asm lea        edx,  [edx + 24]                                          \
-  }
+    __asm punpcklwd  xmm0, xmm2 /* BGRR first 4 pixels */                      \
+    __asm punpckhwd  xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB24 */   \
+    __asm pshufb     xmm0, xmm5 /* Pack first 8 and last 4 bytes. */           \
+    __asm pshufb     xmm1, xmm6 /* Pack first 12 bytes. */                     \
+    __asm palignr    xmm1, xmm0, 12 /* last 4 bytes of xmm0 + 12 xmm1 */       \
+    __asm movq       qword ptr 0[edx], xmm0 /* First 8 bytes */                \
+    __asm movdqu     8[edx], xmm1 /* Last 16 bytes */                          \
+    __asm lea        edx,  [edx + 24]}
 
 // Store 8 RGB565 values.
-#define STORERGB565 __asm {                                                    \
-    /* Weave into RRGB */                                                      \
-    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
-    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+#define STORERGB565 \
+  __asm {/* Weave into RRGB */                                                 \
+    __asm punpcklbw  xmm0, xmm1 /* BG */                                       \
+    __asm punpcklbw  xmm2, xmm2 /* RR */                                       \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
-    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
-    /* RRGB -> RGB565 */                                                       \
-    __asm movdqa     xmm3, xmm0    /* B  first 4 pixels of argb */             \
-    __asm movdqa     xmm2, xmm0    /* G */                                     \
-    __asm pslld      xmm0, 8       /* R */                                     \
-    __asm psrld      xmm3, 3       /* B */                                     \
-    __asm psrld      xmm2, 5       /* G */                                     \
-    __asm psrad      xmm0, 16      /* R */                                     \
-    __asm pand       xmm3, xmm5    /* B */                                     \
-    __asm pand       xmm2, xmm6    /* G */                                     \
-    __asm pand       xmm0, xmm7    /* R */                                     \
-    __asm por        xmm3, xmm2    /* BG */                                    \
-    __asm por        xmm0, xmm3    /* BGR */                                   \
-    __asm movdqa     xmm3, xmm1    /* B  next 4 pixels of argb */              \
-    __asm movdqa     xmm2, xmm1    /* G */                                     \
-    __asm pslld      xmm1, 8       /* R */                                     \
-    __asm psrld      xmm3, 3       /* B */                                     \
-    __asm psrld      xmm2, 5       /* G */                                     \
-    __asm psrad      xmm1, 16      /* R */                                     \
-    __asm pand       xmm3, xmm5    /* B */                                     \
-    __asm pand       xmm2, xmm6    /* G */                                     \
-    __asm pand       xmm1, xmm7    /* R */                                     \
-    __asm por        xmm3, xmm2    /* BG */                                    \
-    __asm por        xmm1, xmm3    /* BGR */                                   \
+    __asm punpcklwd  xmm0, xmm2 /* BGRR first 4 pixels */                      \
+    __asm punpckhwd  xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB565 */  \
+    __asm movdqa     xmm3, xmm0 /* B  first 4 pixels of argb */                \
+    __asm movdqa     xmm2, xmm0 /* G */                                        \
+    __asm pslld      xmm0, 8 /* R */                                           \
+    __asm psrld      xmm3, 3 /* B */                                           \
+    __asm psrld      xmm2, 5 /* G */                                           \
+    __asm psrad      xmm0, 16 /* R */                                          \
+    __asm pand       xmm3, xmm5 /* B */                                        \
+    __asm pand       xmm2, xmm6 /* G */                                        \
+    __asm pand       xmm0, xmm7 /* R */                                        \
+    __asm por        xmm3, xmm2 /* BG */                                       \
+    __asm por        xmm0, xmm3 /* BGR */                                      \
+    __asm movdqa     xmm3, xmm1 /* B  next 4 pixels of argb */                 \
+    __asm movdqa     xmm2, xmm1 /* G */                                        \
+    __asm pslld      xmm1, 8 /* R */                                           \
+    __asm psrld      xmm3, 3 /* B */                                           \
+    __asm psrld      xmm2, 5 /* G */                                           \
+    __asm psrad      xmm1, 16 /* R */                                          \
+    __asm pand       xmm3, xmm5 /* B */                                        \
+    __asm pand       xmm2, xmm6 /* G */                                        \
+    __asm pand       xmm1, xmm7 /* R */                                        \
+    __asm por        xmm3, xmm2 /* BG */                                       \
+    __asm por        xmm1, xmm3 /* BGR */                                      \
     __asm packssdw   xmm0, xmm1                                                \
-    __asm movdqu     0[edx], xmm0  /* store 8 pixels of RGB565 */              \
-    __asm lea        edx, [edx + 16]                                           \
-  }
+    __asm movdqu     0[edx], xmm0 /* store 8 pixels of RGB565 */               \
+    __asm lea        edx, [edx + 16]}
 
 // 8 pixels.
 // 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked)
-void I444ToARGBRow_SSSE3(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+__declspec(naked) void I444ToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
     mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
-    pcmpeqb    xmm5, xmm5            // generate 0xffffffff for alpha
+    pcmpeqb    xmm5, xmm5  // generate 0xffffffff for alpha
 
  convertloop:
     READYUV444
@@ -2662,20 +2712,60 @@ void I444ToARGBRow_SSSE3(const uint8* y_buf,
 }
 
 // 8 pixels.
+// 8 UV values, mixed with 8 Y and 8A producing 8 ARGB (32 bytes).
+__declspec(naked) void I444AlphaToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    const uint8_t* a_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebx
+    push       ebp
+    mov        eax, [esp + 16 + 4]  // Y
+    mov        esi, [esp + 16 + 8]  // U
+    mov        edi, [esp + 16 + 12]  // V
+    mov        ebp, [esp + 16 + 16]  // A
+    mov        edx, [esp + 16 + 20]  // argb
+    mov        ebx, [esp + 16 + 24]  // yuvconstants
+    mov        ecx, [esp + 16 + 28]  // width
+    sub        edi, esi
+
+ convertloop:
+    READYUVA444
+    YUVTORGB(ebx)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        ebp
+    pop        ebx
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes).
-__declspec(naked)
-void I422ToRGB24Row_SSSE3(const uint8* y_buf,
-                          const uint8* u_buf,
-                          const uint8* v_buf,
-                          uint8* dst_rgb24,
-                          const struct YuvConstants* yuvconstants,
-                          int width) {
+__declspec(naked) void I422ToRGB24Row_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_rgb24,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
@@ -2699,38 +2789,33 @@ void I422ToRGB24Row_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB565 (16 bytes).
-__declspec(naked)
-void I422ToRGB565Row_SSSE3(const uint8* y_buf,
-                           const uint8* u_buf,
-                           const uint8* v_buf,
-                           uint8* rgb565_buf,
-                           const struct YuvConstants* yuvconstants,
-                           int width) {
+// 8 pixels.
+// 8 UV values, mixed with 8 Y producing 8 RGB24 (24 bytes).
+__declspec(naked) void I444ToRGB24Row_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_rgb24,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
     mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
-    pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
-    psrld      xmm5, 27
-    pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
-    psrld      xmm6, 26
-    pslld      xmm6, 5
-    pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
-    pslld      xmm7, 11
+    movdqa     xmm5, xmmword ptr kShuffleMaskARGBToRGB24_0
+    movdqa     xmm6, xmmword ptr kShuffleMaskARGBToRGB24
 
  convertloop:
-    READYUV422
+    READYUV444
     YUVTORGB(ebx)
-    STORERGB565
+    STORERGB24
 
     sub        ecx, 8
     jg         convertloop
@@ -2742,32 +2827,38 @@ void I422ToRGB565Row_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked)
-void I422ToARGBRow_SSSE3(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+// 8 pixels
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB565 (16 bytes).
+__declspec(naked) void I422ToRGB565Row_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* rgb565_buf,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
     mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pcmpeqb    xmm5, xmm5  // generate mask 0x0000001f
+    psrld      xmm5, 27
+    pcmpeqb    xmm6, xmm6  // generate mask 0x000007e0
+    psrld      xmm6, 26
+    pslld      xmm6, 5
+    pcmpeqb    xmm7, xmm7  // generate mask 0xfffff800
+    pslld      xmm7, 11
 
  convertloop:
     READYUV422
     YUVTORGB(ebx)
-    STOREARGB
+    STORERGB565
 
     sub        ecx, 8
     jg         convertloop
@@ -2780,38 +2871,35 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
 }
 
 // 8 pixels.
-// 4 UV values upsampled to 8 UV, mixed with 8 Y and 8 A producing 8 ARGB.
-__declspec(naked)
-void I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
-                              const uint8* u_buf,
-                              const uint8* v_buf,
-                              const uint8* a_buf,
-                              uint8* dst_argb,
-                              const struct YuvConstants* yuvconstants,
-                              int width) {
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) void I422ToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    push       ebp
-    mov        eax, [esp + 16 + 4]   // Y
-    mov        esi, [esp + 16 + 8]   // U
-    mov        edi, [esp + 16 + 12]  // V
-    mov        ebp, [esp + 16 + 16]  // A
-    mov        edx, [esp + 16 + 20]  // argb
-    mov        ebx, [esp + 16 + 24]  // yuvconstants
-    mov        ecx, [esp + 16 + 28]  // width
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ebx, [esp + 12 + 20]  // yuvconstants
+    mov        ecx, [esp + 12 + 24]  // width
     sub        edi, esi
+    pcmpeqb    xmm5, xmm5  // generate 0xffffffff for alpha
 
  convertloop:
-    READYUVA422
+    READYUV422
     YUVTORGB(ebx)
     STOREARGB
 
     sub        ecx, 8
     jg         convertloop
 
-    pop        ebp
     pop        ebx
     pop        edi
     pop        esi
@@ -2820,32 +2908,32 @@ void I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
 }
 
 // 8 pixels.
-// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-// Similar to I420 but duplicate UV once more.
-__declspec(naked)
-void I411ToARGBRow_SSSE3(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+// 4 UV values upsampled to 8 UV, mixed with 8 Y and 8 A producing 8 ARGB.
+__declspec(naked) void I422AlphaToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    const uint8_t* a_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
     push       ebp
-    mov        eax, [esp + 16 + 4]   // Y
-    mov        esi, [esp + 16 + 8]   // U
+    mov        eax, [esp + 16 + 4]  // Y
+    mov        esi, [esp + 16 + 8]  // U
     mov        edi, [esp + 16 + 12]  // V
-    mov        edx, [esp + 16 + 16]  // abgr
-    mov        ebp, [esp + 16 + 20]  // yuvconstants
-    mov        ecx, [esp + 16 + 24]  // width
+    mov        ebp, [esp + 16 + 16]  // A
+    mov        edx, [esp + 16 + 20]  // argb
+    mov        ebx, [esp + 16 + 24]  // yuvconstants
+    mov        ecx, [esp + 16 + 28]  // width
     sub        edi, esi
-    pcmpeqb    xmm5, xmm5            // generate 0xffffffff for alpha
 
  convertloop:
-    READYUV411_EBX
-    YUVTORGB(ebp)
+    READYUVA422
+    YUVTORGB(ebx)
     STOREARGB
 
     sub        ecx, 8
@@ -2861,21 +2949,21 @@ void I411ToARGBRow_SSSE3(const uint8* y_buf,
 
 // 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked)
-void NV12ToARGBRow_SSSE3(const uint8* y_buf,
-                         const uint8* uv_buf,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+__declspec(naked) void NV12ToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* uv_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       ebx
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // UV
+    mov        eax, [esp + 8 + 4]  // Y
+    mov        esi, [esp + 8 + 8]  // UV
     mov        edx, [esp + 8 + 12]  // argb
     mov        ebx, [esp + 8 + 16]  // yuvconstants
     mov        ecx, [esp + 8 + 20]  // width
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pcmpeqb    xmm5, xmm5  // generate 0xffffffff for alpha
 
  convertloop:
     READNV12
@@ -2893,21 +2981,21 @@ void NV12ToARGBRow_SSSE3(const uint8* y_buf,
 
 // 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
-__declspec(naked)
-void NV21ToARGBRow_SSSE3(const uint8* y_buf,
-                         const uint8* vu_buf,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+__declspec(naked) void NV21ToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* vu_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       ebx
-    mov        eax, [esp + 8 + 4]   // Y
-    mov        esi, [esp + 8 + 8]   // VU
+    mov        eax, [esp + 8 + 4]  // Y
+    mov        esi, [esp + 8 + 8]  // VU
     mov        edx, [esp + 8 + 12]  // argb
     mov        ebx, [esp + 8 + 16]  // yuvconstants
     mov        ecx, [esp + 8 + 20]  // width
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pcmpeqb    xmm5, xmm5  // generate 0xffffffff for alpha
 
  convertloop:
     READNV21
@@ -2925,18 +3013,18 @@ void NV21ToARGBRow_SSSE3(const uint8* y_buf,
 
 // 8 pixels.
 // 4 YUY2 values with 8 Y and 4 UV producing 8 ARGB (32 bytes).
-__declspec(naked)
-void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+__declspec(naked) void YUY2ToARGBRow_SSSE3(
+    const uint8_t* src_yuy2,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       ebx
-    mov        eax, [esp + 4 + 4]   // yuy2
-    mov        edx, [esp + 4 + 8]   // argb
+    mov        eax, [esp + 4 + 4]  // yuy2
+    mov        edx, [esp + 4 + 8]  // argb
     mov        ebx, [esp + 4 + 12]  // yuvconstants
     mov        ecx, [esp + 4 + 16]  // width
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pcmpeqb    xmm5, xmm5  // generate 0xffffffff for alpha
 
  convertloop:
     READYUY2
@@ -2953,18 +3041,18 @@ void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
 
 // 8 pixels.
 // 4 UYVY values with 8 Y and 4 UV producing 8 ARGB (32 bytes).
-__declspec(naked)
-void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
-                         uint8* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+__declspec(naked) void UYVYToARGBRow_SSSE3(
+    const uint8_t* src_uyvy,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       ebx
-    mov        eax, [esp + 4 + 4]   // uyvy
-    mov        edx, [esp + 4 + 8]   // argb
+    mov        eax, [esp + 4 + 4]  // uyvy
+    mov        edx, [esp + 4 + 8]  // argb
     mov        ebx, [esp + 4 + 12]  // yuvconstants
     mov        ecx, [esp + 4 + 16]  // width
-    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pcmpeqb    xmm5, xmm5  // generate 0xffffffff for alpha
 
  convertloop:
     READUYVY
@@ -2979,19 +3067,19 @@ void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
   }
 }
 
-__declspec(naked)
-void I422ToRGBARow_SSSE3(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* dst_rgba,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
+__declspec(naked) void I422ToRGBARow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    uint8_t* dst_rgba,
+    const struct YuvConstants* yuvconstants,
+    int width) {
   __asm {
     push       esi
     push       edi
     push       ebx
-    mov        eax, [esp + 12 + 4]   // Y
-    mov        esi, [esp + 12 + 8]   // U
+    mov        eax, [esp + 12 + 4]  // Y
+    mov        esi, [esp + 12 + 8]  // U
     mov        edi, [esp + 12 + 12]  // V
     mov        edx, [esp + 12 + 16]  // argb
     mov        ebx, [esp + 12 + 20]  // yuvconstants
@@ -3014,41 +3102,42 @@ void I422ToRGBARow_SSSE3(const uint8* y_buf,
 }
 #endif  // HAS_I422TOARGBROW_SSSE3
 
+// I400ToARGBRow_SSE2 is disabled due to new yuvconstant parameter
 #ifdef HAS_I400TOARGBROW_SSE2
 // 8 pixels of Y converted to 8 pixels of ARGB (32 bytes).
-__declspec(naked)
-void I400ToARGBRow_SSE2(const uint8* y_buf,
-                        uint8* rgb_buf,
-                        int width) {
+__declspec(naked) void I400ToARGBRow_SSE2(const uint8_t* y_buf,
+                                          uint8_t* rgb_buf,
+                                          const struct YuvConstants*,
+                                          int width) {
   __asm {
-    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    mov        eax, 0x4a354a35  // 4a35 = 18997 = round(1.164 * 64 * 256)
     movd       xmm2, eax
     pshufd     xmm2, xmm2,0
-    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    mov        eax, 0x04880488  // 0488 = 1160 = round(1.164 * 64 * 16)
     movd       xmm3, eax
     pshufd     xmm3, xmm3, 0
-    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pcmpeqb    xmm4, xmm4  // generate mask 0xff000000
     pslld      xmm4, 24
 
-    mov        eax, [esp + 4]       // Y
-    mov        edx, [esp + 8]       // rgb
-    mov        ecx, [esp + 12]      // width
+    mov        eax, [esp + 4]  // Y
+    mov        edx, [esp + 8]  // rgb
+    mov        ecx, [esp + 12]  // width
 
  convertloop:
-    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+        // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
     movq       xmm0, qword ptr [eax]
     lea        eax, [eax + 8]
-    punpcklbw  xmm0, xmm0           // Y.Y
+    punpcklbw  xmm0, xmm0  // Y.Y
     pmulhuw    xmm0, xmm2
     psubusw    xmm0, xmm3
     psrlw      xmm0, 6
-    packuswb   xmm0, xmm0           // G
+    packuswb   xmm0, xmm0        // G
 
-    // Step 2: Weave into ARGB
-    punpcklbw  xmm0, xmm0           // GG
+        // Step 2: Weave into ARGB
+    punpcklbw  xmm0, xmm0  // GG
     movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm0           // BGRA first 4 pixels
-    punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
+    punpcklwd  xmm0, xmm0  // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm1  // BGRA next 4 pixels
     por        xmm0, xmm4
     por        xmm1, xmm4
     movdqu     [edx], xmm0
@@ -3064,41 +3153,41 @@ void I400ToARGBRow_SSE2(const uint8* y_buf,
 #ifdef HAS_I400TOARGBROW_AVX2
 // 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
 // note: vpunpcklbw mutates and vpackuswb unmutates.
-__declspec(naked)
-void I400ToARGBRow_AVX2(const uint8* y_buf,
-                        uint8* rgb_buf,
-                        int width) {
+__declspec(naked) void I400ToARGBRow_AVX2(const uint8_t* y_buf,
+                                          uint8_t* rgb_buf,
+                                          const struct YuvConstants*,
+                                          int width) {
   __asm {
-    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    mov        eax, 0x4a354a35  // 4a35 = 18997 = round(1.164 * 64 * 256)
     vmovd      xmm2, eax
     vbroadcastss ymm2, xmm2
-    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    mov        eax, 0x04880488  // 0488 = 1160 = round(1.164 * 64 * 16)
     vmovd      xmm3, eax
     vbroadcastss ymm3, xmm3
-    vpcmpeqb   ymm4, ymm4, ymm4     // generate mask 0xff000000
+    vpcmpeqb   ymm4, ymm4, ymm4  // generate mask 0xff000000
     vpslld     ymm4, ymm4, 24
 
-    mov        eax, [esp + 4]       // Y
-    mov        edx, [esp + 8]       // rgb
-    mov        ecx, [esp + 12]      // width
+    mov        eax, [esp + 4]  // Y
+    mov        edx, [esp + 8]  // rgb
+    mov        ecx, [esp + 12]  // width
 
  convertloop:
-    // Step 1: Scale Y contriportbution to 16 G values. G = (y - 16) * 1.164
+        // Step 1: Scale Y contriportbution to 16 G values. G = (y - 16) * 1.164
     vmovdqu    xmm0, [eax]
     lea        eax, [eax + 16]
-    vpermq     ymm0, ymm0, 0xd8           // vpunpcklbw mutates
-    vpunpcklbw ymm0, ymm0, ymm0           // Y.Y
+    vpermq     ymm0, ymm0, 0xd8  // vpunpcklbw mutates
+    vpunpcklbw ymm0, ymm0, ymm0  // Y.Y
     vpmulhuw   ymm0, ymm0, ymm2
     vpsubusw   ymm0, ymm0, ymm3
     vpsrlw     ymm0, ymm0, 6
-    vpackuswb  ymm0, ymm0, ymm0           // G.  still mutated: 3120
+    vpackuswb  ymm0, ymm0, ymm0        // G.  still mutated: 3120
 
-    // TODO(fbarchard): Weave alpha with unpack.
-    // Step 2: Weave into ARGB
-    vpunpcklbw ymm1, ymm0, ymm0           // GG - mutates
+        // TODO(fbarchard): Weave alpha with unpack.
+        // Step 2: Weave into ARGB
+    vpunpcklbw ymm1, ymm0, ymm0  // GG - mutates
     vpermq     ymm1, ymm1, 0xd8
-    vpunpcklwd ymm0, ymm1, ymm1           // GGGG first 8 pixels
-    vpunpckhwd ymm1, ymm1, ymm1           // GGGG next 8 pixels
+    vpunpcklwd ymm0, ymm1, ymm1  // GGGG first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm1  // GGGG next 8 pixels
     vpor       ymm0, ymm0, ymm4
     vpor       ymm1, ymm1, ymm4
     vmovdqu    [edx], ymm0
@@ -3114,16 +3203,16 @@ void I400ToARGBRow_AVX2(const uint8* y_buf,
 
 #ifdef HAS_MIRRORROW_SSSE3
 // Shuffle table for reversing the bytes.
-static const uvec8 kShuffleMirror = {
-  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
-};
+static const uvec8 kShuffleMirror = {15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u,
+                                     7u,  6u,  5u,  4u,  3u,  2u,  1u, 0u};
 
 // TODO(fbarchard): Replace lea with -16 offset.
-__declspec(naked)
-void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void MirrorRow_SSSE3(const uint8_t* src,
+                                       uint8_t* dst,
+                                       int width) {
   __asm {
-    mov       eax, [esp + 4]   // src
-    mov       edx, [esp + 8]   // dst
+    mov       eax, [esp + 4]  // src
+    mov       edx, [esp + 8]  // dst
     mov       ecx, [esp + 12]  // width
     movdqa    xmm5, xmmword ptr kShuffleMirror
 
@@ -3140,11 +3229,12 @@ void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
 #endif  // HAS_MIRRORROW_SSSE3
 
 #ifdef HAS_MIRRORROW_AVX2
-__declspec(naked)
-void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void MirrorRow_AVX2(const uint8_t* src,
+                                      uint8_t* dst,
+                                      int width) {
   __asm {
-    mov       eax, [esp + 4]   // src
-    mov       edx, [esp + 8]   // dst
+    mov       eax, [esp + 4]  // src
+    mov       edx, [esp + 8]  // dst
     mov       ecx, [esp + 12]  // width
     vbroadcastf128 ymm5, xmmword ptr kShuffleMirror
 
@@ -3162,19 +3252,19 @@ void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_MIRRORROW_AVX2
 
-#ifdef HAS_MIRRORUVROW_SSSE3
+#ifdef HAS_MIRRORSPLITUVROW_SSSE3
 // Shuffle table for reversing the bytes of UV channels.
-static const uvec8 kShuffleMirrorUV = {
-  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
-};
+static const uvec8 kShuffleMirrorUV = {14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u,
+                                       15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u};
 
-__declspec(naked)
-void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
-                       int width) {
+__declspec(naked) void MirrorSplitUVRow_SSSE3(const uint8_t* src,
+                                              uint8_t* dst_u,
+                                              uint8_t* dst_v,
+                                              int width) {
   __asm {
     push      edi
-    mov       eax, [esp + 4 + 4]   // src
-    mov       edx, [esp + 4 + 8]   // dst_u
+    mov       eax, [esp + 4 + 4]  // src
+    mov       edx, [esp + 4 + 8]  // dst_u
     mov       edi, [esp + 4 + 12]  // dst_v
     mov       ecx, [esp + 4 + 16]  // width
     movdqa    xmm1, xmmword ptr kShuffleMirrorUV
@@ -3195,14 +3285,15 @@ void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
     ret
   }
 }
-#endif  // HAS_MIRRORUVROW_SSSE3
+#endif  // HAS_MIRRORSPLITUVROW_SSSE3
 
 #ifdef HAS_ARGBMIRRORROW_SSE2
-__declspec(naked)
-void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void ARGBMirrorRow_SSE2(const uint8_t* src,
+                                          uint8_t* dst,
+                                          int width) {
   __asm {
-    mov       eax, [esp + 4]   // src
-    mov       edx, [esp + 8]   // dst
+    mov       eax, [esp + 4]  // src
+    mov       edx, [esp + 8]  // dst
     mov       ecx, [esp + 12]  // width
     lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
 
@@ -3221,15 +3312,14 @@ void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
 
 #ifdef HAS_ARGBMIRRORROW_AVX2
 // Shuffle table for reversing the bytes.
-static const ulvec32 kARGBShuffleMirror_AVX2 = {
-  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
-};
+static const ulvec32 kARGBShuffleMirror_AVX2 = {7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u};
 
-__declspec(naked)
-void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void ARGBMirrorRow_AVX2(const uint8_t* src,
+                                          uint8_t* dst,
+                                          int width) {
   __asm {
-    mov       eax, [esp + 4]   // src
-    mov       edx, [esp + 8]   // dst
+    mov       eax, [esp + 4]  // src
+    mov       edx, [esp + 8]  // dst
     mov       ecx, [esp + 12]  // width
     vmovdqu   ymm5, ymmword ptr kARGBShuffleMirror_AVX2
 
@@ -3246,16 +3336,17 @@ void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
 #endif  // HAS_ARGBMIRRORROW_AVX2
 
 #ifdef HAS_SPLITUVROW_SSE2
-__declspec(naked)
-void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) {
+__declspec(naked) void SplitUVRow_SSE2(const uint8_t* src_uv,
+                                       uint8_t* dst_u,
+                                       uint8_t* dst_v,
+                                       int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_uv
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    mov        eax, [esp + 4 + 4]  // src_uv
+    mov        edx, [esp + 4 + 8]  // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0x00ff00ff
     psrlw      xmm5, 8
     sub        edi, edx
 
@@ -3265,10 +3356,10 @@ void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
     lea        eax,  [eax + 32]
     movdqa     xmm2, xmm0
     movdqa     xmm3, xmm1
-    pand       xmm0, xmm5   // even bytes
+    pand       xmm0, xmm5  // even bytes
     pand       xmm1, xmm5
     packuswb   xmm0, xmm1
-    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm2, 8  // odd bytes
     psrlw      xmm3, 8
     packuswb   xmm2, xmm3
     movdqu     [edx], xmm0
@@ -3285,16 +3376,17 @@ void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
 #endif  // HAS_SPLITUVROW_SSE2
 
 #ifdef HAS_SPLITUVROW_AVX2
-__declspec(naked)
-void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) {
+__declspec(naked) void SplitUVRow_AVX2(const uint8_t* src_uv,
+                                       uint8_t* dst_u,
+                                       uint8_t* dst_v,
+                                       int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_uv
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // width
-    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    mov        eax, [esp + 4 + 4]  // src_uv
+    mov        edx, [esp + 4 + 8]  // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
@@ -3302,9 +3394,9 @@ void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     lea        eax,  [eax + 64]
-    vpsrlw     ymm2, ymm0, 8      // odd bytes
+    vpsrlw     ymm2, ymm0, 8  // odd bytes
     vpsrlw     ymm3, ymm1, 8
-    vpand      ymm0, ymm0, ymm5   // even bytes
+    vpand      ymm0, ymm0, ymm5  // even bytes
     vpand      ymm1, ymm1, ymm5
     vpackuswb  ymm0, ymm0, ymm1
     vpackuswb  ymm2, ymm2, ymm3
@@ -3324,24 +3416,25 @@ void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
 #endif  // HAS_SPLITUVROW_AVX2
 
 #ifdef HAS_MERGEUVROW_SSE2
-__declspec(naked)
-void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
-                     int width) {
+__declspec(naked) void MergeUVRow_SSE2(const uint8_t* src_u,
+                                       const uint8_t* src_v,
+                                       uint8_t* dst_uv,
+                                       int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_u
-    mov        edx, [esp + 4 + 8]    // src_v
-    mov        edi, [esp + 4 + 12]   // dst_uv
-    mov        ecx, [esp + 4 + 16]   // width
+    mov        eax, [esp + 4 + 4]  // src_u
+    mov        edx, [esp + 4 + 8]  // src_v
+    mov        edi, [esp + 4 + 12]  // dst_uv
+    mov        ecx, [esp + 4 + 16]  // width
     sub        edx, eax
 
   convertloop:
-    movdqu     xmm0, [eax]      // read 16 U's
+    movdqu     xmm0, [eax]  // read 16 U's
     movdqu     xmm1, [eax + edx]  // and 16 V's
     lea        eax,  [eax + 16]
     movdqa     xmm2, xmm0
-    punpcklbw  xmm0, xmm1       // first 8 UV pairs
-    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    punpcklbw  xmm0, xmm1  // first 8 UV pairs
+    punpckhbw  xmm2, xmm1  // next 8 UV pairs
     movdqu     [edi], xmm0
     movdqu     [edi + 16], xmm2
     lea        edi, [edi + 32]
@@ -3355,24 +3448,25 @@ void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
 #endif  //  HAS_MERGEUVROW_SSE2
 
 #ifdef HAS_MERGEUVROW_AVX2
-__declspec(naked)
-void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
-                     int width) {
+__declspec(naked) void MergeUVRow_AVX2(const uint8_t* src_u,
+                                       const uint8_t* src_v,
+                                       uint8_t* dst_uv,
+                                       int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_u
-    mov        edx, [esp + 4 + 8]    // src_v
-    mov        edi, [esp + 4 + 12]   // dst_uv
-    mov        ecx, [esp + 4 + 16]   // width
+    mov        eax, [esp + 4 + 4]  // src_u
+    mov        edx, [esp + 4 + 8]  // src_v
+    mov        edi, [esp + 4 + 12]  // dst_uv
+    mov        ecx, [esp + 4 + 16]  // width
     sub        edx, eax
 
   convertloop:
-    vmovdqu    ymm0, [eax]           // read 32 U's
-    vmovdqu    ymm1, [eax + edx]     // and 32 V's
+    vmovdqu    ymm0, [eax]  // read 32 U's
+    vmovdqu    ymm1, [eax + edx]  // and 32 V's
     lea        eax,  [eax + 32]
-    vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
-    vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
-    vextractf128 [edi], ymm2, 0       // bytes 0..15
+    vpunpcklbw ymm2, ymm0, ymm1  // low 16 UV pairs. mutated qqword 0,2
+    vpunpckhbw ymm0, ymm0, ymm1  // high 16 UV pairs. mutated qqword 1,3
+    vextractf128 [edi], ymm2, 0  // bytes 0..15
     vextractf128 [edi + 16], ymm0, 0  // bytes 16..31
     vextractf128 [edi + 32], ymm2, 1  // bytes 32..47
     vextractf128 [edi + 48], ymm0, 1  // bytes 47..63
@@ -3388,13 +3482,14 @@ void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
 #endif  //  HAS_MERGEUVROW_AVX2
 
 #ifdef HAS_COPYROW_SSE2
-// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time.
-__declspec(naked)
-void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+// CopyRow copys 'width' bytes using a 16 byte load/store, 32 bytes at time.
+__declspec(naked) void CopyRow_SSE2(const uint8_t* src,
+                                    uint8_t* dst,
+                                    int width) {
   __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
+    mov        eax, [esp + 4]  // src
+    mov        edx, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
     test       eax, 15
     jne        convertloopu
     test       edx, 15
@@ -3426,13 +3521,14 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
 #endif  // HAS_COPYROW_SSE2
 
 #ifdef HAS_COPYROW_AVX
-// CopyRow copys 'count' bytes using a 32 byte load/store, 64 bytes at time.
-__declspec(naked)
-void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+// CopyRow copys 'width' bytes using a 32 byte load/store, 64 bytes at time.
+__declspec(naked) void CopyRow_AVX(const uint8_t* src,
+                                   uint8_t* dst,
+                                   int width) {
   __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
+    mov        eax, [esp + 4]  // src
+    mov        edx, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
 
   convertloop:
     vmovdqu    ymm0, [eax]
@@ -3451,14 +3547,15 @@ void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
 #endif  // HAS_COPYROW_AVX
 
 // Multiple of 1.
-__declspec(naked)
-void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+__declspec(naked) void CopyRow_ERMS(const uint8_t* src,
+                                    uint8_t* dst,
+                                    int width) {
   __asm {
     mov        eax, esi
     mov        edx, edi
-    mov        esi, [esp + 4]   // src
-    mov        edi, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
+    mov        esi, [esp + 4]  // src
+    mov        edi, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
     rep movsb
     mov        edi, edx
     mov        esi, eax
@@ -3468,15 +3565,16 @@ void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
 
 #ifdef HAS_ARGBCOPYALPHAROW_SSE2
 // width in pixels
-__declspec(naked)
-void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void ARGBCopyAlphaRow_SSE2(const uint8_t* src,
+                                             uint8_t* dst,
+                                             int width) {
   __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
-    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    mov        eax, [esp + 4]  // src
+    mov        edx, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm0, xmm0  // generate mask 0xff000000
     pslld      xmm0, 24
-    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    pcmpeqb    xmm1, xmm1  // generate mask 0x00ffffff
     psrld      xmm1, 8
 
   convertloop:
@@ -3504,14 +3602,15 @@ void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
 
 #ifdef HAS_ARGBCOPYALPHAROW_AVX2
 // width in pixels
-__declspec(naked)
-void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void ARGBCopyAlphaRow_AVX2(const uint8_t* src,
+                                             uint8_t* dst,
+                                             int width) {
   __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
+    mov        eax, [esp + 4]  // src
+    mov        edx, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
     vpcmpeqb   ymm0, ymm0, ymm0
-    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+    vpsrld     ymm0, ymm0, 8  // generate mask 0x00ffffff
 
   convertloop:
     vmovdqu    ymm1, [eax]
@@ -3533,11 +3632,12 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
 
 #ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
 // width in pixels
-__declspec(naked)
-void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
+__declspec(naked) void ARGBExtractAlphaRow_SSE2(const uint8_t* src_argb,
+                                                uint8_t* dst_a,
+                                                int width) {
   __asm {
-    mov        eax, [esp + 4]   // src_argb
-    mov        edx, [esp + 8]   // dst_a
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_a
     mov        ecx, [esp + 12]  // width
 
   extractloop:
@@ -3558,17 +3658,54 @@ void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
 }
 #endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
+// width in pixels
+__declspec(naked) void ARGBExtractAlphaRow_AVX2(const uint8_t* src_argb,
+                                                uint8_t* dst_a,
+                                                int width) {
+  __asm {
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_a
+    mov        ecx, [esp + 12]  // width
+    vmovdqa    ymm4, ymmword ptr kPermdARGBToY_AVX
+
+  extractloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpsrld     ymm0, ymm0, 24
+    vpsrld     ymm1, ymm1, 24
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    lea        eax, [eax + 128]
+    vpackssdw  ymm0, ymm0, ymm1  // mutates
+    vpsrld     ymm2, ymm2, 24
+    vpsrld     ymm3, ymm3, 24
+    vpackssdw  ymm2, ymm2, ymm3  // mutates
+    vpackuswb  ymm0, ymm0, ymm2  // mutates
+    vpermd     ymm0, ymm4, ymm0  // unmutate
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         extractloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_AVX2
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
-__declspec(naked)
-void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src,
+                                                uint8_t* dst,
+                                                int width) {
   __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
-    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    mov        eax, [esp + 4]  // src
+    mov        edx, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm0, xmm0  // generate mask 0xff000000
     pslld      xmm0, 24
-    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    pcmpeqb    xmm1, xmm1  // generate mask 0x00ffffff
     psrld      xmm1, 8
 
   convertloop:
@@ -3598,14 +3735,15 @@ void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
 
 #ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
 // width in pixels
-__declspec(naked)
-void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+__declspec(naked) void ARGBCopyYToAlphaRow_AVX2(const uint8_t* src,
+                                                uint8_t* dst,
+                                                int width) {
   __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
+    mov        eax, [esp + 4]  // src
+    mov        edx, [esp + 8]  // dst
+    mov        ecx, [esp + 12]  // width
     vpcmpeqb   ymm0, ymm0, ymm0
-    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+    vpsrld     ymm0, ymm0, 8  // generate mask 0x00ffffff
 
   convertloop:
     vpmovzxbd  ymm1, qword ptr [eax]
@@ -3628,17 +3766,16 @@ void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
 #endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
 
 #ifdef HAS_SETROW_X86
-// Write 'count' bytes using an 8 bit value repeated.
-// Count should be multiple of 4.
-__declspec(naked)
-void SetRow_X86(uint8* dst, uint8 v8, int count) {
+// Write 'width' bytes using an 8 bit value repeated.
+// width should be multiple of 4.
+__declspec(naked) void SetRow_X86(uint8_t* dst, uint8_t v8, int width) {
   __asm {
-    movzx      eax, byte ptr [esp + 8]    // v8
+    movzx      eax, byte ptr [esp + 8]  // v8
     mov        edx, 0x01010101  // Duplicate byte to all bytes.
-    mul        edx              // overwrites edx with upper part of result.
+    mul        edx  // overwrites edx with upper part of result.
     mov        edx, edi
-    mov        edi, [esp + 4]   // dst
-    mov        ecx, [esp + 12]  // count
+    mov        edi, [esp + 4]  // dst
+    mov        ecx, [esp + 12]  // width
     shr        ecx, 2
     rep stosd
     mov        edi, edx
@@ -3646,28 +3783,28 @@ void SetRow_X86(uint8* dst, uint8 v8, int count) {
   }
 }
 
-// Write 'count' bytes using an 8 bit value repeated.
-__declspec(naked)
-void SetRow_ERMS(uint8* dst, uint8 v8, int count) {
+// Write 'width' bytes using an 8 bit value repeated.
+__declspec(naked) void SetRow_ERMS(uint8_t* dst, uint8_t v8, int width) {
   __asm {
     mov        edx, edi
-    mov        edi, [esp + 4]   // dst
-    mov        eax, [esp + 8]   // v8
-    mov        ecx, [esp + 12]  // count
+    mov        edi, [esp + 4]  // dst
+    mov        eax, [esp + 8]  // v8
+    mov        ecx, [esp + 12]  // width
     rep stosb
     mov        edi, edx
     ret
   }
 }
 
-// Write 'count' 32 bit values.
-__declspec(naked)
-void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count) {
+// Write 'width' 32 bit values.
+__declspec(naked) void ARGBSetRow_X86(uint8_t* dst_argb,
+                                      uint32_t v32,
+                                      int width) {
   __asm {
     mov        edx, edi
-    mov        edi, [esp + 4]   // dst
-    mov        eax, [esp + 8]   // v32
-    mov        ecx, [esp + 12]  // count
+    mov        edi, [esp + 4]  // dst
+    mov        eax, [esp + 8]  // v32
+    mov        ecx, [esp + 12]  // width
     rep stosd
     mov        edi, edx
     ret
@@ -3676,12 +3813,13 @@ void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count) {
 #endif  // HAS_SETROW_X86
 
 #ifdef HAS_YUY2TOYROW_AVX2
-__declspec(naked)
-void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int width) {
+__declspec(naked) void YUY2ToYRow_AVX2(const uint8_t* src_yuy2,
+                                       uint8_t* dst_y,
+                                       int width) {
   __asm {
-    mov        eax, [esp + 4]    // src_yuy2
-    mov        edx, [esp + 8]    // dst_y
-    mov        ecx, [esp + 12]   // width
+    mov        eax, [esp + 4]  // src_yuy2
+    mov        edx, [esp + 8]  // dst_y
+    mov        ecx, [esp + 12]  // width
     vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
 
@@ -3689,9 +3827,9 @@ void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int width) {
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     lea        eax,  [eax + 64]
-    vpand      ymm0, ymm0, ymm5   // even bytes are Y
+    vpand      ymm0, ymm0, ymm5  // even bytes are Y
     vpand      ymm1, ymm1, ymm5
-    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpackuswb  ymm0, ymm0, ymm1  // mutates.
     vpermq     ymm0, ymm0, 0xd8
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -3702,18 +3840,20 @@ void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int width) {
   }
 }
 
-__declspec(naked)
-void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void YUY2ToUVRow_AVX2(const uint8_t* src_yuy2,
+                                        int stride_yuy2,
+                                        uint8_t* dst_u,
+                                        uint8_t* dst_v,
+                                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_yuy2
-    mov        esi, [esp + 8 + 8]    // stride_yuy2
-    mov        edx, [esp + 8 + 12]   // dst_u
-    mov        edi, [esp + 8 + 16]   // dst_v
-    mov        ecx, [esp + 8 + 20]   // width
-    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    mov        eax, [esp + 8 + 4]  // src_yuy2
+    mov        esi, [esp + 8 + 8]  // stride_yuy2
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
@@ -3723,18 +3863,18 @@ void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
     vpavgb     ymm0, ymm0, [eax + esi]
     vpavgb     ymm1, ymm1, [eax + esi + 32]
     lea        eax,  [eax + 64]
-    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm0, ymm0, 8  // YUYV -> UVUV
     vpsrlw     ymm1, ymm1, 8
-    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpackuswb  ymm0, ymm0, ymm1  // mutates.
     vpermq     ymm0, ymm0, 0xd8
     vpand      ymm1, ymm0, ymm5  // U
-    vpsrlw     ymm0, ymm0, 8     // V
+    vpsrlw     ymm0, ymm0, 8  // V
     vpackuswb  ymm1, ymm1, ymm1  // mutates.
     vpackuswb  ymm0, ymm0, ymm0  // mutates.
     vpermq     ymm1, ymm1, 0xd8
     vpermq     ymm0, ymm0, 0xd8
     vextractf128 [edx], ymm1, 0  // U
-    vextractf128 [edx + edi], ymm0, 0 // V
+    vextractf128 [edx + edi], ymm0, 0  // V
     lea        edx, [edx + 16]
     sub        ecx, 32
     jg         convertloop
@@ -3746,16 +3886,17 @@ void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
   }
 }
 
-__declspec(naked)
-void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void YUY2ToUV422Row_AVX2(const uint8_t* src_yuy2,
+                                           uint8_t* dst_u,
+                                           uint8_t* dst_v,
+                                           int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_yuy2
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // width
-    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    mov        eax, [esp + 4 + 4]  // src_yuy2
+    mov        edx, [esp + 4 + 8]  // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
@@ -3763,18 +3904,18 @@ void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     lea        eax,  [eax + 64]
-    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm0, ymm0, 8  // YUYV -> UVUV
     vpsrlw     ymm1, ymm1, 8
-    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpackuswb  ymm0, ymm0, ymm1  // mutates.
     vpermq     ymm0, ymm0, 0xd8
     vpand      ymm1, ymm0, ymm5  // U
-    vpsrlw     ymm0, ymm0, 8     // V
+    vpsrlw     ymm0, ymm0, 8  // V
     vpackuswb  ymm1, ymm1, ymm1  // mutates.
     vpackuswb  ymm0, ymm0, ymm0  // mutates.
     vpermq     ymm1, ymm1, 0xd8
     vpermq     ymm0, ymm0, 0xd8
     vextractf128 [edx], ymm1, 0  // U
-    vextractf128 [edx + edi], ymm0, 0 // V
+    vextractf128 [edx + edi], ymm0, 0  // V
     lea        edx, [edx + 16]
     sub        ecx, 32
     jg         convertloop
@@ -3785,21 +3926,21 @@ void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
   }
 }
 
-__declspec(naked)
-void UYVYToYRow_AVX2(const uint8* src_uyvy,
-                     uint8* dst_y, int width) {
+__declspec(naked) void UYVYToYRow_AVX2(const uint8_t* src_uyvy,
+                                       uint8_t* dst_y,
+                                       int width) {
   __asm {
-    mov        eax, [esp + 4]    // src_uyvy
-    mov        edx, [esp + 8]    // dst_y
-    mov        ecx, [esp + 12]   // width
+    mov        eax, [esp + 4]  // src_uyvy
+    mov        edx, [esp + 8]  // dst_y
+    mov        ecx, [esp + 12]  // width
 
   convertloop:
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     lea        eax,  [eax + 64]
-    vpsrlw     ymm0, ymm0, 8      // odd bytes are Y
+    vpsrlw     ymm0, ymm0, 8  // odd bytes are Y
     vpsrlw     ymm1, ymm1, 8
-    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpackuswb  ymm0, ymm0, ymm1  // mutates.
     vpermq     ymm0, ymm0, 0xd8
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -3810,18 +3951,20 @@ void UYVYToYRow_AVX2(const uint8* src_uyvy,
   }
 }
 
-__declspec(naked)
-void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void UYVYToUVRow_AVX2(const uint8_t* src_uyvy,
+                                        int stride_uyvy,
+                                        uint8_t* dst_u,
+                                        uint8_t* dst_v,
+                                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_yuy2
-    mov        esi, [esp + 8 + 8]    // stride_yuy2
-    mov        edx, [esp + 8 + 12]   // dst_u
-    mov        edi, [esp + 8 + 16]   // dst_v
-    mov        ecx, [esp + 8 + 20]   // width
-    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    mov        eax, [esp + 8 + 4]  // src_yuy2
+    mov        esi, [esp + 8 + 8]  // stride_yuy2
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
@@ -3831,18 +3974,18 @@ void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
     vpavgb     ymm0, ymm0, [eax + esi]
     vpavgb     ymm1, ymm1, [eax + esi + 32]
     lea        eax,  [eax + 64]
-    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm0, ymm0, ymm5  // UYVY -> UVUV
     vpand      ymm1, ymm1, ymm5
-    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpackuswb  ymm0, ymm0, ymm1  // mutates.
     vpermq     ymm0, ymm0, 0xd8
     vpand      ymm1, ymm0, ymm5  // U
-    vpsrlw     ymm0, ymm0, 8     // V
+    vpsrlw     ymm0, ymm0, 8  // V
     vpackuswb  ymm1, ymm1, ymm1  // mutates.
     vpackuswb  ymm0, ymm0, ymm0  // mutates.
     vpermq     ymm1, ymm1, 0xd8
     vpermq     ymm0, ymm0, 0xd8
     vextractf128 [edx], ymm1, 0  // U
-    vextractf128 [edx + edi], ymm0, 0 // V
+    vextractf128 [edx + edi], ymm0, 0  // V
     lea        edx, [edx + 16]
     sub        ecx, 32
     jg         convertloop
@@ -3854,16 +3997,17 @@ void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
   }
 }
 
-__declspec(naked)
-void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void UYVYToUV422Row_AVX2(const uint8_t* src_uyvy,
+                                           uint8_t* dst_u,
+                                           uint8_t* dst_v,
+                                           int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_yuy2
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // width
-    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    mov        eax, [esp + 4 + 4]  // src_yuy2
+    mov        edx, [esp + 4 + 8]  // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
     vpsrlw     ymm5, ymm5, 8
     sub        edi, edx
 
@@ -3871,18 +4015,18 @@ void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
     vmovdqu    ymm0, [eax]
     vmovdqu    ymm1, [eax + 32]
     lea        eax,  [eax + 64]
-    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm0, ymm0, ymm5  // UYVY -> UVUV
     vpand      ymm1, ymm1, ymm5
-    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpackuswb  ymm0, ymm0, ymm1  // mutates.
     vpermq     ymm0, ymm0, 0xd8
     vpand      ymm1, ymm0, ymm5  // U
-    vpsrlw     ymm0, ymm0, 8     // V
+    vpsrlw     ymm0, ymm0, 8  // V
     vpackuswb  ymm1, ymm1, ymm1  // mutates.
     vpackuswb  ymm0, ymm0, ymm0  // mutates.
     vpermq     ymm1, ymm1, 0xd8
     vpermq     ymm0, ymm0, 0xd8
     vextractf128 [edx], ymm1, 0  // U
-    vextractf128 [edx + edi], ymm0, 0 // V
+    vextractf128 [edx + edi], ymm0, 0  // V
     lea        edx, [edx + 16]
     sub        ecx, 32
     jg         convertloop
@@ -3895,21 +4039,21 @@ void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
 #endif  // HAS_YUY2TOYROW_AVX2
 
 #ifdef HAS_YUY2TOYROW_SSE2
-__declspec(naked)
-void YUY2ToYRow_SSE2(const uint8* src_yuy2,
-                     uint8* dst_y, int width) {
-  __asm {
-    mov        eax, [esp + 4]    // src_yuy2
-    mov        edx, [esp + 8]    // dst_y
-    mov        ecx, [esp + 12]   // width
-    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+__declspec(naked) void YUY2ToYRow_SSE2(const uint8_t* src_yuy2,
+                                       uint8_t* dst_y,
+                                       int width) {
+  __asm {
+    mov        eax, [esp + 4]  // src_yuy2
+    mov        edx, [esp + 8]  // dst_y
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0x00ff00ff
     psrlw      xmm5, 8
 
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
-    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm0, xmm5  // even bytes are Y
     pand       xmm1, xmm5
     packuswb   xmm0, xmm1
     movdqu     [edx], xmm0
@@ -3920,18 +4064,20 @@ void YUY2ToYRow_SSE2(const uint8* src_yuy2,
   }
 }
 
-__declspec(naked)
-void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
-                      uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void YUY2ToUVRow_SSE2(const uint8_t* src_yuy2,
+                                        int stride_yuy2,
+                                        uint8_t* dst_u,
+                                        uint8_t* dst_v,
+                                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_yuy2
-    mov        esi, [esp + 8 + 8]    // stride_yuy2
-    mov        edx, [esp + 8 + 12]   // dst_u
-    mov        edi, [esp + 8 + 16]   // dst_v
-    mov        ecx, [esp + 8 + 20]   // width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    mov        eax, [esp + 8 + 4]  // src_yuy2
+    mov        esi, [esp + 8 + 8]  // stride_yuy2
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0x00ff00ff
     psrlw      xmm5, 8
     sub        edi, edx
 
@@ -3943,13 +4089,13 @@ void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
     lea        eax,  [eax + 32]
     pavgb      xmm0, xmm2
     pavgb      xmm1, xmm3
-    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm0, 8  // YUYV -> UVUV
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
     movdqa     xmm1, xmm0
     pand       xmm0, xmm5  // U
     packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
+    psrlw      xmm1, 8  // V
     packuswb   xmm1, xmm1
     movq       qword ptr [edx], xmm0
     movq       qword ptr [edx + edi], xmm1
@@ -3963,16 +4109,17 @@ void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
   }
 }
 
-__declspec(naked)
-void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
-                         uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void YUY2ToUV422Row_SSE2(const uint8_t* src_yuy2,
+                                           uint8_t* dst_u,
+                                           uint8_t* dst_v,
+                                           int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_yuy2
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    mov        eax, [esp + 4 + 4]  // src_yuy2
+    mov        edx, [esp + 4 + 8]  // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0x00ff00ff
     psrlw      xmm5, 8
     sub        edi, edx
 
@@ -3980,13 +4127,13 @@ void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
-    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm0, 8  // YUYV -> UVUV
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
     movdqa     xmm1, xmm0
     pand       xmm0, xmm5  // U
     packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
+    psrlw      xmm1, 8  // V
     packuswb   xmm1, xmm1
     movq       qword ptr [edx], xmm0
     movq       qword ptr [edx + edi], xmm1
@@ -3999,19 +4146,19 @@ void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
   }
 }
 
-__declspec(naked)
-void UYVYToYRow_SSE2(const uint8* src_uyvy,
-                     uint8* dst_y, int width) {
+__declspec(naked) void UYVYToYRow_SSE2(const uint8_t* src_uyvy,
+                                       uint8_t* dst_y,
+                                       int width) {
   __asm {
-    mov        eax, [esp + 4]    // src_uyvy
-    mov        edx, [esp + 8]    // dst_y
-    mov        ecx, [esp + 12]   // width
+    mov        eax, [esp + 4]  // src_uyvy
+    mov        edx, [esp + 8]  // dst_y
+    mov        ecx, [esp + 12]  // width
 
   convertloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
-    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm0, 8  // odd bytes are Y
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
     movdqu     [edx], xmm0
@@ -4022,18 +4169,20 @@ void UYVYToYRow_SSE2(const uint8* src_uyvy,
   }
 }
 
-__declspec(naked)
-void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
-                      uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void UYVYToUVRow_SSE2(const uint8_t* src_uyvy,
+                                        int stride_uyvy,
+                                        uint8_t* dst_u,
+                                        uint8_t* dst_v,
+                                        int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_yuy2
-    mov        esi, [esp + 8 + 8]    // stride_yuy2
-    mov        edx, [esp + 8 + 12]   // dst_u
-    mov        edi, [esp + 8 + 16]   // dst_v
-    mov        ecx, [esp + 8 + 20]   // width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    mov        eax, [esp + 8 + 4]  // src_yuy2
+    mov        esi, [esp + 8 + 8]  // stride_yuy2
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0x00ff00ff
     psrlw      xmm5, 8
     sub        edi, edx
 
@@ -4045,13 +4194,13 @@ void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
     lea        eax,  [eax + 32]
     pavgb      xmm0, xmm2
     pavgb      xmm1, xmm3
-    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm0, xmm5  // UYVY -> UVUV
     pand       xmm1, xmm5
     packuswb   xmm0, xmm1
     movdqa     xmm1, xmm0
     pand       xmm0, xmm5  // U
     packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
+    psrlw      xmm1, 8  // V
     packuswb   xmm1, xmm1
     movq       qword ptr [edx], xmm0
     movq       qword ptr [edx + edi], xmm1
@@ -4065,16 +4214,17 @@ void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
   }
 }
 
-__declspec(naked)
-void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
-                         uint8* dst_u, uint8* dst_v, int width) {
+__declspec(naked) void UYVYToUV422Row_SSE2(const uint8_t* src_uyvy,
+                                           uint8_t* dst_u,
+                                           uint8_t* dst_v,
+                                           int width) {
   __asm {
     push       edi
-    mov        eax, [esp + 4 + 4]    // src_yuy2
-    mov        edx, [esp + 4 + 8]    // dst_u
-    mov        edi, [esp + 4 + 12]   // dst_v
-    mov        ecx, [esp + 4 + 16]   // width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    mov        eax, [esp + 4 + 4]  // src_yuy2
+    mov        edx, [esp + 4 + 8]  // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5  // generate mask 0x00ff00ff
     psrlw      xmm5, 8
     sub        edi, edx
 
@@ -4082,13 +4232,13 @@ void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
-    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm0, xmm5  // UYVY -> UVUV
     pand       xmm1, xmm5
     packuswb   xmm0, xmm1
     movdqa     xmm1, xmm0
     pand       xmm0, xmm5  // U
     packuswb   xmm0, xmm0
-    psrlw      xmm1, 8     // V
+    psrlw      xmm1, 8  // V
     packuswb   xmm1, xmm1
     movq       qword ptr [edx], xmm0
     movq       qword ptr [edx + edi], xmm1
@@ -4108,13 +4258,15 @@ void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
 // =((A2*C2)+(B2*(255-C2))+255)/256
 // signed version of math
 // =(((A2-128)*C2)+((B2-128)*(255-C2))+32768+127)/256
-__declspec(naked)
-void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
-                         const uint8* alpha, uint8* dst, int width) {
+__declspec(naked) void BlendPlaneRow_SSSE3(const uint8_t* src0,
+                                           const uint8_t* src1,
+                                           const uint8_t* alpha,
+                                           uint8_t* dst,
+                                           int width) {
   __asm {
     push       esi
     push       edi
-    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    pcmpeqb    xmm5, xmm5  // generate mask 0xff00ff00
     psllw      xmm5, 8
     mov        eax, 0x80808080  // 128 for biasing image to signed.
     movd       xmm6, eax
@@ -4123,8 +4275,8 @@ void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
     mov        eax, 0x807f807f  // 32768 + 127 for unbias and round.
     movd       xmm7, eax
     pshufd     xmm7, xmm7, 0x00
-    mov        eax, [esp + 8 + 4]   // src0
-    mov        edx, [esp + 8 + 8]   // src1
+    mov        eax, [esp + 8 + 4]  // src0
+    mov        edx, [esp + 8 + 8]  // src1
     mov        esi, [esp + 8 + 12]  // alpha
     mov        edi, [esp + 8 + 16]  // dst
     mov        ecx, [esp + 8 + 20]  // width
@@ -4132,17 +4284,17 @@ void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
     sub        edx, esi
     sub        edi, esi
 
-    // 8 pixel loop.
+        // 8 pixel loop.
   convertloop8:
-    movq       xmm0, qword ptr [esi]        // alpha
+    movq       xmm0, qword ptr [esi]  // alpha
     punpcklbw  xmm0, xmm0
-    pxor       xmm0, xmm5         // a, 255-a
+    pxor       xmm0, xmm5  // a, 255-a
     movq       xmm1, qword ptr [eax + esi]  // src0
     movq       xmm2, qword ptr [edx + esi]  // src1
     punpcklbw  xmm1, xmm2
-    psubb      xmm1, xmm6         // bias src0/1 - 128
+    psubb      xmm1, xmm6  // bias src0/1 - 128
     pmaddubsw  xmm0, xmm1
-    paddw      xmm0, xmm7         // unbias result - 32768 and round.
+    paddw      xmm0, xmm7  // unbias result - 32768 and round.
     psrlw      xmm0, 8
     packuswb   xmm0, xmm0
     movq       qword ptr [edi + esi], xmm0
@@ -4163,13 +4315,15 @@ void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
 // =((A2*C2)+(B2*(255-C2))+255)/256
 // signed version of math
 // =(((A2-128)*C2)+((B2-128)*(255-C2))+32768+127)/256
-__declspec(naked)
-void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
-                         const uint8* alpha, uint8* dst, int width) {
+__declspec(naked) void BlendPlaneRow_AVX2(const uint8_t* src0,
+                                          const uint8_t* src1,
+                                          const uint8_t* alpha,
+                                          uint8_t* dst,
+                                          int width) {
   __asm {
     push        esi
     push        edi
-    vpcmpeqb    ymm5, ymm5, ymm5       // generate mask 0xff00ff00
+    vpcmpeqb    ymm5, ymm5, ymm5  // generate mask 0xff00ff00
     vpsllw      ymm5, ymm5, 8
     mov         eax, 0x80808080  // 128 for biasing image to signed.
     vmovd       xmm6, eax
@@ -4177,8 +4331,8 @@ void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
     mov         eax, 0x807f807f  // 32768 + 127 for unbias and round.
     vmovd       xmm7, eax
     vbroadcastss ymm7, xmm7
-    mov         eax, [esp + 8 + 4]   // src0
-    mov         edx, [esp + 8 + 8]   // src1
+    mov         eax, [esp + 8 + 4]  // src0
+    mov         edx, [esp + 8 + 8]  // src1
     mov         esi, [esp + 8 + 12]  // alpha
     mov         edi, [esp + 8 + 16]  // dst
     mov         ecx, [esp + 8 + 20]  // width
@@ -4186,23 +4340,23 @@ void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
     sub         edx, esi
     sub         edi, esi
 
-    // 32 pixel loop.
+        // 32 pixel loop.
   convertloop32:
-    vmovdqu     ymm0, [esi]        // alpha
-    vpunpckhbw  ymm3, ymm0, ymm0   // 8..15, 24..31
-    vpunpcklbw  ymm0, ymm0, ymm0   // 0..7, 16..23
-    vpxor       ymm3, ymm3, ymm5   // a, 255-a
-    vpxor       ymm0, ymm0, ymm5   // a, 255-a
+    vmovdqu     ymm0, [esi]  // alpha
+    vpunpckhbw  ymm3, ymm0, ymm0  // 8..15, 24..31
+    vpunpcklbw  ymm0, ymm0, ymm0  // 0..7, 16..23
+    vpxor       ymm3, ymm3, ymm5  // a, 255-a
+    vpxor       ymm0, ymm0, ymm5  // a, 255-a
     vmovdqu     ymm1, [eax + esi]  // src0
     vmovdqu     ymm2, [edx + esi]  // src1
     vpunpckhbw  ymm4, ymm1, ymm2
     vpunpcklbw  ymm1, ymm1, ymm2
-    vpsubb      ymm4, ymm4, ymm6   // bias src0/1 - 128
-    vpsubb      ymm1, ymm1, ymm6   // bias src0/1 - 128
+    vpsubb      ymm4, ymm4, ymm6  // bias src0/1 - 128
+    vpsubb      ymm1, ymm1, ymm6  // bias src0/1 - 128
     vpmaddubsw  ymm3, ymm3, ymm4
     vpmaddubsw  ymm0, ymm0, ymm1
-    vpaddw      ymm3, ymm3, ymm7   // unbias result - 32768 and round.
-    vpaddw      ymm0, ymm0, ymm7   // unbias result - 32768 and round.
+    vpaddw      ymm3, ymm3, ymm7  // unbias result - 32768 and round.
+    vpaddw      ymm0, ymm0, ymm7  // unbias result - 32768 and round.
     vpsrlw      ymm3, ymm3, 8
     vpsrlw      ymm0, ymm0, 8
     vpackuswb   ymm0, ymm0, ymm3
@@ -4221,52 +4375,51 @@ void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
 
 #ifdef HAS_ARGBBLENDROW_SSSE3
 // Shuffle table for isolating alpha.
-static const uvec8 kShuffleAlpha = {
-  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
-  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
-};
+static const uvec8 kShuffleAlpha = {3u,  0x80, 3u,  0x80, 7u,  0x80, 7u,  0x80,
+                                    11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80};
 
 // Blend 8 pixels at a time.
-__declspec(naked)
-void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
-                        uint8* dst_argb, int width) {
+__declspec(naked) void ARGBBlendRow_SSSE3(const uint8_t* src_argb,
+                                          const uint8_t* src_argb1,
+                                          uint8_t* dst_argb,
+                                          int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
-    pcmpeqb    xmm7, xmm7       // generate constant 0x0001
+    pcmpeqb    xmm7, xmm7  // generate constant 0x0001
     psrlw      xmm7, 15
-    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    pcmpeqb    xmm6, xmm6  // generate mask 0x00ff00ff
     psrlw      xmm6, 8
-    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    pcmpeqb    xmm5, xmm5  // generate mask 0xff00ff00
     psllw      xmm5, 8
-    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pcmpeqb    xmm4, xmm4  // generate mask 0xff000000
     pslld      xmm4, 24
     sub        ecx, 4
-    jl         convertloop4b    // less than 4 pixels?
+    jl         convertloop4b  // less than 4 pixels?
 
-    // 4 pixel loop.
+        // 4 pixel loop.
   convertloop4:
-    movdqu     xmm3, [eax]      // src argb
+    movdqu     xmm3, [eax]  // src argb
     lea        eax, [eax + 16]
-    movdqa     xmm0, xmm3       // src argb
-    pxor       xmm3, xmm4       // ~alpha
-    movdqu     xmm2, [esi]      // _r_b
-    pshufb     xmm3, xmmword ptr kShuffleAlpha // alpha
-    pand       xmm2, xmm6       // _r_b
-    paddw      xmm3, xmm7       // 256 - alpha
-    pmullw     xmm2, xmm3       // _r_b * alpha
-    movdqu     xmm1, [esi]      // _a_g
+    movdqa     xmm0, xmm3  // src argb
+    pxor       xmm3, xmm4  // ~alpha
+    movdqu     xmm2, [esi]  // _r_b
+    pshufb     xmm3, xmmword ptr kShuffleAlpha  // alpha
+    pand       xmm2, xmm6  // _r_b
+    paddw      xmm3, xmm7  // 256 - alpha
+    pmullw     xmm2, xmm3  // _r_b * alpha
+    movdqu     xmm1, [esi]  // _a_g
     lea        esi, [esi + 16]
-    psrlw      xmm1, 8          // _a_g
-    por        xmm0, xmm4       // set alpha to 255
-    pmullw     xmm1, xmm3       // _a_g * alpha
-    psrlw      xmm2, 8          // _r_b convert to 8 bits again
-    paddusb    xmm0, xmm2       // + src argb
-    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
-    paddusb    xmm0, xmm1       // + src argb
+    psrlw      xmm1, 8  // _a_g
+    por        xmm0, xmm4  // set alpha to 255
+    pmullw     xmm1, xmm3  // _a_g * alpha
+    psrlw      xmm2, 8  // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2  // + src argb
+    pand       xmm1, xmm5  // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1  // + src argb
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -4276,26 +4429,26 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     add        ecx, 4 - 1
     jl         convertloop1b
 
-    // 1 pixel loop.
+            // 1 pixel loop.
   convertloop1:
-    movd       xmm3, [eax]      // src argb
+    movd       xmm3, [eax]  // src argb
     lea        eax, [eax + 4]
-    movdqa     xmm0, xmm3       // src argb
-    pxor       xmm3, xmm4       // ~alpha
-    movd       xmm2, [esi]      // _r_b
-    pshufb     xmm3, xmmword ptr kShuffleAlpha // alpha
-    pand       xmm2, xmm6       // _r_b
-    paddw      xmm3, xmm7       // 256 - alpha
-    pmullw     xmm2, xmm3       // _r_b * alpha
-    movd       xmm1, [esi]      // _a_g
+    movdqa     xmm0, xmm3  // src argb
+    pxor       xmm3, xmm4  // ~alpha
+    movd       xmm2, [esi]  // _r_b
+    pshufb     xmm3, xmmword ptr kShuffleAlpha  // alpha
+    pand       xmm2, xmm6  // _r_b
+    paddw      xmm3, xmm7  // 256 - alpha
+    pmullw     xmm2, xmm3  // _r_b * alpha
+    movd       xmm1, [esi]  // _a_g
     lea        esi, [esi + 4]
-    psrlw      xmm1, 8          // _a_g
-    por        xmm0, xmm4       // set alpha to 255
-    pmullw     xmm1, xmm3       // _a_g * alpha
-    psrlw      xmm2, 8          // _r_b convert to 8 bits again
-    paddusb    xmm0, xmm2       // + src argb
-    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
-    paddusb    xmm0, xmm1       // + src argb
+    psrlw      xmm1, 8  // _a_g
+    por        xmm0, xmm4  // set alpha to 255
+    pmullw     xmm1, xmm3  // _a_g * alpha
+    psrlw      xmm2, 8  // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2  // + src argb
+    pand       xmm1, xmm5  // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1  // + src argb
     movd       [edx], xmm0
     lea        edx, [edx + 4]
     sub        ecx, 1
@@ -4311,41 +4464,42 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
 // Shuffle table duplicating alpha.
 static const uvec8 kShuffleAlpha0 = {
-  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+    3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
 };
 static const uvec8 kShuffleAlpha1 = {
-  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
-  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+    11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+    15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
 };
-__declspec(naked)
-void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+__declspec(naked) void ARGBAttenuateRow_SSSE3(const uint8_t* src_argb,
+                                              uint8_t* dst_argb,
+                                              int width) {
   __asm {
-    mov        eax, [esp + 4]   // src_argb0
-    mov        edx, [esp + 8]   // dst_argb
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
-    pcmpeqb    xmm3, xmm3       // generate mask 0xff000000
+    pcmpeqb    xmm3, xmm3  // generate mask 0xff000000
     pslld      xmm3, 24
     movdqa     xmm4, xmmword ptr kShuffleAlpha0
     movdqa     xmm5, xmmword ptr kShuffleAlpha1
 
  convertloop:
-    movdqu     xmm0, [eax]      // read 4 pixels
-    pshufb     xmm0, xmm4       // isolate first 2 alphas
-    movdqu     xmm1, [eax]      // read 4 pixels
-    punpcklbw  xmm1, xmm1       // first 2 pixel rgbs
-    pmulhuw    xmm0, xmm1       // rgb * a
-    movdqu     xmm1, [eax]      // read 4 pixels
-    pshufb     xmm1, xmm5       // isolate next 2 alphas
-    movdqu     xmm2, [eax]      // read 4 pixels
-    punpckhbw  xmm2, xmm2       // next 2 pixel rgbs
-    pmulhuw    xmm1, xmm2       // rgb * a
-    movdqu     xmm2, [eax]      // mask original alpha
+    movdqu     xmm0, [eax]  // read 4 pixels
+    pshufb     xmm0, xmm4  // isolate first 2 alphas
+    movdqu     xmm1, [eax]  // read 4 pixels
+    punpcklbw  xmm1, xmm1  // first 2 pixel rgbs
+    pmulhuw    xmm0, xmm1  // rgb * a
+    movdqu     xmm1, [eax]  // read 4 pixels
+    pshufb     xmm1, xmm5  // isolate next 2 alphas
+    movdqu     xmm2, [eax]  // read 4 pixels
+    punpckhbw  xmm2, xmm2  // next 2 pixel rgbs
+    pmulhuw    xmm1, xmm2  // rgb * a
+    movdqu     xmm2, [eax]  // mask original alpha
     lea        eax, [eax + 16]
     pand       xmm2, xmm3
     psrlw      xmm0, 8
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
-    por        xmm0, xmm2       // copy original alpha
+    por        xmm0, xmm2  // copy original alpha
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -4358,22 +4512,23 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
-static const uvec8 kShuffleAlpha_AVX2 = {
-  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
-};
-__declspec(naked)
-void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
-  __asm {
-    mov        eax, [esp + 4]   // src_argb0
-    mov        edx, [esp + 8]   // dst_argb
+static const uvec8 kShuffleAlpha_AVX2 = {6u,   7u,   6u,   7u,  6u,  7u,
+                                         128u, 128u, 14u,  15u, 14u, 15u,
+                                         14u,  15u,  128u, 128u};
+__declspec(naked) void ARGBAttenuateRow_AVX2(const uint8_t* src_argb,
+                                             uint8_t* dst_argb,
+                                             int width) {
+  __asm {
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
     vbroadcastf128 ymm4, xmmword ptr kShuffleAlpha_AVX2
-    vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xff000000
     vpslld     ymm5, ymm5, 24
 
  convertloop:
-    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vmovdqu    ymm6, [eax]  // read 8 pixels.
     vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
     vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
     vpshufb    ymm2, ymm0, ymm4  // low 4 alphas
@@ -4398,40 +4553,40 @@ void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
 // Unattenuate 4 pixels at a time.
-__declspec(naked)
-void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
-                             int width) {
+__declspec(naked) void ARGBUnattenuateRow_SSE2(const uint8_t* src_argb,
+                                               uint8_t* dst_argb,
+                                               int width) {
   __asm {
     push       ebx
     push       esi
     push       edi
-    mov        eax, [esp + 12 + 4]   // src_argb
-    mov        edx, [esp + 12 + 8]   // dst_argb
+    mov        eax, [esp + 12 + 4]  // src_argb
+    mov        edx, [esp + 12 + 8]  // dst_argb
     mov        ecx, [esp + 12 + 12]  // width
     lea        ebx, fixed_invtbl8
 
  convertloop:
-    movdqu     xmm0, [eax]      // read 4 pixels
+    movdqu     xmm0, [eax]  // read 4 pixels
     movzx      esi, byte ptr [eax + 3]  // first alpha
     movzx      edi, byte ptr [eax + 7]  // second alpha
-    punpcklbw  xmm0, xmm0       // first 2
+    punpcklbw  xmm0, xmm0  // first 2
     movd       xmm2, dword ptr [ebx + esi * 4]
     movd       xmm3, dword ptr [ebx + edi * 4]
-    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words.  1, a, a, a
-    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    pshuflw    xmm2, xmm2, 040h  // first 4 inv_alpha words.  1, a, a, a
+    pshuflw    xmm3, xmm3, 040h  // next 4 inv_alpha words
     movlhps    xmm2, xmm3
-    pmulhuw    xmm0, xmm2       // rgb * a
+    pmulhuw    xmm0, xmm2  // rgb * a
 
-    movdqu     xmm1, [eax]      // read 4 pixels
+    movdqu     xmm1, [eax]  // read 4 pixels
     movzx      esi, byte ptr [eax + 11]  // third alpha
     movzx      edi, byte ptr [eax + 15]  // forth alpha
-    punpckhbw  xmm1, xmm1       // next 2
+    punpckhbw  xmm1, xmm1  // next 2
     movd       xmm2, dword ptr [ebx + esi * 4]
     movd       xmm3, dword ptr [ebx + edi * 4]
-    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words
-    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    pshuflw    xmm2, xmm2, 040h  // first 4 inv_alpha words
+    pshuflw    xmm3, xmm3, 040h  // next 4 inv_alpha words
     movlhps    xmm2, xmm3
-    pmulhuw    xmm1, xmm2       // rgb * a
+    pmulhuw    xmm1, xmm2  // rgb * a
     lea        eax, [eax + 16]
     packuswb   xmm0, xmm1
     movdqu     [edx], xmm0
@@ -4450,25 +4605,24 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
 static const uvec8 kUnattenShuffleAlpha_AVX2 = {
-  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
-};
+    0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u};
 // TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
 // USE_GATHER is not on by default, due to being a slow instruction.
 #ifdef USE_GATHER
-__declspec(naked)
-void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
-                             int width) {
+__declspec(naked) void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
+                                               uint8_t* dst_argb,
+                                               int width) {
   __asm {
-    mov        eax, [esp + 4]   // src_argb0
-    mov        edx, [esp + 8]   // dst_argb
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
     vbroadcastf128 ymm4, xmmword ptr kUnattenShuffleAlpha_AVX2
 
  convertloop:
-    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vmovdqu    ymm6, [eax]  // read 8 pixels.
     vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
-    vpsrld     ymm2, ymm6, 24    // alpha in low 8 bits.
+    vpsrld     ymm2, ymm6, 24  // alpha in low 8 bits.
     vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
     vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
     vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5  // ymm5 cleared.  1, a
@@ -4488,50 +4642,50 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     ret
   }
 }
-#else  // USE_GATHER
-__declspec(naked)
-void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
-                             int width) {
+#else   // USE_GATHER
+__declspec(naked) void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
+                                               uint8_t* dst_argb,
+                                               int width) {
   __asm {
 
     push       ebx
     push       esi
     push       edi
-    mov        eax, [esp + 12 + 4]   // src_argb
-    mov        edx, [esp + 12 + 8]   // dst_argb
+    mov        eax, [esp + 12 + 4]  // src_argb
+    mov        edx, [esp + 12 + 8]  // dst_argb
     mov        ecx, [esp + 12 + 12]  // width
     sub        edx, eax
     lea        ebx, fixed_invtbl8
     vbroadcastf128 ymm5, xmmword ptr kUnattenShuffleAlpha_AVX2
 
  convertloop:
-    // replace VPGATHER
-    movzx      esi, byte ptr [eax + 3]                 // alpha0
-    movzx      edi, byte ptr [eax + 7]                 // alpha1
+        // replace VPGATHER
+    movzx      esi, byte ptr [eax + 3]  // alpha0
+    movzx      edi, byte ptr [eax + 7]  // alpha1
     vmovd      xmm0, dword ptr [ebx + esi * 4]  // [1,a0]
     vmovd      xmm1, dword ptr [ebx + edi * 4]  // [1,a1]
-    movzx      esi, byte ptr [eax + 11]                // alpha2
-    movzx      edi, byte ptr [eax + 15]                // alpha3
-    vpunpckldq xmm6, xmm0, xmm1                        // [1,a1,1,a0]
+    movzx      esi, byte ptr [eax + 11]  // alpha2
+    movzx      edi, byte ptr [eax + 15]  // alpha3
+    vpunpckldq xmm6, xmm0, xmm1  // [1,a1,1,a0]
     vmovd      xmm2, dword ptr [ebx + esi * 4]  // [1,a2]
     vmovd      xmm3, dword ptr [ebx + edi * 4]  // [1,a3]
-    movzx      esi, byte ptr [eax + 19]                // alpha4
-    movzx      edi, byte ptr [eax + 23]                // alpha5
-    vpunpckldq xmm7, xmm2, xmm3                        // [1,a3,1,a2]
+    movzx      esi, byte ptr [eax + 19]  // alpha4
+    movzx      edi, byte ptr [eax + 23]  // alpha5
+    vpunpckldq xmm7, xmm2, xmm3  // [1,a3,1,a2]
     vmovd      xmm0, dword ptr [ebx + esi * 4]  // [1,a4]
     vmovd      xmm1, dword ptr [ebx + edi * 4]  // [1,a5]
-    movzx      esi, byte ptr [eax + 27]                // alpha6
-    movzx      edi, byte ptr [eax + 31]                // alpha7
-    vpunpckldq xmm0, xmm0, xmm1                        // [1,a5,1,a4]
+    movzx      esi, byte ptr [eax + 27]  // alpha6
+    movzx      edi, byte ptr [eax + 31]  // alpha7
+    vpunpckldq xmm0, xmm0, xmm1  // [1,a5,1,a4]
     vmovd      xmm2, dword ptr [ebx + esi * 4]  // [1,a6]
     vmovd      xmm3, dword ptr [ebx + edi * 4]  // [1,a7]
-    vpunpckldq xmm2, xmm2, xmm3                        // [1,a7,1,a6]
-    vpunpcklqdq xmm3, xmm6, xmm7                       // [1,a3,1,a2,1,a1,1,a0]
-    vpunpcklqdq xmm0, xmm0, xmm2                       // [1,a7,1,a6,1,a5,1,a4]
-    vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
+    vpunpckldq xmm2, xmm2, xmm3  // [1,a7,1,a6]
+    vpunpcklqdq xmm3, xmm6, xmm7  // [1,a3,1,a2,1,a1,1,a0]
+    vpunpcklqdq xmm0, xmm0, xmm2  // [1,a7,1,a6,1,a5,1,a4]
+    vinserti128 ymm3, ymm3, xmm0, 1                // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
     // end of VPGATHER
 
-    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vmovdqu    ymm6, [eax]  // read 8 pixels.
     vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
     vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
     vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
@@ -4540,7 +4694,7 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
     vpshufb    ymm3, ymm3, ymm5  // replicate high 4 alphas
     vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
     vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
-    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vpackuswb  ymm0, ymm0, ymm1             // unmutated.
     vmovdqu    [eax + edx], ymm0
     lea        eax, [eax + 32]
     sub        ecx, 8
@@ -4558,12 +4712,13 @@ void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBGRAYROW_SSSE3
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels.
-__declspec(naked)
-void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+__declspec(naked) void ARGBGrayRow_SSSE3(const uint8_t* src_argb,
+                                         uint8_t* dst_argb,
+                                         int width) {
   __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_argb */
-    mov        ecx, [esp + 12]  /* width */
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_argb */
+    mov        ecx, [esp + 12] /* width */
     movdqa     xmm4, xmmword ptr kARGBToYJ
     movdqa     xmm5, xmmword ptr kAddYJ64
 
@@ -4575,20 +4730,20 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
     phaddw     xmm0, xmm1
     paddw      xmm0, xmm5  // Add .5 for rounding.
     psrlw      xmm0, 7
-    packuswb   xmm0, xmm0   // 8 G bytes
+    packuswb   xmm0, xmm0  // 8 G bytes
     movdqu     xmm2, [eax]  // A
     movdqu     xmm3, [eax + 16]
     lea        eax, [eax + 32]
     psrld      xmm2, 24
     psrld      xmm3, 24
     packuswb   xmm2, xmm3
-    packuswb   xmm2, xmm2   // 8 A bytes
-    movdqa     xmm3, xmm0   // Weave into GG, GA, then GGGA
-    punpcklbw  xmm0, xmm0   // 8 GG words
-    punpcklbw  xmm3, xmm2   // 8 GA words
+    packuswb   xmm2, xmm2  // 8 A bytes
+    movdqa     xmm3, xmm0  // Weave into GG, GA, then GGGA
+    punpcklbw  xmm0, xmm0  // 8 GG words
+    punpcklbw  xmm3, xmm2  // 8 GA words
     movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm3   // GGGA first 4
-    punpckhwd  xmm1, xmm3   // GGGA next 4
+    punpcklwd  xmm0, xmm3  // GGGA first 4
+    punpckhwd  xmm1, xmm3  // GGGA next 4
     movdqu     [edx], xmm0
     movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
@@ -4604,24 +4759,20 @@ void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
 //    g = (r * 45 + g * 88 + b * 22) >> 7
 //    r = (r * 50 + g * 98 + b * 24) >> 7
 // Constant for ARGB color to sepia tone.
-static const vec8 kARGBToSepiaB = {
-  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
-};
+static const vec8 kARGBToSepiaB = {17, 68, 35, 0, 17, 68, 35, 0,
+                                   17, 68, 35, 0, 17, 68, 35, 0};
 
-static const vec8 kARGBToSepiaG = {
-  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
-};
+static const vec8 kARGBToSepiaG = {22, 88, 45, 0, 22, 88, 45, 0,
+                                   22, 88, 45, 0, 22, 88, 45, 0};
 
-static const vec8 kARGBToSepiaR = {
-  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
-};
+static const vec8 kARGBToSepiaR = {24, 98, 50, 0, 24, 98, 50, 0,
+                                   24, 98, 50, 0, 24, 98, 50, 0};
 
 // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
-__declspec(naked)
-void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+__declspec(naked) void ARGBSepiaRow_SSSE3(uint8_t* dst_argb, int width) {
   __asm {
-    mov        eax, [esp + 4]   /* dst_argb */
-    mov        ecx, [esp + 8]   /* width */
+    mov        eax, [esp + 4] /* dst_argb */
+    mov        ecx, [esp + 8] /* width */
     movdqa     xmm2, xmmword ptr kARGBToSepiaB
     movdqa     xmm3, xmmword ptr kARGBToSepiaG
     movdqa     xmm4, xmmword ptr kARGBToSepiaR
@@ -4633,32 +4784,32 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
     pmaddubsw  xmm6, xmm2
     phaddw     xmm0, xmm6
     psrlw      xmm0, 7
-    packuswb   xmm0, xmm0   // 8 B values
+    packuswb   xmm0, xmm0  // 8 B values
     movdqu     xmm5, [eax]  // G
     movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm5, xmm3
     pmaddubsw  xmm1, xmm3
     phaddw     xmm5, xmm1
     psrlw      xmm5, 7
-    packuswb   xmm5, xmm5   // 8 G values
-    punpcklbw  xmm0, xmm5   // 8 BG values
+    packuswb   xmm5, xmm5  // 8 G values
+    punpcklbw  xmm0, xmm5  // 8 BG values
     movdqu     xmm5, [eax]  // R
     movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm5, xmm4
     pmaddubsw  xmm1, xmm4
     phaddw     xmm5, xmm1
     psrlw      xmm5, 7
-    packuswb   xmm5, xmm5   // 8 R values
+    packuswb   xmm5, xmm5  // 8 R values
     movdqu     xmm6, [eax]  // A
     movdqu     xmm1, [eax + 16]
     psrld      xmm6, 24
     psrld      xmm1, 24
     packuswb   xmm6, xmm1
-    packuswb   xmm6, xmm6   // 8 A values
-    punpcklbw  xmm5, xmm6   // 8 RA values
-    movdqa     xmm1, xmm0   // Weave BG, RA together
-    punpcklwd  xmm0, xmm5   // BGRA first 4
-    punpckhwd  xmm1, xmm5   // BGRA next 4
+    packuswb   xmm6, xmm6  // 8 A values
+    punpcklbw  xmm5, xmm6  // 8 RA values
+    movdqa     xmm1, xmm0  // Weave BG, RA together
+    punpcklwd  xmm0, xmm5  // BGRA first 4
+    punpckhwd  xmm1, xmm5  // BGRA next 4
     movdqu     [eax], xmm0
     movdqu     [eax + 16], xmm1
     lea        eax, [eax + 32]
@@ -4674,19 +4825,20 @@ void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
 // Same as Sepia except matrix is provided.
 // TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R
 // and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd.
-__declspec(naked)
-void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                              const int8* matrix_argb, int width) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_argb */
-    mov        ecx, [esp + 12]  /* matrix_argb */
+__declspec(naked) void ARGBColorMatrixRow_SSSE3(const uint8_t* src_argb,
+                                                uint8_t* dst_argb,
+                                                const int8_t* matrix_argb,
+                                                int width) {
+  __asm {
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_argb */
+    mov        ecx, [esp + 12] /* matrix_argb */
     movdqu     xmm5, [ecx]
     pshufd     xmm2, xmm5, 0x00
     pshufd     xmm3, xmm5, 0x55
     pshufd     xmm4, xmm5, 0xaa
     pshufd     xmm5, xmm5, 0xff
-    mov        ecx, [esp + 16]  /* width */
+    mov        ecx, [esp + 16] /* width */
 
  convertloop:
     movdqu     xmm0, [eax]  // B
@@ -4697,31 +4849,31 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
     movdqu     xmm1, [eax + 16]
     pmaddubsw  xmm6, xmm3
     pmaddubsw  xmm1, xmm3
-    phaddsw    xmm0, xmm7   // B
-    phaddsw    xmm6, xmm1   // G
-    psraw      xmm0, 6      // B
-    psraw      xmm6, 6      // G
-    packuswb   xmm0, xmm0   // 8 B values
-    packuswb   xmm6, xmm6   // 8 G values
-    punpcklbw  xmm0, xmm6   // 8 BG values
+    phaddsw    xmm0, xmm7  // B
+    phaddsw    xmm6, xmm1  // G
+    psraw      xmm0, 6  // B
+    psraw      xmm6, 6  // G
+    packuswb   xmm0, xmm0  // 8 B values
+    packuswb   xmm6, xmm6  // 8 G values
+    punpcklbw  xmm0, xmm6  // 8 BG values
     movdqu     xmm1, [eax]  // R
     movdqu     xmm7, [eax + 16]
     pmaddubsw  xmm1, xmm4
     pmaddubsw  xmm7, xmm4
-    phaddsw    xmm1, xmm7   // R
+    phaddsw    xmm1, xmm7  // R
     movdqu     xmm6, [eax]  // A
     movdqu     xmm7, [eax + 16]
     pmaddubsw  xmm6, xmm5
     pmaddubsw  xmm7, xmm5
-    phaddsw    xmm6, xmm7   // A
-    psraw      xmm1, 6      // R
-    psraw      xmm6, 6      // A
-    packuswb   xmm1, xmm1   // 8 R values
-    packuswb   xmm6, xmm6   // 8 A values
-    punpcklbw  xmm1, xmm6   // 8 RA values
-    movdqa     xmm6, xmm0   // Weave BG, RA together
-    punpcklwd  xmm0, xmm1   // BGRA first 4
-    punpckhwd  xmm6, xmm1   // BGRA next 4
+    phaddsw    xmm6, xmm7  // A
+    psraw      xmm1, 6  // R
+    psraw      xmm6, 6  // A
+    packuswb   xmm1, xmm1  // 8 R values
+    packuswb   xmm6, xmm6  // 8 A values
+    punpcklbw  xmm1, xmm6  // 8 RA values
+    movdqa     xmm6, xmm0  // Weave BG, RA together
+    punpcklwd  xmm0, xmm1  // BGRA first 4
+    punpckhwd  xmm6, xmm1  // BGRA next 4
     movdqu     [edx], xmm0
     movdqu     [edx + 16], xmm6
     lea        eax, [eax + 32]
@@ -4735,15 +4887,17 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBQUANTIZEROW_SSE2
 // Quantize 4 ARGB pixels (16 bytes).
-__declspec(naked)
-void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
-                          int interval_offset, int width) {
-  __asm {
-    mov        eax, [esp + 4]    /* dst_argb */
-    movd       xmm2, [esp + 8]   /* scale */
-    movd       xmm3, [esp + 12]  /* interval_size */
-    movd       xmm4, [esp + 16]  /* interval_offset */
-    mov        ecx, [esp + 20]   /* width */
+__declspec(naked) void ARGBQuantizeRow_SSE2(uint8_t* dst_argb,
+                                            int scale,
+                                            int interval_size,
+                                            int interval_offset,
+                                            int width) {
+  __asm {
+    mov        eax, [esp + 4] /* dst_argb */
+    movd       xmm2, [esp + 8] /* scale */
+    movd       xmm3, [esp + 12] /* interval_size */
+    movd       xmm4, [esp + 16] /* interval_offset */
+    mov        ecx, [esp + 20] /* width */
     pshuflw    xmm2, xmm2, 040h
     pshufd     xmm2, xmm2, 044h
     pshuflw    xmm3, xmm3, 040h
@@ -4756,16 +4910,16 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
 
  convertloop:
     movdqu     xmm0, [eax]  // read 4 pixels
-    punpcklbw  xmm0, xmm5   // first 2 pixels
-    pmulhuw    xmm0, xmm2   // pixel * scale >> 16
+    punpcklbw  xmm0, xmm5  // first 2 pixels
+    pmulhuw    xmm0, xmm2  // pixel * scale >> 16
     movdqu     xmm1, [eax]  // read 4 pixels
-    punpckhbw  xmm1, xmm5   // next 2 pixels
+    punpckhbw  xmm1, xmm5  // next 2 pixels
     pmulhuw    xmm1, xmm2
-    pmullw     xmm0, xmm3   // * interval_size
+    pmullw     xmm0, xmm3  // * interval_size
     movdqu     xmm7, [eax]  // read 4 pixels
     pmullw     xmm1, xmm3
-    pand       xmm7, xmm6   // mask alpha
-    paddw      xmm0, xmm4   // + interval_size / 2
+    pand       xmm7, xmm6  // mask alpha
+    paddw      xmm0, xmm4  // + interval_size / 2
     paddw      xmm1, xmm4
     packuswb   xmm0, xmm1
     por        xmm0, xmm7
@@ -4780,25 +4934,26 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
 
 #ifdef HAS_ARGBSHADEROW_SSE2
 // Shade 4 pixels at a time by specified value.
-__declspec(naked)
-void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value) {
+__declspec(naked) void ARGBShadeRow_SSE2(const uint8_t* src_argb,
+                                         uint8_t* dst_argb,
+                                         int width,
+                                         uint32_t value) {
   __asm {
-    mov        eax, [esp + 4]   // src_argb
-    mov        edx, [esp + 8]   // dst_argb
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     movd       xmm2, [esp + 16]  // value
     punpcklbw  xmm2, xmm2
     punpcklqdq xmm2, xmm2
 
  convertloop:
-    movdqu     xmm0, [eax]      // read 4 pixels
+    movdqu     xmm0, [eax]  // read 4 pixels
     lea        eax, [eax + 16]
     movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm0       // first 2
-    punpckhbw  xmm1, xmm1       // next 2
-    pmulhuw    xmm0, xmm2       // argb * value
-    pmulhuw    xmm1, xmm2       // argb * value
+    punpcklbw  xmm0, xmm0  // first 2
+    punpckhbw  xmm1, xmm1  // next 2
+    pmulhuw    xmm0, xmm2  // argb * value
+    pmulhuw    xmm1, xmm2  // argb * value
     psrlw      xmm0, 8
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
@@ -4814,28 +4969,29 @@ void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
 
 #ifdef HAS_ARGBMULTIPLYROW_SSE2
 // Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
-__declspec(naked)
-void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
+__declspec(naked) void ARGBMultiplyRow_SSE2(const uint8_t* src_argb,
+                                            const uint8_t* src_argb1,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
     pxor       xmm5, xmm5  // constant 0
 
  convertloop:
-    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
-    movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm0, [eax]  // read 4 pixels from src_argb
+    movdqu     xmm2, [esi]  // read 4 pixels from src_argb1
     movdqu     xmm1, xmm0
     movdqu     xmm3, xmm2
-    punpcklbw  xmm0, xmm0         // first 2
-    punpckhbw  xmm1, xmm1         // next 2
-    punpcklbw  xmm2, xmm5         // first 2
-    punpckhbw  xmm3, xmm5         // next 2
-    pmulhuw    xmm0, xmm2         // src_argb0 * src_argb1 first 2
-    pmulhuw    xmm1, xmm3         // src_argb0 * src_argb1 next 2
+    punpcklbw  xmm0, xmm0  // first 2
+    punpckhbw  xmm1, xmm1  // next 2
+    punpcklbw  xmm2, xmm5  // first 2
+    punpckhbw  xmm3, xmm5  // next 2
+    pmulhuw    xmm0, xmm2  // src_argb * src_argb1 first 2
+    pmulhuw    xmm1, xmm3  // src_argb * src_argb1 next 2
     lea        eax, [eax + 16]
     lea        esi, [esi + 16]
     packuswb   xmm0, xmm1
@@ -4853,13 +5009,14 @@ void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
 #ifdef HAS_ARGBADDROW_SSE2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
 // TODO(fbarchard): Port this to posix, neon and other math functions.
-__declspec(naked)
-void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
-                     uint8* dst_argb, int width) {
+__declspec(naked) void ARGBAddRow_SSE2(const uint8_t* src_argb,
+                                       const uint8_t* src_argb1,
+                                       uint8_t* dst_argb,
+                                       int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
@@ -4867,11 +5024,11 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     jl         convertloop49
 
  convertloop4:
-    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm0, [eax]  // read 4 pixels from src_argb
     lea        eax, [eax + 16]
-    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, [esi]  // read 4 pixels from src_argb1
     lea        esi, [esi + 16]
-    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    paddusb    xmm0, xmm1  // src_argb + src_argb1
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -4882,11 +5039,11 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
     jl         convertloop19
 
  convertloop1:
-    movd       xmm0, [eax]        // read 1 pixels from src_argb0
+    movd       xmm0, [eax]  // read 1 pixels from src_argb
     lea        eax, [eax + 4]
-    movd       xmm1, [esi]        // read 1 pixels from src_argb1
+    movd       xmm1, [esi]  // read 1 pixels from src_argb1
     lea        esi, [esi + 4]
-    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    paddusb    xmm0, xmm1  // src_argb + src_argb1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
     sub        ecx, 1
@@ -4901,22 +5058,23 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBSUBTRACTROW_SSE2
 // Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
-__declspec(naked)
-void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
+__declspec(naked) void ARGBSubtractRow_SSE2(const uint8_t* src_argb,
+                                            const uint8_t* src_argb1,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
  convertloop:
-    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm0, [eax]  // read 4 pixels from src_argb
     lea        eax, [eax + 16]
-    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, [esi]  // read 4 pixels from src_argb1
     lea        esi, [esi + 16]
-    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    psubusb    xmm0, xmm1  // src_argb - src_argb1
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -4930,28 +5088,29 @@ void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBMULTIPLYROW_AVX2
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-__declspec(naked)
-void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
+__declspec(naked) void ARGBMultiplyRow_AVX2(const uint8_t* src_argb,
+                                            const uint8_t* src_argb1,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
-    vpxor      ymm5, ymm5, ymm5     // constant 0
+    vpxor      ymm5, ymm5, ymm5  // constant 0
 
  convertloop:
-    vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
+    vmovdqu    ymm1, [eax]  // read 8 pixels from src_argb
     lea        eax, [eax + 32]
-    vmovdqu    ymm3, [esi]        // read 8 pixels from src_argb1
+    vmovdqu    ymm3, [esi]  // read 8 pixels from src_argb1
     lea        esi, [esi + 32]
-    vpunpcklbw ymm0, ymm1, ymm1   // low 4
-    vpunpckhbw ymm1, ymm1, ymm1   // high 4
-    vpunpcklbw ymm2, ymm3, ymm5   // low 4
-    vpunpckhbw ymm3, ymm3, ymm5   // high 4
-    vpmulhuw   ymm0, ymm0, ymm2   // src_argb0 * src_argb1 low 4
-    vpmulhuw   ymm1, ymm1, ymm3   // src_argb0 * src_argb1 high 4
+    vpunpcklbw ymm0, ymm1, ymm1  // low 4
+    vpunpckhbw ymm1, ymm1, ymm1  // high 4
+    vpunpcklbw ymm2, ymm3, ymm5  // low 4
+    vpunpckhbw ymm3, ymm3, ymm5  // high 4
+    vpmulhuw   ymm0, ymm0, ymm2  // src_argb * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3  // src_argb * src_argb1 high 4
     vpackuswb  ymm0, ymm0, ymm1
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -4967,20 +5126,21 @@ void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBADDROW_AVX2
 // Add 2 rows of ARGB pixels together, 8 pixels at a time.
-__declspec(naked)
-void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
-                     uint8* dst_argb, int width) {
+__declspec(naked) void ARGBAddRow_AVX2(const uint8_t* src_argb,
+                                       const uint8_t* src_argb1,
+                                       uint8_t* dst_argb,
+                                       int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
  convertloop:
-    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    vmovdqu    ymm0, [eax]  // read 8 pixels from src_argb
     lea        eax, [eax + 32]
-    vpaddusb   ymm0, ymm0, [esi]        // add 8 pixels from src_argb1
+    vpaddusb   ymm0, ymm0, [esi]  // add 8 pixels from src_argb1
     lea        esi, [esi + 32]
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -4996,20 +5156,21 @@ void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBSUBTRACTROW_AVX2
 // Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
-__declspec(naked)
-void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
-                          uint8* dst_argb, int width) {
+__declspec(naked) void ARGBSubtractRow_AVX2(const uint8_t* src_argb,
+                                            const uint8_t* src_argb1,
+                                            uint8_t* dst_argb,
+                                            int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_argb0
-    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
  convertloop:
-    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    vmovdqu    ymm0, [eax]  // read 8 pixels from src_argb
     lea        eax, [eax + 32]
-    vpsubusb   ymm0, ymm0, [esi]        // src_argb0 - src_argb1
+    vpsubusb   ymm0, ymm0, [esi]  // src_argb - src_argb1
     lea        esi, [esi + 32]
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -5028,14 +5189,16 @@ void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
 // -1  0  1
 // -2  0  2
 // -1  0  1
-__declspec(naked)
-void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
-                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+__declspec(naked) void SobelXRow_SSE2(const uint8_t* src_y0,
+                                      const uint8_t* src_y1,
+                                      const uint8_t* src_y2,
+                                      uint8_t* dst_sobelx,
+                                      int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   // src_y0
-    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        eax, [esp + 8 + 4]  // src_y0
+    mov        esi, [esp + 8 + 8]  // src_y1
     mov        edi, [esp + 8 + 12]  // src_y2
     mov        edx, [esp + 8 + 16]  // dst_sobelx
     mov        ecx, [esp + 8 + 20]  // width
@@ -5045,17 +5208,17 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     pxor       xmm5, xmm5  // constant 0
 
  convertloop:
-    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
-    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm0, qword ptr [eax]  // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]  // read 8 pixels from src_y0[2]
     punpcklbw  xmm0, xmm5
     punpcklbw  xmm1, xmm5
     psubw      xmm0, xmm1
-    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm1, qword ptr [eax + esi]  // read 8 pixels from src_y1[0]
     movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
     punpcklbw  xmm1, xmm5
     punpcklbw  xmm2, xmm5
     psubw      xmm1, xmm2
-    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm2, qword ptr [eax + edi]  // read 8 pixels from src_y2[0]
     movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
     punpcklbw  xmm2, xmm5
     punpcklbw  xmm3, xmm5
@@ -5063,7 +5226,7 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     paddw      xmm0, xmm2
     paddw      xmm0, xmm1
     paddw      xmm0, xmm1
-    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    pxor       xmm1, xmm1  // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
     psubw      xmm1, xmm0
     pmaxsw     xmm0, xmm1
     packuswb   xmm0, xmm0
@@ -5084,13 +5247,14 @@ void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
 // -1 -2 -1
 //  0  0  0
 //  1  2  1
-__declspec(naked)
-void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
-                    uint8* dst_sobely, int width) {
+__declspec(naked) void SobelYRow_SSE2(const uint8_t* src_y0,
+                                      const uint8_t* src_y1,
+                                      uint8_t* dst_sobely,
+                                      int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_y0
-    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        eax, [esp + 4 + 4]  // src_y0
+    mov        esi, [esp + 4 + 8]  // src_y1
     mov        edx, [esp + 4 + 12]  // dst_sobely
     mov        ecx, [esp + 4 + 16]  // width
     sub        esi, eax
@@ -5098,17 +5262,17 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     pxor       xmm5, xmm5  // constant 0
 
  convertloop:
-    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
-    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm0, qword ptr [eax]  // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]  // read 8 pixels from src_y1[0]
     punpcklbw  xmm0, xmm5
     punpcklbw  xmm1, xmm5
     psubw      xmm0, xmm1
-    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm1, qword ptr [eax + 1]  // read 8 pixels from src_y0[1]
     movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
     punpcklbw  xmm1, xmm5
     punpcklbw  xmm2, xmm5
     psubw      xmm1, xmm2
-    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm2, qword ptr [eax + 2]  // read 8 pixels from src_y0[2]
     movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
     punpcklbw  xmm2, xmm5
     punpcklbw  xmm3, xmm5
@@ -5116,7 +5280,7 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
     paddw      xmm0, xmm2
     paddw      xmm0, xmm1
     paddw      xmm0, xmm1
-    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    pxor       xmm1, xmm1  // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
     psubw      xmm1, xmm0
     pmaxsw     xmm0, xmm1
     packuswb   xmm0, xmm0
@@ -5137,36 +5301,37 @@ void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
 // R = Sobel
 // G = Sobel
 // B = Sobel
-__declspec(naked)
-void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                   uint8* dst_argb, int width) {
+__declspec(naked) void SobelRow_SSE2(const uint8_t* src_sobelx,
+                                     const uint8_t* src_sobely,
+                                     uint8_t* dst_argb,
+                                     int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_sobelx
-    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        eax, [esp + 4 + 4]  // src_sobelx
+    mov        esi, [esp + 4 + 8]  // src_sobely
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
     sub        esi, eax
-    pcmpeqb    xmm5, xmm5           // alpha 255
-    pslld      xmm5, 24             // 0xff000000
+    pcmpeqb    xmm5, xmm5  // alpha 255
+    pslld      xmm5, 24  // 0xff000000
 
  convertloop:
-    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
-    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    movdqu     xmm0, [eax]  // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]  // read 16 pixels src_sobely
     lea        eax, [eax + 16]
-    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
-    movdqa     xmm2, xmm0             // GG
-    punpcklbw  xmm2, xmm0             // First 8
-    punpckhbw  xmm0, xmm0             // Next 8
-    movdqa     xmm1, xmm2             // GGGG
-    punpcklwd  xmm1, xmm2             // First 4
-    punpckhwd  xmm2, xmm2             // Next 4
-    por        xmm1, xmm5             // GGGA
+    paddusb    xmm0, xmm1  // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0  // GG
+    punpcklbw  xmm2, xmm0  // First 8
+    punpckhbw  xmm0, xmm0  // Next 8
+    movdqa     xmm1, xmm2  // GGGG
+    punpcklwd  xmm1, xmm2  // First 4
+    punpckhwd  xmm2, xmm2  // Next 4
+    por        xmm1, xmm5  // GGGA
     por        xmm2, xmm5
-    movdqa     xmm3, xmm0             // GGGG
-    punpcklwd  xmm3, xmm0             // Next 4
-    punpckhwd  xmm0, xmm0             // Last 4
-    por        xmm3, xmm5             // GGGA
+    movdqa     xmm3, xmm0  // GGGG
+    punpcklwd  xmm3, xmm0  // Next 4
+    punpckhwd  xmm0, xmm0  // Last 4
+    por        xmm3, xmm5  // GGGA
     por        xmm0, xmm5
     movdqu     [edx], xmm1
     movdqu     [edx + 16], xmm2
@@ -5184,22 +5349,23 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 
 #ifdef HAS_SOBELTOPLANEROW_SSE2
 // Adds Sobel X and Sobel Y and stores Sobel into a plane.
-__declspec(naked)
-void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                          uint8* dst_y, int width) {
+__declspec(naked) void SobelToPlaneRow_SSE2(const uint8_t* src_sobelx,
+                                            const uint8_t* src_sobely,
+                                            uint8_t* dst_y,
+                                            int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_sobelx
-    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        eax, [esp + 4 + 4]  // src_sobelx
+    mov        esi, [esp + 4 + 8]  // src_sobely
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
     sub        esi, eax
 
  convertloop:
-    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
-    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    movdqu     xmm0, [eax]  // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]  // read 16 pixels src_sobely
     lea        eax, [eax + 16]
-    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    paddusb    xmm0, xmm1  // sobel = sobelx + sobely
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 16
@@ -5217,36 +5383,37 @@ void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 // R = Sobel X
 // G = Sobel
 // B = Sobel Y
-__declspec(naked)
-void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
-                     uint8* dst_argb, int width) {
+__declspec(naked) void SobelXYRow_SSE2(const uint8_t* src_sobelx,
+                                       const uint8_t* src_sobely,
+                                       uint8_t* dst_argb,
+                                       int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   // src_sobelx
-    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        eax, [esp + 4 + 4]  // src_sobelx
+    mov        esi, [esp + 4 + 8]  // src_sobely
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
     sub        esi, eax
-    pcmpeqb    xmm5, xmm5           // alpha 255
+    pcmpeqb    xmm5, xmm5  // alpha 255
 
  convertloop:
-    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
-    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    movdqu     xmm0, [eax]  // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]  // read 16 pixels src_sobely
     lea        eax, [eax + 16]
     movdqa     xmm2, xmm0
-    paddusb    xmm2, xmm1             // sobel = sobelx + sobely
-    movdqa     xmm3, xmm0             // XA
+    paddusb    xmm2, xmm1  // sobel = sobelx + sobely
+    movdqa     xmm3, xmm0  // XA
     punpcklbw  xmm3, xmm5
     punpckhbw  xmm0, xmm5
-    movdqa     xmm4, xmm1             // YS
+    movdqa     xmm4, xmm1  // YS
     punpcklbw  xmm4, xmm2
     punpckhbw  xmm1, xmm2
-    movdqa     xmm6, xmm4             // YSXA
-    punpcklwd  xmm6, xmm3             // First 4
-    punpckhwd  xmm4, xmm3             // Next 4
-    movdqa     xmm7, xmm1             // YSXA
-    punpcklwd  xmm7, xmm0             // Next 4
-    punpckhwd  xmm1, xmm0             // Last 4
+    movdqa     xmm6, xmm4  // YSXA
+    punpcklwd  xmm6, xmm3  // First 4
+    punpckhwd  xmm4, xmm3  // Next 4
+    movdqa     xmm7, xmm1  // YSXA
+    punpcklwd  xmm7, xmm0  // Next 4
+    punpckhwd  xmm1, xmm0  // Last 4
     movdqu     [edx], xmm6
     movdqu     [edx + 16], xmm4
     movdqu     [edx + 32], xmm7
@@ -5275,8 +5442,11 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 // count is number of averaged pixels to produce.
 // Does 4 pixels at a time.
 // This function requires alignment on accumulation buffer pointers.
-void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
-                                    int width, int area, uint8* dst,
+void CumulativeSumToAverageRow_SSE2(const int32_t* topleft,
+                                    const int32_t* botleft,
+                                    int width,
+                                    int area,
+                                    uint8_t* dst,
                                     int count) {
   __asm {
     mov        eax, topleft  // eax topleft
@@ -5294,18 +5464,18 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     cmp        area, 128  // 128 pixels will not overflow 15 bits.
     ja         l4
 
-    pshufd     xmm5, xmm5, 0        // area
-    pcmpeqb    xmm6, xmm6           // constant of 65536.0 - 1 = 65535.0
+    pshufd     xmm5, xmm5, 0  // area
+    pcmpeqb    xmm6, xmm6  // constant of 65536.0 - 1 = 65535.0
     psrld      xmm6, 16
     cvtdq2ps   xmm6, xmm6
-    addps      xmm5, xmm6           // (65536.0 + area - 1)
-    mulps      xmm5, xmm4           // (65536.0 + area - 1) * 1 / area
-    cvtps2dq   xmm5, xmm5           // 0.16 fixed point
-    packssdw   xmm5, xmm5           // 16 bit shorts
+    addps      xmm5, xmm6  // (65536.0 + area - 1)
+    mulps      xmm5, xmm4  // (65536.0 + area - 1) * 1 / area
+    cvtps2dq   xmm5, xmm5  // 0.16 fixed point
+    packssdw   xmm5, xmm5  // 16 bit shorts
 
-    // 4 pixel loop small blocks.
+        // 4 pixel loop small blocks.
   s4:
-    // top left
+        // top left
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     movdqu     xmm2, [eax + 32]
@@ -5345,9 +5515,9 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
 
     jmp        l4b
 
-    // 4 pixel loop
+            // 4 pixel loop
   l4:
-    // top left
+        // top left
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     movdqu     xmm2, [eax + 32]
@@ -5373,7 +5543,7 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     paddd      xmm3, [esi + edx * 4 + 48]
     lea        esi, [esi + 64]
 
-    cvtdq2ps   xmm0, xmm0   // Average = Sum * 1 / Area
+    cvtdq2ps   xmm0, xmm0  // Average = Sum * 1 / Area
     cvtdq2ps   xmm1, xmm1
     mulps      xmm0, xmm4
     mulps      xmm1, xmm4
@@ -5397,7 +5567,7 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
     add        ecx, 4 - 1
     jl         l1b
 
-    // 1 pixel loop
+            // 1 pixel loop
   l1:
     movdqu     xmm0, [eax]
     psubd      xmm0, [eax + edx * 4]
@@ -5422,8 +5592,10 @@ void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
 #ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
 // Creates a table of cumulative sums where each value is a sum of all values
 // above and to the left of the value.
-void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
-                                  const int32* previous_cumsum, int width) {
+void ComputeCumulativeSumRow_SSE2(const uint8_t* row,
+                                  int32_t* cumsum,
+                                  const int32_t* previous_cumsum,
+                                  int width) {
   __asm {
     mov        eax, row
     mov        edx, cumsum
@@ -5437,7 +5609,7 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     test       edx, 15
     jne        l4b
 
-    // 4 pixel loop
+        // 4 pixel loop
   l4:
     movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
     lea        eax, [eax + 16]
@@ -5483,9 +5655,9 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
     add        ecx, 4 - 1
     jl         l1b
 
-    // 1 pixel loop
+            // 1 pixel loop
   l1:
-    movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel
     lea        eax, [eax + 4]
     punpcklbw  xmm2, xmm1
     punpcklwd  xmm2, xmm1
@@ -5505,10 +5677,11 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
 
 #ifdef HAS_ARGBAFFINEROW_SSE2
 // Copy ARGB pixels from source image with slope to a row of destination.
-__declspec(naked)
-LIBYUV_API
-void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
-                        uint8* dst_argb, const float* uv_dudv, int width) {
+__declspec(naked) LIBYUV_API void ARGBAffineRow_SSE2(const uint8_t* src_argb,
+                                                     int src_argb_stride,
+                                                     uint8_t* dst_argb,
+                                                     const float* uv_dudv,
+                                                     int width) {
   __asm {
     push       esi
     push       edi
@@ -5519,46 +5692,46 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     movq       xmm2, qword ptr [ecx]  // uv
     movq       xmm7, qword ptr [ecx + 8]  // dudv
     mov        ecx, [esp + 28]  // width
-    shl        esi, 16          // 4, stride
+    shl        esi, 16  // 4, stride
     add        esi, 4
     movd       xmm5, esi
     sub        ecx, 4
     jl         l4b
 
-    // setup for 4 pixel loop
+        // setup for 4 pixel loop
     pshufd     xmm7, xmm7, 0x44  // dup dudv
     pshufd     xmm5, xmm5, 0  // dup 4, stride
-    movdqa     xmm0, xmm2    // x0, y0, x1, y1
+    movdqa     xmm0, xmm2  // x0, y0, x1, y1
     addps      xmm0, xmm7
     movlhps    xmm2, xmm0
     movdqa     xmm4, xmm7
-    addps      xmm4, xmm4    // dudv *= 2
-    movdqa     xmm3, xmm2    // x2, y2, x3, y3
+    addps      xmm4, xmm4  // dudv *= 2
+    movdqa     xmm3, xmm2  // x2, y2, x3, y3
     addps      xmm3, xmm4
-    addps      xmm4, xmm4    // dudv *= 4
+    addps      xmm4, xmm4  // dudv *= 4
 
-    // 4 pixel loop
+        // 4 pixel loop
   l4:
-    cvttps2dq  xmm0, xmm2    // x, y float to int first 2
-    cvttps2dq  xmm1, xmm3    // x, y float to int next 2
-    packssdw   xmm0, xmm1    // x, y as 8 shorts
-    pmaddwd    xmm0, xmm5    // offsets = x * 4 + y * stride.
+    cvttps2dq  xmm0, xmm2  // x, y float to int first 2
+    cvttps2dq  xmm1, xmm3  // x, y float to int next 2
+    packssdw   xmm0, xmm1  // x, y as 8 shorts
+    pmaddwd    xmm0, xmm5  // offsets = x * 4 + y * stride.
     movd       esi, xmm0
     pshufd     xmm0, xmm0, 0x39  // shift right
     movd       edi, xmm0
     pshufd     xmm0, xmm0, 0x39  // shift right
     movd       xmm1, [eax + esi]  // read pixel 0
     movd       xmm6, [eax + edi]  // read pixel 1
-    punpckldq  xmm1, xmm6     // combine pixel 0 and 1
-    addps      xmm2, xmm4    // x, y += dx, dy first 2
+    punpckldq  xmm1, xmm6  // combine pixel 0 and 1
+    addps      xmm2, xmm4  // x, y += dx, dy first 2
     movq       qword ptr [edx], xmm1
     movd       esi, xmm0
     pshufd     xmm0, xmm0, 0x39  // shift right
     movd       edi, xmm0
     movd       xmm6, [eax + esi]  // read pixel 2
     movd       xmm0, [eax + edi]  // read pixel 3
-    punpckldq  xmm6, xmm0     // combine pixel 2 and 3
-    addps      xmm3, xmm4    // x, y += dx, dy next 2
+    punpckldq  xmm6, xmm0  // combine pixel 2 and 3
+    addps      xmm3, xmm4  // x, y += dx, dy next 2
     movq       qword ptr 8[edx], xmm6
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -5568,12 +5741,12 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
     add        ecx, 4 - 1
     jl         l1b
 
-    // 1 pixel loop
+            // 1 pixel loop
   l1:
-    cvttps2dq  xmm0, xmm2    // x, y float to int
-    packssdw   xmm0, xmm0    // x, y as shorts
-    pmaddwd    xmm0, xmm5    // offset = x * 4 + y * stride
-    addps      xmm2, xmm7    // x, y += dx, dy
+    cvttps2dq  xmm0, xmm2  // x, y float to int
+    packssdw   xmm0, xmm0  // x, y as shorts
+    pmaddwd    xmm0, xmm5  // offset = x * 4 + y * stride
+    addps      xmm2, xmm7  // x, y += dx, dy
     movd       esi, xmm0
     movd       xmm0, [eax + esi]  // copy a pixel
     movd       [edx], xmm0
@@ -5590,15 +5763,16 @@ void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
 
 #ifdef HAS_INTERPOLATEROW_AVX2
 // Bilinear filter 32x2 -> 32x1
-__declspec(naked)
-void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
-                         ptrdiff_t src_stride, int dst_width,
-                         int source_y_fraction) {
+__declspec(naked) void InterpolateRow_AVX2(uint8_t* dst_ptr,
+                                           const uint8_t* src_ptr,
+                                           ptrdiff_t src_stride,
+                                           int dst_width,
+                                           int source_y_fraction) {
   __asm {
     push       esi
     push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edi, [esp + 8 + 4]  // dst_ptr
+    mov        esi, [esp + 8 + 8]  // src_ptr
     mov        edx, [esp + 8 + 12]  // src_stride
     mov        ecx, [esp + 8 + 16]  // dst_width
     mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
@@ -5607,7 +5781,7 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
     je         xloop100  // 0 / 256.  Blend 100 / 0.
     sub        edi, esi
     cmp        eax, 128
-    je         xloop50   // 128 /256 is 0.50.  Blend 50 / 50.
+    je         xloop50  // 128 /256 is 0.50.  Blend 50 / 50.
 
     vmovd      xmm0, eax  // high fraction 0..255
     neg        eax
@@ -5634,14 +5808,14 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
     vpaddw     ymm0, ymm0, ymm4
     vpsrlw     ymm1, ymm1, 8
     vpsrlw     ymm0, ymm0, 8
-    vpackuswb  ymm0, ymm0, ymm1  // unmutates
+    vpackuswb  ymm0, ymm0, ymm1            // unmutates
     vmovdqu    [esi + edi], ymm0
     lea        esi, [esi + 32]
     sub        ecx, 32
     jg         xloop
     jmp        xloop99
 
-   // Blend 50 / 50.
+        // Blend 50 / 50.
  xloop50:
    vmovdqu    ymm0, [esi]
    vpavgb     ymm0, ymm0, [esi + edx]
@@ -5651,7 +5825,7 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
    jg         xloop50
    jmp        xloop99
 
-   // Blend 100 / 0 - Copy row unchanged.
+        // Blend 100 / 0 - Copy row unchanged.
  xloop100:
    rep movsb
 
@@ -5666,25 +5840,26 @@ void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
 
 // Bilinear filter 16x2 -> 16x1
 // TODO(fbarchard): Consider allowing 256 using memcpy.
-__declspec(naked)
-void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                          ptrdiff_t src_stride, int dst_width,
-                          int source_y_fraction) {
+__declspec(naked) void InterpolateRow_SSSE3(uint8_t* dst_ptr,
+                                            const uint8_t* src_ptr,
+                                            ptrdiff_t src_stride,
+                                            int dst_width,
+                                            int source_y_fraction) {
   __asm {
     push       esi
     push       edi
 
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edi, [esp + 8 + 4]  // dst_ptr
+    mov        esi, [esp + 8 + 8]  // src_ptr
     mov        edx, [esp + 8 + 12]  // src_stride
     mov        ecx, [esp + 8 + 16]  // dst_width
     mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
     sub        edi, esi
-    // Dispatch to specialized filters if applicable.
+        // Dispatch to specialized filters if applicable.
     cmp        eax, 0
     je         xloop100  // 0 /256.  Blend 100 / 0.
     cmp        eax, 128
-    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    je         xloop50  // 128 / 256 is 0.50.  Blend 50 / 50.
 
     movd       xmm0, eax  // high fraction 0..255
     neg        eax
@@ -5703,7 +5878,7 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     movdqu     xmm1, xmm0
     punpcklbw  xmm0, xmm2
     punpckhbw  xmm1, xmm2
-    psubb      xmm0, xmm4  // bias image by -128
+    psubb      xmm0, xmm4            // bias image by -128
     psubb      xmm1, xmm4
     movdqa     xmm2, xmm5
     movdqa     xmm3, xmm5
@@ -5720,7 +5895,7 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     jg         xloop
     jmp        xloop99
 
-    // Blend 50 / 50.
+        // Blend 50 / 50.
   xloop50:
     movdqu     xmm0, [esi]
     movdqu     xmm1, [esi + edx]
@@ -5731,7 +5906,7 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
     jg         xloop50
     jmp        xloop99
 
-    // Blend 100 / 0 - Copy row unchanged.
+        // Blend 100 / 0 - Copy row unchanged.
   xloop100:
     movdqu     xmm0, [esi]
     movdqu     [esi + edi], xmm0
@@ -5747,15 +5922,16 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
 }
 
 // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
-__declspec(naked)
-void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                          const uint8* shuffler, int width) {
-  __asm {
-    mov        eax, [esp + 4]    // src_argb
-    mov        edx, [esp + 8]    // dst_argb
-    mov        ecx, [esp + 12]   // shuffler
+__declspec(naked) void ARGBShuffleRow_SSSE3(const uint8_t* src_argb,
+                                            uint8_t* dst_argb,
+                                            const uint8_t* shuffler,
+                                            int width) {
+  __asm {
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_argb
+    mov        ecx, [esp + 12]  // shuffler
     movdqu     xmm5, [ecx]
-    mov        ecx, [esp + 16]   // width
+    mov        ecx, [esp + 16]  // width
 
   wloop:
     movdqu     xmm0, [eax]
@@ -5773,15 +5949,16 @@ void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
 }
 
 #ifdef HAS_ARGBSHUFFLEROW_AVX2
-__declspec(naked)
-void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width) {
+__declspec(naked) void ARGBShuffleRow_AVX2(const uint8_t* src_argb,
+                                           uint8_t* dst_argb,
+                                           const uint8_t* shuffler,
+                                           int width) {
   __asm {
-    mov        eax, [esp + 4]     // src_argb
-    mov        edx, [esp + 8]     // dst_argb
-    mov        ecx, [esp + 12]    // shuffler
-    vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
-    mov        ecx, [esp + 16]    // width
+    mov        eax, [esp + 4]  // src_argb
+    mov        edx, [esp + 8]  // dst_argb
+    mov        ecx, [esp + 12]  // shuffler
+    vbroadcastf128 ymm5, [ecx]  // same shuffle in high as low.
+    mov        ecx, [esp + 16]  // width
 
   wloop:
     vmovdqu    ymm0, [eax]
@@ -5801,152 +5978,36 @@ void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
 }
 #endif  // HAS_ARGBSHUFFLEROW_AVX2
 
-__declspec(naked)
-void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
-                         const uint8* shuffler, int width) {
-  __asm {
-    push       ebx
-    push       esi
-    mov        eax, [esp + 8 + 4]    // src_argb
-    mov        edx, [esp + 8 + 8]    // dst_argb
-    mov        esi, [esp + 8 + 12]   // shuffler
-    mov        ecx, [esp + 8 + 16]   // width
-    pxor       xmm5, xmm5
-
-    mov        ebx, [esi]   // shuffler
-    cmp        ebx, 0x03000102
-    je         shuf_3012
-    cmp        ebx, 0x00010203
-    je         shuf_0123
-    cmp        ebx, 0x00030201
-    je         shuf_0321
-    cmp        ebx, 0x02010003
-    je         shuf_2103
-
-  // TODO(fbarchard): Use one source pointer and 3 offsets.
-  shuf_any1:
-    movzx      ebx, byte ptr [esi]
-    movzx      ebx, byte ptr [eax + ebx]
-    mov        [edx], bl
-    movzx      ebx, byte ptr [esi + 1]
-    movzx      ebx, byte ptr [eax + ebx]
-    mov        [edx + 1], bl
-    movzx      ebx, byte ptr [esi + 2]
-    movzx      ebx, byte ptr [eax + ebx]
-    mov        [edx + 2], bl
-    movzx      ebx, byte ptr [esi + 3]
-    movzx      ebx, byte ptr [eax + ebx]
-    mov        [edx + 3], bl
-    lea        eax, [eax + 4]
-    lea        edx, [edx + 4]
-    sub        ecx, 1
-    jg         shuf_any1
-    jmp        shuf99
-
-  shuf_0123:
-    movdqu     xmm0, [eax]
-    lea        eax, [eax + 16]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm5
-    punpckhbw  xmm1, xmm5
-    pshufhw    xmm0, xmm0, 01Bh   // 1B = 00011011 = 0x0123 = BGRAToARGB
-    pshuflw    xmm0, xmm0, 01Bh
-    pshufhw    xmm1, xmm1, 01Bh
-    pshuflw    xmm1, xmm1, 01Bh
-    packuswb   xmm0, xmm1
-    movdqu     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 4
-    jg         shuf_0123
-    jmp        shuf99
-
-  shuf_0321:
-    movdqu     xmm0, [eax]
-    lea        eax, [eax + 16]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm5
-    punpckhbw  xmm1, xmm5
-    pshufhw    xmm0, xmm0, 039h   // 39 = 00111001 = 0x0321 = RGBAToARGB
-    pshuflw    xmm0, xmm0, 039h
-    pshufhw    xmm1, xmm1, 039h
-    pshuflw    xmm1, xmm1, 039h
-    packuswb   xmm0, xmm1
-    movdqu     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 4
-    jg         shuf_0321
-    jmp        shuf99
-
-  shuf_2103:
-    movdqu     xmm0, [eax]
-    lea        eax, [eax + 16]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm5
-    punpckhbw  xmm1, xmm5
-    pshufhw    xmm0, xmm0, 093h   // 93 = 10010011 = 0x2103 = ARGBToRGBA
-    pshuflw    xmm0, xmm0, 093h
-    pshufhw    xmm1, xmm1, 093h
-    pshuflw    xmm1, xmm1, 093h
-    packuswb   xmm0, xmm1
-    movdqu     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 4
-    jg         shuf_2103
-    jmp        shuf99
-
-  shuf_3012:
-    movdqu     xmm0, [eax]
-    lea        eax, [eax + 16]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm5
-    punpckhbw  xmm1, xmm5
-    pshufhw    xmm0, xmm0, 0C6h   // C6 = 11000110 = 0x3012 = ABGRToARGB
-    pshuflw    xmm0, xmm0, 0C6h
-    pshufhw    xmm1, xmm1, 0C6h
-    pshuflw    xmm1, xmm1, 0C6h
-    packuswb   xmm0, xmm1
-    movdqu     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 4
-    jg         shuf_3012
-
-  shuf99:
-    pop        esi
-    pop        ebx
-    ret
-  }
-}
-
 // YUY2 - Macro-pixel = 2 image pixels
 // Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4....
 
 // UYVY - Macro-pixel = 2 image pixels
 // U0Y0V0Y1
 
-__declspec(naked)
-void I422ToYUY2Row_SSE2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_frame, int width) {
+__declspec(naked) void I422ToYUY2Row_SSE2(const uint8_t* src_y,
+                                          const uint8_t* src_u,
+                                          const uint8_t* src_v,
+                                          uint8_t* dst_frame,
+                                          int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_y
-    mov        esi, [esp + 8 + 8]    // src_u
-    mov        edx, [esp + 8 + 12]   // src_v
-    mov        edi, [esp + 8 + 16]   // dst_frame
-    mov        ecx, [esp + 8 + 20]   // width
+    mov        eax, [esp + 8 + 4]  // src_y
+    mov        esi, [esp + 8 + 8]  // src_u
+    mov        edx, [esp + 8 + 12]  // src_v
+    mov        edi, [esp + 8 + 16]  // dst_frame
+    mov        ecx, [esp + 8 + 20]  // width
     sub        edx, esi
 
   convertloop:
-    movq       xmm2, qword ptr [esi] // U
-    movq       xmm3, qword ptr [esi + edx] // V
+    movq       xmm2, qword ptr [esi]  // U
+    movq       xmm3, qword ptr [esi + edx]  // V
     lea        esi, [esi + 8]
-    punpcklbw  xmm2, xmm3 // UV
-    movdqu     xmm0, [eax] // Y
+    punpcklbw  xmm2, xmm3  // UV
+    movdqu     xmm0, [eax]  // Y
     lea        eax, [eax + 16]
     movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm2 // YUYV
+    punpcklbw  xmm0, xmm2  // YUYV
     punpckhbw  xmm1, xmm2
     movdqu     [edi], xmm0
     movdqu     [edi + 16], xmm1
@@ -5960,30 +6021,30 @@ void I422ToYUY2Row_SSE2(const uint8* src_y,
   }
 }
 
-__declspec(naked)
-void I422ToUYVYRow_SSE2(const uint8* src_y,
-                        const uint8* src_u,
-                        const uint8* src_v,
-                        uint8* dst_frame, int width) {
+__declspec(naked) void I422ToUYVYRow_SSE2(const uint8_t* src_y,
+                                          const uint8_t* src_u,
+                                          const uint8_t* src_v,
+                                          uint8_t* dst_frame,
+                                          int width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_y
-    mov        esi, [esp + 8 + 8]    // src_u
-    mov        edx, [esp + 8 + 12]   // src_v
-    mov        edi, [esp + 8 + 16]   // dst_frame
-    mov        ecx, [esp + 8 + 20]   // width
+    mov        eax, [esp + 8 + 4]  // src_y
+    mov        esi, [esp + 8 + 8]  // src_u
+    mov        edx, [esp + 8 + 12]  // src_v
+    mov        edi, [esp + 8 + 16]  // dst_frame
+    mov        ecx, [esp + 8 + 20]  // width
     sub        edx, esi
 
   convertloop:
-    movq       xmm2, qword ptr [esi] // U
-    movq       xmm3, qword ptr [esi + edx] // V
+    movq       xmm2, qword ptr [esi]  // U
+    movq       xmm3, qword ptr [esi + edx]  // V
     lea        esi, [esi + 8]
-    punpcklbw  xmm2, xmm3 // UV
-    movdqu     xmm0, [eax] // Y
+    punpcklbw  xmm2, xmm3  // UV
+    movdqu     xmm0, [eax]  // Y
     movdqa     xmm1, xmm2
     lea        eax, [eax + 16]
-    punpcklbw  xmm1, xmm0 // UYVY
+    punpcklbw  xmm1, xmm0  // UYVY
     punpckhbw  xmm2, xmm0
     movdqu     [edi], xmm1
     movdqu     [edi + 16], xmm2
@@ -5998,22 +6059,22 @@ void I422ToUYVYRow_SSE2(const uint8* src_y,
 }
 
 #ifdef HAS_ARGBPOLYNOMIALROW_SSE2
-__declspec(naked)
-void ARGBPolynomialRow_SSE2(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
-                            int width) {
+__declspec(naked) void ARGBPolynomialRow_SSE2(const uint8_t* src_argb,
+                                              uint8_t* dst_argb,
+                                              const float* poly,
+                                              int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   /* src_argb */
-    mov        edx, [esp + 4 + 8]   /* dst_argb */
-    mov        esi, [esp + 4 + 12]  /* poly */
-    mov        ecx, [esp + 4 + 16]  /* width */
+    mov        eax, [esp + 4 + 4] /* src_argb */
+    mov        edx, [esp + 4 + 8] /* dst_argb */
+    mov        esi, [esp + 4 + 12] /* poly */
+    mov        ecx, [esp + 4 + 16] /* width */
     pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
 
-    // 2 pixel loop.
+        // 2 pixel loop.
  convertloop:
-//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
-//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
+        //    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+        //    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
     movq       xmm0, qword ptr [eax]  // BGRABGRA
     lea        eax, [eax + 8]
     punpcklbw  xmm0, xmm3
@@ -6057,25 +6118,25 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
 #endif  // HAS_ARGBPOLYNOMIALROW_SSE2
 
 #ifdef HAS_ARGBPOLYNOMIALROW_AVX2
-__declspec(naked)
-void ARGBPolynomialRow_AVX2(const uint8* src_argb,
-                            uint8* dst_argb, const float* poly,
-                            int width) {
-  __asm {
-    mov        eax, [esp + 4]   /* src_argb */
-    mov        edx, [esp + 8]   /* dst_argb */
-    mov        ecx, [esp + 12]   /* poly */
-    vbroadcastf128 ymm4, [ecx]       // C0
+__declspec(naked) void ARGBPolynomialRow_AVX2(const uint8_t* src_argb,
+                                              uint8_t* dst_argb,
+                                              const float* poly,
+                                              int width) {
+  __asm {
+    mov        eax, [esp + 4] /* src_argb */
+    mov        edx, [esp + 8] /* dst_argb */
+    mov        ecx, [esp + 12] /* poly */
+    vbroadcastf128 ymm4, [ecx]  // C0
     vbroadcastf128 ymm5, [ecx + 16]  // C1
     vbroadcastf128 ymm6, [ecx + 32]  // C2
     vbroadcastf128 ymm7, [ecx + 48]  // C3
-    mov        ecx, [esp + 16]  /* width */
+    mov        ecx, [esp + 16] /* width */
 
     // 2 pixel loop.
  convertloop:
     vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
     lea         eax, [eax + 8]
-    vcvtdq2ps   ymm0, ymm0        // X 8 floats
+    vcvtdq2ps   ymm0, ymm0  // X 8 floats
     vmulps      ymm2, ymm0, ymm0  // X * X
     vmulps      ymm3, ymm0, ymm7  // C3 * X
     vfmadd132ps ymm0, ymm4, ymm5  // result = C0 + C1 * X
@@ -6095,16 +6156,125 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_AVX2
 
+#ifdef HAS_HALFFLOATROW_SSE2
+static float kExpBias = 1.9259299444e-34f;
+__declspec(naked) void HalfFloatRow_SSE2(const uint16_t* src,
+                                         uint16_t* dst,
+                                         float scale,
+                                         int width) {
+  __asm {
+    mov        eax, [esp + 4] /* src */
+    mov        edx, [esp + 8] /* dst */
+    movd       xmm4, dword ptr [esp + 12] /* scale */
+    mov        ecx, [esp + 16] /* width */
+    mulss      xmm4, kExpBias
+    pshufd     xmm4, xmm4, 0
+    pxor       xmm5, xmm5
+    sub        edx, eax
+
+        // 8 pixel loop.
+ convertloop:
+    movdqu      xmm2, xmmword ptr [eax]  // 8 shorts
+    add         eax, 16
+    movdqa      xmm3, xmm2
+    punpcklwd   xmm2, xmm5
+    cvtdq2ps    xmm2, xmm2  // convert 8 ints to floats
+    punpckhwd   xmm3, xmm5
+    cvtdq2ps    xmm3, xmm3
+    mulps       xmm2, xmm4
+    mulps       xmm3, xmm4
+    psrld       xmm2, 13
+    psrld       xmm3, 13
+    packssdw    xmm2, xmm3
+    movdqu      [eax + edx - 16], xmm2
+    sub         ecx, 8
+    jg          convertloop
+    ret
+  }
+}
+#endif  // HAS_HALFFLOATROW_SSE2
+
+#ifdef HAS_HALFFLOATROW_AVX2
+__declspec(naked) void HalfFloatRow_AVX2(const uint16_t* src,
+                                         uint16_t* dst,
+                                         float scale,
+                                         int width) {
+  __asm {
+    mov        eax, [esp + 4] /* src */
+    mov        edx, [esp + 8] /* dst */
+    movd       xmm4, dword ptr [esp + 12] /* scale */
+    mov        ecx, [esp + 16] /* width */
+
+    vmulss     xmm4, xmm4, kExpBias
+    vbroadcastss ymm4, xmm4
+    vpxor      ymm5, ymm5, ymm5
+    sub        edx, eax
+
+        // 16 pixel loop.
+ convertloop:
+    vmovdqu     ymm2, [eax]  // 16 shorts
+    add         eax, 32
+    vpunpckhwd  ymm3, ymm2, ymm5  // convert 16 shorts to 16 ints
+    vpunpcklwd  ymm2, ymm2, ymm5
+    vcvtdq2ps   ymm3, ymm3  // convert 16 ints to floats
+    vcvtdq2ps   ymm2, ymm2
+    vmulps      ymm3, ymm3, ymm4  // scale to adjust exponent for 5 bit range.
+    vmulps      ymm2, ymm2, ymm4
+    vpsrld      ymm3, ymm3, 13  // float convert to 8 half floats truncate
+    vpsrld      ymm2, ymm2, 13
+    vpackssdw   ymm2, ymm2, ymm3
+    vmovdqu     [eax + edx - 32], ymm2
+    sub         ecx, 16
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_HALFFLOATROW_AVX2
+
+#ifdef HAS_HALFFLOATROW_F16C
+__declspec(naked) void HalfFloatRow_F16C(const uint16_t* src,
+                                         uint16_t* dst,
+                                         float scale,
+                                         int width) {
+  __asm {
+    mov        eax, [esp + 4] /* src */
+    mov        edx, [esp + 8] /* dst */
+    vbroadcastss ymm4, [esp + 12] /* scale */
+    mov        ecx, [esp + 16] /* width */
+    sub        edx, eax
+
+        // 16 pixel loop.
+ convertloop:
+    vpmovzxwd   ymm2, xmmword ptr [eax]  // 8 shorts -> 8 ints
+    vpmovzxwd   ymm3, xmmword ptr [eax + 16]  // 8 more shorts
+    add         eax, 32
+    vcvtdq2ps   ymm2, ymm2  // convert 8 ints to floats
+    vcvtdq2ps   ymm3, ymm3
+    vmulps      ymm2, ymm2, ymm4  // scale to normalized range 0 to 1
+    vmulps      ymm3, ymm3, ymm4
+    vcvtps2ph   xmm2, ymm2, 3  // float convert to 8 half floats truncate
+    vcvtps2ph   xmm3, ymm3, 3
+    vmovdqu     [eax + edx + 32], xmm2
+    vmovdqu     [eax + edx + 32 + 16], xmm3
+    sub         ecx, 16
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_HALFFLOATROW_F16C
+
 #ifdef HAS_ARGBCOLORTABLEROW_X86
 // Tranform ARGB pixels with color table.
-__declspec(naked)
-void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
-                           int width) {
+__declspec(naked) void ARGBColorTableRow_X86(uint8_t* dst_argb,
+                                             const uint8_t* table_argb,
+                                             int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   /* dst_argb */
-    mov        esi, [esp + 4 + 8]   /* table_argb */
-    mov        ecx, [esp + 4 + 12]  /* width */
+    mov        eax, [esp + 4 + 4] /* dst_argb */
+    mov        esi, [esp + 4 + 8] /* table_argb */
+    mov        ecx, [esp + 4 + 12] /* width */
 
     // 1 pixel loop.
   convertloop:
@@ -6131,13 +6301,14 @@ void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
 
 #ifdef HAS_RGBCOLORTABLEROW_X86
 // Tranform RGB pixels with color table.
-__declspec(naked)
-void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+__declspec(naked) void RGBColorTableRow_X86(uint8_t* dst_argb,
+                                            const uint8_t* table_argb,
+                                            int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]   /* dst_argb */
-    mov        esi, [esp + 4 + 8]   /* table_argb */
-    mov        ecx, [esp + 4 + 12]  /* width */
+    mov        eax, [esp + 4 + 4] /* dst_argb */
+    mov        esi, [esp + 4 + 8] /* table_argb */
+    mov        ecx, [esp + 4 + 12] /* width */
 
     // 1 pixel loop.
   convertloop:
@@ -6162,27 +6333,28 @@ void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
 
 #ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
 // Tranform RGB pixels with luma table.
-__declspec(naked)
-void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
-                                 int width,
-                                 const uint8* luma, uint32 lumacoeff) {
+__declspec(naked) void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb,
+                                                   uint8_t* dst_argb,
+                                                   int width,
+                                                   const uint8_t* luma,
+                                                   uint32_t lumacoeff) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]   /* src_argb */
-    mov        edi, [esp + 8 + 8]   /* dst_argb */
-    mov        ecx, [esp + 8 + 12]  /* width */
+    mov        eax, [esp + 8 + 4] /* src_argb */
+    mov        edi, [esp + 8 + 8] /* dst_argb */
+    mov        ecx, [esp + 8 + 12] /* width */
     movd       xmm2, dword ptr [esp + 8 + 16]  // luma table
     movd       xmm3, dword ptr [esp + 8 + 20]  // lumacoeff
     pshufd     xmm2, xmm2, 0
     pshufd     xmm3, xmm3, 0
-    pcmpeqb    xmm4, xmm4        // generate mask 0xff00ff00
+    pcmpeqb    xmm4, xmm4  // generate mask 0xff00ff00
     psllw      xmm4, 8
     pxor       xmm5, xmm5
 
-    // 4 pixel loop.
+        // 4 pixel loop.
   convertloop:
-    movdqu     xmm0, xmmword ptr [eax]      // generate luma ptr
+    movdqu     xmm0, xmmword ptr [eax]  // generate luma ptr
     pmaddubsw  xmm0, xmm3
     phaddw     xmm0, xmm0
     pand       xmm0, xmm4  // mask out low bits
diff --git a/media/libyuv/source/scale.cc b/media/libyuv/source/scale.cc
index 36e3fe5281..830754ce64 100644
--- a/media/libyuv/source/scale.cc
+++ b/media/libyuv/source/scale.cc
@@ -17,6 +17,7 @@
 #include "libyuv/planar_functions.h"  // For CopyPlane
 #include "libyuv/row.h"
 #include "libyuv/scale_row.h"
+#include "libyuv/scale_uv.h"  // For UVScale
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -28,22 +29,31 @@ static __inline int Abs(int v) {
 }
 
 #define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
 
 // Scale plane, 1/2
 // This is an optimized version for scaling down a plane to 1/2 of
 // its original size.
 
-static void ScalePlaneDown2(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
+static void ScalePlaneDown2(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
                             enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) =
-      filtering == kFilterNone ? ScaleRowDown2_C :
-      (filtering == kFilterLinear ? ScaleRowDown2Linear_C : ScaleRowDown2Box_C);
-  int row_stride = src_stride << 1;
+  void (*ScaleRowDown2)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                        uint8_t* dst_ptr, int dst_width) =
+      filtering == kFilterNone
+          ? ScaleRowDown2_C
+          : (filtering == kFilterLinear ? ScaleRowDown2Linear_C
+                                        : ScaleRowDown2Box_C);
+  int row_stride = src_stride * 2;
+  (void)src_width;
+  (void)src_height;
   if (!filtering) {
     src_ptr += src_stride;  // Point to odd rows.
     src_stride = 0;
@@ -51,46 +61,78 @@ static void ScalePlaneDown2(int src_width, int src_height,
 
 #if defined(HAS_SCALEROWDOWN2_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_NEON :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_NEON :
-        ScaleRowDown2Box_Any_NEON);
+    ScaleRowDown2 =
+        filtering == kFilterNone
+            ? ScaleRowDown2_Any_NEON
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_NEON
+                                          : ScaleRowDown2Box_Any_NEON);
     if (IS_ALIGNED(dst_width, 16)) {
-      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON :
-          (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON :
-          ScaleRowDown2Box_NEON);
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON
+                                               : (filtering == kFilterLinear
+                                                      ? ScaleRowDown2Linear_NEON
+                                                      : ScaleRowDown2Box_NEON);
     }
   }
 #endif
 #if defined(HAS_SCALEROWDOWN2_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_SSSE3 :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_SSSE3 :
-        ScaleRowDown2Box_Any_SSSE3);
+    ScaleRowDown2 =
+        filtering == kFilterNone
+            ? ScaleRowDown2_Any_SSSE3
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_SSSE3
+                                          : ScaleRowDown2Box_Any_SSSE3);
     if (IS_ALIGNED(dst_width, 16)) {
-      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSSE3 :
-          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSSE3 :
-          ScaleRowDown2Box_SSSE3);
+      ScaleRowDown2 =
+          filtering == kFilterNone
+              ? ScaleRowDown2_SSSE3
+              : (filtering == kFilterLinear ? ScaleRowDown2Linear_SSSE3
+                                            : ScaleRowDown2Box_SSSE3);
     }
   }
 #endif
 #if defined(HAS_SCALEROWDOWN2_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_AVX2 :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_AVX2 :
-        ScaleRowDown2Box_Any_AVX2);
+    ScaleRowDown2 =
+        filtering == kFilterNone
+            ? ScaleRowDown2_Any_AVX2
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_AVX2
+                                          : ScaleRowDown2Box_Any_AVX2);
     if (IS_ALIGNED(dst_width, 32)) {
-      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_AVX2 :
-          (filtering == kFilterLinear ? ScaleRowDown2Linear_AVX2 :
-          ScaleRowDown2Box_AVX2);
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_AVX2
+                                               : (filtering == kFilterLinear
+                                                      ? ScaleRowDown2Linear_AVX2
+                                                      : ScaleRowDown2Box_AVX2);
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN2_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(src_ptr, 4) &&
-      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    ScaleRowDown2 = filtering ?
-        ScaleRowDown2Box_DSPR2 : ScaleRowDown2_DSPR2;
+#if defined(HAS_SCALEROWDOWN2_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleRowDown2 =
+        filtering == kFilterNone
+            ? ScaleRowDown2_Any_MSA
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_MSA
+                                          : ScaleRowDown2Box_Any_MSA);
+    if (IS_ALIGNED(dst_width, 32)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_MSA
+                                               : (filtering == kFilterLinear
+                                                      ? ScaleRowDown2Linear_MSA
+                                                      : ScaleRowDown2Box_MSA);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleRowDown2 =
+        filtering == kFilterNone
+            ? ScaleRowDown2_Any_LSX
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_LSX
+                                          : ScaleRowDown2Box_Any_LSX);
+    if (IS_ALIGNED(dst_width, 32)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_LSX
+                                               : (filtering == kFilterLinear
+                                                      ? ScaleRowDown2Linear_LSX
+                                                      : ScaleRowDown2Box_LSX);
+    }
   }
 #endif
 
@@ -105,18 +147,25 @@ static void ScalePlaneDown2(int src_width, int src_height,
   }
 }
 
-static void ScalePlaneDown2_16(int src_width, int src_height,
-                               int dst_width, int dst_height,
-                               int src_stride, int dst_stride,
-                               const uint16* src_ptr, uint16* dst_ptr,
+static void ScalePlaneDown2_16(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
                                enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown2)(const uint16* src_ptr, ptrdiff_t src_stride,
-                        uint16* dst_ptr, int dst_width) =
-    filtering == kFilterNone ? ScaleRowDown2_16_C :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C :
-        ScaleRowDown2Box_16_C);
-  int row_stride = src_stride << 1;
+  void (*ScaleRowDown2)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                        uint16_t* dst_ptr, int dst_width) =
+      filtering == kFilterNone
+          ? ScaleRowDown2_16_C
+          : (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C
+                                        : ScaleRowDown2Box_16_C);
+  int row_stride = src_stride * 2;
+  (void)src_width;
+  (void)src_height;
   if (!filtering) {
     src_ptr += src_stride;  // Point to odd rows.
     src_stride = 0;
@@ -124,23 +173,17 @@ static void ScalePlaneDown2_16(int src_width, int src_height,
 
 #if defined(HAS_SCALEROWDOWN2_16_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON :
-        ScaleRowDown2_16_NEON;
+    ScaleRowDown2 =
+        filtering ? ScaleRowDown2Box_16_NEON : ScaleRowDown2_16_NEON;
   }
 #endif
 #if defined(HAS_SCALEROWDOWN2_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
-        ScaleRowDown2Box_16_SSE2);
-  }
-#endif
-#if defined(HAS_SCALEROWDOWN2_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(src_ptr, 4) &&
-      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    ScaleRowDown2 = filtering ?
-        ScaleRowDown2Box_16_DSPR2 : ScaleRowDown2_16_DSPR2;
+    ScaleRowDown2 =
+        filtering == kFilterNone
+            ? ScaleRowDown2_16_SSE2
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2
+                                          : ScaleRowDown2Box_16_SSE2);
   }
 #endif
 
@@ -155,28 +198,79 @@ static void ScalePlaneDown2_16(int src_width, int src_height,
   }
 }
 
+void ScalePlaneDown2_16To8(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint16_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           int scale,
+                           enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                        uint8_t* dst_ptr, int dst_width, int scale) =
+      (src_width & 1)
+          ? (filtering == kFilterNone
+                 ? ScaleRowDown2_16To8_Odd_C
+                 : (filtering == kFilterLinear ? ScaleRowDown2Linear_16To8_Odd_C
+                                               : ScaleRowDown2Box_16To8_Odd_C))
+          : (filtering == kFilterNone
+                 ? ScaleRowDown2_16To8_C
+                 : (filtering == kFilterLinear ? ScaleRowDown2Linear_16To8_C
+                                               : ScaleRowDown2Box_16To8_C));
+  int row_stride = src_stride * 2;
+  (void)dst_height;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < src_height / 2; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width, scale);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+  if (src_height & 1) {
+    if (!filtering) {
+      src_ptr -= src_stride;  // Point to last row.
+    }
+    ScaleRowDown2(src_ptr, 0, dst_ptr, dst_width, scale);
+  }
+}
+
 // Scale plane, 1/4
 // This is an optimized version for scaling down a plane to 1/4 of
 // its original size.
 
-static void ScalePlaneDown4(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
+static void ScalePlaneDown4(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
                             enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) =
+  void (*ScaleRowDown4)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                        uint8_t* dst_ptr, int dst_width) =
       filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
-  int row_stride = src_stride << 2;
+  int row_stride = src_stride * 4;
+  (void)src_width;
+  (void)src_height;
   if (!filtering) {
     src_ptr += src_stride * 2;  // Point to row 2.
     src_stride = 0;
   }
 #if defined(HAS_SCALEROWDOWN4_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ScaleRowDown4 = filtering ?
-        ScaleRowDown4Box_Any_NEON : ScaleRowDown4_Any_NEON;
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_Any_NEON : ScaleRowDown4_Any_NEON;
     if (IS_ALIGNED(dst_width, 8)) {
       ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
     }
@@ -184,8 +278,8 @@ static void ScalePlaneDown4(int src_width, int src_height,
 #endif
 #if defined(HAS_SCALEROWDOWN4_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ScaleRowDown4 = filtering ?
-        ScaleRowDown4Box_Any_SSSE3 : ScaleRowDown4_Any_SSSE3;
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_Any_SSSE3 : ScaleRowDown4_Any_SSSE3;
     if (IS_ALIGNED(dst_width, 8)) {
       ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSSE3 : ScaleRowDown4_SSSE3;
     }
@@ -193,19 +287,29 @@ static void ScalePlaneDown4(int src_width, int src_height,
 #endif
 #if defined(HAS_SCALEROWDOWN4_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ScaleRowDown4 = filtering ?
-        ScaleRowDown4Box_Any_AVX2 : ScaleRowDown4_Any_AVX2;
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_Any_AVX2 : ScaleRowDown4_Any_AVX2;
     if (IS_ALIGNED(dst_width, 16)) {
       ScaleRowDown4 = filtering ? ScaleRowDown4Box_AVX2 : ScaleRowDown4_AVX2;
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN4_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(row_stride, 4) &&
-      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    ScaleRowDown4 = filtering ?
-        ScaleRowDown4Box_DSPR2 : ScaleRowDown4_DSPR2;
+#if defined(HAS_SCALEROWDOWN4_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_Any_MSA : ScaleRowDown4_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_MSA : ScaleRowDown4_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_Any_LSX : ScaleRowDown4_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_LSX : ScaleRowDown4_LSX;
+    }
   }
 #endif
 
@@ -219,38 +323,36 @@ static void ScalePlaneDown4(int src_width, int src_height,
   }
 }
 
-static void ScalePlaneDown4_16(int src_width, int src_height,
-                               int dst_width, int dst_height,
-                               int src_stride, int dst_stride,
-                               const uint16* src_ptr, uint16* dst_ptr,
+static void ScalePlaneDown4_16(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
                                enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown4)(const uint16* src_ptr, ptrdiff_t src_stride,
-                        uint16* dst_ptr, int dst_width) =
+  void (*ScaleRowDown4)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                        uint16_t* dst_ptr, int dst_width) =
       filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C;
-  int row_stride = src_stride << 2;
+  int row_stride = src_stride * 4;
+  (void)src_width;
+  (void)src_height;
   if (!filtering) {
     src_ptr += src_stride * 2;  // Point to row 2.
     src_stride = 0;
   }
 #if defined(HAS_SCALEROWDOWN4_16_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON :
-        ScaleRowDown4_16_NEON;
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_16_NEON : ScaleRowDown4_16_NEON;
   }
 #endif
 #if defined(HAS_SCALEROWDOWN4_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 :
-        ScaleRowDown4_16_SSE2;
-  }
-#endif
-#if defined(HAS_SCALEROWDOWN4_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(row_stride, 4) &&
-      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    ScaleRowDown4 = filtering ?
-        ScaleRowDown4Box_16_DSPR2 : ScaleRowDown4_16_DSPR2;
+    ScaleRowDown4 =
+        filtering ? ScaleRowDown4Box_16_SSE2 : ScaleRowDown4_16_SSE2;
   }
 #endif
 
@@ -265,18 +367,23 @@ static void ScalePlaneDown4_16(int src_width, int src_height,
 }
 
 // Scale plane down, 3/4
-
-static void ScalePlaneDown34(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr,
+static void ScalePlaneDown34(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
                              enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
-  void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_0)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                           uint8_t* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                           uint8_t* dst_ptr, int dst_width);
   const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  (void)src_width;
+  (void)src_height;
   assert(dst_width % 3 == 0);
   if (!filtering) {
     ScaleRowDown34_0 = ScaleRowDown34_C;
@@ -305,6 +412,46 @@ static void ScalePlaneDown34(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEROWDOWN34_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_MSA;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_MSA;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_MSA;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_MSA;
+    }
+    if (dst_width % 48 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_MSA;
+        ScaleRowDown34_1 = ScaleRowDown34_MSA;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_MSA;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_MSA;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_LSX;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_LSX;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_LSX;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_LSX;
+    }
+    if (dst_width % 48 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_LSX;
+        ScaleRowDown34_1 = ScaleRowDown34_LSX;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_LSX;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_LSX;
+      }
+    }
+  }
+#endif
 #if defined(HAS_SCALEROWDOWN34_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     if (!filtering) {
@@ -325,19 +472,6 @@ static void ScalePlaneDown34(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN34_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_DSPR2;
-      ScaleRowDown34_1 = ScaleRowDown34_DSPR2;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Box_DSPR2;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Box_DSPR2;
-    }
-  }
-#endif
 
   for (y = 0; y < dst_height - 2; y += 3) {
     ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
@@ -346,8 +480,7 @@ static void ScalePlaneDown34(int src_width, int src_height,
     ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
     src_ptr += src_stride;
     dst_ptr += dst_stride;
-    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
-                     dst_ptr, dst_width);
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride, dst_ptr, dst_width);
     src_ptr += src_stride * 2;
     dst_ptr += dst_stride;
   }
@@ -363,17 +496,23 @@ static void ScalePlaneDown34(int src_width, int src_height,
   }
 }
 
-static void ScalePlaneDown34_16(int src_width, int src_height,
-                                int dst_width, int dst_height,
-                                int src_stride, int dst_stride,
-                                const uint16* src_ptr, uint16* dst_ptr,
+static void ScalePlaneDown34_16(int src_width,
+                                int src_height,
+                                int dst_width,
+                                int dst_height,
+                                int src_stride,
+                                int dst_stride,
+                                const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
                                 enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown34_0)(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst_ptr, int dst_width);
-  void (*ScaleRowDown34_1)(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_0)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                           uint16_t* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                           uint16_t* dst_ptr, int dst_width);
   const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  (void)src_width;
+  (void)src_height;
   assert(dst_width % 3 == 0);
   if (!filtering) {
     ScaleRowDown34_0 = ScaleRowDown34_16_C;
@@ -404,19 +543,6 @@ static void ScalePlaneDown34_16(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN34_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_16_DSPR2;
-      ScaleRowDown34_1 = ScaleRowDown34_16_DSPR2;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_DSPR2;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_DSPR2;
-    }
-  }
-#endif
 
   for (y = 0; y < dst_height - 2; y += 3) {
     ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
@@ -425,8 +551,7 @@ static void ScalePlaneDown34_16(int src_width, int src_height,
     ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
     src_ptr += src_stride;
     dst_ptr += dst_stride;
-    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
-                     dst_ptr, dst_width);
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride, dst_ptr, dst_width);
     src_ptr += src_stride * 2;
     dst_ptr += dst_stride;
   }
@@ -442,7 +567,6 @@ static void ScalePlaneDown34_16(int src_width, int src_height,
   }
 }
 
-
 // Scale plane, 3/8
 // This is an optimized version for scaling down a plane to 3/8
 // of its original size.
@@ -458,18 +582,24 @@ static void ScalePlaneDown34_16(int src_width, int src_height,
 // ggghhhii
 // Boxes are 3x3, 2x3, 3x2 and 2x2
 
-static void ScalePlaneDown38(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr,
+static void ScalePlaneDown38(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
                              enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
-  void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_3)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                           uint8_t* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                           uint8_t* dst_ptr, int dst_width);
   const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
   assert(dst_width % 3 == 0);
+  (void)src_width;
+  (void)src_height;
   if (!filtering) {
     ScaleRowDown38_3 = ScaleRowDown38_C;
     ScaleRowDown38_2 = ScaleRowDown38_C;
@@ -517,16 +647,43 @@ static void ScalePlaneDown38(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN38_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && (dst_width % 12 == 0) &&
-      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+#if defined(HAS_SCALEROWDOWN38_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_MSA;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_MSA;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_MSA;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_MSA;
+    }
+    if (dst_width % 12 == 0) {
+      if (!filtering) {
+        ScaleRowDown38_3 = ScaleRowDown38_MSA;
+        ScaleRowDown38_2 = ScaleRowDown38_MSA;
+      } else {
+        ScaleRowDown38_3 = ScaleRowDown38_3_Box_MSA;
+        ScaleRowDown38_2 = ScaleRowDown38_2_Box_MSA;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
     if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_DSPR2;
-      ScaleRowDown38_2 = ScaleRowDown38_DSPR2;
+      ScaleRowDown38_3 = ScaleRowDown38_Any_LSX;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_LSX;
     } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Box_DSPR2;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Box_DSPR2;
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_LSX;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_LSX;
+    }
+    if (dst_width % 12 == 0) {
+      if (!filtering) {
+        ScaleRowDown38_3 = ScaleRowDown38_LSX;
+        ScaleRowDown38_2 = ScaleRowDown38_LSX;
+      } else {
+        ScaleRowDown38_3 = ScaleRowDown38_3_Box_LSX;
+        ScaleRowDown38_2 = ScaleRowDown38_2_Box_LSX;
+      }
     }
   }
 #endif
@@ -554,17 +711,23 @@ static void ScalePlaneDown38(int src_width, int src_height,
   }
 }
 
-static void ScalePlaneDown38_16(int src_width, int src_height,
-                                int dst_width, int dst_height,
-                                int src_stride, int dst_stride,
-                                const uint16* src_ptr, uint16* dst_ptr,
+static void ScalePlaneDown38_16(int src_width,
+                                int src_height,
+                                int dst_width,
+                                int dst_height,
+                                int src_stride,
+                                int dst_stride,
+                                const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
                                 enum FilterMode filtering) {
   int y;
-  void (*ScaleRowDown38_3)(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst_ptr, int dst_width);
-  void (*ScaleRowDown38_2)(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_3)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                           uint16_t* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                           uint16_t* dst_ptr, int dst_width);
   const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  (void)src_width;
+  (void)src_height;
   assert(dst_width % 3 == 0);
   if (!filtering) {
     ScaleRowDown38_3 = ScaleRowDown38_16_C;
@@ -595,19 +758,6 @@ static void ScalePlaneDown38_16(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN38_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && (dst_width % 12 == 0) &&
-      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_16_DSPR2;
-      ScaleRowDown38_2 = ScaleRowDown38_16_DSPR2;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_DSPR2;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_DSPR2;
-    }
-  }
-#endif
 
   for (y = 0; y < dst_height - 2; y += 3) {
     ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
@@ -634,8 +784,8 @@ static void ScalePlaneDown38_16(int src_width, int src_height,
 
 #define MIN1(x) ((x) < 1 ? 1 : (x))
 
-static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
-  uint32 sum = 0u;
+static __inline uint32_t SumPixels(int iboxwidth, const uint16_t* src_ptr) {
+  uint32_t sum = 0u;
   int x;
   assert(iboxwidth > 0);
   for (x = 0; x < iboxwidth; ++x) {
@@ -644,8 +794,8 @@ static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
   return sum;
 }
 
-static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) {
-  uint32 sum = 0u;
+static __inline uint32_t SumPixels_16(int iboxwidth, const uint32_t* src_ptr) {
+  uint32_t sum = 0u;
   int x;
   assert(iboxwidth > 0);
   for (x = 0; x < iboxwidth; ++x) {
@@ -654,8 +804,12 @@ static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) {
   return sum;
 }
 
-static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
-                            const uint16* src_ptr, uint8* dst_ptr) {
+static void ScaleAddCols2_C(int dst_width,
+                            int boxheight,
+                            int x,
+                            int dx,
+                            const uint16_t* src_ptr,
+                            uint8_t* dst_ptr) {
   int i;
   int scaletbl[2];
   int minboxwidth = dx >> 16;
@@ -666,13 +820,18 @@ static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
     int ix = x >> 16;
     x += dx;
     boxwidth = MIN1((x >> 16) - ix);
-    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) *
-        scaletbl[boxwidth - minboxwidth] >> 16;
+    *dst_ptr++ = (uint8_t)(SumPixels(boxwidth, src_ptr + ix) *
+                               scaletbl[boxwidth - minboxwidth] >>
+                           16);
   }
 }
 
-static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx,
-                               const uint32* src_ptr, uint16* dst_ptr) {
+static void ScaleAddCols2_16_C(int dst_width,
+                               int boxheight,
+                               int x,
+                               int dx,
+                               const uint32_t* src_ptr,
+                               uint16_t* dst_ptr) {
   int i;
   int scaletbl[2];
   int minboxwidth = dx >> 16;
@@ -684,34 +843,48 @@ static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx,
     x += dx;
     boxwidth = MIN1((x >> 16) - ix);
     *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + ix) *
-        scaletbl[boxwidth - minboxwidth]  >> 16;
+                     scaletbl[boxwidth - minboxwidth] >>
+                 16;
   }
 }
 
-static void ScaleAddCols0_C(int dst_width, int boxheight, int x, int,
-                            const uint16* src_ptr, uint8* dst_ptr) {
+static void ScaleAddCols0_C(int dst_width,
+                            int boxheight,
+                            int x,
+                            int dx,
+                            const uint16_t* src_ptr,
+                            uint8_t* dst_ptr) {
   int scaleval = 65536 / boxheight;
   int i;
+  (void)dx;
   src_ptr += (x >> 16);
   for (i = 0; i < dst_width; ++i) {
-    *dst_ptr++ = src_ptr[i] * scaleval >> 16;
+    *dst_ptr++ = (uint8_t)(src_ptr[i] * scaleval >> 16);
   }
 }
 
-static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx,
-                            const uint16* src_ptr, uint8* dst_ptr) {
+static void ScaleAddCols1_C(int dst_width,
+                            int boxheight,
+                            int x,
+                            int dx,
+                            const uint16_t* src_ptr,
+                            uint8_t* dst_ptr) {
   int boxwidth = MIN1(dx >> 16);
   int scaleval = 65536 / (boxwidth * boxheight);
   int i;
   x >>= 16;
   for (i = 0; i < dst_width; ++i) {
-    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
+    *dst_ptr++ = (uint8_t)(SumPixels(boxwidth, src_ptr + x) * scaleval >> 16);
     x += boxwidth;
   }
 }
 
-static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx,
-                               const uint32* src_ptr, uint16* dst_ptr) {
+static void ScaleAddCols1_16_C(int dst_width,
+                               int boxheight,
+                               int x,
+                               int dx,
+                               const uint32_t* src_ptr,
+                               uint16_t* dst_ptr) {
   int boxwidth = MIN1(dx >> 16);
   int scaleval = 65536 / (boxwidth * boxheight);
   int i;
@@ -728,10 +901,14 @@ static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx,
 // one pixel of destination using fixed point (16.16) to step
 // through source, sampling a box of pixel with simple
 // averaging.
-static void ScalePlaneBox(int src_width, int src_height,
-                          int dst_width, int dst_height,
-                          int src_stride, int dst_stride,
-                          const uint8* src_ptr, uint8* dst_ptr) {
+static void ScalePlaneBox(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          int src_stride,
+                          int dst_stride,
+                          const uint8_t* src_ptr,
+                          uint8_t* dst_ptr) {
   int j, k;
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
@@ -739,18 +916,18 @@ static void ScalePlaneBox(int src_width, int src_height,
   int dx = 0;
   int dy = 0;
   const int max_y = (src_height << 16);
-  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
   {
-    // Allocate a row buffer of uint16.
+    // Allocate a row buffer of uint16_t.
     align_buffer_64(row16, src_width * 2);
     void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
-        const uint16* src_ptr, uint8* dst_ptr) =
-        (dx & 0xffff) ? ScaleAddCols2_C:
-        ((dx != 0x10000) ? ScaleAddCols1_C : ScaleAddCols0_C);
-    void (*ScaleAddRow)(const uint8* src_ptr, uint16* dst_ptr, int src_width) =
-        ScaleAddRow_C;
+                         const uint16_t* src_ptr, uint8_t* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_C
+                      : ((dx != 0x10000) ? ScaleAddCols1_C : ScaleAddCols0_C);
+    void (*ScaleAddRow)(const uint8_t* src_ptr, uint16_t* dst_ptr,
+                        int src_width) = ScaleAddRow_C;
 #if defined(HAS_SCALEADDROW_SSE2)
     if (TestCpuFlag(kCpuHasSSE2)) {
       ScaleAddRow = ScaleAddRow_Any_SSE2;
@@ -775,11 +952,27 @@ static void ScalePlaneBox(int src_width, int src_height,
       }
     }
 #endif
+#if defined(HAS_SCALEADDROW_MSA)
+    if (TestCpuFlag(kCpuHasMSA)) {
+      ScaleAddRow = ScaleAddRow_Any_MSA;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_MSA;
+      }
+    }
+#endif
+#if defined(HAS_SCALEADDROW_LSX)
+    if (TestCpuFlag(kCpuHasLSX)) {
+      ScaleAddRow = ScaleAddRow_Any_LSX;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_LSX;
+      }
+    }
+#endif
 
     for (j = 0; j < dst_height; ++j) {
       int boxheight;
       int iy = y >> 16;
-      const uint8* src = src_ptr + iy * src_stride;
+      const uint8_t* src = src_ptr + iy * (int64_t)src_stride;
       y += dy;
       if (y > max_y) {
         y = max_y;
@@ -787,20 +980,24 @@ static void ScalePlaneBox(int src_width, int src_height,
       boxheight = MIN1((y >> 16) - iy);
       memset(row16, 0, src_width * 2);
       for (k = 0; k < boxheight; ++k) {
-        ScaleAddRow(src, (uint16 *)(row16), src_width);
+        ScaleAddRow(src, (uint16_t*)(row16), src_width);
         src += src_stride;
       }
-      ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16), dst_ptr);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint16_t*)(row16), dst_ptr);
       dst_ptr += dst_stride;
     }
     free_aligned_buffer_64(row16);
   }
 }
 
-static void ScalePlaneBox_16(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint16* src_ptr, uint16* dst_ptr) {
+static void ScalePlaneBox_16(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint16_t* src_ptr,
+                             uint16_t* dst_ptr) {
   int j, k;
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
@@ -808,17 +1005,17 @@ static void ScalePlaneBox_16(int src_width, int src_height,
   int dx = 0;
   int dy = 0;
   const int max_y = (src_height << 16);
-  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
   {
-    // Allocate a row buffer of uint32.
+    // Allocate a row buffer of uint32_t.
     align_buffer_64(row32, src_width * 4);
     void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
-        const uint32* src_ptr, uint16* dst_ptr) =
-        (dx & 0xffff) ? ScaleAddCols2_16_C: ScaleAddCols1_16_C;
-    void (*ScaleAddRow)(const uint16* src_ptr, uint32* dst_ptr, int src_width) =
-        ScaleAddRow_16_C;
+                         const uint32_t* src_ptr, uint16_t* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_16_C : ScaleAddCols1_16_C;
+    void (*ScaleAddRow)(const uint16_t* src_ptr, uint32_t* dst_ptr,
+                        int src_width) = ScaleAddRow_16_C;
 
 #if defined(HAS_SCALEADDROW_16_SSE2)
     if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(src_width, 16)) {
@@ -829,7 +1026,7 @@ static void ScalePlaneBox_16(int src_width, int src_height,
     for (j = 0; j < dst_height; ++j) {
       int boxheight;
       int iy = y >> 16;
-      const uint16* src = src_ptr + iy * src_stride;
+      const uint16_t* src = src_ptr + iy * (int64_t)src_stride;
       y += dy;
       if (y > max_y) {
         y = max_y;
@@ -837,10 +1034,10 @@ static void ScalePlaneBox_16(int src_width, int src_height,
       boxheight = MIN1((y >> 16) - iy);
       memset(row32, 0, src_width * 4);
       for (k = 0; k < boxheight; ++k) {
-        ScaleAddRow(src, (uint32 *)(row32), src_width);
+        ScaleAddRow(src, (uint32_t*)(row32), src_width);
         src += src_stride;
       }
-      ScaleAddCols(dst_width, boxheight, x, dx, (uint32*)(row32), dst_ptr);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint32_t*)(row32), dst_ptr);
       dst_ptr += dst_stride;
     }
     free_aligned_buffer_64(row32);
@@ -848,10 +1045,14 @@ static void ScalePlaneBox_16(int src_width, int src_height,
 }
 
 // Scale plane down with bilinear interpolation.
-void ScalePlaneBilinearDown(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
+void ScalePlaneBilinearDown(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
                             enum FilterMode filtering) {
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
@@ -864,14 +1065,14 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
 
   const int max_y = (src_height - 1) << 16;
   int j;
-  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
-      int dst_width, int x, int dx) =
+  void (*ScaleFilterCols)(uint8_t* dst_ptr, const uint8_t* src_ptr,
+                          int dst_width, int x, int dx) =
       (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C;
-  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_C;
-  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
-             &x, &y, &dx, &dy);
+  void (*InterpolateRow)(uint8_t* dst_ptr, const uint8_t* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_SSSE3)
@@ -898,15 +1099,22 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2)) {
-    InterpolateRow = InterpolateRow_Any_DSPR2;
-    if (IS_ALIGNED(src_width, 4)) {
-      InterpolateRow = InterpolateRow_DSPR2;
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
-
 
 #if defined(HAS_SCALEFILTERCOLS_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
@@ -921,13 +1129,29 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEFILTERCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_LSX;
+    }
+  }
+#endif
   if (y > max_y) {
     y = max_y;
   }
 
   for (j = 0; j < dst_height; ++j) {
     int yi = y >> 16;
-    const uint8* src = src_ptr + yi * src_stride;
+    const uint8_t* src = src_ptr + yi * (int64_t)src_stride;
     if (filtering == kFilterLinear) {
       ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
     } else {
@@ -944,10 +1168,14 @@ void ScalePlaneBilinearDown(int src_width, int src_height,
   free_aligned_buffer_64(row);
 }
 
-void ScalePlaneBilinearDown_16(int src_width, int src_height,
-                               int dst_width, int dst_height,
-                               int src_stride, int dst_stride,
-                               const uint16* src_ptr, uint16* dst_ptr,
+void ScalePlaneBilinearDown_16(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
                                enum FilterMode filtering) {
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
@@ -960,19 +1188,19 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
 
   const int max_y = (src_height - 1) << 16;
   int j;
-  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
-      int dst_width, int x, int dx) =
+  void (*ScaleFilterCols)(uint16_t* dst_ptr, const uint16_t* src_ptr,
+                          int dst_width, int x, int dx) =
       (src_width >= 32768) ? ScaleFilterCols64_16_C : ScaleFilterCols_16_C;
-  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_16_C;
-  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
-             &x, &y, &dx, &dy);
+  void (*InterpolateRow)(uint16_t* dst_ptr, const uint16_t* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    InterpolateRow = InterpolateRow_16_Any_SSE2;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSE2;
     }
@@ -980,7 +1208,7 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    InterpolateRow = InterpolateRow_16_Any_SSSE3;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSSE3;
     }
@@ -988,7 +1216,7 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    InterpolateRow = InterpolateRow_16_Any_AVX2;
     if (IS_ALIGNED(src_width, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
     }
@@ -996,21 +1224,12 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_16_NEON;
+    InterpolateRow = InterpolateRow_16_Any_NEON;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2)) {
-    InterpolateRow = InterpolateRow_Any_16_DSPR2;
-    if (IS_ALIGNED(src_width, 4)) {
-      InterpolateRow = InterpolateRow_16_DSPR2;
-    }
-  }
-#endif
-
 
 #if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
@@ -1023,13 +1242,13 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
 
   for (j = 0; j < dst_height; ++j) {
     int yi = y >> 16;
-    const uint16* src = src_ptr + yi * src_stride;
+    const uint16_t* src = src_ptr + yi * (int64_t)src_stride;
     if (filtering == kFilterLinear) {
       ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
     } else {
       int yf = (y >> 8) & 255;
-      InterpolateRow((uint16*)row, src, src_stride, src_width, yf);
-      ScaleFilterCols(dst_ptr, (uint16*)row, dst_width, x, dx);
+      InterpolateRow((uint16_t*)row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, (uint16_t*)row, dst_width, x, dx);
     }
     dst_ptr += dst_stride;
     y += dy;
@@ -1041,10 +1260,14 @@ void ScalePlaneBilinearDown_16(int src_width, int src_height,
 }
 
 // Scale up down with bilinear interpolation.
-void ScalePlaneBilinearUp(int src_width, int src_height,
-                          int dst_width, int dst_height,
-                          int src_stride, int dst_stride,
-                          const uint8* src_ptr, uint8* dst_ptr,
+void ScalePlaneBilinearUp(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          int src_stride,
+                          int dst_stride,
+                          const uint8_t* src_ptr,
+                          uint8_t* dst_ptr,
                           enum FilterMode filtering) {
   int j;
   // Initial source x/y coordinate and step values as 16.16 fixed point.
@@ -1053,14 +1276,14 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   int dx = 0;
   int dy = 0;
   const int max_y = (src_height - 1) << 16;
-  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_C;
-  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
-      int dst_width, int x, int dx) =
+  void (*InterpolateRow)(uint8_t* dst_ptr, const uint8_t* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleFilterCols)(uint8_t* dst_ptr, const uint8_t* src_ptr,
+                          int dst_width, int x, int dx) =
       filtering ? ScaleFilterCols_C : ScaleCols_C;
-  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_SSSE3)
@@ -1087,14 +1310,6 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2)) {
-    InterpolateRow = InterpolateRow_Any_DSPR2;
-    if (IS_ALIGNED(dst_width, 4)) {
-      InterpolateRow = InterpolateRow_DSPR2;
-    }
-  }
-#endif
 
   if (filtering && src_width >= 32768) {
     ScaleFilterCols = ScaleFilterCols64_C;
@@ -1112,6 +1327,22 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_LSX)
+  if (filtering && TestCpuFlag(kCpuHasLSX) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_LSX;
+    }
+  }
+#endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleFilterCols = ScaleColsUp2_C;
 #if defined(HAS_SCALECOLS_SSE2)
@@ -1126,14 +1357,14 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   }
   {
     int yi = y >> 16;
-    const uint8* src = src_ptr + yi * src_stride;
+    const uint8_t* src = src_ptr + yi * (int64_t)src_stride;
 
     // Allocate 2 row buffers.
-    const int kRowSize = (dst_width + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (dst_width + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 
-    uint8* rowptr = row;
-    int rowstride = kRowSize;
+    uint8_t* rowptr = row;
+    int rowstride = row_size;
     int lasty = yi;
 
     ScaleFilterCols(rowptr, src, dst_width, x, dx);
@@ -1141,7 +1372,9 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
       src += src_stride;
     }
     ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
-    src += src_stride;
+    if (src_height > 2) {
+      src += src_stride;
+    }
 
     for (j = 0; j < dst_height; ++j) {
       yi = y >> 16;
@@ -1149,14 +1382,16 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
         if (y > max_y) {
           y = max_y;
           yi = y >> 16;
-          src = src_ptr + yi * src_stride;
+          src = src_ptr + yi * (int64_t)src_stride;
         }
         if (yi != lasty) {
           ScaleFilterCols(rowptr, src, dst_width, x, dx);
           rowptr += rowstride;
           rowstride = -rowstride;
           lasty = yi;
-          src += src_stride;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
         }
       }
       if (filtering == kFilterLinear) {
@@ -1172,10 +1407,335 @@ void ScalePlaneBilinearUp(int src_width, int src_height,
   }
 }
 
-void ScalePlaneBilinearUp_16(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint16* src_ptr, uint16* dst_ptr,
+// Scale plane, horizontally up by 2 times.
+// Uses linear filter horizontally, nearest vertically.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// This is used to scale U and V planes of I422 to I444.
+void ScalePlaneUp2_Linear(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          int src_stride,
+                          int dst_stride,
+                          const uint8_t* src_ptr,
+                          uint8_t* dst_ptr) {
+  void (*ScaleRowUp)(const uint8_t* src_ptr, uint8_t* dst_ptr, int dst_width) =
+      ScaleRowUp2_Linear_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_ptr + ((src_height - 1) / 2) * (int64_t)src_stride, dst_ptr,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_ptr + (y >> 16) * (int64_t)src_stride, dst_ptr, dst_width);
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale plane, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// This is used to scale U and V planes of I420 to I444.
+void ScalePlaneUp2_Bilinear(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_ptr,
+                            uint8_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                      uint8_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    // TODO(fbarchard): Test performance of writing one row of destination at a
+    // time.
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+// Scale at most 14 bit plane, horizontally up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// stride is in count of uint16_t.
+// This is used to scale U and V planes of I210 to I410 and I212 to I412.
+void ScalePlaneUp2_12_Linear(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint16_t* src_ptr,
+                             uint16_t* dst_ptr) {
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_ptr + ((src_height - 1) / 2) * (int64_t)src_stride, dst_ptr,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_ptr + (y >> 16) * (int64_t)src_stride, dst_ptr, dst_width);
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale at most 12 bit plane, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// stride is in count of uint16_t.
+// This is used to scale U and V planes of I010 to I410 and I012 to I412.
+void ScalePlaneUp2_12_Bilinear(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+void ScalePlaneUp2_16_Linear(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint16_t* src_ptr,
+                             uint16_t* dst_ptr) {
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_16_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_ptr + ((src_height - 1) / 2) * (int64_t)src_stride, dst_ptr,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_ptr + (y >> 16) * (int64_t)src_stride, dst_ptr, dst_width);
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+void ScalePlaneUp2_16_Bilinear(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_16_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+void ScalePlaneBilinearUp_16(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
                              enum FilterMode filtering) {
   int j;
   // Initial source x/y coordinate and step values as 16.16 fixed point.
@@ -1184,19 +1744,19 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
   int dx = 0;
   int dy = 0;
   const int max_y = (src_height - 1) << 16;
-  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_16_C;
-  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
-      int dst_width, int x, int dx) =
+  void (*InterpolateRow)(uint16_t* dst_ptr, const uint16_t* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_16_C;
+  void (*ScaleFilterCols)(uint16_t* dst_ptr, const uint16_t* src_ptr,
+                          int dst_width, int x, int dx) =
       filtering ? ScaleFilterCols_16_C : ScaleCols_16_C;
-  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
 
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    InterpolateRow = InterpolateRow_16_Any_SSE2;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSE2;
     }
@@ -1204,7 +1764,7 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    InterpolateRow = InterpolateRow_16_Any_SSSE3;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSSE3;
     }
@@ -1212,7 +1772,7 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    InterpolateRow = InterpolateRow_16_Any_AVX2;
     if (IS_ALIGNED(dst_width, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
     }
@@ -1220,20 +1780,12 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_16_NEON;
+    InterpolateRow = InterpolateRow_16_Any_NEON;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2)) {
-    InterpolateRow = InterpolateRow_Any_16_DSPR2;
-    if (IS_ALIGNED(dst_width, 4)) {
-      InterpolateRow = InterpolateRow_16_DSPR2;
-    }
-  }
-#endif
 
   if (filtering && src_width >= 32768) {
     ScaleFilterCols = ScaleFilterCols64_16_C;
@@ -1251,20 +1803,19 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
     }
 #endif
   }
-
   if (y > max_y) {
     y = max_y;
   }
   {
     int yi = y >> 16;
-    const uint16* src = src_ptr + yi * src_stride;
+    const uint16_t* src = src_ptr + yi * (int64_t)src_stride;
 
     // Allocate 2 row buffers.
-    const int kRowSize = (dst_width + 31) & ~31;
-    align_buffer_64(row, kRowSize * 4);
+    const int row_size = (dst_width + 31) & ~31;
+    align_buffer_64(row, row_size * 4);
 
-    uint16* rowptr = (uint16*)row;
-    int rowstride = kRowSize;
+    uint16_t* rowptr = (uint16_t*)row;
+    int rowstride = row_size;
     int lasty = yi;
 
     ScaleFilterCols(rowptr, src, dst_width, x, dx);
@@ -1272,7 +1823,9 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
       src += src_stride;
     }
     ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
-    src += src_stride;
+    if (src_height > 2) {
+      src += src_stride;
+    }
 
     for (j = 0; j < dst_height; ++j) {
       yi = y >> 16;
@@ -1280,14 +1833,16 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
         if (y > max_y) {
           y = max_y;
           yi = y >> 16;
-          src = src_ptr + yi * src_stride;
+          src = src_ptr + yi * (int64_t)src_stride;
         }
         if (yi != lasty) {
           ScaleFilterCols(rowptr, src, dst_width, x, dx);
           rowptr += rowstride;
           rowstride = -rowstride;
           lasty = yi;
-          src += src_stride;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
         }
       }
       if (filtering == kFilterLinear) {
@@ -1308,20 +1863,24 @@ void ScalePlaneBilinearUp_16(int src_width, int src_height,
 // of x and dx is the integer part of the source position and
 // the lower 16 bits are the fixed decimal part.
 
-static void ScalePlaneSimple(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr) {
+static void ScalePlaneSimple(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint8_t* src_ptr,
+                             uint8_t* dst_ptr) {
   int i;
-  void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr,
-      int dst_width, int x, int dx) = ScaleCols_C;
+  void (*ScaleCols)(uint8_t* dst_ptr, const uint8_t* src_ptr, int dst_width,
+                    int x, int dx) = ScaleCols_C;
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
   int y = 0;
   int dx = 0;
   int dy = 0;
-  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
 
   if (src_width * 2 == dst_width && x < 0x8000) {
@@ -1334,26 +1893,31 @@ static void ScalePlaneSimple(int src_width, int src_height,
   }
 
   for (i = 0; i < dst_height; ++i) {
-    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, dst_width, x, dx);
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * (int64_t)src_stride, dst_width, x,
+              dx);
     dst_ptr += dst_stride;
     y += dy;
   }
 }
 
-static void ScalePlaneSimple_16(int src_width, int src_height,
-                                int dst_width, int dst_height,
-                                int src_stride, int dst_stride,
-                                const uint16* src_ptr, uint16* dst_ptr) {
+static void ScalePlaneSimple_16(int src_width,
+                                int src_height,
+                                int dst_width,
+                                int dst_height,
+                                int src_stride,
+                                int dst_stride,
+                                const uint16_t* src_ptr,
+                                uint16_t* dst_ptr) {
   int i;
-  void (*ScaleCols)(uint16* dst_ptr, const uint16* src_ptr,
-      int dst_width, int x, int dx) = ScaleCols_16_C;
+  void (*ScaleCols)(uint16_t* dst_ptr, const uint16_t* src_ptr, int dst_width,
+                    int x, int dx) = ScaleCols_16_C;
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
   int y = 0;
   int dx = 0;
   int dy = 0;
-  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
 
   if (src_width * 2 == dst_width && x < 0x8000) {
@@ -1366,8 +1930,8 @@ static void ScalePlaneSimple_16(int src_width, int src_height,
   }
 
   for (i = 0; i < dst_height; ++i) {
-    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
-              dst_width, x, dx);
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * (int64_t)src_stride, dst_width, x,
+              dx);
     dst_ptr += dst_stride;
     y += dy;
   }
@@ -1375,24 +1939,26 @@ static void ScalePlaneSimple_16(int src_width, int src_height,
 
 // Scale a plane.
 // This function dispatches to a specialized scaler based on scale factor.
-
 LIBYUV_API
-void ScalePlane(const uint8* src, int src_stride,
-                int src_width, int src_height,
-                uint8* dst, int dst_stride,
-                int dst_width, int dst_height,
+void ScalePlane(const uint8_t* src,
+                int src_stride,
+                int src_width,
+                int src_height,
+                uint8_t* dst,
+                int dst_stride,
+                int dst_width,
+                int dst_height,
                 enum FilterMode filtering) {
   // Simplify filtering when possible.
-  filtering = ScaleFilterReduce(src_width, src_height,
-                                dst_width, dst_height, filtering);
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
 
   // Negative height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;
-    src = src + (src_height - 1) * src_stride;
+    src = src + (src_height - 1) * (int64_t)src_stride;
     src_stride = -src_stride;
   }
-
   // Use specialized scales to improve performance for common resolutions.
   // For example, all the 1/2 scalings will use ScalePlaneDown2()
   if (dst_width == src_width && dst_height == src_height) {
@@ -1401,48 +1967,64 @@ void ScalePlane(const uint8* src, int src_stride,
     return;
   }
   if (dst_width == src_width && filtering != kFilterBox) {
-    int dy = FixedDiv(src_height, dst_height);
+    int dy = 0;
+    int y = 0;
+    // When scaling down, use the center 2 rows to filter.
+    // When scaling up, last row of destination uses the last 2 source rows.
+    if (dst_height <= src_height) {
+      dy = FixedDiv(src_height, dst_height);
+      y = CENTERSTART(dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (src_height > 1 && dst_height > 1) {
+      dy = FixedDiv1(src_height, dst_height);
+    }
     // Arbitrary scale vertically, but unscaled horizontally.
-    ScalePlaneVertical(src_height,
-                       dst_width, dst_height,
-                       src_stride, dst_stride, src, dst,
-                       0, 0, dy, 1, filtering);
+    ScalePlaneVertical(src_height, dst_width, dst_height, src_stride,
+                       dst_stride, src, dst, 0, y, dy, /*bpp=*/1, filtering);
     return;
   }
   if (dst_width <= Abs(src_width) && dst_height <= src_height) {
     // Scale down.
-    if (4 * dst_width == 3 * src_width &&
-        4 * dst_height == 3 * src_height) {
+    if (4 * dst_width == 3 * src_width && 4 * dst_height == 3 * src_height) {
       // optimized, 3/4
-      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src, dst, filtering);
+      ScalePlaneDown34(src_width, src_height, dst_width, dst_height, src_stride,
+                       dst_stride, src, dst, filtering);
       return;
     }
     if (2 * dst_width == src_width && 2 * dst_height == src_height) {
       // optimized, 1/2
-      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
+      ScalePlaneDown2(src_width, src_height, dst_width, dst_height, src_stride,
+                      dst_stride, src, dst, filtering);
       return;
     }
     // 3/8 rounded up for odd sized chroma height.
-    if (8 * dst_width == 3 * src_width &&
-        dst_height == ((src_height * 3 + 7) / 8)) {
+    if (8 * dst_width == 3 * src_width && 8 * dst_height == 3 * src_height) {
       // optimized, 3/8
-      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src, dst, filtering);
+      ScalePlaneDown38(src_width, src_height, dst_width, dst_height, src_stride,
+                       dst_stride, src, dst, filtering);
       return;
     }
     if (4 * dst_width == src_width && 4 * dst_height == src_height &&
         (filtering == kFilterBox || filtering == kFilterNone)) {
       // optimized, 1/4
-      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
+      ScalePlaneDown4(src_width, src_height, dst_width, dst_height, src_stride,
+                      dst_stride, src, dst, filtering);
       return;
     }
   }
   if (filtering == kFilterBox && dst_height * 2 < src_height) {
-    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
-                  src_stride, dst_stride, src, dst);
+    ScalePlaneBox(src_width, src_height, dst_width, dst_height, src_stride,
+                  dst_stride, src, dst);
+    return;
+  }
+  if ((dst_width + 1) / 2 == src_width && filtering == kFilterLinear) {
+    ScalePlaneUp2_Linear(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst);
+    return;
+  }
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScalePlaneUp2_Bilinear(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst);
     return;
   }
   if (filtering && dst_height > src_height) {
@@ -1455,27 +2037,30 @@ void ScalePlane(const uint8* src, int src_stride,
                            src_stride, dst_stride, src, dst, filtering);
     return;
   }
-  ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
-                   src_stride, dst_stride, src, dst);
+  ScalePlaneSimple(src_width, src_height, dst_width, dst_height, src_stride,
+                   dst_stride, src, dst);
 }
 
 LIBYUV_API
-void ScalePlane_16(const uint16* src, int src_stride,
-                  int src_width, int src_height,
-                  uint16* dst, int dst_stride,
-                  int dst_width, int dst_height,
-                  enum FilterMode filtering) {
+void ScalePlane_16(const uint16_t* src,
+                   int src_stride,
+                   int src_width,
+                   int src_height,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int dst_width,
+                   int dst_height,
+                   enum FilterMode filtering) {
   // Simplify filtering when possible.
-  filtering = ScaleFilterReduce(src_width, src_height,
-                                dst_width, dst_height, filtering);
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
 
   // Negative height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;
-    src = src + (src_height - 1) * src_stride;
+    src = src + (src_height - 1) * (int64_t)src_stride;
     src_stride = -src_stride;
   }
-
   // Use specialized scales to improve performance for common resolutions.
   // For example, all the 1/2 scalings will use ScalePlaneDown2()
   if (dst_width == src_width && dst_height == src_height) {
@@ -1483,19 +2068,28 @@ void ScalePlane_16(const uint16* src, int src_stride,
     CopyPlane_16(src, src_stride, dst, dst_stride, dst_width, dst_height);
     return;
   }
-  if (dst_width == src_width) {
-    int dy = FixedDiv(src_height, dst_height);
-    // Arbitrary scale vertically, but unscaled vertically.
-    ScalePlaneVertical_16(src_height,
-                          dst_width, dst_height,
-                          src_stride, dst_stride, src, dst,
-                          0, 0, dy, 1, filtering);
+  if (dst_width == src_width && filtering != kFilterBox) {
+    int dy = 0;
+    int y = 0;
+    // When scaling down, use the center 2 rows to filter.
+    // When scaling up, last row of destination uses the last 2 source rows.
+    if (dst_height <= src_height) {
+      dy = FixedDiv(src_height, dst_height);
+      y = CENTERSTART(dy, -32768);  // Subtract 0.5 (32768) to center filter.
+      // When scaling up, ensure the last row of destination uses the last
+      // source. Avoid divide by zero for dst_height but will do no scaling
+      // later.
+    } else if (src_height > 1 && dst_height > 1) {
+      dy = FixedDiv1(src_height, dst_height);
+    }
+    // Arbitrary scale vertically, but unscaled horizontally.
+    ScalePlaneVertical_16(src_height, dst_width, dst_height, src_stride,
+                          dst_stride, src, dst, 0, y, dy, /*bpp=*/1, filtering);
     return;
   }
   if (dst_width <= Abs(src_width) && dst_height <= src_height) {
     // Scale down.
-    if (4 * dst_width == 3 * src_width &&
-        4 * dst_height == 3 * src_height) {
+    if (4 * dst_width == 3 * src_width && 4 * dst_height == 3 * src_height) {
       // optimized, 3/4
       ScalePlaneDown34_16(src_width, src_height, dst_width, dst_height,
                           src_stride, dst_stride, src, dst, filtering);
@@ -1508,15 +2102,14 @@ void ScalePlane_16(const uint16* src, int src_stride,
       return;
     }
     // 3/8 rounded up for odd sized chroma height.
-    if (8 * dst_width == 3 * src_width &&
-        dst_height == ((src_height * 3 + 7) / 8)) {
+    if (8 * dst_width == 3 * src_width && 8 * dst_height == 3 * src_height) {
       // optimized, 3/8
       ScalePlaneDown38_16(src_width, src_height, dst_width, dst_height,
                           src_stride, dst_stride, src, dst, filtering);
       return;
     }
     if (4 * dst_width == src_width && 4 * dst_height == src_height &&
-               filtering != kFilterBilinear) {
+        (filtering == kFilterBox || filtering == kFilterNone)) {
       // optimized, 1/4
       ScalePlaneDown4_16(src_width, src_height, dst_width, dst_height,
                          src_stride, dst_stride, src, dst, filtering);
@@ -1524,8 +2117,19 @@ void ScalePlane_16(const uint16* src, int src_stride,
     }
   }
   if (filtering == kFilterBox && dst_height * 2 < src_height) {
-    ScalePlaneBox_16(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src, dst);
+    ScalePlaneBox_16(src_width, src_height, dst_width, dst_height, src_stride,
+                     dst_stride, src, dst);
+    return;
+  }
+  if ((dst_width + 1) / 2 == src_width && filtering == kFilterLinear) {
+    ScalePlaneUp2_16_Linear(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst);
+    return;
+  }
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScalePlaneUp2_16_Bilinear(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst);
     return;
   }
   if (filtering && dst_height > src_height) {
@@ -1538,132 +2142,435 @@ void ScalePlane_16(const uint16* src, int src_stride,
                               src_stride, dst_stride, src, dst, filtering);
     return;
   }
-  ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst);
+  ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height, src_stride,
+                      dst_stride, src, dst);
+}
+
+LIBYUV_API
+void ScalePlane_12(const uint16_t* src,
+                   int src_stride,
+                   int src_width,
+                   int src_height,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int dst_width,
+                   int dst_height,
+                   enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * (int64_t)src_stride;
+    src_stride = -src_stride;
+  }
+
+  if ((dst_width + 1) / 2 == src_width && filtering == kFilterLinear) {
+    ScalePlaneUp2_12_Linear(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst);
+    return;
+  }
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScalePlaneUp2_12_Bilinear(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst);
+    return;
+  }
+
+  ScalePlane_16(src, src_stride, src_width, src_height, dst, dst_stride,
+                dst_width, dst_height, filtering);
 }
 
 // Scale an I420 image.
 // This function in turn calls a scaling function for each plane.
 
 LIBYUV_API
-int I420Scale(const uint8* src_y, int src_stride_y,
-              const uint8* src_u, int src_stride_u,
-              const uint8* src_v, int src_stride_v,
-              int src_width, int src_height,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int dst_width, int dst_height,
+int I420Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
               enum FilterMode filtering) {
   int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
   int src_halfheight = SUBSAMPLE(src_height, 1, 1);
   int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
   int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
-  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
-      src_width > 32768 || src_height > 32768 ||
-      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
     return -1;
   }
 
-  ScalePlane(src_y, src_stride_y, src_width, src_height,
-             dst_y, dst_stride_y, dst_width, dst_height,
-             filtering);
-  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
-             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
-             filtering);
-  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
-             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
-             filtering);
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+             dst_width, dst_height, filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight, dst_u,
+             dst_stride_u, dst_halfwidth, dst_halfheight, filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight, dst_v,
+             dst_stride_v, dst_halfwidth, dst_halfheight, filtering);
   return 0;
 }
 
 LIBYUV_API
-int I420Scale_16(const uint16* src_y, int src_stride_y,
-                 const uint16* src_u, int src_stride_u,
-                 const uint16* src_v, int src_stride_v,
-                 int src_width, int src_height,
-                 uint16* dst_y, int dst_stride_y,
-                 uint16* dst_u, int dst_stride_u,
-                 uint16* dst_v, int dst_stride_v,
-                 int dst_width, int dst_height,
+int I420Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
                  enum FilterMode filtering) {
   int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
   int src_halfheight = SUBSAMPLE(src_height, 1, 1);
   int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
   int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
-  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
-      src_width > 32768 || src_height > 32768 ||
-      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
     return -1;
   }
 
-  ScalePlane_16(src_y, src_stride_y, src_width, src_height,
-                dst_y, dst_stride_y, dst_width, dst_height,
-                filtering);
-  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight,
-                dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
-                filtering);
-  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight,
-                dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
-                filtering);
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight, dst_u,
+                dst_stride_u, dst_halfwidth, dst_halfheight, filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight, dst_v,
+                dst_stride_v, dst_halfwidth, dst_halfheight, filtering);
   return 0;
 }
 
-// Deprecated api
 LIBYUV_API
-int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
-          int src_stride_y, int src_stride_u, int src_stride_v,
-          int src_width, int src_height,
-          uint8* dst_y, uint8* dst_u, uint8* dst_v,
-          int dst_stride_y, int dst_stride_u, int dst_stride_v,
-          int dst_width, int dst_height,
-          LIBYUV_BOOL interpolate) {
-  return I420Scale(src_y, src_stride_y,
-                   src_u, src_stride_u,
-                   src_v, src_stride_v,
-                   src_width, src_height,
-                   dst_y, dst_stride_y,
-                   dst_u, dst_stride_u,
-                   dst_v, dst_stride_v,
-                   dst_width, dst_height,
-                   interpolate ? kFilterBox : kFilterNone);
+int I420Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_12(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_u, src_stride_u, src_halfwidth, src_halfheight, dst_u,
+                dst_stride_u, dst_halfwidth, dst_halfheight, filtering);
+  ScalePlane_12(src_v, src_stride_v, src_halfwidth, src_halfheight, dst_v,
+                dst_stride_v, dst_halfwidth, dst_halfheight, filtering);
+  return 0;
 }
 
-// Deprecated api
+// Scale an I444 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I444Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering) {
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+             dst_width, dst_height, filtering);
+  ScalePlane(src_u, src_stride_u, src_width, src_height, dst_u, dst_stride_u,
+             dst_width, dst_height, filtering);
+  ScalePlane(src_v, src_stride_v, src_width, src_height, dst_v, dst_stride_v,
+             dst_width, dst_height, filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I444Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_16(src_u, src_stride_u, src_width, src_height, dst_u, dst_stride_u,
+                dst_width, dst_height, filtering);
+  ScalePlane_16(src_v, src_stride_v, src_width, src_height, dst_v, dst_stride_v,
+                dst_width, dst_height, filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I444Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_12(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_u, src_stride_u, src_width, src_height, dst_u, dst_stride_u,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_v, src_stride_v, src_width, src_height, dst_v, dst_stride_v,
+                dst_width, dst_height, filtering);
+  return 0;
+}
+
+// Scale an I422 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I422Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+             dst_width, dst_height, filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_height, dst_u,
+             dst_stride_u, dst_halfwidth, dst_height, filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_height, dst_v,
+             dst_stride_v, dst_halfwidth, dst_height, filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I422Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_height, dst_u,
+                dst_stride_u, dst_halfwidth, dst_height, filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_height, dst_v,
+                dst_stride_v, dst_halfwidth, dst_height, filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I422Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_12(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_u, src_stride_u, src_halfwidth, src_height, dst_u,
+                dst_stride_u, dst_halfwidth, dst_height, filtering);
+  ScalePlane_12(src_v, src_stride_v, src_halfwidth, src_height, dst_v,
+                dst_stride_v, dst_halfwidth, dst_height, filtering);
+  return 0;
+}
+
+// Scale an NV12 image.
+// This function in turn calls a scaling function for each plane.
+
 LIBYUV_API
-int ScaleOffset(const uint8* src, int src_width, int src_height,
-                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
-                LIBYUV_BOOL interpolate) {
-  // Chroma requires offset to multiple of 2.
-  int dst_yoffset_even = dst_yoffset & ~1;
+int NV12Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_uv,
+              int src_stride_uv,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_uv,
+              int dst_stride_uv,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering) {
   int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
   int src_halfheight = SUBSAMPLE(src_height, 1, 1);
   int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
   int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
-  int aheight = dst_height - dst_yoffset_even * 2;  // actual output height
-  const uint8* src_y = src;
-  const uint8* src_u = src + src_width * src_height;
-  const uint8* src_v = src + src_width * src_height +
-                             src_halfwidth * src_halfheight;
-  uint8* dst_y = dst + dst_yoffset_even * dst_width;
-  uint8* dst_u = dst + dst_width * dst_height +
-                 (dst_yoffset_even >> 1) * dst_halfwidth;
-  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
-                 (dst_yoffset_even >> 1) * dst_halfwidth;
-  if (!src || src_width <= 0 || src_height <= 0 ||
-      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 ||
-      dst_yoffset_even >= dst_height) {
+
+  if (!src_y || !src_uv || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_uv ||
+      dst_width <= 0 || dst_height <= 0) {
     return -1;
   }
-  return I420Scale(src_y, src_width,
-                   src_u, src_halfwidth,
-                   src_v, src_halfwidth,
-                   src_width, src_height,
-                   dst_y, dst_width,
-                   dst_u, dst_halfwidth,
-                   dst_v, dst_halfwidth,
-                   dst_width, aheight,
-                   interpolate ? kFilterBox : kFilterNone);
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+             dst_width, dst_height, filtering);
+  UVScale(src_uv, src_stride_uv, src_halfwidth, src_halfheight, dst_uv,
+          dst_stride_uv, dst_halfwidth, dst_halfheight, filtering);
+  return 0;
+}
+
+// Deprecated api
+LIBYUV_API
+int Scale(const uint8_t* src_y,
+          const uint8_t* src_u,
+          const uint8_t* src_v,
+          int src_stride_y,
+          int src_stride_u,
+          int src_stride_v,
+          int src_width,
+          int src_height,
+          uint8_t* dst_y,
+          uint8_t* dst_u,
+          uint8_t* dst_v,
+          int dst_stride_y,
+          int dst_stride_u,
+          int dst_stride_v,
+          int dst_width,
+          int dst_height,
+          LIBYUV_BOOL interpolate) {
+  return I420Scale(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                   src_stride_v, src_width, src_height, dst_y, dst_stride_y,
+                   dst_u, dst_stride_u, dst_v, dst_stride_v, dst_width,
+                   dst_height, interpolate ? kFilterBox : kFilterNone);
 }
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/scale_any.cc b/media/libyuv/source/scale_any.cc
index ed76a9e4c0..317041f806 100644
--- a/media/libyuv/source/scale_any.cc
+++ b/media/libyuv/source/scale_any.cc
@@ -8,6 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <string.h>  // For memset/memcpy
+
 #include "libyuv/scale.h"
 #include "libyuv/scale_row.h"
 
@@ -18,186 +20,558 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
-#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                            \
-    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
-                 int dst_width, int x, int dx) {                               \
-      int n = dst_width & ~MASK;                                               \
-      if (n > 0) {                                                             \
-        TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                                 \
-      }                                                                        \
-      TERP_C(dst_ptr + n * BPP, src_ptr,                                       \
-             dst_width & MASK, x + n * dx, dx);                                \
-    }
-
-#ifdef HAS_SCALEFILTERCOLS_NEON
-CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
-#endif
-#ifdef HAS_SCALEARGBCOLS_NEON
-CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
-#endif
-#ifdef HAS_SCALEARGBFILTERCOLS_NEON
-CANY(ScaleARGBFilterCols_Any_NEON, ScaleARGBFilterCols_NEON,
-     ScaleARGBFilterCols_C, 4, 3)
-#endif
-#undef CANY
-
 // Fixed scale down.
+// Mask may be non-power of 2, so use MOD
 #define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
-    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
-                 uint8* dst_ptr, int dst_width) {                              \
-      int r = (int)((unsigned int)dst_width % (MASK + 1));                     \
-      int n = dst_width - r;                                                   \
-      if (n > 0) {                                                             \
-        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
-      }                                                                        \
-      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
-                     dst_ptr + n * BPP, r);                                    \
-    }
+  void NAMEANY(const uint8_t* src_ptr, ptrdiff_t src_stride, uint8_t* dst_ptr, \
+               int dst_width) {                                                \
+    int r = (int)((unsigned int)dst_width % (MASK + 1)); /* NOLINT */          \
+    int n = dst_width - r;                                                     \
+    if (n > 0) {                                                               \
+      SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                      \
+    }                                                                          \
+    SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                   \
+                   dst_ptr + n * BPP, r);                                      \
+  }
 
 // Fixed scale down for odd source width.  Used by I420Blend subsampling.
 // Since dst_width is (width + 1) / 2, this function scales one less pixel
 // and copies the last pixel.
 #define SDODD(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
-    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
-                 uint8* dst_ptr, int dst_width) {                              \
-      int r = (int)((unsigned int)(dst_width - 1) % (MASK + 1));               \
-      int n = dst_width - r;                                                   \
-      if (n > 0) {                                                             \
-        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
-      }                                                                        \
-      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
-                     dst_ptr + n * BPP, r);                                    \
-    }
+  void NAMEANY(const uint8_t* src_ptr, ptrdiff_t src_stride, uint8_t* dst_ptr, \
+               int dst_width) {                                                \
+    int r = (int)((unsigned int)(dst_width - 1) % (MASK + 1)); /* NOLINT */    \
+    int n = (dst_width - 1) - r;                                               \
+    if (n > 0) {                                                               \
+      SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                      \
+    }                                                                          \
+    SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                   \
+                   dst_ptr + n * BPP, r + 1);                                  \
+  }
 
 #ifdef HAS_SCALEROWDOWN2_SSSE3
 SDANY(ScaleRowDown2_Any_SSSE3, ScaleRowDown2_SSSE3, ScaleRowDown2_C, 2, 1, 15)
-SDANY(ScaleRowDown2Linear_Any_SSSE3, ScaleRowDown2Linear_SSSE3,
-      ScaleRowDown2Linear_C, 2, 1, 15)
-SDANY(ScaleRowDown2Box_Any_SSSE3, ScaleRowDown2Box_SSSE3, ScaleRowDown2Box_C,
-      2, 1, 15)
-SDODD(ScaleRowDown2Box_Odd_SSSE3, ScaleRowDown2Box_SSSE3,
-      ScaleRowDown2Box_Odd_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_SSSE3,
+      ScaleRowDown2Linear_SSSE3,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      15)
+SDANY(ScaleRowDown2Box_Any_SSSE3,
+      ScaleRowDown2Box_SSSE3,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      15)
+SDODD(ScaleRowDown2Box_Odd_SSSE3,
+      ScaleRowDown2Box_SSSE3,
+      ScaleRowDown2Box_Odd_C,
+      2,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEUVROWDOWN2BOX_SSSE3
+SDANY(ScaleUVRowDown2Box_Any_SSSE3,
+      ScaleUVRowDown2Box_SSSE3,
+      ScaleUVRowDown2Box_C,
+      2,
+      2,
+      3)
+#endif
+#ifdef HAS_SCALEUVROWDOWN2BOX_AVX2
+SDANY(ScaleUVRowDown2Box_Any_AVX2,
+      ScaleUVRowDown2Box_AVX2,
+      ScaleUVRowDown2Box_C,
+      2,
+      2,
+      7)
 #endif
 #ifdef HAS_SCALEROWDOWN2_AVX2
 SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31)
-SDANY(ScaleRowDown2Linear_Any_AVX2, ScaleRowDown2Linear_AVX2,
-      ScaleRowDown2Linear_C, 2, 1, 31)
-SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C,
-      2, 1, 31)
-SDODD(ScaleRowDown2Box_Odd_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_Odd_C,
-      2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_AVX2,
+      ScaleRowDown2Linear_AVX2,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      31)
+SDANY(ScaleRowDown2Box_Any_AVX2,
+      ScaleRowDown2Box_AVX2,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      31)
+SDODD(ScaleRowDown2Box_Odd_AVX2,
+      ScaleRowDown2Box_AVX2,
+      ScaleRowDown2Box_Odd_C,
+      2,
+      1,
+      31)
 #endif
 #ifdef HAS_SCALEROWDOWN2_NEON
 SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15)
-SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON,
-      ScaleRowDown2Linear_C, 2, 1, 15)
-SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON,
-      ScaleRowDown2Box_C, 2, 1, 15)
-SDODD(ScaleRowDown2Box_Odd_NEON, ScaleRowDown2Box_NEON,
-      ScaleRowDown2Box_Odd_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_NEON,
+      ScaleRowDown2Linear_NEON,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      15)
+SDANY(ScaleRowDown2Box_Any_NEON,
+      ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      15)
+SDODD(ScaleRowDown2Box_Odd_NEON,
+      ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_Odd_C,
+      2,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEUVROWDOWN2BOX_NEON
+SDANY(ScaleUVRowDown2Box_Any_NEON,
+      ScaleUVRowDown2Box_NEON,
+      ScaleUVRowDown2Box_C,
+      2,
+      2,
+      7)
+#endif
+
+#ifdef HAS_SCALEROWDOWN2_MSA
+SDANY(ScaleRowDown2_Any_MSA, ScaleRowDown2_MSA, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_MSA,
+      ScaleRowDown2Linear_MSA,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      31)
+SDANY(ScaleRowDown2Box_Any_MSA,
+      ScaleRowDown2Box_MSA,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      31)
+#endif
+#ifdef HAS_SCALEROWDOWN2_LSX
+SDANY(ScaleRowDown2_Any_LSX, ScaleRowDown2_LSX, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_LSX,
+      ScaleRowDown2Linear_LSX,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      31)
+SDANY(ScaleRowDown2Box_Any_LSX,
+      ScaleRowDown2Box_LSX,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      31)
 #endif
 #ifdef HAS_SCALEROWDOWN4_SSSE3
 SDANY(ScaleRowDown4_Any_SSSE3, ScaleRowDown4_SSSE3, ScaleRowDown4_C, 4, 1, 7)
-SDANY(ScaleRowDown4Box_Any_SSSE3, ScaleRowDown4Box_SSSE3, ScaleRowDown4Box_C,
-      4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_SSSE3,
+      ScaleRowDown4Box_SSSE3,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      7)
 #endif
 #ifdef HAS_SCALEROWDOWN4_AVX2
 SDANY(ScaleRowDown4_Any_AVX2, ScaleRowDown4_AVX2, ScaleRowDown4_C, 4, 1, 15)
-SDANY(ScaleRowDown4Box_Any_AVX2, ScaleRowDown4Box_AVX2, ScaleRowDown4Box_C,
-      4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_AVX2,
+      ScaleRowDown4Box_AVX2,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      15)
 #endif
 #ifdef HAS_SCALEROWDOWN4_NEON
 SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7)
-SDANY(ScaleRowDown4Box_Any_NEON, ScaleRowDown4Box_NEON, ScaleRowDown4Box_C,
-      4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_NEON,
+      ScaleRowDown4Box_NEON,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      7)
+#endif
+#ifdef HAS_SCALEROWDOWN4_MSA
+SDANY(ScaleRowDown4_Any_MSA, ScaleRowDown4_MSA, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_MSA,
+      ScaleRowDown4Box_MSA,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEROWDOWN4_LSX
+SDANY(ScaleRowDown4_Any_LSX, ScaleRowDown4_LSX, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_LSX,
+      ScaleRowDown4Box_LSX,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      15)
 #endif
 #ifdef HAS_SCALEROWDOWN34_SSSE3
-SDANY(ScaleRowDown34_Any_SSSE3, ScaleRowDown34_SSSE3,
-      ScaleRowDown34_C, 4 / 3, 1, 23)
-SDANY(ScaleRowDown34_0_Box_Any_SSSE3, ScaleRowDown34_0_Box_SSSE3,
-      ScaleRowDown34_0_Box_C, 4 / 3, 1, 23)
-SDANY(ScaleRowDown34_1_Box_Any_SSSE3, ScaleRowDown34_1_Box_SSSE3,
-      ScaleRowDown34_1_Box_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_Any_SSSE3,
+      ScaleRowDown34_SSSE3,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_0_Box_Any_SSSE3,
+      ScaleRowDown34_0_Box_SSSE3,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_1_Box_Any_SSSE3,
+      ScaleRowDown34_1_Box_SSSE3,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      23)
 #endif
 #ifdef HAS_SCALEROWDOWN34_NEON
-SDANY(ScaleRowDown34_Any_NEON, ScaleRowDown34_NEON,
-      ScaleRowDown34_C, 4 / 3, 1, 23)
-SDANY(ScaleRowDown34_0_Box_Any_NEON, ScaleRowDown34_0_Box_NEON,
-      ScaleRowDown34_0_Box_C, 4 / 3, 1, 23)
-SDANY(ScaleRowDown34_1_Box_Any_NEON, ScaleRowDown34_1_Box_NEON,
-      ScaleRowDown34_1_Box_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_Any_NEON,
+      ScaleRowDown34_NEON,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_0_Box_Any_NEON,
+      ScaleRowDown34_0_Box_NEON,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_1_Box_Any_NEON,
+      ScaleRowDown34_1_Box_NEON,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      23)
+#endif
+#ifdef HAS_SCALEROWDOWN34_MSA
+SDANY(ScaleRowDown34_Any_MSA,
+      ScaleRowDown34_MSA,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_0_Box_Any_MSA,
+      ScaleRowDown34_0_Box_MSA,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_1_Box_Any_MSA,
+      ScaleRowDown34_1_Box_MSA,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      47)
+#endif
+#ifdef HAS_SCALEROWDOWN34_LSX
+SDANY(ScaleRowDown34_Any_LSX,
+      ScaleRowDown34_LSX,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_0_Box_Any_LSX,
+      ScaleRowDown34_0_Box_LSX,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_1_Box_Any_LSX,
+      ScaleRowDown34_1_Box_LSX,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      47)
 #endif
 #ifdef HAS_SCALEROWDOWN38_SSSE3
-SDANY(ScaleRowDown38_Any_SSSE3, ScaleRowDown38_SSSE3,
-      ScaleRowDown38_C, 8 / 3, 1, 11)
-SDANY(ScaleRowDown38_3_Box_Any_SSSE3, ScaleRowDown38_3_Box_SSSE3,
-      ScaleRowDown38_3_Box_C, 8 / 3, 1, 5)
-SDANY(ScaleRowDown38_2_Box_Any_SSSE3, ScaleRowDown38_2_Box_SSSE3,
-      ScaleRowDown38_2_Box_C, 8 / 3, 1, 5)
+SDANY(ScaleRowDown38_Any_SSSE3,
+      ScaleRowDown38_SSSE3,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_SSSE3,
+      ScaleRowDown38_3_Box_SSSE3,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      5)
+SDANY(ScaleRowDown38_2_Box_Any_SSSE3,
+      ScaleRowDown38_2_Box_SSSE3,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      5)
 #endif
 #ifdef HAS_SCALEROWDOWN38_NEON
-SDANY(ScaleRowDown38_Any_NEON, ScaleRowDown38_NEON,
-      ScaleRowDown38_C, 8 / 3, 1, 11)
-SDANY(ScaleRowDown38_3_Box_Any_NEON, ScaleRowDown38_3_Box_NEON,
-      ScaleRowDown38_3_Box_C, 8 / 3, 1, 11)
-SDANY(ScaleRowDown38_2_Box_Any_NEON, ScaleRowDown38_2_Box_NEON,
-      ScaleRowDown38_2_Box_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_Any_NEON,
+      ScaleRowDown38_NEON,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_NEON,
+      ScaleRowDown38_3_Box_NEON,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_2_Box_Any_NEON,
+      ScaleRowDown38_2_Box_NEON,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      11)
+#endif
+#ifdef HAS_SCALEROWDOWN38_MSA
+SDANY(ScaleRowDown38_Any_MSA,
+      ScaleRowDown38_MSA,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_MSA,
+      ScaleRowDown38_3_Box_MSA,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_2_Box_Any_MSA,
+      ScaleRowDown38_2_Box_MSA,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      11)
+#endif
+#ifdef HAS_SCALEROWDOWN38_LSX
+SDANY(ScaleRowDown38_Any_LSX,
+      ScaleRowDown38_LSX,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_LSX,
+      ScaleRowDown38_3_Box_LSX,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_2_Box_Any_LSX,
+      ScaleRowDown38_2_Box_LSX,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      11)
 #endif
 
 #ifdef HAS_SCALEARGBROWDOWN2_SSE2
-SDANY(ScaleARGBRowDown2_Any_SSE2, ScaleARGBRowDown2_SSE2,
-      ScaleARGBRowDown2_C, 2, 4, 3)
-SDANY(ScaleARGBRowDown2Linear_Any_SSE2, ScaleARGBRowDown2Linear_SSE2,
-      ScaleARGBRowDown2Linear_C, 2, 4, 3)
-SDANY(ScaleARGBRowDown2Box_Any_SSE2, ScaleARGBRowDown2Box_SSE2,
-      ScaleARGBRowDown2Box_C, 2, 4, 3)
+SDANY(ScaleARGBRowDown2_Any_SSE2,
+      ScaleARGBRowDown2_SSE2,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Linear_Any_SSE2,
+      ScaleARGBRowDown2Linear_SSE2,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Box_Any_SSE2,
+      ScaleARGBRowDown2Box_SSE2,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      3)
 #endif
 #ifdef HAS_SCALEARGBROWDOWN2_NEON
-SDANY(ScaleARGBRowDown2_Any_NEON, ScaleARGBRowDown2_NEON,
-      ScaleARGBRowDown2_C, 2, 4, 7)
-SDANY(ScaleARGBRowDown2Linear_Any_NEON, ScaleARGBRowDown2Linear_NEON,
-      ScaleARGBRowDown2Linear_C, 2, 4, 7)
-SDANY(ScaleARGBRowDown2Box_Any_NEON, ScaleARGBRowDown2Box_NEON,
-      ScaleARGBRowDown2Box_C, 2, 4, 7)
+SDANY(ScaleARGBRowDown2_Any_NEON,
+      ScaleARGBRowDown2_NEON,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      7)
+SDANY(ScaleARGBRowDown2Linear_Any_NEON,
+      ScaleARGBRowDown2Linear_NEON,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      7)
+SDANY(ScaleARGBRowDown2Box_Any_NEON,
+      ScaleARGBRowDown2Box_NEON,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      7)
+#endif
+#ifdef HAS_SCALEARGBROWDOWN2_MSA
+SDANY(ScaleARGBRowDown2_Any_MSA,
+      ScaleARGBRowDown2_MSA,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Linear_Any_MSA,
+      ScaleARGBRowDown2Linear_MSA,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Box_Any_MSA,
+      ScaleARGBRowDown2Box_MSA,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWN2_LSX
+SDANY(ScaleARGBRowDown2_Any_LSX,
+      ScaleARGBRowDown2_LSX,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Linear_Any_LSX,
+      ScaleARGBRowDown2Linear_LSX,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Box_Any_LSX,
+      ScaleARGBRowDown2Box_LSX,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      3)
 #endif
 #undef SDANY
 
 // Scale down by even scale factor.
-#define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK)          \
-    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, int src_stepx,    \
-                 uint8* dst_ptr, int dst_width) {                              \
-      int r = (int)((unsigned int)dst_width % (MASK + 1));                     \
-      int n = dst_width - r;                                                   \
-      if (n > 0) {                                                             \
-        SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n);         \
-      }                                                                        \
-      SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride,              \
-                     src_stepx, dst_ptr + n * BPP, r);                         \
-    }
+#define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK)       \
+  void NAMEANY(const uint8_t* src_ptr, ptrdiff_t src_stride, int src_stepx, \
+               uint8_t* dst_ptr, int dst_width) {                           \
+    int r = dst_width & MASK;                                               \
+    int n = dst_width & ~MASK;                                              \
+    if (n > 0) {                                                            \
+      SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n);        \
+    }                                                                       \
+    SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride, src_stepx,  \
+                   dst_ptr + n * BPP, r);                                   \
+  }
 
 #ifdef HAS_SCALEARGBROWDOWNEVEN_SSE2
-SDAANY(ScaleARGBRowDownEven_Any_SSE2, ScaleARGBRowDownEven_SSE2,
-       ScaleARGBRowDownEven_C, 4, 3)
-SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2, ScaleARGBRowDownEvenBox_SSE2,
-       ScaleARGBRowDownEvenBox_C, 4, 3)
+SDAANY(ScaleARGBRowDownEven_Any_SSE2,
+       ScaleARGBRowDownEven_SSE2,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2,
+       ScaleARGBRowDownEvenBox_SSE2,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
 #endif
 #ifdef HAS_SCALEARGBROWDOWNEVEN_NEON
-SDAANY(ScaleARGBRowDownEven_Any_NEON, ScaleARGBRowDownEven_NEON,
-       ScaleARGBRowDownEven_C, 4, 3)
-SDAANY(ScaleARGBRowDownEvenBox_Any_NEON, ScaleARGBRowDownEvenBox_NEON,
-       ScaleARGBRowDownEvenBox_C, 4, 3)
+SDAANY(ScaleARGBRowDownEven_Any_NEON,
+       ScaleARGBRowDownEven_NEON,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_NEON,
+       ScaleARGBRowDownEvenBox_NEON,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWNEVEN_MSA
+SDAANY(ScaleARGBRowDownEven_Any_MSA,
+       ScaleARGBRowDownEven_MSA,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_MSA,
+       ScaleARGBRowDownEvenBox_MSA,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWNEVEN_LSX
+SDAANY(ScaleARGBRowDownEven_Any_LSX,
+       ScaleARGBRowDownEven_LSX,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_LSX,
+       ScaleARGBRowDownEvenBox_LSX,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
+#endif
+#ifdef HAS_SCALEUVROWDOWNEVEN_NEON
+SDAANY(ScaleUVRowDownEven_Any_NEON,
+       ScaleUVRowDownEven_NEON,
+       ScaleUVRowDownEven_C,
+       2,
+       3)
+#endif
+
+#ifdef SASIMDONLY
+// This also works and uses memcpy and SIMD instead of C, but is slower on ARM
+
+// Add rows box filter scale down.  Using macro from row_any
+#define SAROW(NAMEANY, ANY_SIMD, SBPP, BPP, MASK)                      \
+  void NAMEANY(const uint8_t* src_ptr, uint16_t* dst_ptr, int width) { \
+    SIMD_ALIGNED(uint16_t dst_temp[32]);                               \
+    SIMD_ALIGNED(uint8_t src_temp[32]);                                \
+    memset(dst_temp, 0, 32 * 2); /* for msan */                        \
+    int r = width & MASK;                                              \
+    int n = width & ~MASK;                                             \
+    if (n > 0) {                                                       \
+      ANY_SIMD(src_ptr, dst_ptr, n);                                   \
+    }                                                                  \
+    memcpy(src_temp, src_ptr + n * SBPP, r * SBPP);                    \
+    memcpy(dst_temp, dst_ptr + n * BPP, r * BPP);                      \
+    ANY_SIMD(src_temp, dst_temp, MASK + 1);                            \
+    memcpy(dst_ptr + n * BPP, dst_temp, r * BPP);                      \
+  }
+
+#ifdef HAS_SCALEADDROW_SSE2
+SAROW(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, 1, 2, 15)
+#endif
+#ifdef HAS_SCALEADDROW_AVX2
+SAROW(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, 1, 2, 31)
+#endif
+#ifdef HAS_SCALEADDROW_NEON
+SAROW(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, 1, 2, 15)
+#endif
+#ifdef HAS_SCALEADDROW_MSA
+SAROW(ScaleAddRow_Any_MSA, ScaleAddRow_MSA, 1, 2, 15)
+#endif
+#ifdef HAS_SCALEADDROW_LSX
+SAROW(ScaleAddRow_Any_LSX, ScaleAddRow_LSX, 1, 2, 15)
 #endif
+#undef SAANY
+
+#else
 
 // Add rows box filter scale down.
-#define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK)                  \
-  void NAMEANY(const uint8* src_ptr, uint16* dst_ptr, int src_width) {         \
-      int n = src_width & ~MASK;                                               \
-      if (n > 0) {                                                             \
-        SCALEADDROW_SIMD(src_ptr, dst_ptr, n);                                 \
-      }                                                                        \
-      SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK);               \
-    }
+#define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK)              \
+  void NAMEANY(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) { \
+    int n = src_width & ~MASK;                                             \
+    if (n > 0) {                                                           \
+      SCALEADDROW_SIMD(src_ptr, dst_ptr, n);                               \
+    }                                                                      \
+    SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK);             \
+  }
 
 #ifdef HAS_SCALEADDROW_SSE2
 SAANY(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, ScaleAddRow_C, 15)
@@ -208,14 +582,481 @@ SAANY(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, ScaleAddRow_C, 31)
 #ifdef HAS_SCALEADDROW_NEON
 SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15)
 #endif
+#ifdef HAS_SCALEADDROW_MSA
+SAANY(ScaleAddRow_Any_MSA, ScaleAddRow_MSA, ScaleAddRow_C, 15)
+#endif
+#ifdef HAS_SCALEADDROW_LSX
+SAANY(ScaleAddRow_Any_LSX, ScaleAddRow_LSX, ScaleAddRow_C, 15)
+#endif
 #undef SAANY
 
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
+#endif  // SASIMDONLY
+
+// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
+#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                            \
+  void NAMEANY(uint8_t* dst_ptr, const uint8_t* src_ptr, int dst_width, int x, \
+               int dx) {                                                       \
+    int r = dst_width & MASK;                                                  \
+    int n = dst_width & ~MASK;                                                 \
+    if (n > 0) {                                                               \
+      TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                                   \
+    }                                                                          \
+    TERP_C(dst_ptr + n * BPP, src_ptr, r, x + n * dx, dx);                     \
+  }
+
+#ifdef HAS_SCALEFILTERCOLS_NEON
+CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
+#endif
+#ifdef HAS_SCALEFILTERCOLS_MSA
+CANY(ScaleFilterCols_Any_MSA, ScaleFilterCols_MSA, ScaleFilterCols_C, 1, 15)
+#endif
+#ifdef HAS_SCALEFILTERCOLS_LSX
+CANY(ScaleFilterCols_Any_LSX, ScaleFilterCols_LSX, ScaleFilterCols_C, 1, 15)
+#endif
+#ifdef HAS_SCALEARGBCOLS_NEON
+CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
+#endif
+#ifdef HAS_SCALEARGBCOLS_MSA
+CANY(ScaleARGBCols_Any_MSA, ScaleARGBCols_MSA, ScaleARGBCols_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBCOLS_LSX
+CANY(ScaleARGBCols_Any_LSX, ScaleARGBCols_LSX, ScaleARGBCols_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_NEON
+CANY(ScaleARGBFilterCols_Any_NEON,
+     ScaleARGBFilterCols_NEON,
+     ScaleARGBFilterCols_C,
+     4,
+     3)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_MSA
+CANY(ScaleARGBFilterCols_Any_MSA,
+     ScaleARGBFilterCols_MSA,
+     ScaleARGBFilterCols_C,
+     4,
+     7)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_LSX
+CANY(ScaleARGBFilterCols_Any_LSX,
+     ScaleARGBFilterCols_LSX,
+     ScaleARGBFilterCols_C,
+     4,
+     7)
+#endif
+#undef CANY
+
+// Scale up horizontally 2 times using linear filter.
+#define SUH2LANY(NAME, SIMD, C, MASK, PTYPE)                       \
+  void NAME(const PTYPE* src_ptr, PTYPE* dst_ptr, int dst_width) { \
+    int work_width = (dst_width - 1) & ~1;                         \
+    int r = work_width & MASK;                                     \
+    int n = work_width & ~MASK;                                    \
+    dst_ptr[0] = src_ptr[0];                                       \
+    if (work_width > 0) {                                          \
+      if (n != 0) {                                                \
+        SIMD(src_ptr, dst_ptr + 1, n);                             \
+      }                                                            \
+      C(src_ptr + (n / 2), dst_ptr + n + 1, r);                    \
+    }                                                              \
+    dst_ptr[dst_width - 1] = src_ptr[(dst_width - 1) / 2];         \
+  }
+
+// Even the C versions need to be wrapped, because boundary pixels have to
+// be handled differently
+
+SUH2LANY(ScaleRowUp2_Linear_Any_C,
+         ScaleRowUp2_Linear_C,
+         ScaleRowUp2_Linear_C,
+         0,
+         uint8_t)
+
+SUH2LANY(ScaleRowUp2_Linear_16_Any_C,
+         ScaleRowUp2_Linear_16_C,
+         ScaleRowUp2_Linear_16_C,
+         0,
+         uint16_t)
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSE2
+SUH2LANY(ScaleRowUp2_Linear_Any_SSE2,
+         ScaleRowUp2_Linear_SSE2,
+         ScaleRowUp2_Linear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSSE3
+SUH2LANY(ScaleRowUp2_Linear_Any_SSSE3,
+         ScaleRowUp2_Linear_SSSE3,
+         ScaleRowUp2_Linear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_SSSE3
+SUH2LANY(ScaleRowUp2_Linear_12_Any_SSSE3,
+         ScaleRowUp2_Linear_12_SSSE3,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_SSE2
+SUH2LANY(ScaleRowUp2_Linear_16_Any_SSE2,
+         ScaleRowUp2_Linear_16_SSE2,
+         ScaleRowUp2_Linear_16_C,
+         7,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_AVX2
+SUH2LANY(ScaleRowUp2_Linear_Any_AVX2,
+         ScaleRowUp2_Linear_AVX2,
+         ScaleRowUp2_Linear_C,
+         31,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_AVX2
+SUH2LANY(ScaleRowUp2_Linear_12_Any_AVX2,
+         ScaleRowUp2_Linear_12_AVX2,
+         ScaleRowUp2_Linear_16_C,
+         31,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_AVX2
+SUH2LANY(ScaleRowUp2_Linear_16_Any_AVX2,
+         ScaleRowUp2_Linear_16_AVX2,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_NEON
+SUH2LANY(ScaleRowUp2_Linear_Any_NEON,
+         ScaleRowUp2_Linear_NEON,
+         ScaleRowUp2_Linear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_NEON
+SUH2LANY(ScaleRowUp2_Linear_12_Any_NEON,
+         ScaleRowUp2_Linear_12_NEON,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_NEON
+SUH2LANY(ScaleRowUp2_Linear_16_Any_NEON,
+         ScaleRowUp2_Linear_16_NEON,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#undef SUH2LANY
+
+// Scale up 2 times using bilinear filter.
+// This function produces 2 rows at a time.
+#define SU2BLANY(NAME, SIMD, C, MASK, PTYPE)                              \
+  void NAME(const PTYPE* src_ptr, ptrdiff_t src_stride, PTYPE* dst_ptr,   \
+            ptrdiff_t dst_stride, int dst_width) {                        \
+    int work_width = (dst_width - 1) & ~1;                                \
+    int r = work_width & MASK;                                            \
+    int n = work_width & ~MASK;                                           \
+    const PTYPE* sa = src_ptr;                                            \
+    const PTYPE* sb = src_ptr + src_stride;                               \
+    PTYPE* da = dst_ptr;                                                  \
+    PTYPE* db = dst_ptr + dst_stride;                                     \
+    da[0] = (3 * sa[0] + sb[0] + 2) >> 2;                                 \
+    db[0] = (sa[0] + 3 * sb[0] + 2) >> 2;                                 \
+    if (work_width > 0) {                                                 \
+      if (n != 0) {                                                       \
+        SIMD(sa, sb - sa, da + 1, db - da, n);                            \
+      }                                                                   \
+      C(sa + (n / 2), sb - sa, da + n + 1, db - da, r);                   \
+    }                                                                     \
+    da[dst_width - 1] =                                                   \
+        (3 * sa[(dst_width - 1) / 2] + sb[(dst_width - 1) / 2] + 2) >> 2; \
+    db[dst_width - 1] =                                                   \
+        (sa[(dst_width - 1) / 2] + 3 * sb[(dst_width - 1) / 2] + 2) >> 2; \
+  }
+
+SU2BLANY(ScaleRowUp2_Bilinear_Any_C,
+         ScaleRowUp2_Bilinear_C,
+         ScaleRowUp2_Bilinear_C,
+         0,
+         uint8_t)
+
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_C,
+         ScaleRowUp2_Bilinear_16_C,
+         ScaleRowUp2_Bilinear_16_C,
+         0,
+         uint16_t)
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSE2
+SU2BLANY(ScaleRowUp2_Bilinear_Any_SSE2,
+         ScaleRowUp2_Bilinear_SSE2,
+         ScaleRowUp2_Bilinear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+SU2BLANY(ScaleRowUp2_Bilinear_12_Any_SSSE3,
+         ScaleRowUp2_Bilinear_12_SSSE3,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_SSE2
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_SSE2,
+         ScaleRowUp2_Bilinear_16_SSE2,
+         ScaleRowUp2_Bilinear_16_C,
+         7,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSSE3
+SU2BLANY(ScaleRowUp2_Bilinear_Any_SSSE3,
+         ScaleRowUp2_Bilinear_SSSE3,
+         ScaleRowUp2_Bilinear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_AVX2
+SU2BLANY(ScaleRowUp2_Bilinear_Any_AVX2,
+         ScaleRowUp2_Bilinear_AVX2,
+         ScaleRowUp2_Bilinear_C,
+         31,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_AVX2
+SU2BLANY(ScaleRowUp2_Bilinear_12_Any_AVX2,
+         ScaleRowUp2_Bilinear_12_AVX2,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_AVX2
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_AVX2,
+         ScaleRowUp2_Bilinear_16_AVX2,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_NEON
+SU2BLANY(ScaleRowUp2_Bilinear_Any_NEON,
+         ScaleRowUp2_Bilinear_NEON,
+         ScaleRowUp2_Bilinear_C,
+         15,
+         uint8_t)
 #endif
 
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_NEON
+SU2BLANY(ScaleRowUp2_Bilinear_12_Any_NEON,
+         ScaleRowUp2_Bilinear_12_NEON,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_NEON
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_NEON,
+         ScaleRowUp2_Bilinear_16_NEON,
+         ScaleRowUp2_Bilinear_16_C,
+         7,
+         uint16_t)
+#endif
+
+#undef SU2BLANY
+
+// Scale bi-planar plane up horizontally 2 times using linear filter.
+#define SBUH2LANY(NAME, SIMD, C, MASK, PTYPE)                         \
+  void NAME(const PTYPE* src_ptr, PTYPE* dst_ptr, int dst_width) {    \
+    int work_width = (dst_width - 1) & ~1;                            \
+    int r = work_width & MASK;                                        \
+    int n = work_width & ~MASK;                                       \
+    dst_ptr[0] = src_ptr[0];                                          \
+    dst_ptr[1] = src_ptr[1];                                          \
+    if (work_width > 0) {                                             \
+      if (n != 0) {                                                   \
+        SIMD(src_ptr, dst_ptr + 2, n);                                \
+      }                                                               \
+      C(src_ptr + n, dst_ptr + 2 * n + 2, r);                         \
+    }                                                                 \
+    dst_ptr[2 * dst_width - 2] = src_ptr[((dst_width + 1) & ~1) - 2]; \
+    dst_ptr[2 * dst_width - 1] = src_ptr[((dst_width + 1) & ~1) - 1]; \
+  }
+
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_C,
+          ScaleUVRowUp2_Linear_C,
+          ScaleUVRowUp2_Linear_C,
+          0,
+          uint8_t)
+
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_C,
+          ScaleUVRowUp2_Linear_16_C,
+          ScaleUVRowUp2_Linear_16_C,
+          0,
+          uint16_t)
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_SSSE3
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_SSSE3,
+          ScaleUVRowUp2_Linear_SSSE3,
+          ScaleUVRowUp2_Linear_C,
+          7,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_AVX2
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_AVX2,
+          ScaleUVRowUp2_Linear_AVX2,
+          ScaleUVRowUp2_Linear_C,
+          15,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_SSE41,
+          ScaleUVRowUp2_Linear_16_SSE41,
+          ScaleUVRowUp2_Linear_16_C,
+          3,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_AVX2,
+          ScaleUVRowUp2_Linear_16_AVX2,
+          ScaleUVRowUp2_Linear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_NEON
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_NEON,
+          ScaleUVRowUp2_Linear_NEON,
+          ScaleUVRowUp2_Linear_C,
+          15,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_NEON
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_NEON,
+          ScaleUVRowUp2_Linear_16_NEON,
+          ScaleUVRowUp2_Linear_16_C,
+          15,
+          uint16_t)
+#endif
 
+#undef SBUH2LANY
 
+// Scale bi-planar plane up 2 times using bilinear filter.
+// This function produces 2 rows at a time.
+#define SBU2BLANY(NAME, SIMD, C, MASK, PTYPE)                           \
+  void NAME(const PTYPE* src_ptr, ptrdiff_t src_stride, PTYPE* dst_ptr, \
+            ptrdiff_t dst_stride, int dst_width) {                      \
+    int work_width = (dst_width - 1) & ~1;                              \
+    int r = work_width & MASK;                                          \
+    int n = work_width & ~MASK;                                         \
+    const PTYPE* sa = src_ptr;                                          \
+    const PTYPE* sb = src_ptr + src_stride;                             \
+    PTYPE* da = dst_ptr;                                                \
+    PTYPE* db = dst_ptr + dst_stride;                                   \
+    da[0] = (3 * sa[0] + sb[0] + 2) >> 2;                               \
+    db[0] = (sa[0] + 3 * sb[0] + 2) >> 2;                               \
+    da[1] = (3 * sa[1] + sb[1] + 2) >> 2;                               \
+    db[1] = (sa[1] + 3 * sb[1] + 2) >> 2;                               \
+    if (work_width > 0) {                                               \
+      if (n != 0) {                                                     \
+        SIMD(sa, sb - sa, da + 2, db - da, n);                          \
+      }                                                                 \
+      C(sa + n, sb - sa, da + 2 * n + 2, db - da, r);                   \
+    }                                                                   \
+    da[2 * dst_width - 2] = (3 * sa[((dst_width + 1) & ~1) - 2] +       \
+                             sb[((dst_width + 1) & ~1) - 2] + 2) >>     \
+                            2;                                          \
+    db[2 * dst_width - 2] = (sa[((dst_width + 1) & ~1) - 2] +           \
+                             3 * sb[((dst_width + 1) & ~1) - 2] + 2) >> \
+                            2;                                          \
+    da[2 * dst_width - 1] = (3 * sa[((dst_width + 1) & ~1) - 1] +       \
+                             sb[((dst_width + 1) & ~1) - 1] + 2) >>     \
+                            2;                                          \
+    db[2 * dst_width - 1] = (sa[((dst_width + 1) & ~1) - 1] +           \
+                             3 * sb[((dst_width + 1) & ~1) - 1] + 2) >> \
+                            2;                                          \
+  }
 
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_C,
+          ScaleUVRowUp2_Bilinear_C,
+          ScaleUVRowUp2_Bilinear_C,
+          0,
+          uint8_t)
 
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_C,
+          ScaleUVRowUp2_Bilinear_16_C,
+          ScaleUVRowUp2_Bilinear_16_C,
+          0,
+          uint16_t)
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_SSSE3
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_SSSE3,
+          ScaleUVRowUp2_Bilinear_SSSE3,
+          ScaleUVRowUp2_Bilinear_C,
+          7,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_AVX2
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_AVX2,
+          ScaleUVRowUp2_Bilinear_AVX2,
+          ScaleUVRowUp2_Bilinear_C,
+          15,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_SSE41,
+          ScaleUVRowUp2_Bilinear_16_SSE41,
+          ScaleUVRowUp2_Bilinear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_AVX2,
+          ScaleUVRowUp2_Bilinear_16_AVX2,
+          ScaleUVRowUp2_Bilinear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_NEON
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_NEON,
+          ScaleUVRowUp2_Bilinear_NEON,
+          ScaleUVRowUp2_Bilinear_C,
+          7,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_NEON
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_NEON,
+          ScaleUVRowUp2_Bilinear_16_NEON,
+          ScaleUVRowUp2_Bilinear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#undef SBU2BLANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/scale_argb.cc b/media/libyuv/source/scale_argb.cc
index 17f51ae9bf..3e6f54776a 100644
--- a/media/libyuv/source/scale_argb.cc
+++ b/media/libyuv/source/scale_argb.cc
@@ -30,49 +30,100 @@ static __inline int Abs(int v) {
 // ScaleARGB ARGB, 1/2
 // This is an optimized version for scaling down a ARGB to 1/2 of
 // its original size.
-static void ScaleARGBDown2(int src_width, int src_height,
-                           int dst_width, int dst_height,
-                           int src_stride, int dst_stride,
-                           const uint8* src_argb, uint8* dst_argb,
-                           int x, int dx, int y, int dy,
+static void ScaleARGBDown2(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int x,
+                           int dx,
+                           int y,
+                           int dy,
                            enum FilterMode filtering) {
   int j;
   int row_stride = src_stride * (dy >> 16);
-  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
-                            uint8* dst_argb, int dst_width) =
-    filtering == kFilterNone ? ScaleARGBRowDown2_C :
-        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
-        ScaleARGBRowDown2Box_C);
-  assert(dx == 65536 * 2);  // Test scale factor of 2.
+  void (*ScaleARGBRowDown2)(const uint8_t* src_argb, ptrdiff_t src_stride,
+                            uint8_t* dst_argb, int dst_width) =
+      filtering == kFilterNone
+          ? ScaleARGBRowDown2_C
+          : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C
+                                        : ScaleARGBRowDown2Box_C);
+  (void)src_width;
+  (void)src_height;
+  (void)dx;
+  assert(dx == 65536 * 2);      // Test scale factor of 2.
   assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
   // Advance to odd row, even column.
   if (filtering == kFilterBilinear) {
-    src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+    src_argb += (y >> 16) * (intptr_t)src_stride + (x >> 16) * 4;
   } else {
-    src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
+    src_argb += (y >> 16) * (intptr_t)src_stride + ((x >> 16) - 1) * 4;
   }
 
 #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_SSE2 :
-        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_SSE2 :
-        ScaleARGBRowDown2Box_Any_SSE2);
+    ScaleARGBRowDown2 =
+        filtering == kFilterNone
+            ? ScaleARGBRowDown2_Any_SSE2
+            : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_SSE2
+                                          : ScaleARGBRowDown2Box_Any_SSE2);
     if (IS_ALIGNED(dst_width, 4)) {
-      ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
-          (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
-          ScaleARGBRowDown2Box_SSE2);
+      ScaleARGBRowDown2 =
+          filtering == kFilterNone
+              ? ScaleARGBRowDown2_SSE2
+              : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2
+                                            : ScaleARGBRowDown2Box_SSE2);
     }
   }
 #endif
 #if defined(HAS_SCALEARGBROWDOWN2_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_NEON :
-        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_NEON :
-        ScaleARGBRowDown2Box_Any_NEON);
+    ScaleARGBRowDown2 =
+        filtering == kFilterNone
+            ? ScaleARGBRowDown2_Any_NEON
+            : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_NEON
+                                          : ScaleARGBRowDown2Box_Any_NEON);
     if (IS_ALIGNED(dst_width, 8)) {
-      ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_NEON :
-          (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_NEON :
-          ScaleARGBRowDown2Box_NEON);
+      ScaleARGBRowDown2 =
+          filtering == kFilterNone
+              ? ScaleARGBRowDown2_NEON
+              : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_NEON
+                                            : ScaleARGBRowDown2Box_NEON);
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBRowDown2 =
+        filtering == kFilterNone
+            ? ScaleARGBRowDown2_Any_MSA
+            : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_MSA
+                                          : ScaleARGBRowDown2Box_Any_MSA);
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDown2 =
+          filtering == kFilterNone
+              ? ScaleARGBRowDown2_MSA
+              : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_MSA
+                                            : ScaleARGBRowDown2Box_MSA);
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBRowDown2 =
+        filtering == kFilterNone
+            ? ScaleARGBRowDown2_Any_LSX
+            : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_LSX
+                                          : ScaleARGBRowDown2Box_Any_LSX);
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDown2 =
+          filtering == kFilterNone
+              ? ScaleARGBRowDown2_LSX
+              : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_LSX
+                                            : ScaleARGBRowDown2Box_LSX);
     }
   }
 #endif
@@ -90,21 +141,32 @@ static void ScaleARGBDown2(int src_width, int src_height,
 // ScaleARGB ARGB, 1/4
 // This is an optimized version for scaling down a ARGB to 1/4 of
 // its original size.
-static void ScaleARGBDown4Box(int src_width, int src_height,
-                              int dst_width, int dst_height,
-                              int src_stride, int dst_stride,
-                              const uint8* src_argb, uint8* dst_argb,
-                              int x, int dx, int y, int dy) {
+static void ScaleARGBDown4Box(int src_width,
+                              int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint8_t* src_argb,
+                              uint8_t* dst_argb,
+                              int x,
+                              int dx,
+                              int y,
+                              int dy) {
   int j;
   // Allocate 2 rows of ARGB.
-  const int kRowSize = (dst_width * 2 * 4 + 31) & ~31;
-  align_buffer_64(row, kRowSize * 2);
+  const int row_size = (dst_width * 2 * 4 + 31) & ~31;
+  align_buffer_64(row, row_size * 2);
   int row_stride = src_stride * (dy >> 16);
-  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
-    uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C;
+  void (*ScaleARGBRowDown2)(const uint8_t* src_argb, ptrdiff_t src_stride,
+                            uint8_t* dst_argb, int dst_width) =
+      ScaleARGBRowDown2Box_C;
   // Advance to odd row, even column.
-  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
-  assert(dx == 65536 * 4);  // Test scale factor of 4.
+  src_argb += (y >> 16) * (intptr_t)src_stride + (x >> 16) * 4;
+  (void)src_width;
+  (void)src_height;
+  (void)dx;
+  assert(dx == 65536 * 4);      // Test scale factor of 4.
   assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
 #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
@@ -125,9 +187,9 @@ static void ScaleARGBDown4Box(int src_width, int src_height,
 
   for (j = 0; j < dst_height; ++j) {
     ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2);
-    ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride,
-                      row + kRowSize, dst_width * 2);
-    ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width);
+    ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride, row + row_size,
+                      dst_width * 2);
+    ScaleARGBRowDown2(row, row_size, dst_argb, dst_width);
     src_argb += row_stride;
     dst_argb += dst_stride;
   }
@@ -137,38 +199,67 @@ static void ScaleARGBDown4Box(int src_width, int src_height,
 // ScaleARGB ARGB Even
 // This is an optimized version for scaling down a ARGB to even
 // multiple of its original size.
-static void ScaleARGBDownEven(int src_width, int src_height,
-                              int dst_width, int dst_height,
-                              int src_stride, int dst_stride,
-                              const uint8* src_argb, uint8* dst_argb,
-                              int x, int dx, int y, int dy,
+static void ScaleARGBDownEven(int src_width,
+                              int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint8_t* src_argb,
+                              uint8_t* dst_argb,
+                              int x,
+                              int dx,
+                              int y,
+                              int dy,
                               enum FilterMode filtering) {
   int j;
   int col_step = dx >> 16;
-  int row_stride = (dy >> 16) * src_stride;
-  void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride,
-                               int src_step, uint8* dst_argb, int dst_width) =
+  ptrdiff_t row_stride = (ptrdiff_t)((dy >> 16) * (intptr_t)src_stride);
+  void (*ScaleARGBRowDownEven)(const uint8_t* src_argb, ptrdiff_t src_stride,
+                               int src_step, uint8_t* dst_argb, int dst_width) =
       filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
+  (void)src_width;
+  (void)src_height;
   assert(IS_ALIGNED(src_width, 2));
   assert(IS_ALIGNED(src_height, 2));
-  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  src_argb += (y >> 16) * (intptr_t)src_stride + (x >> 16) * 4;
 #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_SSE2 :
-        ScaleARGBRowDownEven_Any_SSE2;
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_SSE2
+                                     : ScaleARGBRowDownEven_Any_SSE2;
     if (IS_ALIGNED(dst_width, 4)) {
-      ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
-          ScaleARGBRowDownEven_SSE2;
+      ScaleARGBRowDownEven =
+          filtering ? ScaleARGBRowDownEvenBox_SSE2 : ScaleARGBRowDownEven_SSE2;
     }
   }
 #endif
 #if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_NEON :
-        ScaleARGBRowDownEven_Any_NEON;
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_NEON
+                                     : ScaleARGBRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDownEven =
+          filtering ? ScaleARGBRowDownEvenBox_NEON : ScaleARGBRowDownEven_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_MSA
+                                     : ScaleARGBRowDownEven_Any_MSA;
     if (IS_ALIGNED(dst_width, 4)) {
-      ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
-          ScaleARGBRowDownEven_NEON;
+      ScaleARGBRowDownEven =
+          filtering ? ScaleARGBRowDownEvenBox_MSA : ScaleARGBRowDownEven_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_LSX
+                                     : ScaleARGBRowDownEven_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDownEven =
+          filtering ? ScaleARGBRowDownEvenBox_LSX : ScaleARGBRowDownEven_LSX;
     }
   }
 #endif
@@ -184,25 +275,32 @@ static void ScaleARGBDownEven(int src_width, int src_height,
 }
 
 // Scale ARGB down with bilinear interpolation.
-static void ScaleARGBBilinearDown(int src_width, int src_height,
-                                  int dst_width, int dst_height,
-                                  int src_stride, int dst_stride,
-                                  const uint8* src_argb, uint8* dst_argb,
-                                  int x, int dx, int y, int dy,
+static void ScaleARGBBilinearDown(int src_width,
+                                  int src_height,
+                                  int dst_width,
+                                  int dst_height,
+                                  int src_stride,
+                                  int dst_stride,
+                                  const uint8_t* src_argb,
+                                  uint8_t* dst_argb,
+                                  int x,
+                                  int dx,
+                                  int y,
+                                  int dy,
                                   enum FilterMode filtering) {
   int j;
-  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_C;
-  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
-      int dst_width, int x, int dx) =
+  void (*InterpolateRow)(uint8_t* dst_argb, const uint8_t* src_argb,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8_t* dst_argb, const uint8_t* src_argb,
+                              int dst_width, int x, int dx) =
       (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C;
-  int64 xlast = x + (int64)(dst_width - 1) * dx;
-  int64 xl = (dx >= 0) ? x : xlast;
-  int64 xr = (dx >= 0) ? xlast : x;
+  int64_t xlast = x + (int64_t)(dst_width - 1) * dx;
+  int64_t xl = (dx >= 0) ? x : xlast;
+  int64_t xr = (dx >= 0) ? xlast : x;
   int clip_src_width;
-  xl = (xl >> 16) & ~3;  // Left edge aligned.
-  xr = (xr >> 16) + 1;  // Right most pixel used.  Bilinear uses 2 pixels.
+  xl = (xl >> 16) & ~3;    // Left edge aligned.
+  xr = (xr >> 16) + 1;     // Right most pixel used.  Bilinear uses 2 pixels.
   xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
   if (xr > src_width) {
     xr = src_width;
@@ -234,12 +332,19 @@ static void ScaleARGBBilinearDown(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
-    InterpolateRow = InterpolateRow_Any_DSPR2;
-    if (IS_ALIGNED(clip_src_width, 4)) {
-      InterpolateRow = InterpolateRow_DSPR2;
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -256,6 +361,22 @@ static void ScaleARGBBilinearDown(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_LSX;
+    }
+  }
+#endif
   // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
   // Allocate a row of ARGB.
   {
@@ -267,7 +388,7 @@ static void ScaleARGBBilinearDown(int src_width, int src_height,
     }
     for (j = 0; j < dst_height; ++j) {
       int yi = y >> 16;
-      const uint8* src = src_argb + yi * src_stride;
+      const uint8_t* src = src_argb + yi * (intptr_t)src_stride;
       if (filtering == kFilterLinear) {
         ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
       } else {
@@ -286,18 +407,25 @@ static void ScaleARGBBilinearDown(int src_width, int src_height,
 }
 
 // Scale ARGB up with bilinear interpolation.
-static void ScaleARGBBilinearUp(int src_width, int src_height,
-                                int dst_width, int dst_height,
-                                int src_stride, int dst_stride,
-                                const uint8* src_argb, uint8* dst_argb,
-                                int x, int dx, int y, int dy,
+static void ScaleARGBBilinearUp(int src_width,
+                                int src_height,
+                                int dst_width,
+                                int dst_height,
+                                int src_stride,
+                                int dst_stride,
+                                const uint8_t* src_argb,
+                                uint8_t* dst_argb,
+                                int x,
+                                int dx,
+                                int y,
+                                int dy,
                                 enum FilterMode filtering) {
   int j;
-  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_C;
-  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
-      int dst_width, int x, int dx) =
+  void (*InterpolateRow)(uint8_t* dst_argb, const uint8_t* src_argb,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8_t* dst_argb, const uint8_t* src_argb,
+                              int dst_width, int x, int dx) =
       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
   const int max_y = (src_height - 1) << 16;
 #if defined(HAS_INTERPOLATEROW_SSSE3)
@@ -324,15 +452,25 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
-    InterpolateRow = InterpolateRow_DSPR2;
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
   }
 #endif
   if (src_width >= 32768) {
-    ScaleARGBFilterCols = filtering ?
-        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+    ScaleARGBFilterCols =
+        filtering ? ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
   }
 #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
   if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
@@ -347,6 +485,22 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_LSX)
+  if (filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_LSX;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
@@ -360,6 +514,22 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBCOLS_MSA)
+  if (!filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_LSX)
+  if (!filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_LSX;
+    }
+  }
+#endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
 #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
@@ -375,14 +545,14 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
 
   {
     int yi = y >> 16;
-    const uint8* src = src_argb + yi * src_stride;
+    const uint8_t* src = src_argb + yi * (intptr_t)src_stride;
 
     // Allocate 2 rows of ARGB.
-    const int kRowSize = (dst_width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
+    const int row_size = (dst_width * 4 + 31) & ~31;
+    align_buffer_64(row, row_size * 2);
 
-    uint8* rowptr = row;
-    int rowstride = kRowSize;
+    uint8_t* rowptr = row;
+    int rowstride = row_size;
     int lasty = yi;
 
     ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
@@ -390,7 +560,9 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
       src += src_stride;
     }
     ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
-    src += src_stride;
+    if (src_height > 2) {
+      src += src_stride;
+    }
 
     for (j = 0; j < dst_height; ++j) {
       yi = y >> 16;
@@ -398,14 +570,16 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
         if (y > max_y) {
           y = max_y;
           yi = y >> 16;
-          src = src_argb + yi * src_stride;
+          src = src_argb + yi * (intptr_t)src_stride;
         }
         if (yi != lasty) {
           ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
           rowptr += rowstride;
           rowstride = -rowstride;
           lasty = yi;
-          src += src_stride;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
         }
       }
       if (filtering == kFilterLinear) {
@@ -423,24 +597,27 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
 
 #ifdef YUVSCALEUP
 // Scale YUV to ARGB up with bilinear interpolation.
-static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
-                                     int dst_width, int dst_height,
+static void ScaleYUVToARGBBilinearUp(int src_width,
+                                     int src_height,
+                                     int dst_width,
+                                     int dst_height,
                                      int src_stride_y,
                                      int src_stride_u,
                                      int src_stride_v,
                                      int dst_stride_argb,
-                                     const uint8* src_y,
-                                     const uint8* src_u,
-                                     const uint8* src_v,
-                                     uint8* dst_argb,
-                                     int x, int dx, int y, int dy,
+                                     const uint8_t* src_y,
+                                     const uint8_t* src_u,
+                                     const uint8_t* src_v,
+                                     uint8_t* dst_argb,
+                                     int x,
+                                     int dx,
+                                     int y,
+                                     int dy,
                                      enum FilterMode filtering) {
   int j;
-  void (*I422ToARGBRow)(const uint8* y_buf,
-                        const uint8* u_buf,
-                        const uint8* v_buf,
-                        uint8* rgb_buf,
-                        int width) = I422ToARGBRow_C;
+  void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf, int width) =
+      I422ToARGBRow_C;
 #if defined(HAS_I422TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
@@ -457,6 +634,15 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(src_width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
@@ -465,19 +651,26 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_I422TOARGBROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(src_width, 4) &&
-      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
-      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
-      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
-    I422ToARGBRow = I422ToARGBRow_DSPR2;
+#if defined(HAS_I422TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGBRow = I422ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(src_width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
   }
 #endif
 
-  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_C;
+  void (*InterpolateRow)(uint8_t* dst_argb, const uint8_t* src_argb,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
 #if defined(HAS_INTERPOLATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
@@ -502,19 +695,29 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
-    InterpolateRow = InterpolateRow_DSPR2;
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
   }
 #endif
 
-  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
-      int dst_width, int x, int dx) =
+  void (*ScaleARGBFilterCols)(uint8_t* dst_argb, const uint8_t* src_argb,
+                              int dst_width, int x, int dx) =
       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
   if (src_width >= 32768) {
-    ScaleARGBFilterCols = filtering ?
-        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+    ScaleARGBFilterCols =
+        filtering ? ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
   }
 #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
   if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
@@ -529,6 +732,22 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_LSX)
+  if (filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_LSX;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
@@ -542,6 +761,22 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBCOLS_MSA)
+  if (!filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_LSX)
+  if (!filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_LSX;
+    }
+  }
+#endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
 #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
@@ -558,19 +793,19 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
   const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
   int yi = y >> 16;
   int uv_yi = yi >> kYShift;
-  const uint8* src_row_y = src_y + yi * src_stride_y;
-  const uint8* src_row_u = src_u + uv_yi * src_stride_u;
-  const uint8* src_row_v = src_v + uv_yi * src_stride_v;
+  const uint8_t* src_row_y = src_y + yi * (intptr_t)src_stride_y;
+  const uint8_t* src_row_u = src_u + uv_yi * (intptr_t)src_stride_u;
+  const uint8_t* src_row_v = src_v + uv_yi * (intptr_t)src_stride_v;
 
   // Allocate 2 rows of ARGB.
-  const int kRowSize = (dst_width * 4 + 31) & ~31;
-  align_buffer_64(row, kRowSize * 2);
+  const int row_size = (dst_width * 4 + 31) & ~31;
+  align_buffer_64(row, row_size * 2);
 
   // Allocate 1 row of ARGB for source conversion.
   align_buffer_64(argb_row, src_width * 4);
 
-  uint8* rowptr = row;
-  int rowstride = kRowSize;
+  uint8_t* rowptr = row;
+  int rowstride = row_size;
   int lasty = yi;
 
   // TODO(fbarchard): Convert first 2 rows of YUV to ARGB.
@@ -598,9 +833,9 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
         y = max_y;
         yi = y >> 16;
         uv_yi = yi >> kYShift;
-        src_row_y = src_y + yi * src_stride_y;
-        src_row_u = src_u + uv_yi * src_stride_u;
-        src_row_v = src_v + uv_yi * src_stride_v;
+        src_row_y = src_y + yi * (intptr_t)src_stride_y;
+        src_row_u = src_u + uv_yi * (intptr_t)src_stride_u;
+        src_row_v = src_v + uv_yi * (intptr_t)src_stride_v;
       }
       if (yi != lasty) {
         // TODO(fbarchard): Convert the clipped region of row.
@@ -635,15 +870,23 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
 // of x and dx is the integer part of the source position and
 // the lower 16 bits are the fixed decimal part.
 
-static void ScaleARGBSimple(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_argb, uint8* dst_argb,
-                            int x, int dx, int y, int dy) {
+static void ScaleARGBSimple(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            int x,
+                            int dx,
+                            int y,
+                            int dy) {
   int j;
-  void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
-      int dst_width, int x, int dx) =
+  void (*ScaleARGBCols)(uint8_t* dst_argb, const uint8_t* src_argb,
+                        int dst_width, int x, int dx) =
       (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
+  (void)src_height;
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBCols = ScaleARGBCols_SSE2;
@@ -657,6 +900,22 @@ static void ScaleARGBSimple(int src_width, int src_height,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBCols = ScaleARGBCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBCols = ScaleARGBCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBCols = ScaleARGBCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBCols = ScaleARGBCols_LSX;
+    }
+  }
+#endif
   if (src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBCols = ScaleARGBColsUp2_C;
 #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
@@ -667,7 +926,7 @@ static void ScaleARGBSimple(int src_width, int src_height,
   }
 
   for (j = 0; j < dst_height; ++j) {
-    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride,
+    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * (intptr_t)src_stride,
                   dst_width, x, dx);
     dst_argb += dst_stride;
     y += dy;
@@ -677,11 +936,18 @@ static void ScaleARGBSimple(int src_width, int src_height,
 // ScaleARGB a ARGB.
 // This function in turn calls a scaling function
 // suitable for handling the desired resolutions.
-static void ScaleARGB(const uint8* src, int src_stride,
-                      int src_width, int src_height,
-                      uint8* dst, int dst_stride,
-                      int dst_width, int dst_height,
-                      int clip_x, int clip_y, int clip_width, int clip_height,
+static void ScaleARGB(const uint8_t* src,
+                      int src_stride,
+                      int src_width,
+                      int src_height,
+                      uint8_t* dst,
+                      int dst_stride,
+                      int dst_width,
+                      int dst_height,
+                      int clip_x,
+                      int clip_y,
+                      int clip_width,
+                      int clip_height,
                       enum FilterMode filtering) {
   // Initial source x/y coordinate and step values as 16.16 fixed point.
   int x = 0;
@@ -690,29 +956,28 @@ static void ScaleARGB(const uint8* src, int src_stride,
   int dy = 0;
   // ARGB does not support box filter yet, but allow the user to pass it.
   // Simplify filtering when possible.
-  filtering = ScaleFilterReduce(src_width, src_height,
-                                dst_width, dst_height,
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
                                 filtering);
 
   // Negative src_height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;
-    src = src + (src_height - 1) * src_stride;
+    src = src + (src_height - 1) * (intptr_t)src_stride;
     src_stride = -src_stride;
   }
-  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
-             &x, &y, &dx, &dy);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
   src_width = Abs(src_width);
   if (clip_x) {
-    int64 clipf = (int64)(clip_x) * dx;
+    int64_t clipf = (int64_t)(clip_x)*dx;
     x += (clipf & 0xffff);
     src += (clipf >> 16) * 4;
     dst += clip_x * 4;
   }
   if (clip_y) {
-    int64 clipf = (int64)(clip_y) * dy;
+    int64_t clipf = (int64_t)(clip_y)*dy;
     y += (clipf & 0xffff);
-    src += (clipf >> 16) * src_stride;
+    src += (clipf >> 16) * (intptr_t)src_stride;
     dst += clip_y * dst_stride;
   }
 
@@ -725,24 +990,20 @@ static void ScaleARGB(const uint8* src, int src_stride,
       if (!(dx & 0x10000) && !(dy & 0x10000)) {
         if (dx == 0x20000) {
           // Optimized 1/2 downsample.
-          ScaleARGBDown2(src_width, src_height,
-                         clip_width, clip_height,
-                         src_stride, dst_stride, src, dst,
-                         x, dx, y, dy, filtering);
+          ScaleARGBDown2(src_width, src_height, clip_width, clip_height,
+                         src_stride, dst_stride, src, dst, x, dx, y, dy,
+                         filtering);
           return;
         }
         if (dx == 0x40000 && filtering == kFilterBox) {
           // Optimized 1/4 box downsample.
-          ScaleARGBDown4Box(src_width, src_height,
-                            clip_width, clip_height,
-                            src_stride, dst_stride, src, dst,
-                            x, dx, y, dy);
+          ScaleARGBDown4Box(src_width, src_height, clip_width, clip_height,
+                            src_stride, dst_stride, src, dst, x, dx, y, dy);
           return;
         }
-        ScaleARGBDownEven(src_width, src_height,
-                          clip_width, clip_height,
-                          src_stride, dst_stride, src, dst,
-                          x, dx, y, dy, filtering);
+        ScaleARGBDownEven(src_width, src_height, clip_width, clip_height,
+                          src_stride, dst_stride, src, dst, x, dx, y, dy,
+                          filtering);
         return;
       }
       // Optimized odd scale down. ie 3, 5, 7, 9x.
@@ -750,105 +1011,114 @@ static void ScaleARGB(const uint8* src, int src_stride,
         filtering = kFilterNone;
         if (dx == 0x10000 && dy == 0x10000) {
           // Straight copy.
-          ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
-                   dst, dst_stride, clip_width, clip_height);
+          ARGBCopy(src + (y >> 16) * (intptr_t)src_stride + (x >> 16) * 4,
+                   src_stride, dst, dst_stride, clip_width, clip_height);
           return;
         }
       }
     }
   }
   if (dx == 0x10000 && (x & 0xffff) == 0) {
-    // Arbitrary scale vertically, but unscaled vertically.
-    ScalePlaneVertical(src_height,
-                       clip_width, clip_height,
-                       src_stride, dst_stride, src, dst,
-                       x, y, dy, 4, filtering);
+    // Arbitrary scale vertically, but unscaled horizontally.
+    ScalePlaneVertical(src_height, clip_width, clip_height, src_stride,
+                       dst_stride, src, dst, x, y, dy, /*bpp=*/4, filtering);
     return;
   }
   if (filtering && dy < 65536) {
-    ScaleARGBBilinearUp(src_width, src_height,
-                        clip_width, clip_height,
-                        src_stride, dst_stride, src, dst,
-                        x, dx, y, dy, filtering);
+    ScaleARGBBilinearUp(src_width, src_height, clip_width, clip_height,
+                        src_stride, dst_stride, src, dst, x, dx, y, dy,
+                        filtering);
     return;
   }
   if (filtering) {
-    ScaleARGBBilinearDown(src_width, src_height,
-                          clip_width, clip_height,
-                          src_stride, dst_stride, src, dst,
-                          x, dx, y, dy, filtering);
+    ScaleARGBBilinearDown(src_width, src_height, clip_width, clip_height,
+                          src_stride, dst_stride, src, dst, x, dx, y, dy,
+                          filtering);
     return;
   }
-  ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
-                  src_stride, dst_stride, src, dst,
-                  x, dx, y, dy);
+  ScaleARGBSimple(src_width, src_height, clip_width, clip_height, src_stride,
+                  dst_stride, src, dst, x, dx, y, dy);
 }
 
 LIBYUV_API
-int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
-                  int src_width, int src_height,
-                  uint8* dst_argb, int dst_stride_argb,
-                  int dst_width, int dst_height,
-                  int clip_x, int clip_y, int clip_width, int clip_height,
+int ARGBScaleClip(const uint8_t* src_argb,
+                  int src_stride_argb,
+                  int src_width,
+                  int src_height,
+                  uint8_t* dst_argb,
+                  int dst_stride_argb,
+                  int dst_width,
+                  int dst_height,
+                  int clip_x,
+                  int clip_y,
+                  int clip_width,
+                  int clip_height,
                   enum FilterMode filtering) {
-  if (!src_argb || src_width == 0 || src_height == 0 ||
-      !dst_argb || dst_width <= 0 || dst_height <= 0 ||
-      clip_x < 0 || clip_y < 0 ||
+  if (!src_argb || src_width == 0 || src_height == 0 || !dst_argb ||
+      dst_width <= 0 || dst_height <= 0 || clip_x < 0 || clip_y < 0 ||
       clip_width > 32768 || clip_height > 32768 ||
       (clip_x + clip_width) > dst_width ||
       (clip_y + clip_height) > dst_height) {
     return -1;
   }
-  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
-            dst_argb, dst_stride_argb, dst_width, dst_height,
-            clip_x, clip_y, clip_width, clip_height, filtering);
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height, dst_argb,
+            dst_stride_argb, dst_width, dst_height, clip_x, clip_y, clip_width,
+            clip_height, filtering);
   return 0;
 }
 
 // Scale an ARGB image.
 LIBYUV_API
-int ARGBScale(const uint8* src_argb, int src_stride_argb,
-              int src_width, int src_height,
-              uint8* dst_argb, int dst_stride_argb,
-              int dst_width, int dst_height,
+int ARGBScale(const uint8_t* src_argb,
+              int src_stride_argb,
+              int src_width,
+              int src_height,
+              uint8_t* dst_argb,
+              int dst_stride_argb,
+              int dst_width,
+              int dst_height,
               enum FilterMode filtering) {
-  if (!src_argb || src_width == 0 || src_height == 0 ||
-      src_width > 32768 || src_height > 32768 ||
-      !dst_argb || dst_width <= 0 || dst_height <= 0) {
+  if (!src_argb || src_width == 0 || src_height == 0 || src_width > 32768 ||
+      src_height > 32768 || !dst_argb || dst_width <= 0 || dst_height <= 0) {
     return -1;
   }
-  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
-            dst_argb, dst_stride_argb, dst_width, dst_height,
-            0, 0, dst_width, dst_height, filtering);
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height, dst_argb,
+            dst_stride_argb, dst_width, dst_height, 0, 0, dst_width, dst_height,
+            filtering);
   return 0;
 }
 
 // Scale with YUV conversion to ARGB and clipping.
 LIBYUV_API
-int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
-                       const uint8* src_u, int src_stride_u,
-                       const uint8* src_v, int src_stride_v,
-                       uint32 src_fourcc,
-                       int src_width, int src_height,
-                       uint8* dst_argb, int dst_stride_argb,
-                       uint32 dst_fourcc,
-                       int dst_width, int dst_height,
-                       int clip_x, int clip_y, int clip_width, int clip_height,
+int YUVToARGBScaleClip(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint32_t src_fourcc,
+                       int src_width,
+                       int src_height,
+                       uint8_t* dst_argb,
+                       int dst_stride_argb,
+                       uint32_t dst_fourcc,
+                       int dst_width,
+                       int dst_height,
+                       int clip_x,
+                       int clip_y,
+                       int clip_width,
+                       int clip_height,
                        enum FilterMode filtering) {
-  uint8* argb_buffer = (uint8*)malloc(src_width * src_height * 4);
+  uint8_t* argb_buffer = (uint8_t*)malloc(src_width * src_height * 4);
   int r;
-  I420ToARGB(src_y, src_stride_y,
-             src_u, src_stride_u,
-             src_v, src_stride_v,
-             argb_buffer, src_width * 4,
-             src_width, src_height);
-
-  r = ARGBScaleClip(argb_buffer, src_width * 4,
-                    src_width, src_height,
-                    dst_argb, dst_stride_argb,
-                    dst_width, dst_height,
-                    clip_x, clip_y, clip_width, clip_height,
-                    filtering);
+  (void)src_fourcc;  // TODO(fbarchard): implement and/or assert.
+  (void)dst_fourcc;
+  I420ToARGB(src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+             argb_buffer, src_width * 4, src_width, src_height);
+
+  r = ARGBScaleClip(argb_buffer, src_width * 4, src_width, src_height, dst_argb,
+                    dst_stride_argb, dst_width, dst_height, clip_x, clip_y,
+                    clip_width, clip_height, filtering);
   free(argb_buffer);
   return r;
 }
diff --git a/media/libyuv/source/scale_common.cc b/media/libyuv/source/scale_common.cc
index 3507aa4d9f..da9ca713ec 100644
--- a/media/libyuv/source/scale_common.cc
+++ b/media/libyuv/source/scale_common.cc
@@ -23,14 +23,36 @@ namespace libyuv {
 extern "C" {
 #endif
 
+#ifdef __cplusplus
+#define STATIC_CAST(type, expr) static_cast<type>(expr)
+#else
+#define STATIC_CAST(type, expr) (type)(expr)
+#endif
+
+// TODO(fbarchard): make clamp255 preserve negative values.
+static __inline int32_t clamp255(int32_t v) {
+  return (-(v >= 255) | v) & 255;
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+#define C16TO8(v, scale) clamp255(((v) * (scale)) >> 16)
+
 static __inline int Abs(int v) {
   return v >= 0 ? v : -v;
 }
 
 // CPU agnostic row functions
-void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                     uint8* dst, int dst_width) {
+void ScaleRowDown2_C(const uint8_t* src_ptr,
+                     ptrdiff_t src_stride,
+                     uint8_t* dst,
+                     int dst_width) {
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = src_ptr[1];
     dst[1] = src_ptr[3];
@@ -42,9 +64,12 @@ void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                        uint16* dst, int dst_width) {
+void ScaleRowDown2_16_C(const uint16_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint16_t* dst,
+                        int dst_width) {
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = src_ptr[1];
     dst[1] = src_ptr[3];
@@ -56,10 +81,57 @@ void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width) {
-  const uint8* s = src_ptr;
+void ScaleRowDown2_16To8_C(const uint16_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width,
+                           int scale) {
+  int x;
+  (void)src_stride;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8(src_ptr[1], scale));
+    dst[1] = STATIC_CAST(uint8_t, C16TO8(src_ptr[3], scale));
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8(src_ptr[1], scale));
+  }
+}
+
+void ScaleRowDown2_16To8_Odd_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst,
+                               int dst_width,
+                               int scale) {
+  int x;
+  (void)src_stride;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+  dst_width -= 1;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8(src_ptr[1], scale));
+    dst[1] = STATIC_CAST(uint8_t, C16TO8(src_ptr[3], scale));
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8(src_ptr[1], scale));
+    dst += 1;
+    src_ptr += 2;
+  }
+  dst[0] = STATIC_CAST(uint8_t, C16TO8(src_ptr[0], scale));
+}
+
+void ScaleRowDown2Linear_C(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width) {
+  const uint8_t* s = src_ptr;
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = (s[0] + s[1] + 1) >> 1;
     dst[1] = (s[2] + s[3] + 1) >> 1;
@@ -71,10 +143,13 @@ void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                              uint16* dst, int dst_width) {
-  const uint16* s = src_ptr;
+void ScaleRowDown2Linear_16_C(const uint16_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint16_t* dst,
+                              int dst_width) {
+  const uint16_t* s = src_ptr;
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = (s[0] + s[1] + 1) >> 1;
     dst[1] = (s[2] + s[3] + 1) >> 1;
@@ -86,10 +161,58 @@ void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width) {
-  const uint8* s = src_ptr;
-  const uint8* t = src_ptr + src_stride;
+void ScaleRowDown2Linear_16To8_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst,
+                                 int dst_width,
+                                 int scale) {
+  const uint16_t* s = src_ptr;
+  int x;
+  (void)src_stride;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8((s[0] + s[1] + 1) >> 1, scale));
+    dst[1] = STATIC_CAST(uint8_t, C16TO8((s[2] + s[3] + 1) >> 1, scale));
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8((s[0] + s[1] + 1) >> 1, scale));
+  }
+}
+
+void ScaleRowDown2Linear_16To8_Odd_C(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst,
+                                     int dst_width,
+                                     int scale) {
+  const uint16_t* s = src_ptr;
+  int x;
+  (void)src_stride;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+  dst_width -= 1;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8((s[0] + s[1] + 1) >> 1, scale));
+    dst[1] = STATIC_CAST(uint8_t, C16TO8((s[2] + s[3] + 1) >> 1, scale));
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = STATIC_CAST(uint8_t, C16TO8((s[0] + s[1] + 1) >> 1, scale));
+    dst += 1;
+    s += 2;
+  }
+  dst[0] = STATIC_CAST(uint8_t, C16TO8(s[0], scale));
+}
+
+void ScaleRowDown2Box_C(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
   int x;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
@@ -103,10 +226,12 @@ void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown2Box_Odd_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width) {
-  const uint8* s = src_ptr;
-  const uint8* t = src_ptr + src_stride;
+void ScaleRowDown2Box_Odd_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst,
+                            int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
   int x;
   dst_width -= 1;
   for (x = 0; x < dst_width - 1; x += 2) {
@@ -125,10 +250,12 @@ void ScaleRowDown2Box_Odd_C(const uint8* src_ptr, ptrdiff_t src_stride,
   dst[0] = (s[0] + t[0] + 1) >> 1;
 }
 
-void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst, int dst_width) {
-  const uint16* s = src_ptr;
-  const uint16* t = src_ptr + src_stride;
+void ScaleRowDown2Box_16_C(const uint16_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint16_t* dst,
+                           int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
   int x;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
@@ -142,9 +269,67 @@ void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                     uint8* dst, int dst_width) {
+void ScaleRowDown2Box_16To8_C(const uint16_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst,
+                              int dst_width,
+                              int scale) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+  int x;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = STATIC_CAST(uint8_t,
+                         C16TO8((s[0] + s[1] + t[0] + t[1] + 2) >> 2, scale));
+    dst[1] = STATIC_CAST(uint8_t,
+                         C16TO8((s[2] + s[3] + t[2] + t[3] + 2) >> 2, scale));
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = STATIC_CAST(uint8_t,
+                         C16TO8((s[0] + s[1] + t[0] + t[1] + 2) >> 2, scale));
+  }
+}
+
+void ScaleRowDown2Box_16To8_Odd_C(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst,
+                                  int dst_width,
+                                  int scale) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+  int x;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+  dst_width -= 1;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = STATIC_CAST(uint8_t,
+                         C16TO8((s[0] + s[1] + t[0] + t[1] + 2) >> 2, scale));
+    dst[1] = STATIC_CAST(uint8_t,
+                         C16TO8((s[2] + s[3] + t[2] + t[3] + 2) >> 2, scale));
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = STATIC_CAST(uint8_t,
+                         C16TO8((s[0] + s[1] + t[0] + t[1] + 2) >> 2, scale));
+    dst += 1;
+    s += 2;
+    t += 2;
+  }
+  dst[0] = STATIC_CAST(uint8_t, C16TO8((s[0] + t[0] + 1) >> 1, scale));
+}
+
+void ScaleRowDown4_C(const uint8_t* src_ptr,
+                     ptrdiff_t src_stride,
+                     uint8_t* dst,
+                     int dst_width) {
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = src_ptr[2];
     dst[1] = src_ptr[6];
@@ -156,9 +341,12 @@ void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                        uint16* dst, int dst_width) {
+void ScaleRowDown4_16_C(const uint16_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint16_t* dst,
+                        int dst_width) {
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = src_ptr[2];
     dst[1] = src_ptr[6];
@@ -170,81 +358,88 @@ void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width) {
+void ScaleRowDown4Box_C(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
   intptr_t stride = src_stride;
   int x;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
-             src_ptr[stride + 0] + src_ptr[stride + 1] +
-             src_ptr[stride + 2] + src_ptr[stride + 3] +
-             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
-             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
-             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
-             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
-             8) >> 4;
+              src_ptr[stride + 0] + src_ptr[stride + 1] + src_ptr[stride + 2] +
+              src_ptr[stride + 3] + src_ptr[stride * 2 + 0] +
+              src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2] +
+              src_ptr[stride * 2 + 3] + src_ptr[stride * 3 + 0] +
+              src_ptr[stride * 3 + 1] + src_ptr[stride * 3 + 2] +
+              src_ptr[stride * 3 + 3] + 8) >>
+             4;
     dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
-             src_ptr[stride + 4] + src_ptr[stride + 5] +
-             src_ptr[stride + 6] + src_ptr[stride + 7] +
-             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
-             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
-             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
-             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
-             8) >> 4;
+              src_ptr[stride + 4] + src_ptr[stride + 5] + src_ptr[stride + 6] +
+              src_ptr[stride + 7] + src_ptr[stride * 2 + 4] +
+              src_ptr[stride * 2 + 5] + src_ptr[stride * 2 + 6] +
+              src_ptr[stride * 2 + 7] + src_ptr[stride * 3 + 4] +
+              src_ptr[stride * 3 + 5] + src_ptr[stride * 3 + 6] +
+              src_ptr[stride * 3 + 7] + 8) >>
+             4;
     dst += 2;
     src_ptr += 8;
   }
   if (dst_width & 1) {
     dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
-             src_ptr[stride + 0] + src_ptr[stride + 1] +
-             src_ptr[stride + 2] + src_ptr[stride + 3] +
-             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
-             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
-             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
-             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
-             8) >> 4;
+              src_ptr[stride + 0] + src_ptr[stride + 1] + src_ptr[stride + 2] +
+              src_ptr[stride + 3] + src_ptr[stride * 2 + 0] +
+              src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2] +
+              src_ptr[stride * 2 + 3] + src_ptr[stride * 3 + 0] +
+              src_ptr[stride * 3 + 1] + src_ptr[stride * 3 + 2] +
+              src_ptr[stride * 3 + 3] + 8) >>
+             4;
   }
 }
 
-void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                           uint16* dst, int dst_width) {
+void ScaleRowDown4Box_16_C(const uint16_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint16_t* dst,
+                           int dst_width) {
   intptr_t stride = src_stride;
   int x;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
-             src_ptr[stride + 0] + src_ptr[stride + 1] +
-             src_ptr[stride + 2] + src_ptr[stride + 3] +
-             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
-             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
-             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
-             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
-             8) >> 4;
+              src_ptr[stride + 0] + src_ptr[stride + 1] + src_ptr[stride + 2] +
+              src_ptr[stride + 3] + src_ptr[stride * 2 + 0] +
+              src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2] +
+              src_ptr[stride * 2 + 3] + src_ptr[stride * 3 + 0] +
+              src_ptr[stride * 3 + 1] + src_ptr[stride * 3 + 2] +
+              src_ptr[stride * 3 + 3] + 8) >>
+             4;
     dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
-             src_ptr[stride + 4] + src_ptr[stride + 5] +
-             src_ptr[stride + 6] + src_ptr[stride + 7] +
-             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
-             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
-             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
-             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
-             8) >> 4;
+              src_ptr[stride + 4] + src_ptr[stride + 5] + src_ptr[stride + 6] +
+              src_ptr[stride + 7] + src_ptr[stride * 2 + 4] +
+              src_ptr[stride * 2 + 5] + src_ptr[stride * 2 + 6] +
+              src_ptr[stride * 2 + 7] + src_ptr[stride * 3 + 4] +
+              src_ptr[stride * 3 + 5] + src_ptr[stride * 3 + 6] +
+              src_ptr[stride * 3 + 7] + 8) >>
+             4;
     dst += 2;
     src_ptr += 8;
   }
   if (dst_width & 1) {
     dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
-             src_ptr[stride + 0] + src_ptr[stride + 1] +
-             src_ptr[stride + 2] + src_ptr[stride + 3] +
-             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
-             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
-             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
-             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
-             8) >> 4;
+              src_ptr[stride + 0] + src_ptr[stride + 1] + src_ptr[stride + 2] +
+              src_ptr[stride + 3] + src_ptr[stride * 2 + 0] +
+              src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2] +
+              src_ptr[stride * 2 + 3] + src_ptr[stride * 3 + 0] +
+              src_ptr[stride * 3 + 1] + src_ptr[stride * 3 + 2] +
+              src_ptr[stride * 3 + 3] + 8) >>
+             4;
   }
 }
 
-void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                      uint8* dst, int dst_width) {
+void ScaleRowDown34_C(const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint8_t* dst,
+                      int dst_width) {
   int x;
+  (void)src_stride;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (x = 0; x < dst_width; x += 3) {
     dst[0] = src_ptr[0];
@@ -255,9 +450,12 @@ void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                         uint16* dst, int dst_width) {
+void ScaleRowDown34_16_C(const uint16_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint16_t* dst,
+                         int dst_width) {
   int x;
+  (void)src_stride;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (x = 0; x < dst_width; x += 3) {
     dst[0] = src_ptr[0];
@@ -269,19 +467,21 @@ void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
 }
 
 // Filter rows 0 and 1 together, 3 : 1
-void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* d, int dst_width) {
-  const uint8* s = src_ptr;
-  const uint8* t = src_ptr + src_stride;
+void ScaleRowDown34_0_Box_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* d,
+                            int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
   int x;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (x = 0; x < dst_width; x += 3) {
-    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    uint8_t a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8_t a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8_t a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8_t b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8_t b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8_t b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
     d[0] = (a0 * 3 + b0 + 2) >> 2;
     d[1] = (a1 * 3 + b1 + 2) >> 2;
     d[2] = (a2 * 3 + b2 + 2) >> 2;
@@ -291,19 +491,21 @@ void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                               uint16* d, int dst_width) {
-  const uint16* s = src_ptr;
-  const uint16* t = src_ptr + src_stride;
+void ScaleRowDown34_0_Box_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* d,
+                               int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
   int x;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (x = 0; x < dst_width; x += 3) {
-    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    uint16_t a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16_t a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16_t a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16_t b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16_t b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16_t b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
     d[0] = (a0 * 3 + b0 + 2) >> 2;
     d[1] = (a1 * 3 + b1 + 2) >> 2;
     d[2] = (a2 * 3 + b2 + 2) >> 2;
@@ -314,19 +516,21 @@ void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
 }
 
 // Filter rows 1 and 2 together, 1 : 1
-void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* d, int dst_width) {
-  const uint8* s = src_ptr;
-  const uint8* t = src_ptr + src_stride;
+void ScaleRowDown34_1_Box_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* d,
+                            int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
   int x;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (x = 0; x < dst_width; x += 3) {
-    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    uint8_t a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8_t a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8_t a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8_t b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8_t b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8_t b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
     d[0] = (a0 + b0 + 1) >> 1;
     d[1] = (a1 + b1 + 1) >> 1;
     d[2] = (a2 + b2 + 1) >> 1;
@@ -336,19 +540,21 @@ void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                               uint16* d, int dst_width) {
-  const uint16* s = src_ptr;
-  const uint16* t = src_ptr + src_stride;
+void ScaleRowDown34_1_Box_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* d,
+                               int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
   int x;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (x = 0; x < dst_width; x += 3) {
-    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    uint16_t a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16_t a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16_t a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16_t b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16_t b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16_t b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
     d[0] = (a0 + b0 + 1) >> 1;
     d[1] = (a1 + b1 + 1) >> 1;
     d[2] = (a2 + b2 + 1) >> 1;
@@ -358,9 +564,101 @@ void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
   }
 }
 
+// Sample position: (O is src sample position, X is dst sample position)
+//
+//      v dst_ptr at here           v stop at here
+//  X O X   X O X   X O X   X O X   X O X
+//    ^ src_ptr at here
+void ScaleRowUp2_Linear_C(const uint8_t* src_ptr,
+                          uint8_t* dst_ptr,
+                          int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[2 * x + 0] = (src_ptr[x + 0] * 3 + src_ptr[x + 1] * 1 + 2) >> 2;
+    dst_ptr[2 * x + 1] = (src_ptr[x + 0] * 1 + src_ptr[x + 1] * 3 + 2) >> 2;
+  }
+}
+
+// Sample position: (O is src sample position, X is dst sample position)
+//
+//    src_ptr at here
+//  X v X   X   X   X   X   X   X   X   X
+//    O       O       O       O       O
+//  X   X   X   X   X   X   X   X   X   X
+//      ^ dst_ptr at here           ^ stop at here
+//  X   X   X   X   X   X   X   X   X   X
+//    O       O       O       O       O
+//  X   X   X   X   X   X   X   X   X   X
+void ScaleRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            ptrdiff_t dst_stride,
+                            int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  uint8_t* d = dst_ptr;
+  uint8_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[2 * x + 0] =
+        (s[x + 0] * 9 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 1 + 8) >> 4;
+    d[2 * x + 1] =
+        (s[x + 0] * 3 + s[x + 1] * 9 + t[x + 0] * 1 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 0] =
+        (s[x + 0] * 3 + s[x + 1] * 1 + t[x + 0] * 9 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 1] =
+        (s[x + 0] * 1 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 9 + 8) >> 4;
+  }
+}
+
+// Only suitable for at most 14 bit range.
+void ScaleRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[2 * x + 0] = (src_ptr[x + 0] * 3 + src_ptr[x + 1] * 1 + 2) >> 2;
+    dst_ptr[2 * x + 1] = (src_ptr[x + 0] * 1 + src_ptr[x + 1] * 3 + 2) >> 2;
+  }
+}
+
+// Only suitable for at most 12bit range.
+void ScaleRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+  uint16_t* d = dst_ptr;
+  uint16_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[2 * x + 0] =
+        (s[x + 0] * 9 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 1 + 8) >> 4;
+    d[2 * x + 1] =
+        (s[x + 0] * 3 + s[x + 1] * 9 + t[x + 0] * 1 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 0] =
+        (s[x + 0] * 3 + s[x + 1] * 1 + t[x + 0] * 9 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 1] =
+        (s[x + 0] * 1 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 9 + 8) >> 4;
+  }
+}
+
 // Scales a single row of pixels using point sampling.
-void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
-                 int dst_width, int x, int dx) {
+void ScaleCols_C(uint8_t* dst_ptr,
+                 const uint8_t* src_ptr,
+                 int dst_width,
+                 int x,
+                 int dx) {
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst_ptr[0] = src_ptr[x >> 16];
@@ -374,8 +672,11 @@ void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
   }
 }
 
-void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                    int dst_width, int x, int dx) {
+void ScaleCols_16_C(uint16_t* dst_ptr,
+                    const uint16_t* src_ptr,
+                    int dst_width,
+                    int x,
+                    int dx) {
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst_ptr[0] = src_ptr[x >> 16];
@@ -390,9 +691,14 @@ void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
 }
 
 // Scales a single row of pixels up by 2x using point sampling.
-void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
-                    int dst_width, int x, int dx) {
+void ScaleColsUp2_C(uint8_t* dst_ptr,
+                    const uint8_t* src_ptr,
+                    int dst_width,
+                    int x,
+                    int dx) {
   int j;
+  (void)x;
+  (void)dx;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst_ptr[1] = dst_ptr[0] = src_ptr[0];
     src_ptr += 1;
@@ -403,9 +709,14 @@ void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
   }
 }
 
-void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                       int dst_width, int x, int dx) {
+void ScaleColsUp2_16_C(uint16_t* dst_ptr,
+                       const uint16_t* src_ptr,
+                       int dst_width,
+                       int x,
+                       int dx) {
   int j;
+  (void)x;
+  (void)dx;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst_ptr[1] = dst_ptr[0] = src_ptr[0];
     src_ptr += 1;
@@ -418,16 +729,19 @@ void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
 
 // (1-f)a + fb can be replaced with a + f(b-a)
 #if defined(__arm__) || defined(__aarch64__)
-#define BLENDER(a, b, f) (uint8)((int)(a) + \
-    ((((int)((f)) * ((int)(b) - (int)(a))) + 0x8000) >> 16))
+#define BLENDER(a, b, f) \
+  (uint8_t)((int)(a) + ((((int)((f)) * ((int)(b) - (int)(a))) + 0x8000) >> 16))
 #else
-// inteluses 7 bit math with rounding.
-#define BLENDER(a, b, f) (uint8)((int)(a) + \
-    (((int)((f) >> 9) * ((int)(b) - (int)(a)) + 0x40) >> 7))
+// Intel uses 7 bit math with rounding.
+#define BLENDER(a, b, f) \
+  (uint8_t)((int)(a) + (((int)((f) >> 9) * ((int)(b) - (int)(a)) + 0x40) >> 7))
 #endif
 
-void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
-                       int dst_width, int x, int dx) {
+void ScaleFilterCols_C(uint8_t* dst_ptr,
+                       const uint8_t* src_ptr,
+                       int dst_width,
+                       int x,
+                       int dx) {
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     int xi = x >> 16;
@@ -450,12 +764,15 @@ void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
   }
 }
 
-void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
-                         int dst_width, int x32, int dx) {
-  int64 x = (int64)(x32);
+void ScaleFilterCols64_C(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x32,
+                         int dx) {
+  int64_t x = (int64_t)(x32);
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
-    int64 xi = x >> 16;
+    int64_t xi = x >> 16;
     int a = src_ptr[xi];
     int b = src_ptr[xi + 1];
     dst_ptr[0] = BLENDER(a, b, x & 0xffff);
@@ -468,7 +785,7 @@ void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
     dst_ptr += 2;
   }
   if (dst_width & 1) {
-    int64 xi = x >> 16;
+    int64_t xi = x >> 16;
     int a = src_ptr[xi];
     int b = src_ptr[xi + 1];
     dst_ptr[0] = BLENDER(a, b, x & 0xffff);
@@ -476,12 +793,17 @@ void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
 }
 #undef BLENDER
 
-// Same as 8 bit arm blender but return is cast to uint16
-#define BLENDER(a, b, f) (uint16)((int)(a) + \
-    ((((int)((f)) * ((int)(b) - (int)(a))) + 0x8000) >> 16))
+// Same as 8 bit arm blender but return is cast to uint16_t
+#define BLENDER(a, b, f) \
+  (uint16_t)(            \
+      (int)(a) +         \
+      (int)((((int64_t)((f)) * ((int64_t)(b) - (int)(a))) + 0x8000) >> 16))
 
-void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                       int dst_width, int x, int dx) {
+void ScaleFilterCols_16_C(uint16_t* dst_ptr,
+                          const uint16_t* src_ptr,
+                          int dst_width,
+                          int x,
+                          int dx) {
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     int xi = x >> 16;
@@ -504,12 +826,15 @@ void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
   }
 }
 
-void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
-                         int dst_width, int x32, int dx) {
-  int64 x = (int64)(x32);
+void ScaleFilterCols64_16_C(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            int dst_width,
+                            int x32,
+                            int dx) {
+  int64_t x = (int64_t)(x32);
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
-    int64 xi = x >> 16;
+    int64_t xi = x >> 16;
     int a = src_ptr[xi];
     int b = src_ptr[xi + 1];
     dst_ptr[0] = BLENDER(a, b, x & 0xffff);
@@ -522,7 +847,7 @@ void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
     dst_ptr += 2;
   }
   if (dst_width & 1) {
-    int64 xi = x >> 16;
+    int64_t xi = x >> 16;
     int a = src_ptr[xi];
     int b = src_ptr[xi + 1];
     dst_ptr[0] = BLENDER(a, b, x & 0xffff);
@@ -530,9 +855,12 @@ void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
 }
 #undef BLENDER
 
-void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                      uint8* dst, int dst_width) {
+void ScaleRowDown38_C(const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint8_t* dst,
+                      int dst_width) {
   int x;
+  (void)src_stride;
   assert(dst_width % 3 == 0);
   for (x = 0; x < dst_width; x += 3) {
     dst[0] = src_ptr[0];
@@ -543,9 +871,12 @@ void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
-void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                         uint16* dst, int dst_width) {
+void ScaleRowDown38_16_C(const uint16_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint16_t* dst,
+                         int dst_width) {
   int x;
+  (void)src_stride;
   assert(dst_width % 3 == 0);
   for (x = 0; x < dst_width; x += 3) {
     dst[0] = src_ptr[0];
@@ -557,100 +888,118 @@ void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
 }
 
 // 8x3 -> 3x1
-void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+void ScaleRowDown38_3_Box_C(const uint8_t* src_ptr,
                             ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width) {
+                            uint8_t* dst_ptr,
+                            int dst_width) {
   intptr_t stride = src_stride;
   int i;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (i = 0; i < dst_width; i += 3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[stride + 0] + src_ptr[stride + 1] +
-        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
-        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
-        (65536 / 9) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[stride + 3] + src_ptr[stride + 4] +
-        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
-        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
-        (65536 / 9) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[stride + 6] + src_ptr[stride + 7] +
-        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
-        (65536 / 6) >> 16;
+    dst_ptr[0] =
+        (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[stride + 0] +
+         src_ptr[stride + 1] + src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+         src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+            (65536 / 9) >>
+        16;
+    dst_ptr[1] =
+        (src_ptr[3] + src_ptr[4] + src_ptr[5] + src_ptr[stride + 3] +
+         src_ptr[stride + 4] + src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+         src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+            (65536 / 9) >>
+        16;
+    dst_ptr[2] =
+        (src_ptr[6] + src_ptr[7] + src_ptr[stride + 6] + src_ptr[stride + 7] +
+         src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+            (65536 / 6) >>
+        16;
     src_ptr += 8;
     dst_ptr += 3;
   }
 }
 
-void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+void ScaleRowDown38_3_Box_16_C(const uint16_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint16* dst_ptr, int dst_width) {
+                               uint16_t* dst_ptr,
+                               int dst_width) {
   intptr_t stride = src_stride;
   int i;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (i = 0; i < dst_width; i += 3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[stride + 0] + src_ptr[stride + 1] +
-        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
-        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
-        (65536 / 9) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[stride + 3] + src_ptr[stride + 4] +
-        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
-        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
-        (65536 / 9) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[stride + 6] + src_ptr[stride + 7] +
-        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
-        (65536 / 6) >> 16;
+    dst_ptr[0] =
+        (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[stride + 0] +
+         src_ptr[stride + 1] + src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+         src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+            (65536u / 9u) >>
+        16;
+    dst_ptr[1] =
+        (src_ptr[3] + src_ptr[4] + src_ptr[5] + src_ptr[stride + 3] +
+         src_ptr[stride + 4] + src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+         src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+            (65536u / 9u) >>
+        16;
+    dst_ptr[2] =
+        (src_ptr[6] + src_ptr[7] + src_ptr[stride + 6] + src_ptr[stride + 7] +
+         src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+            (65536u / 6u) >>
+        16;
     src_ptr += 8;
     dst_ptr += 3;
   }
 }
 
 // 8x2 -> 3x1
-void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width) {
+void ScaleRowDown38_2_Box_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width) {
   intptr_t stride = src_stride;
   int i;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (i = 0; i < dst_width; i += 3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[stride + 0] + src_ptr[stride + 1] +
-        src_ptr[stride + 2]) * (65536 / 6) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[stride + 3] + src_ptr[stride + 4] +
-        src_ptr[stride + 5]) * (65536 / 6) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[stride + 6] + src_ptr[stride + 7]) *
-        (65536 / 4) >> 16;
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[stride + 0] +
+                  src_ptr[stride + 1] + src_ptr[stride + 2]) *
+                     (65536 / 6) >>
+                 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + src_ptr[stride + 3] +
+                  src_ptr[stride + 4] + src_ptr[stride + 5]) *
+                     (65536 / 6) >>
+                 16;
+    dst_ptr[2] =
+        (src_ptr[6] + src_ptr[7] + src_ptr[stride + 6] + src_ptr[stride + 7]) *
+            (65536 / 4) >>
+        16;
     src_ptr += 8;
     dst_ptr += 3;
   }
 }
 
-void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
-                               uint16* dst_ptr, int dst_width) {
+void ScaleRowDown38_2_Box_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* dst_ptr,
+                               int dst_width) {
   intptr_t stride = src_stride;
   int i;
   assert((dst_width % 3 == 0) && (dst_width > 0));
   for (i = 0; i < dst_width; i += 3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[stride + 0] + src_ptr[stride + 1] +
-        src_ptr[stride + 2]) * (65536 / 6) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[stride + 3] + src_ptr[stride + 4] +
-        src_ptr[stride + 5]) * (65536 / 6) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[stride + 6] + src_ptr[stride + 7]) *
-        (65536 / 4) >> 16;
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[stride + 0] +
+                  src_ptr[stride + 1] + src_ptr[stride + 2]) *
+                     (65536u / 6u) >>
+                 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + src_ptr[stride + 3] +
+                  src_ptr[stride + 4] + src_ptr[stride + 5]) *
+                     (65536u / 6u) >>
+                 16;
+    dst_ptr[2] =
+        (src_ptr[6] + src_ptr[7] + src_ptr[stride + 6] + src_ptr[stride + 7]) *
+            (65536u / 4u) >>
+        16;
     src_ptr += 8;
     dst_ptr += 3;
   }
 }
 
-void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+void ScaleAddRow_C(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
   int x;
   assert(src_width > 0);
   for (x = 0; x < src_width - 1; x += 2) {
@@ -664,7 +1013,9 @@ void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
   }
 }
 
-void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width) {
+void ScaleAddRow_16_C(const uint16_t* src_ptr,
+                      uint32_t* dst_ptr,
+                      int src_width) {
   int x;
   assert(src_width > 0);
   for (x = 0; x < src_width - 1; x += 2) {
@@ -678,13 +1029,16 @@ void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width) {
   }
 }
 
-void ScaleARGBRowDown2_C(const uint8* src_argb,
-                         ptrdiff_t src_stride,
-                         uint8* dst_argb, int dst_width) {
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
+// ARGB scale row functions
 
+void ScaleARGBRowDown2_C(const uint8_t* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_argb,
+                         int dst_width) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = src[1];
     dst[1] = src[3];
@@ -696,10 +1050,12 @@ void ScaleARGBRowDown2_C(const uint8* src_argb,
   }
 }
 
-void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+void ScaleARGBRowDown2Linear_C(const uint8_t* src_argb,
                                ptrdiff_t src_stride,
-                               uint8* dst_argb, int dst_width) {
+                               uint8_t* dst_argb,
+                               int dst_width) {
   int x;
+  (void)src_stride;
   for (x = 0; x < dst_width; ++x) {
     dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
     dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
@@ -710,29 +1066,37 @@ void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
   }
 }
 
-void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
-                            uint8* dst_argb, int dst_width) {
+void ScaleARGBRowDown2Box_C(const uint8_t* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_argb,
+                            int dst_width) {
   int x;
   for (x = 0; x < dst_width; ++x) {
-    dst_argb[0] = (src_argb[0] + src_argb[4] +
-                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
-    dst_argb[1] = (src_argb[1] + src_argb[5] +
-                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
-    dst_argb[2] = (src_argb[2] + src_argb[6] +
-                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
-    dst_argb[3] = (src_argb[3] + src_argb[7] +
-                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    dst_argb[0] = (src_argb[0] + src_argb[4] + src_argb[src_stride] +
+                   src_argb[src_stride + 4] + 2) >>
+                  2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + src_argb[src_stride + 1] +
+                   src_argb[src_stride + 5] + 2) >>
+                  2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + src_argb[src_stride + 2] +
+                   src_argb[src_stride + 6] + 2) >>
+                  2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + src_argb[src_stride + 3] +
+                   src_argb[src_stride + 7] + 2) >>
+                  2;
     src_argb += 8;
     dst_argb += 4;
   }
 }
 
-void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+void ScaleARGBRowDownEven_C(const uint8_t* src_argb,
+                            ptrdiff_t src_stride,
                             int src_stepx,
-                            uint8* dst_argb, int dst_width) {
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
-
+                            uint8_t* dst_argb,
+                            int dst_width) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
+  (void)src_stride;
   int x;
   for (x = 0; x < dst_width - 1; x += 2) {
     dst[0] = src[0];
@@ -745,30 +1109,38 @@ void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
   }
 }
 
-void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+void ScaleARGBRowDownEvenBox_C(const uint8_t* src_argb,
                                ptrdiff_t src_stride,
                                int src_stepx,
-                               uint8* dst_argb, int dst_width) {
+                               uint8_t* dst_argb,
+                               int dst_width) {
   int x;
   for (x = 0; x < dst_width; ++x) {
-    dst_argb[0] = (src_argb[0] + src_argb[4] +
-                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
-    dst_argb[1] = (src_argb[1] + src_argb[5] +
-                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
-    dst_argb[2] = (src_argb[2] + src_argb[6] +
-                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
-    dst_argb[3] = (src_argb[3] + src_argb[7] +
-                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    dst_argb[0] = (src_argb[0] + src_argb[4] + src_argb[src_stride] +
+                   src_argb[src_stride + 4] + 2) >>
+                  2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + src_argb[src_stride + 1] +
+                   src_argb[src_stride + 5] + 2) >>
+                  2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + src_argb[src_stride + 2] +
+                   src_argb[src_stride + 6] + 2) >>
+                  2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + src_argb[src_stride + 3] +
+                   src_argb[src_stride + 7] + 2) >>
+                  2;
     src_argb += src_stepx * 4;
     dst_argb += 4;
   }
 }
 
 // Scales a single row of pixels using point sampling.
-void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
-                     int dst_width, int x, int dx) {
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
+void ScaleARGBCols_C(uint8_t* dst_argb,
+                     const uint8_t* src_argb,
+                     int dst_width,
+                     int x,
+                     int dx) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst[0] = src[x >> 16];
@@ -782,11 +1154,14 @@ void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
   }
 }
 
-void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
-                       int dst_width, int x32, int dx) {
-  int64 x = (int64)(x32);
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
+void ScaleARGBCols64_C(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x32,
+                       int dx) {
+  int64_t x = (int64_t)(x32);
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst[0] = src[x >> 16];
@@ -801,11 +1176,16 @@ void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
 }
 
 // Scales a single row of pixels up by 2x using point sampling.
-void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx) {
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
+void ScaleARGBColsUp2_C(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
   int j;
+  (void)x;
+  (void)dx;
   for (j = 0; j < dst_width - 1; j += 2) {
     dst[1] = dst[0] = src[0];
     src += 1;
@@ -818,23 +1198,371 @@ void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
 
 // TODO(fbarchard): Replace 0x7f ^ f with 128-f.  bug=607.
 // Mimics SSSE3 blender
-#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
-#define BLENDERC(a, b, f, s) (uint32)( \
-    BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
-#define BLENDER(a, b, f) \
-    BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
-    BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b)*f) >> 7
+#define BLENDERC(a, b, f, s) \
+  (uint32_t)(BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f)                                                 \
+  BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | BLENDERC(a, b, f, 8) | \
+      BLENDERC(a, b, f, 0)
+
+void ScaleARGBFilterCols_C(uint8_t* dst_argb,
+                           const uint8_t* src_argb,
+                           int dst_width,
+                           int x,
+                           int dx) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32_t a = src[xi];
+    uint32_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32_t a = src[xi];
+    uint32_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleARGBFilterCols64_C(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x32,
+                             int dx) {
+  int64_t x = (int64_t)(x32);
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64_t xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32_t a = src[xi];
+    uint32_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64_t xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32_t a = src[xi];
+    uint32_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
+// UV scale row functions
+// same as ARGB but 2 channels
+
+void ScaleUVRowDown2_C(const uint8_t* src_uv,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst_uv,
+                       int dst_width) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int x;
+  (void)src_stride;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleUVRowDown2Linear_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_uv,
+                             int dst_width) {
+  int x;
+  (void)src_stride;
+  for (x = 0; x < dst_width; ++x) {
+    dst_uv[0] = (src_uv[0] + src_uv[2] + 1) >> 1;
+    dst_uv[1] = (src_uv[1] + src_uv[3] + 1) >> 1;
+    src_uv += 4;
+    dst_uv += 2;
+  }
+}
+
+void ScaleUVRowDown2Box_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_uv,
+                          int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_uv[0] = (src_uv[0] + src_uv[2] + src_uv[src_stride] +
+                 src_uv[src_stride + 2] + 2) >>
+                2;
+    dst_uv[1] = (src_uv[1] + src_uv[3] + src_uv[src_stride + 1] +
+                 src_uv[src_stride + 3] + 2) >>
+                2;
+    src_uv += 4;
+    dst_uv += 2;
+  }
+}
+
+void ScaleUVRowDownEven_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          int src_stepx,
+                          uint8_t* dst_uv,
+                          int dst_width) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  (void)src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
 
-void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
-                           int dst_width, int x, int dx) {
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
+void ScaleUVRowDownEvenBox_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             int src_stepx,
+                             uint8_t* dst_uv,
+                             int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_uv[0] = (src_uv[0] + src_uv[2] + src_uv[src_stride] +
+                 src_uv[src_stride + 2] + 2) >>
+                2;
+    dst_uv[1] = (src_uv[1] + src_uv[3] + src_uv[src_stride + 1] +
+                 src_uv[src_stride + 3] + 2) >>
+                2;
+    src_uv += src_stepx * 2;
+    dst_uv += 2;
+  }
+}
+
+void ScaleUVRowUp2_Linear_C(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[4 * x + 0] =
+        (src_ptr[2 * x + 0] * 3 + src_ptr[2 * x + 2] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 1] =
+        (src_ptr[2 * x + 1] * 3 + src_ptr[2 * x + 3] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 2] =
+        (src_ptr[2 * x + 0] * 1 + src_ptr[2 * x + 2] * 3 + 2) >> 2;
+    dst_ptr[4 * x + 3] =
+        (src_ptr[2 * x + 1] * 1 + src_ptr[2 * x + 3] * 3 + 2) >> 2;
+  }
+}
+
+void ScaleUVRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              ptrdiff_t dst_stride,
+                              int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  uint8_t* d = dst_ptr;
+  uint8_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[4 * x + 0] = (s[2 * x + 0] * 9 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 1 + 8) >>
+                   4;
+    d[4 * x + 1] = (s[2 * x + 1] * 9 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 1 + 8) >>
+                   4;
+    d[4 * x + 2] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 9 + t[2 * x + 0] * 1 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    d[4 * x + 3] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 9 + t[2 * x + 1] * 1 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 0] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 1 + t[2 * x + 0] * 9 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    e[4 * x + 1] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 1 + t[2 * x + 1] * 9 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 2] = (s[2 * x + 0] * 1 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 9 + 8) >>
+                   4;
+    e[4 * x + 3] = (s[2 * x + 1] * 1 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 9 + 8) >>
+                   4;
+  }
+}
+
+void ScaleUVRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
+                               int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[4 * x + 0] =
+        (src_ptr[2 * x + 0] * 3 + src_ptr[2 * x + 2] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 1] =
+        (src_ptr[2 * x + 1] * 3 + src_ptr[2 * x + 3] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 2] =
+        (src_ptr[2 * x + 0] * 1 + src_ptr[2 * x + 2] * 3 + 2) >> 2;
+    dst_ptr[4 * x + 3] =
+        (src_ptr[2 * x + 1] * 1 + src_ptr[2 * x + 3] * 3 + 2) >> 2;
+  }
+}
+
+void ScaleUVRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint16_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+  uint16_t* d = dst_ptr;
+  uint16_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[4 * x + 0] = (s[2 * x + 0] * 9 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 1 + 8) >>
+                   4;
+    d[4 * x + 1] = (s[2 * x + 1] * 9 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 1 + 8) >>
+                   4;
+    d[4 * x + 2] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 9 + t[2 * x + 0] * 1 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    d[4 * x + 3] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 9 + t[2 * x + 1] * 1 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 0] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 1 + t[2 * x + 0] * 9 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    e[4 * x + 1] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 1 + t[2 * x + 1] * 9 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 2] = (s[2 * x + 0] * 1 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 9 + 8) >>
+                   4;
+    e[4 * x + 3] = (s[2 * x + 1] * 1 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 9 + 8) >>
+                   4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleUVCols_C(uint8_t* dst_uv,
+                   const uint8_t* src_uv,
+                   int dst_width,
+                   int x,
+                   int dx) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleUVCols64_C(uint8_t* dst_uv,
+                     const uint8_t* src_uv,
+                     int dst_width,
+                     int x32,
+                     int dx) {
+  int64_t x = (int64_t)(x32);
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleUVColsUp2_C(uint8_t* dst_uv,
+                      const uint8_t* src_uv,
+                      int dst_width,
+                      int x,
+                      int dx) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  (void)x;
+  (void)dx;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// TODO(fbarchard): Replace 0x7f ^ f with 128-f.  bug=607.
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b)*f) >> 7
+#define BLENDERC(a, b, f, s) \
+  (uint16_t)(BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleUVFilterCols_C(uint8_t* dst_uv,
+                         const uint8_t* src_uv,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
     int xi = x >> 16;
     int xf = (x >> 9) & 0x7f;
-    uint32 a = src[xi];
-    uint32 b = src[xi + 1];
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
     dst[0] = BLENDER(a, b, xf);
     x += dx;
     xi = x >> 16;
@@ -848,23 +1576,26 @@ void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
   if (dst_width & 1) {
     int xi = x >> 16;
     int xf = (x >> 9) & 0x7f;
-    uint32 a = src[xi];
-    uint32 b = src[xi + 1];
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
     dst[0] = BLENDER(a, b, xf);
   }
 }
 
-void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
-                             int dst_width, int x32, int dx) {
-  int64 x = (int64)(x32);
-  const uint32* src = (const uint32*)(src_argb);
-  uint32* dst = (uint32*)(dst_argb);
+void ScaleUVFilterCols64_C(uint8_t* dst_uv,
+                           const uint8_t* src_uv,
+                           int dst_width,
+                           int x32,
+                           int dx) {
+  int64_t x = (int64_t)(x32);
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
   int j;
   for (j = 0; j < dst_width - 1; j += 2) {
-    int64 xi = x >> 16;
+    int64_t xi = x >> 16;
     int xf = (x >> 9) & 0x7f;
-    uint32 a = src[xi];
-    uint32 b = src[xi + 1];
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
     dst[0] = BLENDER(a, b, xf);
     x += dx;
     xi = x >> 16;
@@ -876,10 +1607,10 @@ void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
     dst += 2;
   }
   if (dst_width & 1) {
-    int64 xi = x >> 16;
+    int64_t xi = x >> 16;
     int xf = (x >> 9) & 0x7f;
-    uint32 a = src[xi];
-    uint32 b = src[xi + 1];
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
     dst[0] = BLENDER(a, b, xf);
   }
 }
@@ -889,16 +1620,22 @@ void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
 
 // Scale plane vertically with bilinear interpolation.
 void ScalePlaneVertical(int src_height,
-                        int dst_width, int dst_height,
-                        int src_stride, int dst_stride,
-                        const uint8* src_argb, uint8* dst_argb,
-                        int x, int y, int dy,
-                        int bpp, enum FilterMode filtering) {
+                        int dst_width,
+                        int dst_height,
+                        int src_stride,
+                        int dst_stride,
+                        const uint8_t* src_argb,
+                        uint8_t* dst_argb,
+                        int x,
+                        int y,
+                        int dy,
+                        int bpp,  // bytes per pixel. 4 for ARGB.
+                        enum FilterMode filtering) {
   // TODO(fbarchard): Allow higher bpp.
   int dst_width_bytes = dst_width * bpp;
-  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_C;
+  void (*InterpolateRow)(uint8_t* dst_argb, const uint8_t* src_argb,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
   const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
   int j;
   assert(bpp >= 1 && bpp <= 4);
@@ -930,13 +1667,19 @@ void ScalePlaneVertical(int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
-    InterpolateRow = InterpolateRow_Any_DSPR2;
-    if (IS_ALIGNED(dst_width_bytes, 4)) {
-      InterpolateRow = InterpolateRow_DSPR2;
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -948,23 +1691,30 @@ void ScalePlaneVertical(int src_height,
     }
     yi = y >> 16;
     yf = filtering ? ((y >> 8) & 255) : 0;
-    InterpolateRow(dst_argb, src_argb + yi * src_stride,
-                   src_stride, dst_width_bytes, yf);
+    InterpolateRow(dst_argb, src_argb + yi * src_stride, src_stride,
+                   dst_width_bytes, yf);
     dst_argb += dst_stride;
     y += dy;
   }
 }
+
 void ScalePlaneVertical_16(int src_height,
-                           int dst_width, int dst_height,
-                           int src_stride, int dst_stride,
-                           const uint16* src_argb, uint16* dst_argb,
-                           int x, int y, int dy,
-                           int wpp, enum FilterMode filtering) {
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint16_t* src_argb,
+                           uint16_t* dst_argb,
+                           int x,
+                           int y,
+                           int dy,
+                           int wpp, /* words per pixel. normally 1 */
+                           enum FilterMode filtering) {
   // TODO(fbarchard): Allow higher wpp.
   int dst_width_words = dst_width * wpp;
-  void (*InterpolateRow)(uint16* dst_argb, const uint16* src_argb,
-      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
-      InterpolateRow_16_C;
+  void (*InterpolateRow)(uint16_t* dst_argb, const uint16_t* src_argb,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_16_C;
   const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
   int j;
   assert(wpp >= 1 && wpp <= 2);
@@ -974,43 +1724,97 @@ void ScalePlaneVertical_16(int src_height,
   src_argb += (x >> 16) * wpp;
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    InterpolateRow = InterpolateRow_Any_16_SSE2;
-    if (IS_ALIGNED(dst_width_bytes, 16)) {
+    InterpolateRow = InterpolateRow_16_Any_SSE2;
+    if (IS_ALIGNED(dst_width_words, 16)) {
       InterpolateRow = InterpolateRow_16_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_16_SSSE3;
-    if (IS_ALIGNED(dst_width_bytes, 16)) {
+    InterpolateRow = InterpolateRow_16_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_words, 16)) {
       InterpolateRow = InterpolateRow_16_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_16_AVX2;
-    if (IS_ALIGNED(dst_width_bytes, 32)) {
+    InterpolateRow = InterpolateRow_16_Any_AVX2;
+    if (IS_ALIGNED(dst_width_words, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_16_NEON;
-    if (IS_ALIGNED(dst_width_bytes, 16)) {
+    InterpolateRow = InterpolateRow_16_Any_NEON;
+    if (IS_ALIGNED(dst_width_words, 8)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_16_DSPR2)
-  if (TestCpuFlag(kCpuHasDSPR2) &&
-      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
-      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
-    InterpolateRow = InterpolateRow_Any_16_DSPR2;
-    if (IS_ALIGNED(dst_width_bytes, 4)) {
-      InterpolateRow = InterpolateRow_16_DSPR2;
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride, src_stride,
+                   dst_width_words, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+void ScalePlaneVertical_16To8(int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint16_t* src_argb,
+                              uint8_t* dst_argb,
+                              int x,
+                              int y,
+                              int dy,
+                              int wpp, /* words per pixel. normally 1 */
+                              int scale,
+                              enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher wpp.
+  int dst_width_words = dst_width * wpp;
+  // TODO(https://crbug.com/libyuv/931): Add NEON 32 bit and AVX2 versions.
+  void (*InterpolateRow_16To8)(uint8_t* dst_argb, const uint16_t* src_argb,
+                               ptrdiff_t src_stride, int scale, int dst_width,
+                               int source_y_fraction) = InterpolateRow_16To8_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(wpp >= 1 && wpp <= 2);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * wpp;
+
+#if defined(HAS_INTERPOLATEROW_16TO8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow_16To8 = InterpolateRow_16To8_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow_16To8 = InterpolateRow_16To8_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16TO8_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow_16To8 = InterpolateRow_16To8_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow_16To8 = InterpolateRow_16To8_AVX2;
     }
   }
 #endif
@@ -1022,16 +1826,18 @@ void ScalePlaneVertical_16(int src_height,
     }
     yi = y >> 16;
     yf = filtering ? ((y >> 8) & 255) : 0;
-    InterpolateRow(dst_argb, src_argb + yi * src_stride,
-                   src_stride, dst_width_words, yf);
+    InterpolateRow_16To8(dst_argb, src_argb + yi * src_stride, src_stride,
+                         scale, dst_width_words, yf);
     dst_argb += dst_stride;
     y += dy;
   }
 }
 
 // Simplify the filtering based on scale factors.
-enum FilterMode ScaleFilterReduce(int src_width, int src_height,
-                                  int dst_width, int dst_height,
+enum FilterMode ScaleFilterReduce(int src_width,
+                                  int src_height,
+                                  int dst_width,
+                                  int dst_height,
                                   enum FilterMode filtering) {
   if (src_width < 0) {
     src_width = -src_width;
@@ -1040,8 +1846,8 @@ enum FilterMode ScaleFilterReduce(int src_width, int src_height,
     src_height = -src_height;
   }
   if (filtering == kFilterBox) {
-    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
-    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+    // If scaling either axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width || dst_height * 2 >= src_height) {
       filtering = kFilterBilinear;
     }
   }
@@ -1073,22 +1879,26 @@ enum FilterMode ScaleFilterReduce(int src_width, int src_height,
 
 // Divide num by div and return as 16.16 fixed point result.
 int FixedDiv_C(int num, int div) {
-  return (int)(((int64)(num) << 16) / div);
+  return (int)(((int64_t)(num) << 16) / div);
 }
 
-// Divide num by div and return as 16.16 fixed point result.
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
 int FixedDiv1_C(int num, int div) {
-  return (int)((((int64)(num) << 16) - 0x00010001) /
-                          (div - 1));
+  return (int)((((int64_t)(num) << 16) - 0x00010001) / (div - 1));
 }
 
 #define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
 
 // Compute slope values for stepping.
-void ScaleSlope(int src_width, int src_height,
-                int dst_width, int dst_height,
+void ScaleSlope(int src_width,
+                int src_height,
+                int dst_width,
+                int dst_height,
                 enum FilterMode filtering,
-                int* x, int* y, int* dx, int* dy) {
+                int* x,
+                int* y,
+                int* dx,
+                int* dy) {
   assert(x != NULL);
   assert(y != NULL);
   assert(dx != NULL);
@@ -1115,14 +1925,14 @@ void ScaleSlope(int src_width, int src_height,
     if (dst_width <= Abs(src_width)) {
       *dx = FixedDiv(Abs(src_width), dst_width);
       *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
-    } else if (dst_width > 1) {
+    } else if (src_width > 1 && dst_width > 1) {
       *dx = FixedDiv1(Abs(src_width), dst_width);
       *x = 0;
     }
     if (dst_height <= src_height) {
-      *dy = FixedDiv(src_height,  dst_height);
+      *dy = FixedDiv(src_height, dst_height);
       *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
-    } else if (dst_height > 1) {
+    } else if (src_height > 1 && dst_height > 1) {
       *dy = FixedDiv1(src_height, dst_height);
       *y = 0;
     }
@@ -1131,7 +1941,7 @@ void ScaleSlope(int src_width, int src_height,
     if (dst_width <= Abs(src_width)) {
       *dx = FixedDiv(Abs(src_width), dst_width);
       *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
-    } else if (dst_width > 1) {
+    } else if (src_width > 1 && dst_width > 1) {
       *dx = FixedDiv1(Abs(src_width), dst_width);
       *x = 0;
     }
@@ -1153,6 +1963,35 @@ void ScaleSlope(int src_width, int src_height,
 }
 #undef CENTERSTART
 
+// Read 8x2 upsample with filtering and write 16x1.
+// actually reads an extra pixel, so 9x2.
+void ScaleRowUp2_16_C(const uint16_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint16_t* dst,
+                      int dst_width) {
+  const uint16_t* src2 = src_ptr + src_stride;
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    uint16_t p0 = src_ptr[0];
+    uint16_t p1 = src_ptr[1];
+    uint16_t p2 = src2[0];
+    uint16_t p3 = src2[1];
+    dst[0] = (p0 * 9 + p1 * 3 + p2 * 3 + p3 + 8) >> 4;
+    dst[1] = (p0 * 3 + p1 * 9 + p2 + p3 * 3 + 8) >> 4;
+    ++src_ptr;
+    ++src2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    uint16_t p0 = src_ptr[0];
+    uint16_t p1 = src_ptr[1];
+    uint16_t p2 = src2[0];
+    uint16_t p3 = src2[1];
+    dst[0] = (p0 * 9 + p1 * 3 + p2 * 3 + p3 + 8) >> 4;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/media/libyuv/source/scale_gcc.cc b/media/libyuv/source/scale_gcc.cc
index e2f88544b7..17eeffadfb 100644
--- a/media/libyuv/source/scale_gcc.cc
+++ b/media/libyuv/source/scale_gcc.cc
@@ -17,1303 +17,2934 @@ extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 // Offsets for source bytes 0 to 9
-static uvec8 kShuf0 =
-  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf0 = {0,   1,   3,   4,   5,   7,   8,   9,
+                             128, 128, 128, 128, 128, 128, 128, 128};
 
 // Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
-static uvec8 kShuf1 =
-  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf1 = {3,   4,   5,   7,   8,   9,   11,  12,
+                             128, 128, 128, 128, 128, 128, 128, 128};
 
 // Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-static uvec8 kShuf2 =
-  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf2 = {5,   7,   8,   9,   11,  12,  13,  15,
+                             128, 128, 128, 128, 128, 128, 128, 128};
 
 // Offsets for source bytes 0 to 10
-static uvec8 kShuf01 =
-  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+static const uvec8 kShuf01 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
 
 // Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
-static uvec8 kShuf11 =
-  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+static const uvec8 kShuf11 = {2, 3, 4, 5,  5,  6,  6,  7,
+                              8, 9, 9, 10, 10, 11, 12, 13};
 
 // Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-static uvec8 kShuf21 =
-  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+static const uvec8 kShuf21 = {5,  6,  6,  7,  8,  9,  9,  10,
+                              10, 11, 12, 13, 13, 14, 14, 15};
 
 // Coefficients for source bytes 0 to 10
-static uvec8 kMadd01 =
-  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+static const uvec8 kMadd01 = {3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2};
 
 // Coefficients for source bytes 10 to 21
-static uvec8 kMadd11 =
-  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+static const uvec8 kMadd11 = {1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1};
 
 // Coefficients for source bytes 21 to 31
-static uvec8 kMadd21 =
-  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+static const uvec8 kMadd21 = {2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3};
 
 // Coefficients for source bytes 21 to 31
-static vec16 kRound34 =
-  { 2, 2, 2, 2, 2, 2, 2, 2 };
+static const vec16 kRound34 = {2, 2, 2, 2, 2, 2, 2, 2};
 
-static uvec8 kShuf38a =
-  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf38a = {0,   3,   6,   8,   11,  14,  128, 128,
+                               128, 128, 128, 128, 128, 128, 128, 128};
 
-static uvec8 kShuf38b =
-  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+static const uvec8 kShuf38b = {128, 128, 128, 128, 128, 128, 0,   3,
+                               6,   8,   11,  14,  128, 128, 128, 128};
 
 // Arrange words 0,3,6 into 0,1,2
-static uvec8 kShufAc =
-  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAc = {0,   1,   6,   7,   12,  13,  128, 128,
+                              128, 128, 128, 128, 128, 128, 128, 128};
 
 // Arrange words 0,3,6 into 3,4,5
-static uvec8 kShufAc3 =
-  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+static const uvec8 kShufAc3 = {128, 128, 128, 128, 128, 128, 0,   1,
+                               6,   7,   12,  13,  128, 128, 128, 128};
 
 // Scaling values for boxes of 3x3 and 2x3
-static uvec16 kScaleAc33 =
-  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+static const uvec16 kScaleAc33 = {65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9,
+                                  65536 / 9, 65536 / 6, 0,         0};
 
 // Arrange first value for pixels 0,1,2,3,4,5
-static uvec8 kShufAb0 =
-  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAb0 = {0,  128, 3,  128, 6,   128, 8,   128,
+                               11, 128, 14, 128, 128, 128, 128, 128};
 
 // Arrange second value for pixels 0,1,2,3,4,5
-static uvec8 kShufAb1 =
-  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAb1 = {1,  128, 4,  128, 7,   128, 9,   128,
+                               12, 128, 15, 128, 128, 128, 128, 128};
 
 // Arrange third value for pixels 0,1,2,3,4,5
-static uvec8 kShufAb2 =
-  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAb2 = {2,  128, 5,   128, 128, 128, 10,  128,
+                               13, 128, 128, 128, 128, 128, 128, 128};
 
 // Scaling values for boxes of 3x2 and 2x2
-static uvec16 kScaleAb2 =
-  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+static const uvec16 kScaleAb2 = {65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3,
+                                 65536 / 3, 65536 / 2, 0,         0};
 
 // GCC versions of row functions are verbatim conversions from Visual C.
 // Generated using gcc disassembly on Visual C object file:
 // objdump -D yuvscaler.obj >yuvscaler.txt
 
-void ScaleRowDown2_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "psrlw     $0x8,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1"
-  );
+void ScaleRowDown2_SSSE3(const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      // 16 pixel loop.
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1");
 }
 
-void ScaleRowDown2Linear_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "pcmpeqb    %%xmm4,%%xmm4                  \n"
-    "psrlw      $0xf,%%xmm4                    \n"
-    "packuswb   %%xmm4,%%xmm4                  \n"
-    "pxor       %%xmm5,%%xmm5                  \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pmaddubsw  %%xmm4,%%xmm0                  \n"
-    "pmaddubsw  %%xmm4,%%xmm1                  \n"
-    "pavgw      %%xmm5,%%xmm0                  \n"
-    "pavgw      %%xmm5,%%xmm1                  \n"
-    "packuswb   %%xmm1,%%xmm0                  \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
-  );
+void ScaleRowDown2Linear_SSSE3(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm4", "xmm5");
 }
 
-void ScaleRowDown2Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "pcmpeqb    %%xmm4,%%xmm4                  \n"
-    "psrlw      $0xf,%%xmm4                    \n"
-    "packuswb   %%xmm4,%%xmm4                  \n"
-    "pxor       %%xmm5,%%xmm5                  \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pmaddubsw  %%xmm4,%%xmm0                  \n"
-    "pmaddubsw  %%xmm4,%%xmm1                  \n"
-    "pmaddubsw  %%xmm4,%%xmm2                  \n"
-    "pmaddubsw  %%xmm4,%%xmm3                  \n"
-    "paddw      %%xmm2,%%xmm0                  \n"
-    "paddw      %%xmm3,%%xmm1                  \n"
-    "psrlw      $0x1,%%xmm0                    \n"
-    "psrlw      $0x1,%%xmm1                    \n"
-    "pavgw      %%xmm5,%%xmm0                  \n"
-    "pavgw      %%xmm5,%%xmm1                  \n"
-    "packuswb   %%xmm1,%%xmm0                  \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void ScaleRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x1,%%xmm0                   \n"
+      "psrlw       $0x1,%%xmm1                   \n"
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),               // %0
+        "+r"(dst_ptr),               // %1
+        "+r"(dst_width)              // %2
+      : "r"((intptr_t)(src_stride))  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 
 #ifdef HAS_SCALEROWDOWN2_AVX2
-void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    "lea        " MEMLEA(0x40,0) ",%0          \n"
-    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "sub        $0x20,%2                       \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1"
-  );
+void ScaleRowDown2_AVX2(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
-void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
-    "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20, 0) ",%%ymm1 \n"
-    "lea        " MEMLEA(0x40,0) ",%0          \n"
-    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpavgw     %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpavgw     %%ymm5,%%ymm1,%%ymm1           \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "sub        $0x20,%2                       \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
-  );
+void ScaleRowDown2Linear_AVX2(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpavgw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm4", "xmm5");
 }
 
-void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
-    "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    MEMOPREG(vmovdqu,0x00,0,3,1,ymm2)          //  vmovdqu  (%0,%3,1),%%ymm2
-    MEMOPREG(vmovdqu,0x20,0,3,1,ymm3)          //  vmovdqu  0x20(%0,%3,1),%%ymm3
-    "lea        " MEMLEA(0x40,0) ",%0          \n"
-    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-    "vpsrlw     $0x1,%%ymm0,%%ymm0             \n"
-    "vpsrlw     $0x1,%%ymm1,%%ymm1             \n"
-    "vpavgw     %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpavgw     %%ymm5,%%ymm1,%%ymm1           \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "lea        " MEMLEA(0x20,1) ",%1          \n"
-    "sub        $0x20,%2                       \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void ScaleRowDown2Box_AVX2(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x00(%0,%3,1),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%3,1),%%ymm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpsrlw      $0x1,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x1,%%ymm1,%%ymm1            \n"
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpavgw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),               // %0
+        "+r"(dst_ptr),               // %1
+        "+r"(dst_width)              // %2
+      : "r"((intptr_t)(src_stride))  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SCALEROWDOWN2_AVX2
 
-void ScaleRowDown4_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrld     $0x18,%%xmm5                    \n"
-    "pslld     $0x10,%%xmm5                    \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pand      %%xmm5,%%xmm0                   \n"
-    "pand      %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm0                   \n"
-    "psrlw     $0x8,%%xmm0                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
-  );
+void ScaleRowDown4_SSSE3(const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrld       $0x18,%%xmm5                  \n"
+      "pslld       $0x10,%%xmm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void ScaleRowDown4Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
+void ScaleRowDown4Box_SSSE3(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width) {
   intptr_t stridex3;
-  asm volatile (
-    "pcmpeqb    %%xmm4,%%xmm4                  \n"
-    "psrlw      $0xf,%%xmm4                    \n"
-    "movdqa     %%xmm4,%%xmm5                  \n"
-    "packuswb   %%xmm4,%%xmm4                  \n"
-    "psllw      $0x3,%%xmm5                    \n"
-    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
-    "pmaddubsw  %%xmm4,%%xmm0                  \n"
-    "pmaddubsw  %%xmm4,%%xmm1                  \n"
-    "pmaddubsw  %%xmm4,%%xmm2                  \n"
-    "pmaddubsw  %%xmm4,%%xmm3                  \n"
-    "paddw      %%xmm2,%%xmm0                  \n"
-    "paddw      %%xmm3,%%xmm1                  \n"
-    MEMOPREG(movdqu,0x00,0,4,2,xmm2)           //  movdqu  (%0,%4,2),%%xmm2
-    MEMOPREG(movdqu,0x10,0,4,2,xmm3)           //  movdqu  0x10(%0,%4,2),%%xmm3
-    "pmaddubsw  %%xmm4,%%xmm2                  \n"
-    "pmaddubsw  %%xmm4,%%xmm3                  \n"
-    "paddw      %%xmm2,%%xmm0                  \n"
-    "paddw      %%xmm3,%%xmm1                  \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pmaddubsw  %%xmm4,%%xmm2                  \n"
-    "pmaddubsw  %%xmm4,%%xmm3                  \n"
-    "paddw      %%xmm2,%%xmm0                  \n"
-    "paddw      %%xmm3,%%xmm1                  \n"
-    "phaddw     %%xmm1,%%xmm0                  \n"
-    "paddw      %%xmm5,%%xmm0                  \n"
-    "psrlw      $0x4,%%xmm0                    \n"
-    "packuswb   %%xmm0,%%xmm0                  \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x8,1) ",%1            \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width),   // %2
-    "=&r"(stridex3)    // %3
-  : "r"((intptr_t)(src_stride))    // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "psllw       $0x3,%%xmm5                   \n"
+      "lea         0x00(%4,%4,2),%3              \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm3          \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "movdqu      0x00(%0,%4,2),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,2),%%xmm3          \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "phaddw      %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm5,%%xmm0                 \n"
+      "psrlw       $0x4,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),               // %0
+        "+r"(dst_ptr),               // %1
+        "+r"(dst_width),             // %2
+        "=&r"(stridex3)              // %3
+      : "r"((intptr_t)(src_stride))  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-
 #ifdef HAS_SCALEROWDOWN4_AVX2
-void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-    "vpsrld     $0x18,%%ymm5,%%ymm5            \n"
-    "vpslld     $0x10,%%ymm5,%%ymm5            \n"
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    "lea        " MEMLEA(0x40,0) ",%0          \n"
-    "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpand      %%ymm5,%%ymm1,%%ymm1           \n"
-    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%xmm0," MEMACCESS(1) "        \n"
-    "lea        " MEMLEA(0x10,1) ",%1          \n"
-    "sub        $0x10,%2                       \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
-  );
+void ScaleRowDown4_AVX2(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrld      $0x18,%%ymm5,%%ymm5           \n"
+      "vpslld      $0x10,%%ymm5,%%ymm5           \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm5");
 }
 
-void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-    "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
-    "vpsllw     $0x3,%%ymm4,%%ymm5             \n"
-    "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
-    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
-    MEMOPREG(vmovdqu,0x00,0,3,1,ymm2)          //  vmovdqu  (%0,%3,1),%%ymm2
-    MEMOPREG(vmovdqu,0x20,0,3,1,ymm3)          //  vmovdqu  0x20(%0,%3,1),%%ymm3
-    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-    MEMOPREG(vmovdqu,0x00,0,3,2,ymm2)          //  vmovdqu  (%0,%3,2),%%ymm2
-    MEMOPREG(vmovdqu,0x20,0,3,2,ymm3)          //  vmovdqu  0x20(%0,%3,2),%%ymm3
-    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-    MEMOPREG(vmovdqu,0x00,0,4,1,ymm2)          //  vmovdqu  (%0,%4,1),%%ymm2
-    MEMOPREG(vmovdqu,0x20,0,4,1,ymm3)          //  vmovdqu  0x20(%0,%4,1),%%ymm3
-    "lea        " MEMLEA(0x40,0) ",%0          \n"
-    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"
-    "vpsrlw     $0x4,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%xmm0," MEMACCESS(1) "        \n"
-    "lea        " MEMLEA(0x10,1) ",%1          \n"
-    "sub        $0x10,%2                       \n"
-    "jg         1b                             \n"
-    "vzeroupper                                \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"((intptr_t)(src_stride * 3))   // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
-  );
+void ScaleRowDown4Box_AVX2(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpsllw      $0x3,%%ymm4,%%ymm5            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x00(%0,%3,1),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%3,1),%%ymm3          \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     0x00(%0,%3,2),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%3,2),%%ymm3          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     0x00(%0,%4,1),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%4,1),%%ymm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpaddw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpsrlw      $0x4,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                   // %0
+        "+r"(dst_ptr),                   // %1
+        "+r"(dst_width)                  // %2
+      : "r"((intptr_t)(src_stride)),     // %3
+        "r"((intptr_t)(src_stride * 3))  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 #endif  // HAS_SCALEROWDOWN4_AVX2
 
-void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movdqa    %0,%%xmm3                       \n"
-    "movdqa    %1,%%xmm4                       \n"
-    "movdqa    %2,%%xmm5                       \n"
-  :
-  : "m"(kShuf0),  // %0
-    "m"(kShuf1),  // %1
-    "m"(kShuf2)   // %2
-  );
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqa    %%xmm2,%%xmm1                   \n"
-    "palignr   $0x8,%%xmm0,%%xmm1              \n"
-    "pshufb    %%xmm3,%%xmm0                   \n"
-    "pshufb    %%xmm4,%%xmm1                   \n"
-    "pshufb    %%xmm5,%%xmm2                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
-    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x18,1) ",%1           \n"
-    "sub       $0x18,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),   // %0
-    "+r"(dst_ptr),   // %1
-    "+r"(dst_width)  // %2
-  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+void ScaleRowDown34_SSSE3(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_ptr,
+                          int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "movdqa      %0,%%xmm3                     \n"
+      "movdqa      %1,%%xmm4                     \n"
+      "movdqa      %2,%%xmm5                     \n"
+      :
+      : "m"(kShuf0),  // %0
+        "m"(kShuf1),  // %1
+        "m"(kShuf2)   // %2
   );
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm2               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "palignr     $0x8,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm3,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x8(%1)                \n"
+      "movq        %%xmm2,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x18,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
-void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+void ScaleRowDown34_1_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movdqa    %0,%%xmm2                       \n"  // kShuf01
-    "movdqa    %1,%%xmm3                       \n"  // kShuf11
-    "movdqa    %2,%%xmm4                       \n"  // kShuf21
-  :
-  : "m"(kShuf01),  // %0
-    "m"(kShuf11),  // %1
-    "m"(kShuf21)   // %2
+                                uint8_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "movdqa      %0,%%xmm2                     \n"  // kShuf01
+      "movdqa      %1,%%xmm3                     \n"  // kShuf11
+      "movdqa      %2,%%xmm4                     \n"  // kShuf21
+      :
+      : "m"(kShuf01),  // %0
+        "m"(kShuf11),  // %1
+        "m"(kShuf21)   // %2
   );
-  asm volatile (
-    "movdqa    %0,%%xmm5                       \n"  // kMadd01
-    "movdqa    %1,%%xmm0                       \n"  // kMadd11
-    "movdqa    %2,%%xmm1                       \n"  // kRound34
-  :
-  : "m"(kMadd01),  // %0
-    "m"(kMadd11),  // %1
-    "m"(kRound34)  // %2
-  );
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "pshufb    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm5,%%xmm6                   \n"
-    "paddsw    %%xmm1,%%xmm6                   \n"
-    "psrlw     $0x2,%%xmm6                     \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movq      %%xmm6," MEMACCESS(1) "         \n"
-    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
-    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "pshufb    %%xmm3,%%xmm6                   \n"
-    "pmaddubsw %%xmm0,%%xmm6                   \n"
-    "paddsw    %%xmm1,%%xmm6                   \n"
-    "psrlw     $0x2,%%xmm6                     \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3),%%xmm7
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "pshufb    %%xmm4,%%xmm6                   \n"
-    "pmaddubsw %4,%%xmm6                       \n"
-    "paddsw    %%xmm1,%%xmm6                   \n"
-    "psrlw     $0x2,%%xmm6                     \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x18,1) ",%1           \n"
-    "sub       $0x18,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),   // %0
-    "+r"(dst_ptr),   // %1
-    "+r"(dst_width)  // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "m"(kMadd21)     // %4
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  asm volatile(
+      "movdqa      %0,%%xmm5                     \n"  // kMadd01
+      "movdqa      %1,%%xmm0                     \n"  // kMadd11
+      "movdqa      %2,%%xmm1                     \n"  // kRound34
+      :
+      : "m"(kMadd01),  // %0
+        "m"(kMadd11),  // %1
+        "m"(kRound34)  // %2
   );
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm5,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,(%1)                   \n"
+      "movdqu      0x8(%0),%%xmm6                \n"
+      "movdqu      0x8(%0,%3,1),%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm3,%%xmm6                 \n"
+      "pmaddubsw   %%xmm0,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x8(%1)                \n"
+      "movdqu      0x10(%0),%%xmm6               \n"
+      "movdqu      0x10(%0,%3,1),%%xmm7          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm4,%%xmm6                 \n"
+      "pmaddubsw   %4,%%xmm6                     \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x18,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),                // %0
+                 "+r"(dst_ptr),                // %1
+                 "+r"(dst_width)               // %2
+               : "r"((intptr_t)(src_stride)),  // %3
+                 "m"(kMadd21)                  // %4
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6", "xmm7");
 }
 
-void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+void ScaleRowDown34_0_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movdqa    %0,%%xmm2                       \n"  // kShuf01
-    "movdqa    %1,%%xmm3                       \n"  // kShuf11
-    "movdqa    %2,%%xmm4                       \n"  // kShuf21
-  :
-  : "m"(kShuf01),  // %0
-    "m"(kShuf11),  // %1
-    "m"(kShuf21)   // %2
+                                uint8_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "movdqa      %0,%%xmm2                     \n"  // kShuf01
+      "movdqa      %1,%%xmm3                     \n"  // kShuf11
+      "movdqa      %2,%%xmm4                     \n"  // kShuf21
+      :
+      : "m"(kShuf01),  // %0
+        "m"(kShuf11),  // %1
+        "m"(kShuf21)   // %2
   );
-  asm volatile (
-    "movdqa    %0,%%xmm5                       \n"  // kMadd01
-    "movdqa    %1,%%xmm0                       \n"  // kMadd11
-    "movdqa    %2,%%xmm1                       \n"  // kRound34
-  :
-  : "m"(kMadd01),  // %0
-    "m"(kMadd11),  // %1
-    "m"(kRound34)  // %2
+  asm volatile(
+      "movdqa      %0,%%xmm5                     \n"  // kMadd01
+      "movdqa      %1,%%xmm0                     \n"  // kMadd11
+      "movdqa      %2,%%xmm1                     \n"  // kRound34
+      :
+      : "m"(kMadd01),  // %0
+        "m"(kMadd11),  // %1
+        "m"(kRound34)  // %2
   );
 
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3,1),%%xmm7
-    "pavgb     %%xmm6,%%xmm7                   \n"
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "pshufb    %%xmm2,%%xmm6                   \n"
-    "pmaddubsw %%xmm5,%%xmm6                   \n"
-    "paddsw    %%xmm1,%%xmm6                   \n"
-    "psrlw     $0x2,%%xmm6                     \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movq      %%xmm6," MEMACCESS(1) "         \n"
-    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
-    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
-    "pavgb     %%xmm6,%%xmm7                   \n"
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "pshufb    %%xmm3,%%xmm6                   \n"
-    "pmaddubsw %%xmm0,%%xmm6                   \n"
-    "paddsw    %%xmm1,%%xmm6                   \n"
-    "psrlw     $0x2,%%xmm6                     \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
-    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3,1),%%xmm7
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm6,%%xmm7                   \n"
-    "pavgb     %%xmm7,%%xmm6                   \n"
-    "pshufb    %%xmm4,%%xmm6                   \n"
-    "pmaddubsw %4,%%xmm6                       \n"
-    "paddsw    %%xmm1,%%xmm6                   \n"
-    "psrlw     $0x2,%%xmm6                     \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x18,1) ",%1           \n"
-    "sub       $0x18,%2                        \n"
-    "jg        1b                              \n"
-    : "+r"(src_ptr),   // %0
-      "+r"(dst_ptr),   // %1
-      "+r"(dst_width)  // %2
-    : "r"((intptr_t)(src_stride)),  // %3
-      "m"(kMadd21)     // %4
-    : "memory", "cc", NACL_R14
-      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm7          \n"
+      "pavgb       %%xmm6,%%xmm7                 \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm5,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,(%1)                   \n"
+      "movdqu      0x8(%0),%%xmm6                \n"
+      "movdqu      0x8(%0,%3,1),%%xmm7           \n"
+      "pavgb       %%xmm6,%%xmm7                 \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm3,%%xmm6                 \n"
+      "pmaddubsw   %%xmm0,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x8(%1)                \n"
+      "movdqu      0x10(%0),%%xmm6               \n"
+      "movdqu      0x10(%0,%3,1),%%xmm7          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm6,%%xmm7                 \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm4,%%xmm6                 \n"
+      "pmaddubsw   %4,%%xmm6                     \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x18,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),                // %0
+                 "+r"(dst_ptr),                // %1
+                 "+r"(dst_width)               // %2
+               : "r"((intptr_t)(src_stride)),  // %3
+                 "m"(kMadd21)                  // %4
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6", "xmm7");
 }
 
-void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movdqa    %3,%%xmm4                       \n"
-    "movdqa    %4,%%xmm5                       \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "paddusb   %%xmm1,%%xmm0                   \n"
-    "movq      %%xmm0," MEMACCESS(1) "         \n"
-    "movhlps   %%xmm0,%%xmm1                   \n"
-    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
-    "lea       " MEMLEA(0xc,1) ",%1            \n"
-    "sub       $0xc,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),   // %0
-    "+r"(dst_ptr),   // %1
-    "+r"(dst_width)  // %2
-  : "m"(kShuf38a),   // %3
-    "m"(kShuf38b)    // %4
-  : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
-  );
+void ScaleRowDown38_SSSE3(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_ptr,
+                          int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movhlps     %%xmm0,%%xmm1                 \n"
+      "movd        %%xmm1,0x8(%1)                \n"
+      "lea         0xc(%1),%1                    \n"
+      "sub         $0xc,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      : "m"(kShuf38a),   // %3
+        "m"(kShuf38b)    // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5");
 }
 
-void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+void ScaleRowDown38_2_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movdqa    %0,%%xmm2                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm4                       \n"
-    "movdqa    %3,%%xmm5                       \n"
-  :
-  : "m"(kShufAb0),   // %0
-    "m"(kShufAb1),   // %1
-    "m"(kShufAb2),   // %2
-    "m"(kScaleAb2)   // %3
-  );
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm1)           //  movdqu  (%0,%3,1),%%xmm1
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "pavgb     %%xmm1,%%xmm0                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "pshufb    %%xmm2,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm6                   \n"
-    "pshufb    %%xmm3,%%xmm6                   \n"
-    "paddusw   %%xmm6,%%xmm1                   \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "paddusw   %%xmm0,%%xmm1                   \n"
-    "pmulhuw   %%xmm5,%%xmm1                   \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movd      %%xmm1," MEMACCESS(1) "         \n"
-    "psrlq     $0x10,%%xmm1                    \n"
-    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
-    "lea       " MEMLEA(0x6,1) ",%1            \n"
-    "sub       $0x6,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride))  // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+                                uint8_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "movdqa      %0,%%xmm2                     \n"
+      "movdqa      %1,%%xmm3                     \n"
+      "movdqa      %2,%%xmm4                     \n"
+      "movdqa      %3,%%xmm5                     \n"
+      :
+      : "m"(kShufAb0),  // %0
+        "m"(kShufAb1),  // %1
+        "m"(kShufAb2),  // %2
+        "m"(kScaleAb2)  // %3
   );
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "pavgb       %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm6                 \n"
+      "pshufb      %%xmm3,%%xmm6                 \n"
+      "paddusw     %%xmm6,%%xmm1                 \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "paddusw     %%xmm0,%%xmm1                 \n"
+      "pmulhuw     %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movd        %%xmm1,(%1)                   \n"
+      "psrlq       $0x10,%%xmm1                  \n"
+      "movd        %%xmm1,0x2(%1)                \n"
+      "lea         0x6(%1),%1                    \n"
+      "sub         $0x6,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),               // %0
+                 "+r"(dst_ptr),               // %1
+                 "+r"(dst_width)              // %2
+               : "r"((intptr_t)(src_stride))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6");
 }
 
-void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+void ScaleRowDown38_3_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movdqa    %0,%%xmm2                       \n"
-    "movdqa    %1,%%xmm3                       \n"
-    "movdqa    %2,%%xmm4                       \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-  :
-  : "m"(kShufAc),    // %0
-    "m"(kShufAc3),   // %1
-    "m"(kScaleAc33)  // %2
-  );
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm6)           //  movdqu  (%0,%3,1),%%xmm6
-    "movhlps   %%xmm0,%%xmm1                   \n"
-    "movhlps   %%xmm6,%%xmm7                   \n"
-    "punpcklbw %%xmm5,%%xmm0                   \n"
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "punpcklbw %%xmm5,%%xmm6                   \n"
-    "punpcklbw %%xmm5,%%xmm7                   \n"
-    "paddusw   %%xmm6,%%xmm0                   \n"
-    "paddusw   %%xmm7,%%xmm1                   \n"
-    MEMOPREG(movdqu,0x00,0,3,2,xmm6)           //  movdqu  (%0,%3,2),%%xmm6
-    "lea       " MEMLEA(0x10,0) ",%0           \n"
-    "movhlps   %%xmm6,%%xmm7                   \n"
-    "punpcklbw %%xmm5,%%xmm6                   \n"
-    "punpcklbw %%xmm5,%%xmm7                   \n"
-    "paddusw   %%xmm6,%%xmm0                   \n"
-    "paddusw   %%xmm7,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm6                   \n"
-    "psrldq    $0x2,%%xmm0                     \n"
-    "paddusw   %%xmm0,%%xmm6                   \n"
-    "psrldq    $0x2,%%xmm0                     \n"
-    "paddusw   %%xmm0,%%xmm6                   \n"
-    "pshufb    %%xmm2,%%xmm6                   \n"
-    "movdqa    %%xmm1,%%xmm7                   \n"
-    "psrldq    $0x2,%%xmm1                     \n"
-    "paddusw   %%xmm1,%%xmm7                   \n"
-    "psrldq    $0x2,%%xmm1                     \n"
-    "paddusw   %%xmm1,%%xmm7                   \n"
-    "pshufb    %%xmm3,%%xmm7                   \n"
-    "paddusw   %%xmm7,%%xmm6                   \n"
-    "pmulhuw   %%xmm4,%%xmm6                   \n"
-    "packuswb  %%xmm6,%%xmm6                   \n"
-    "movd      %%xmm6," MEMACCESS(1) "         \n"
-    "psrlq     $0x10,%%xmm6                    \n"
-    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
-    "lea       " MEMLEA(0x6,1) ",%1            \n"
-    "sub       $0x6,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+                                uint8_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "movdqa      %0,%%xmm2                     \n"
+      "movdqa      %1,%%xmm3                     \n"
+      "movdqa      %2,%%xmm4                     \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+      :
+      : "m"(kShufAc),    // %0
+        "m"(kShufAc3),   // %1
+        "m"(kScaleAc33)  // %2
   );
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm6          \n"
+      "movhlps     %%xmm0,%%xmm1                 \n"
+      "movhlps     %%xmm6,%%xmm7                 \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm6                 \n"
+      "punpcklbw   %%xmm5,%%xmm7                 \n"
+      "paddusw     %%xmm6,%%xmm0                 \n"
+      "paddusw     %%xmm7,%%xmm1                 \n"
+      "movdqu      0x00(%0,%3,2),%%xmm6          \n"
+      "lea         0x10(%0),%0                   \n"
+      "movhlps     %%xmm6,%%xmm7                 \n"
+      "punpcklbw   %%xmm5,%%xmm6                 \n"
+      "punpcklbw   %%xmm5,%%xmm7                 \n"
+      "paddusw     %%xmm6,%%xmm0                 \n"
+      "paddusw     %%xmm7,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm6                 \n"
+      "psrldq      $0x2,%%xmm0                   \n"
+      "paddusw     %%xmm0,%%xmm6                 \n"
+      "psrldq      $0x2,%%xmm0                   \n"
+      "paddusw     %%xmm0,%%xmm6                 \n"
+      "pshufb      %%xmm2,%%xmm6                 \n"
+      "movdqa      %%xmm1,%%xmm7                 \n"
+      "psrldq      $0x2,%%xmm1                   \n"
+      "paddusw     %%xmm1,%%xmm7                 \n"
+      "psrldq      $0x2,%%xmm1                   \n"
+      "paddusw     %%xmm1,%%xmm7                 \n"
+      "pshufb      %%xmm3,%%xmm7                 \n"
+      "paddusw     %%xmm7,%%xmm6                 \n"
+      "pmulhuw     %%xmm4,%%xmm6                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movd        %%xmm6,(%1)                   \n"
+      "psrlq       $0x10,%%xmm6                  \n"
+      "movd        %%xmm6,0x2(%1)                \n"
+      "lea         0x6(%1),%1                    \n"
+      "sub         $0x6,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),               // %0
+                 "+r"(dst_ptr),               // %1
+                 "+r"(dst_width)              // %2
+               : "r"((intptr_t)(src_stride))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6", "xmm7");
 }
 
-// Reads 16xN bytes and produces 16 shorts at a time.
-void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
-  asm volatile (
-    "pxor      %%xmm5,%%xmm5                   \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
-    "lea       " MEMLEA(0x10,0) ",%0           \n"  // src_ptr += 16
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm1   \n"
-    "movdqa    %%xmm3,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm2                   \n"
-    "punpckhbw %%xmm5,%%xmm3                   \n"
-    "paddusw   %%xmm2,%%xmm0                   \n"
-    "paddusw   %%xmm3,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
-    "lea       " MEMLEA(0x20,1) ",%1           \n"
-    "sub       $0x10,%2                        \n"
-    "jg        1b                              \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(src_width)    // %2
-  :
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+static const uvec8 kLinearShuffleFar = {2,  3,  0, 1, 6,  7,  4,  5,
+                                        10, 11, 8, 9, 14, 15, 12, 13};
+
+static const uvec8 kLinearMadd31 = {3, 1, 1, 3, 3, 1, 1, 3,
+                                    3, 1, 1, 3, 3, 1, 1, 3};
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSE2
+void ScaleRowUp2_Linear_SSE2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  asm volatile(
+      "pxor        %%xmm0,%%xmm0                 \n"  // 0
+      "pcmpeqw     %%xmm6,%%xmm6                 \n"
+      "psrlw       $15,%%xmm6                    \n"
+      "psllw       $1,%%xmm6                     \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm1                   \n"  // 01234567
+      "movq        1(%0),%%xmm2                  \n"  // 12345678
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"  // 0112233445566778
+      "punpcklbw   %%xmm1,%%xmm1                 \n"  // 0011223344556677
+      "punpcklbw   %%xmm2,%%xmm2                 \n"  // 1122334455667788
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "punpcklbw   %%xmm0,%%xmm4                 \n"  // 00112233 (16)
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 11223344 (16)
+      "paddw       %%xmm5,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "paddw       %%xmm6,%%xmm4                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 01122334 (16)
+      "paddw       %%xmm5,%%xmm5                 \n"
+      "paddw       %%xmm4,%%xmm5                 \n"  // 3*near+far+2 (lo)
+      "psrlw       $2,%%xmm5                     \n"  // 3/4*near+1/4*far (lo)
+
+      "punpckhbw   %%xmm0,%%xmm1                 \n"  // 44556677 (16)
+      "punpckhbw   %%xmm0,%%xmm2                 \n"  // 55667788 (16)
+      "paddw       %%xmm2,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm3                 \n"  // 45566778 (16)
+      "paddw       %%xmm6,%%xmm1                 \n"
+      "paddw       %%xmm3,%%xmm3                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"  // 3*near+far+2 (hi)
+      "psrlw       $2,%%xmm1                     \n"  // 3/4*near+1/4*far (hi)
+
+      "packuswb    %%xmm1,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSE2
+void ScaleRowUp2_Bilinear_SSE2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  asm volatile(
+      LABELALIGN
+      "1:                                        \n"
+      "pxor        %%xmm0,%%xmm0                 \n"  // 0
+      // above line
+      "movq        (%0),%%xmm1                   \n"  // 01234567
+      "movq        1(%0),%%xmm2                  \n"  // 12345678
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"  // 0112233445566778
+      "punpcklbw   %%xmm1,%%xmm1                 \n"  // 0011223344556677
+      "punpcklbw   %%xmm2,%%xmm2                 \n"  // 1122334455667788
+
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "punpcklbw   %%xmm0,%%xmm4                 \n"  // 00112233 (16)
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 11223344 (16)
+      "paddw       %%xmm5,%%xmm4                 \n"  // near+far
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 01122334 (16)
+      "paddw       %%xmm5,%%xmm5                 \n"  // 2*near
+      "paddw       %%xmm5,%%xmm4                 \n"  // 3*near+far (1, lo)
+
+      "punpckhbw   %%xmm0,%%xmm1                 \n"  // 44556677 (16)
+      "punpckhbw   %%xmm0,%%xmm2                 \n"  // 55667788 (16)
+      "paddw       %%xmm2,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm3                 \n"  // 45566778 (16)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 2*near
+      "paddw       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      // below line
+      "movq        (%0,%3),%%xmm6                \n"  // 01234567
+      "movq        1(%0,%3),%%xmm2               \n"  // 12345678
+      "movdqa      %%xmm6,%%xmm3                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"  // 0112233445566778
+      "punpcklbw   %%xmm6,%%xmm6                 \n"  // 0011223344556677
+      "punpcklbw   %%xmm2,%%xmm2                 \n"  // 1122334455667788
+
+      "movdqa      %%xmm6,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 00112233 (16)
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "punpcklbw   %%xmm0,%%xmm7                 \n"  // 11223344 (16)
+      "paddw       %%xmm7,%%xmm5                 \n"  // near+far
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "punpcklbw   %%xmm0,%%xmm7                 \n"  // 01122334 (16)
+      "paddw       %%xmm7,%%xmm7                 \n"  // 2*near
+      "paddw       %%xmm7,%%xmm5                 \n"  // 3*near+far (2, lo)
+
+      "punpckhbw   %%xmm0,%%xmm6                 \n"  // 44556677 (16)
+      "punpckhbw   %%xmm0,%%xmm2                 \n"  // 55667788 (16)
+      "paddw       %%xmm6,%%xmm2                 \n"  // near+far
+      "punpckhbw   %%xmm0,%%xmm3                 \n"  // 45566778 (16)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 2*near
+      "paddw       %%xmm3,%%xmm2                 \n"  // 3*near+far (2, hi)
+
+      // xmm4 xmm1
+      // xmm5 xmm2
+      "pcmpeqw     %%xmm0,%%xmm0                 \n"
+      "psrlw       $15,%%xmm0                    \n"
+      "psllw       $3,%%xmm0                     \n"  // all 8
+
+      "movdqa      %%xmm4,%%xmm3                 \n"
+      "movdqa      %%xmm5,%%xmm6                 \n"
+      "paddw       %%xmm3,%%xmm3                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm0,%%xmm6                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm4,%%xmm3                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm6,%%xmm3                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm3                     \n"  // ^ div by 16
+
+      "movdqa      %%xmm1,%%xmm7                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "paddw       %%xmm7,%%xmm7                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm0,%%xmm6                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm1,%%xmm7                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm6,%%xmm7                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm7                     \n"  // ^ div by 16
+
+      "packuswb    %%xmm7,%%xmm3                 \n"
+      "movdqu      %%xmm3,(%1)                   \n"  // save above line
+
+      "movdqa      %%xmm5,%%xmm3                 \n"
+      "paddw       %%xmm0,%%xmm4                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm3,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm4,%%xmm5                 \n"  // 9 3 3 1 + 8 (lo)
+      "psrlw       $4,%%xmm5                     \n"  // ^ div by 16
+
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "paddw       %%xmm0,%%xmm1                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm3,%%xmm2                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm1,%%xmm2                 \n"  // 9 3 3 1 + 8 (hi)
+      "psrlw       $4,%%xmm2                     \n"  // ^ div by 16
+
+      "packuswb    %%xmm2,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4)                \n"  // save below line
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_SSSE3
+void ScaleRowUp2_Linear_12_SSSE3(const uint16_t* src_ptr,
+                                 uint16_t* dst_ptr,
+                                 int dst_width) {
+  asm volatile(
+      "movdqa      %3,%%xmm5                     \n"
+      "pcmpeqw     %%xmm4,%%xmm4                 \n"
+      "psrlw       $15,%%xmm4                    \n"
+      "psllw       $1,%%xmm4                     \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // 01234567 (16)
+      "movdqu      2(%0),%%xmm1                  \n"  // 12345678 (16)
+
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhwd   %%xmm1,%%xmm2                 \n"  // 45566778 (16)
+      "punpcklwd   %%xmm1,%%xmm0                 \n"  // 01122334 (16)
+
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm3                 \n"  // 54657687 (far)
+      "pshufb      %%xmm5,%%xmm1                 \n"  // 10213243 (far)
+
+      "paddw       %%xmm4,%%xmm1                 \n"  // far+2
+      "paddw       %%xmm4,%%xmm3                 \n"  // far+2
+      "paddw       %%xmm0,%%xmm1                 \n"  // near+far+2
+      "paddw       %%xmm2,%%xmm3                 \n"  // near+far+2
+      "paddw       %%xmm0,%%xmm0                 \n"  // 2*near
+      "paddw       %%xmm2,%%xmm2                 \n"  // 2*near
+      "paddw       %%xmm1,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddw       %%xmm3,%%xmm2                 \n"  // 3*near+far+2 (hi)
+
+      "psrlw       $2,%%xmm0                     \n"  // 3/4*near+1/4*far
+      "psrlw       $2,%%xmm2                     \n"  // 3/4*near+1/4*far
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm2,16(%1)                 \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),          // %0
+        "+r"(dst_ptr),          // %1
+        "+r"(dst_width)         // %2
+      : "m"(kLinearShuffleFar)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+void ScaleRowUp2_Bilinear_12_SSSE3(const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint16_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm7,%%xmm7                 \n"
+      "psrlw       $15,%%xmm7                    \n"
+      "psllw       $3,%%xmm7                     \n"  // all 8
+      "movdqa      %5,%%xmm6                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      // above line
+      "movdqu      (%0),%%xmm0                   \n"  // 01234567 (16)
+      "movdqu      2(%0),%%xmm1                  \n"  // 12345678 (16)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhwd   %%xmm1,%%xmm2                 \n"  // 45566778 (16)
+      "punpcklwd   %%xmm1,%%xmm0                 \n"  // 01122334 (16)
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm6,%%xmm3                 \n"  // 54657687 (far)
+      "pshufb      %%xmm6,%%xmm1                 \n"  // 10213243 (far)
+      "paddw       %%xmm0,%%xmm1                 \n"  // near+far
+      "paddw       %%xmm2,%%xmm3                 \n"  // near+far
+      "paddw       %%xmm0,%%xmm0                 \n"  // 2*near
+      "paddw       %%xmm2,%%xmm2                 \n"  // 2*near
+      "paddw       %%xmm1,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddw       %%xmm3,%%xmm2                 \n"  // 3*near+far (1, hi)
+
+      // below line
+      "movdqu      (%0,%3,2),%%xmm1              \n"  // 01234567 (16)
+      "movdqu      2(%0,%3,2),%%xmm4             \n"  // 12345678 (16)
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpckhwd   %%xmm4,%%xmm3                 \n"  // 45566778 (16)
+      "punpcklwd   %%xmm4,%%xmm1                 \n"  // 01122334 (16)
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "pshufb      %%xmm6,%%xmm5                 \n"  // 54657687 (far)
+      "pshufb      %%xmm6,%%xmm4                 \n"  // 10213243 (far)
+      "paddw       %%xmm1,%%xmm4                 \n"  // near+far
+      "paddw       %%xmm3,%%xmm5                 \n"  // near+far
+      "paddw       %%xmm1,%%xmm1                 \n"  // 2*near
+      "paddw       %%xmm3,%%xmm3                 \n"  // 2*near
+      "paddw       %%xmm4,%%xmm1                 \n"  // 3*near+far (2, lo)
+      "paddw       %%xmm5,%%xmm3                 \n"  // 3*near+far (2, hi)
+
+      // xmm0 xmm2
+      // xmm1 xmm3
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm7,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16
+      "movdqu      %%xmm4,(%1)                   \n"
+
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm7,%%xmm5                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm2,%%xmm4                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16
+      "movdqu      %%xmm4,0x10(%1)               \n"
+
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "paddw       %%xmm7,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm4,%%xmm1                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm0,%%xmm1                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrlw       $4,%%xmm1                     \n"  // ^ div by 16
+      "movdqu      %%xmm1,(%1,%4,2)              \n"
+
+      "movdqa      %%xmm3,%%xmm4                 \n"
+      "paddw       %%xmm7,%%xmm2                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm4,%%xmm3                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm2,%%xmm3                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrlw       $4,%%xmm3                     \n"  // ^ div by 16
+      "movdqu      %%xmm3,0x10(%1,%4,2)          \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearShuffleFar)        // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
 
+#ifdef HAS_SCALEROWUP2_LINEAR_16_SSE2
+void ScaleRowUp2_Linear_16_SSE2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "pcmpeqd     %%xmm4,%%xmm4                 \n"
+      "psrld       $31,%%xmm4                    \n"
+      "pslld       $1,%%xmm4                     \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0123 (16b)
+      "movq        2(%0),%%xmm1                  \n"  // 1234 (16b)
+
+      "punpcklwd   %%xmm5,%%xmm0                 \n"  // 0123 (32b)
+      "punpcklwd   %%xmm5,%%xmm1                 \n"  // 1234 (32b)
+
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+
+      "pshufd      $0b10110001,%%xmm2,%%xmm2     \n"  // 1032 (even, far)
+      "pshufd      $0b10110001,%%xmm3,%%xmm3     \n"  // 2143 (odd, far)
+
+      "paddd       %%xmm4,%%xmm2                 \n"  // far+2 (lo)
+      "paddd       %%xmm4,%%xmm3                 \n"  // far+2 (hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far+2 (lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far+2 (hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far+2 (hi)
+
+      "psrld       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrld       $2,%%xmm1                     \n"  // 3/4*near+1/4*far (hi)
+      "packssdw    %%xmm1,%%xmm0                 \n"
+      "pshufd      $0b11011000,%%xmm0,%%xmm0     \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 pixel to 8 pixel
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_SSE2
+void ScaleRowUp2_Bilinear_16_SSE2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "pxor        %%xmm7,%%xmm7                 \n"
+      "pcmpeqd     %%xmm6,%%xmm6                 \n"
+      "psrld       $31,%%xmm6                    \n"
+      "pslld       $3,%%xmm6                     \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0011 (16b, 1u1v)
+      "movq        4(%0),%%xmm1                  \n"  // 1122 (16b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm0                 \n"  // 0011 (near) (32b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm1                 \n"  // 1122 (near) (32b, 1u1v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pshufd      $0b01001110,%%xmm2,%%xmm2     \n"  // 1100 (far) (1, lo)
+      "pshufd      $0b01001110,%%xmm3,%%xmm3     \n"  // 2211 (far) (1, hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far (1, lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far (1, hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (1, lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      "movq        (%0),%%xmm0                   \n"  // 0123 (16b)
+      "movq        2(%0),%%xmm1                  \n"  // 1234 (16b)
+      "punpcklwd   %%xmm7,%%xmm0                 \n"  // 0123 (32b)
+      "punpcklwd   %%xmm7,%%xmm1                 \n"  // 1234 (32b)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pshufd      $0b10110001,%%xmm2,%%xmm2     \n"  // 1032 (even, far)
+      "pshufd      $0b10110001,%%xmm3,%%xmm3     \n"  // 2143 (odd, far)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far (lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far (hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      "movq        (%0,%3,2),%%xmm2              \n"
+      "movq        2(%0,%3,2),%%xmm3             \n"
+      "punpcklwd   %%xmm7,%%xmm2                 \n"  // 0123 (32b)
+      "punpcklwd   %%xmm7,%%xmm3                 \n"  // 1234 (32b)
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "pshufd      $0b10110001,%%xmm4,%%xmm4     \n"  // 1032 (even, far)
+      "pshufd      $0b10110001,%%xmm5,%%xmm5     \n"  // 2143 (odd, far)
+      "paddd       %%xmm2,%%xmm4                 \n"  // near+far (lo)
+      "paddd       %%xmm3,%%xmm5                 \n"  // near+far (hi)
+      "paddd       %%xmm2,%%xmm2                 \n"  // 2*near (lo)
+      "paddd       %%xmm3,%%xmm3                 \n"  // 2*near (hi)
+      "paddd       %%xmm4,%%xmm2                 \n"  // 3*near+far (2, lo)
+      "paddd       %%xmm5,%%xmm3                 \n"  // 3*near+far (2, hi)
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddd       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddd       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddd       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrld       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm2,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddd       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddd       %%xmm2,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddd       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrld       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddd       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (2, hi)
+      "paddd       %%xmm1,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrld       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"  // 6*near+2*far (2, hi)
+      "paddd       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (1, hi)
+      "paddd       %%xmm3,%%xmm2                 \n"  // 9*near+3*far (2, hi)
+      "paddd       %%xmm1,%%xmm2                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrld       $4,%%xmm2                     \n"  // ^ div by 16 (2, hi)
+
+      "packssdw    %%xmm0,%%xmm4                 \n"
+      "pshufd      $0b11011000,%%xmm4,%%xmm4     \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packssdw    %%xmm2,%%xmm5                 \n"
+      "pshufd      $0b11011000,%%xmm5,%%xmm5     \n"
+      "movdqu      %%xmm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 pixel to 8 pixel
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSSE3
+void ScaleRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm4,%%xmm4                 \n"
+      "psrlw       $15,%%xmm4                    \n"
+      "psllw       $1,%%xmm4                     \n"  // all 2
+      "movdqa      %3,%%xmm3                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 01234567
+      "movq        1(%0),%%xmm1                  \n"  // 12345678
+      "punpcklwd   %%xmm0,%%xmm0                 \n"  // 0101232345456767
+      "punpcklwd   %%xmm1,%%xmm1                 \n"  // 1212343456567878
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 4545565667677878
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 0101121223233434
+      "pmaddubsw   %%xmm3,%%xmm2                 \n"  // 3*near+far (hi)
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"  // 3*near+far (lo)
+      "paddw       %%xmm4,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddw       %%xmm4,%%xmm2                 \n"  // 3*near+far+2 (hi)
+      "psrlw       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrlw       $2,%%xmm2                     \n"  // 3/4*near+1/4*far (hi)
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),      // %0
+        "+r"(dst_ptr),      // %1
+        "+r"(dst_width)     // %2
+      : "m"(kLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSSE3
+void ScaleRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                ptrdiff_t dst_stride,
+                                int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm6,%%xmm6                 \n"
+      "psrlw       $15,%%xmm6                    \n"
+      "psllw       $3,%%xmm6                     \n"  // all 8
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 01234567
+      "movq        1(%0),%%xmm1                  \n"  // 12345678
+      "punpcklwd   %%xmm0,%%xmm0                 \n"  // 0101232345456767
+      "punpcklwd   %%xmm1,%%xmm1                 \n"  // 1212343456567878
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 4545565667677878
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 0101121223233434
+      "pmaddubsw   %%xmm7,%%xmm2                 \n"  // 3*near+far (1, hi)
+      "pmaddubsw   %%xmm7,%%xmm0                 \n"  // 3*near+far (1, lo)
+
+      "movq        (%0,%3),%%xmm1                \n"
+      "movq        1(%0,%3),%%xmm4               \n"
+      "punpcklwd   %%xmm1,%%xmm1                 \n"
+      "punpcklwd   %%xmm4,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpckhdq   %%xmm4,%%xmm3                 \n"
+      "punpckldq   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm7,%%xmm3                 \n"  // 3*near+far (2, hi)
+      "pmaddubsw   %%xmm7,%%xmm1                 \n"  // 3*near+far (2, lo)
+
+      // xmm0 xmm2
+      // xmm1 xmm3
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm1,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm1,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrlw       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm2,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm1,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm3,%%xmm1                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm2,%%xmm1                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrlw       $4,%%xmm1                     \n"  // ^ div by 16 (2, hi)
+
+      "packuswb    %%xmm0,%%xmm4                 \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packuswb    %%xmm1,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4)                \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearMadd31)            // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_AVX2
+void ScaleRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $1,%%ymm4,%%ymm4              \n"  // all 2
+      "vbroadcastf128 %3,%%ymm3                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 0123456789ABCDEF
+      "vmovdqu     1(%0),%%xmm1                  \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklwd  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpunpcklwd  %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm3,%%ymm2,%%ymm1          \n"  // 3*near+far (hi)
+      "vpmaddubsw  %%ymm3,%%ymm0,%%ymm0          \n"  // 3*near+far (lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"  // 3*near+far+2 (hi)
+      "vpsrlw      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrlw      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 16 sample to 32 sample
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),      // %0
+        "+r"(dst_ptr),      // %1
+        "+r"(dst_width)     // %2
+      : "m"(kLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_AVX2
+void ScaleRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $15,%%ymm6,%%ymm6             \n"
+      "vpsllw      $3,%%ymm6,%%ymm6              \n"  // all 8
+      "vbroadcastf128 %5,%%ymm7                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 0123456789ABCDEF
+      "vmovdqu     1(%0),%%xmm1                  \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklwd  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpunpcklwd  %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm1          \n"  // 3*near+far (1, hi)
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm0          \n"  // 3*near+far (1, lo)
+
+      "vmovdqu     (%0,%3),%%xmm2                \n"  // 0123456789ABCDEF
+      "vmovdqu     1(%0,%3),%%xmm3               \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm2,%%ymm2     \n"
+      "vpermq      $0b11011000,%%ymm3,%%ymm3     \n"
+      "vpunpcklwd  %%ymm2,%%ymm2,%%ymm2          \n"
+      "vpunpcklwd  %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpunpckhdq  %%ymm3,%%ymm2,%%ymm4          \n"
+      "vpunpckldq  %%ymm3,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm7,%%ymm4,%%ymm3          \n"  // 3*near+far (2, hi)
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm2          \n"  // 3*near+far (2, lo)
+
+      // ymm0 ymm1
+      // ymm2 ymm3
+
+      "vpaddw      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddw      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddw      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrlw      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddw      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddw      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddw      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddw      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrlw      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddw      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddw      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddw      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddw      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddw      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddw      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddw      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrlw      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackuswb   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackuswb   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vmovdqu     %%ymm5,(%1,%4)                \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 16 sample to 32 sample
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearMadd31)            // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_AVX2
+void ScaleRowUp2_Linear_12_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm5                  \n"
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $1,%%ymm4,%%ymm4              \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // 0123456789ABCDEF (16b)
+      "vmovdqu     2(%0),%%ymm1                  \n"  // 123456789ABCDEF0 (16b)
+
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"  // 012389AB4567CDEF
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"  // 12349ABC5678DEF0
+
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm2          \n"  // 899AABBCCDDEEFF0 (near)
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"  // 0112233445566778 (near)
+      "vpshufb     %%ymm5,%%ymm2,%%ymm3          \n"  // 98A9BACBDCEDFE0F (far)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"  // 1021324354657687 (far)
+
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"  // far+2
+      "vpaddw      %%ymm4,%%ymm3,%%ymm3          \n"  // far+2
+      "vpaddw      %%ymm0,%%ymm1,%%ymm1          \n"  // near+far+2
+      "vpaddw      %%ymm2,%%ymm3,%%ymm3          \n"  // near+far+2
+      "vpaddw      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near
+      "vpaddw      %%ymm2,%%ymm2,%%ymm2          \n"  // 2*near
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 3*near+far+2
+      "vpaddw      %%ymm2,%%ymm3,%%ymm2          \n"  // 3*near+far+2
+
+      "vpsrlw      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far
+      "vpsrlw      $2,%%ymm2,%%ymm2              \n"  // 3/4*near+1/4*far
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm2,32(%1)                 \n"
+
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"  // 16 sample to 32 sample
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),          // %0
+        "+r"(dst_ptr),          // %1
+        "+r"(dst_width)         // %2
+      : "m"(kLinearShuffleFar)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_AVX2
+void ScaleRowUp2_Bilinear_12_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $3,%%ymm4,%%ymm4              \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 01234567 (16b)
+      "vmovdqu     2(%0),%%xmm1                  \n"  // 12345678 (16b)
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"  // 0123000045670000
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"  // 1234000056780000
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"  // 0112233445566778 (near)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"  // 1021324354657687 (far)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm1          \n"  // near+far
+      "vpaddw      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near
+      "vpaddw      %%ymm0,%%ymm1,%%ymm2          \n"  // 3*near+far (1)
+
+      "vmovdqu     (%0,%3,2),%%xmm0              \n"  // 01234567 (16b)
+      "vmovdqu     2(%0,%3,2),%%xmm1             \n"  // 12345678 (16b)
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"  // 0123000045670000
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"  // 1234000056780000
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"  // 0112233445566778 (near)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"  // 1021324354657687 (far)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm1          \n"  // near+far
+      "vpaddw      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near
+      "vpaddw      %%ymm0,%%ymm1,%%ymm3          \n"  // 3*near+far (2)
+
+      "vpaddw      %%ymm2,%%ymm2,%%ymm0          \n"  // 6*near+2*far (1)
+      "vpaddw      %%ymm4,%%ymm3,%%ymm1          \n"  // 3*near+far+8 (2)
+      "vpaddw      %%ymm0,%%ymm2,%%ymm0          \n"  // 9*near+3*far (1)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9 3 3 1 + 8 (1)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16
+      "vmovdqu     %%ymm0,(%1)                   \n"  // store above
+
+      "vpaddw      %%ymm3,%%ymm3,%%ymm0          \n"  // 6*near+2*far (2)
+      "vpaddw      %%ymm4,%%ymm2,%%ymm1          \n"  // 3*near+far+8 (1)
+      "vpaddw      %%ymm0,%%ymm3,%%ymm0          \n"  // 9*near+3*far (2)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9 3 3 1 + 8 (2)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16
+      "vmovdqu     %%ymm0,(%1,%4,2)              \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearShuffleFar)        // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_AVX2
+void ScaleRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrld      $31,%%ymm4,%%ymm4             \n"
+      "vpslld      $1,%%ymm4,%%ymm4              \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 01234567 (16b, 1u1v)
+      "vmovdqu     2(%0),%%xmm1                  \n"  // 12345678 (16b, 1u1v)
+
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+
+      "vpshufd     $0b10110001,%%ymm0,%%ymm2     \n"  // 10325476 (lo, far)
+      "vpshufd     $0b10110001,%%ymm1,%%ymm3     \n"  // 21436587 (hi, far)
+
+      "vpaddd      %%ymm4,%%ymm2,%%ymm2          \n"  // far+2 (lo)
+      "vpaddd      %%ymm4,%%ymm3,%%ymm3          \n"  // far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far+2 (hi)
+
+      "vpsrld      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrld      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackusdw   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpshufd     $0b11011000,%%ymm0,%%ymm0     \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 pixel to 16 pixel
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_AVX2
+void ScaleRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrld      $31,%%ymm6,%%ymm6             \n"
+      "vpslld      $3,%%ymm6,%%ymm6              \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 01234567 (16b, 1u1v)
+      "vmovdqu     2(%0),%%xmm1                  \n"  // 12345678 (16b, 1u1v)
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b10110001,%%ymm0,%%ymm2     \n"  // 10325476 (lo, far)
+      "vpshufd     $0b10110001,%%ymm1,%%ymm3     \n"  // 21436587 (hi, far)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far (1, lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far (1, hi)
+
+      "vmovdqu     (%0,%3,2),%%xmm2              \n"  // 01234567 (16b, 1u1v)
+      "vmovdqu     2(%0,%3,2),%%xmm3             \n"  // 12345678 (16b, 1u1v)
+      "vpmovzxwd   %%xmm2,%%ymm2                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm3,%%ymm3                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b10110001,%%ymm2,%%ymm4     \n"  // 10325476 (lo, far)
+      "vpshufd     $0b10110001,%%ymm3,%%ymm5     \n"  // 21436587 (hi, far)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm4          \n"  // near+far (lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm5          \n"  // near+far (hi)
+      "vpaddd      %%ymm2,%%ymm2,%%ymm2          \n"  // 2*near (lo)
+      "vpaddd      %%ymm3,%%ymm3,%%ymm3          \n"  // 2*near (hi)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm2          \n"  // 3*near+far (2, lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm3          \n"  // 3*near+far (2, hi)
+
+      "vpaddd      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddd      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddd      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddd      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrld      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddd      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddd      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddd      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddd      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrld      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddd      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddd      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddd      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrld      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddd      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddd      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddd      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddd      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrld      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackusdw   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vpshufd     $0b11011000,%%ymm4,%%ymm4     \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackusdw   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vpshufd     $0b11011000,%%ymm5,%%ymm5     \n"
+      "vmovdqu     %%ymm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 pixel to 16 pixel
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+void ScaleAddRow_SSE2(const uint8_t* src_ptr,
+                      uint16_t* dst_ptr,
+                      int src_width) {
+      asm volatile("pxor        %%xmm5,%%xmm5                 \n"
+
+               // 16 pixel loop.
+               LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm3                   \n"
+      "lea         0x10(%0),%0                   \n"  // src_ptr += 16
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      0x10(%1),%%xmm1               \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpckhbw   %%xmm5,%%xmm3                 \n"
+      "paddusw     %%xmm2,%%xmm0                 \n"
+      "paddusw     %%xmm3,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(src_width)  // %2
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
+}
 
 #ifdef HAS_SCALEADDROW_AVX2
 // Reads 32 bytes and accumulates to 32 shorts at a time.
-void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
-  asm volatile (
-    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "vmovdqu    " MEMACCESS(0) ",%%ymm3        \n"
-    "lea        " MEMLEA(0x20,0) ",%0          \n"  // src_ptr += 32
-    "vpermq     $0xd8,%%ymm3,%%ymm3            \n"
-    "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
-    "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
-    "vpaddusw   " MEMACCESS(1) ",%%ymm2,%%ymm0 \n"
-    "vpaddusw   " MEMACCESS2(0x20,1) ",%%ymm3,%%ymm1 \n"
-    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
-    "vmovdqu    %%ymm1," MEMACCESS2(0x20,1) "  \n"
-    "lea       " MEMLEA(0x40,1) ",%1           \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(src_width)    // %2
-  :
-  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-  );
+void ScaleAddRow_AVX2(const uint8_t* src_ptr,
+                      uint16_t* dst_ptr,
+                      int src_width) {
+      asm volatile("vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+               LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm3                   \n"
+      "lea         0x20(%0),%0                   \n"  // src_ptr += 32
+      "vpermq      $0xd8,%%ymm3,%%ymm3           \n"
+      "vpunpcklbw  %%ymm5,%%ymm3,%%ymm2          \n"
+      "vpunpckhbw  %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpaddusw    (%1),%%ymm2,%%ymm0            \n"
+      "vpaddusw    0x20(%1),%%ymm3,%%ymm1        \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(src_width)  // %2
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SCALEADDROW_AVX2
 
 // Constant for making pixels signed to avoid pmaddubsw
 // saturation.
-static uvec8 kFsub80 =
-  { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
+static const uvec8 kFsub80 = {0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+                              0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
 
 // Constant for making pixels unsigned and adding .5 for rounding.
-static uvec16 kFadd40 =
-  { 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040 };
+static const uvec16 kFadd40 = {0x4040, 0x4040, 0x4040, 0x4040,
+                               0x4040, 0x4040, 0x4040, 0x4040};
 
 // Bilinear column filtering. SSSE3 version.
-void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                           int dst_width, int x, int dx) {
+void ScaleFilterCols_SSSE3(uint8_t* dst_ptr,
+                           const uint8_t* src_ptr,
+                           int dst_width,
+                           int x,
+                           int dx) {
   intptr_t x0, x1, temp_pixel;
-  asm volatile (
-    "movd      %6,%%xmm2                       \n"
-    "movd      %7,%%xmm3                       \n"
-    "movl      $0x04040000,%k2                 \n"
-    "movd      %k2,%%xmm5                      \n"
-    "pcmpeqb   %%xmm6,%%xmm6                   \n"
-    "psrlw     $0x9,%%xmm6                     \n"  // 0x007f007f
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"
-    "psrlw     $15,%%xmm7                      \n"  // 0x00010001
-
-    "pextrw    $0x1,%%xmm2,%k3                 \n"
-    "subl      $0x2,%5                         \n"
-    "jl        29f                             \n"
-    "movdqa    %%xmm2,%%xmm0                   \n"
-    "paddd     %%xmm3,%%xmm0                   \n"
-    "punpckldq %%xmm0,%%xmm2                   \n"
-    "punpckldq %%xmm3,%%xmm3                   \n"
-    "paddd     %%xmm3,%%xmm3                   \n"
-    "pextrw    $0x3,%%xmm2,%k4                 \n"
-
-    LABELALIGN
-  "2:                                          \n"
-    "movdqa    %%xmm2,%%xmm1                   \n"
-    "paddd     %%xmm3,%%xmm2                   \n"
-    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
-    "movd      %k2,%%xmm0                      \n"
-    "psrlw     $0x9,%%xmm1                     \n"
-    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
-    "movd      %k2,%%xmm4                      \n"
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "punpcklwd %%xmm4,%%xmm0                   \n"
-    "psubb     %8,%%xmm0                       \n"  // make pixels signed.
-    "pxor      %%xmm6,%%xmm1                   \n"  // 128 -f = (f ^ 127 ) + 1
-    "paddusb   %%xmm7,%%xmm1                   \n"
-    "pmaddubsw %%xmm0,%%xmm1                   \n"
-    "pextrw    $0x1,%%xmm2,%k3                 \n"
-    "pextrw    $0x3,%%xmm2,%k4                 \n"
-    "paddw     %9,%%xmm1                       \n"  // make pixels unsigned.
-    "psrlw     $0x7,%%xmm1                     \n"
-    "packuswb  %%xmm1,%%xmm1                   \n"
-    "movd      %%xmm1,%k2                      \n"
-    "mov       %w2," MEMACCESS(0) "            \n"
-    "lea       " MEMLEA(0x2,0) ",%0            \n"
-    "subl      $0x2,%5                         \n"
-    "jge       2b                              \n"
-
-    LABELALIGN
-  "29:                                         \n"
-    "addl      $0x1,%5                         \n"
-    "jl        99f                             \n"
-    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
-    "movd      %k2,%%xmm0                      \n"
-    "psrlw     $0x9,%%xmm2                     \n"
-    "pshufb    %%xmm5,%%xmm2                   \n"
-    "psubb     %8,%%xmm0                       \n"  // make pixels signed.
-    "pxor      %%xmm6,%%xmm2                   \n"
-    "paddusb   %%xmm7,%%xmm2                   \n"
-    "pmaddubsw %%xmm0,%%xmm2                   \n"
-    "paddw     %9,%%xmm2                       \n"  // make pixels unsigned.
-    "psrlw     $0x7,%%xmm2                     \n"
-    "packuswb  %%xmm2,%%xmm2                   \n"
-    "movd      %%xmm2,%k2                      \n"
-    "mov       %b2," MEMACCESS(0) "            \n"
-  "99:                                         \n"
-  : "+r"(dst_ptr),      // %0
-    "+r"(src_ptr),      // %1
-    "=&a"(temp_pixel),  // %2
-    "=&r"(x0),          // %3
-    "=&r"(x1),          // %4
+  asm volatile(
+      "movd        %6,%%xmm2                     \n"
+      "movd        %7,%%xmm3                     \n"
+      "movl        $0x04040000,%k2               \n"
+      "movd        %k2,%%xmm5                    \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrlw       $0x9,%%xmm6                   \n"  // 0x007f007f
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $15,%%xmm7                    \n"  // 0x00010001
+
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "subl        $0x2,%5                       \n"
+      "jl          29f                           \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "paddd       %%xmm3,%%xmm0                 \n"
+      "punpckldq   %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm3,%%xmm3                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
+
+      LABELALIGN
+      "2:                                        \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"
+      "movzwl      0x00(%1,%3,1),%k2             \n"
+      "movd        %k2,%%xmm0                    \n"
+      "psrlw       $0x9,%%xmm1                   \n"
+      "movzwl      0x00(%1,%4,1),%k2             \n"
+      "movd        %k2,%%xmm4                    \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "punpcklwd   %%xmm4,%%xmm0                 \n"
+      "psubb       %8,%%xmm0                     \n"  // make pixels signed.
+      "pxor        %%xmm6,%%xmm1                 \n"  // 128 - f = (f ^ 127 ) +
+                                                      // 1
+      "paddusb     %%xmm7,%%xmm1                 \n"
+      "pmaddubsw   %%xmm0,%%xmm1                 \n"
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
+      "paddw       %9,%%xmm1                     \n"  // make pixels unsigned.
+      "psrlw       $0x7,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movd        %%xmm1,%k2                    \n"
+      "mov         %w2,(%0)                      \n"
+      "lea         0x2(%0),%0                    \n"
+      "subl        $0x2,%5                       \n"
+      "jge         2b                            \n"
+
+      LABELALIGN
+      "29:                                       \n"
+      "addl        $0x1,%5                       \n"
+      "jl          99f                           \n"
+      "movzwl      0x00(%1,%3,1),%k2             \n"
+      "movd        %k2,%%xmm0                    \n"
+      "psrlw       $0x9,%%xmm2                   \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "psubb       %8,%%xmm0                     \n"  // make pixels signed.
+      "pxor        %%xmm6,%%xmm2                 \n"
+      "paddusb     %%xmm7,%%xmm2                 \n"
+      "pmaddubsw   %%xmm0,%%xmm2                 \n"
+      "paddw       %9,%%xmm2                     \n"  // make pixels unsigned.
+      "psrlw       $0x7,%%xmm2                   \n"
+      "packuswb    %%xmm2,%%xmm2                 \n"
+      "movd        %%xmm2,%k2                    \n"
+      "mov         %b2,(%0)                      \n"
+      "99:                                       \n"
+      : "+r"(dst_ptr),      // %0
+        "+r"(src_ptr),      // %1
+        "=&a"(temp_pixel),  // %2
+        "=&r"(x0),          // %3
+        "=&r"(x1),          // %4
 #if defined(__x86_64__)
-    "+rm"(dst_width)    // %5
+        "+rm"(dst_width)  // %5
 #else
-    "+m"(dst_width)    // %5
+        "+m"(dst_width)  // %5
 #endif
-  : "rm"(x),            // %6
-    "rm"(dx),           // %7
+      : "rm"(x),   // %6
+        "rm"(dx),  // %7
 #if defined(__x86_64__)
-    "x"(kFsub80),       // %8
-    "x"(kFadd40)        // %9
+        "x"(kFsub80),  // %8
+        "x"(kFadd40)   // %9
 #else
-    "m"(kFsub80),       // %8
-    "m"(kFadd40)        // %9
+        "m"(kFsub80),    // %8
+        "m"(kFadd40)     // %9
 #endif
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
-  );
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
 // Reads 4 pixels, duplicates them and writes 8 pixels.
 // Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
-void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                       int dst_width, int x, int dx) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklbw %%xmm0,%%xmm0                   \n"
-    "punpckhbw %%xmm1,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "sub       $0x20,%2                         \n"
-    "jg        1b                              \n"
-
-  : "+r"(dst_ptr),     // %0
-    "+r"(src_ptr),     // %1
-    "+r"(dst_width)    // %2
-  :: "memory", "cc", "xmm0", "xmm1"
-  );
+void ScaleColsUp2_SSE2(uint8_t* dst_ptr,
+                       const uint8_t* src_ptr,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  (void)x;
+  (void)dx;
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "movdqu      %%xmm1,0x10(%0)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+
+               : "+r"(dst_ptr),   // %0
+                 "+r"(src_ptr),   // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
-void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+void ScaleARGBRowDown2_SSE2(const uint8_t* src_argb,
                             ptrdiff_t src_stride,
-                            uint8* dst_argb, int dst_width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(dst_width)  // %2
-  :: "memory", "cc", "xmm0", "xmm1"
-  );
+                            uint8_t* dst_argb,
+                            int dst_width) {
+  (void)src_stride;
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "shufps      $0xdd,%%xmm1,%%xmm0           \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_argb),  // %0
+                 "+r"(dst_argb),  // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
-void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+void ScaleARGBRowDown2Linear_SSE2(const uint8_t* src_argb,
                                   ptrdiff_t src_stride,
-                                  uint8* dst_argb, int dst_width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),  // %0
-    "+r"(dst_argb),  // %1
-    "+r"(dst_width)  // %2
-  :: "memory", "cc", "xmm0", "xmm1"
-  );
+                                  uint8_t* dst_argb,
+                                  int dst_width) {
+  (void)src_stride;
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm2           \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_argb),  // %0
+                 "+r"(dst_argb),  // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
-void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+void ScaleARGBRowDown2Box_SSE2(const uint8_t* src_argb,
                                ptrdiff_t src_stride,
-                               uint8* dst_argb, int dst_width) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
-    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
-    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu   (%0,%3,1),%%xmm2
-    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu   0x10(%0,%3,1),%%xmm3
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "sub       $0x4,%2                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),   // %0
-    "+r"(dst_argb),   // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3"
-  );
+                               uint8_t* dst_argb,
+                               int dst_width) {
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm2           \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_argb),              // %0
+                 "+r"(dst_argb),              // %1
+                 "+r"(dst_width)              // %2
+               : "r"((intptr_t)(src_stride))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
 }
 
 // Reads 4 pixels at a time.
 // Alignment requirement: dst_argb 16 byte aligned.
-void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                               int src_stepx, uint8* dst_argb, int dst_width) {
+void ScaleARGBRowDownEven_SSE2(const uint8_t* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8_t* dst_argb,
+                               int dst_width) {
   intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
   intptr_t src_stepx_x12;
-  asm volatile (
-    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
-    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
-    LABELALIGN
-  "1:                                          \n"
-    "movd      " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
-    "punpckldq %%xmm1,%%xmm0                   \n"
-    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
-    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
-    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
-    "punpckldq %%xmm3,%%xmm2                   \n"
-    "punpcklqdq %%xmm2,%%xmm0                  \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),       // %0
-    "+r"(src_stepx_x4),   // %1
-    "+r"(dst_argb),       // %2
-    "+r"(dst_width),      // %3
-    "=&r"(src_stepx_x12)  // %4
-  :: "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3"
-  );
+  (void)src_stride;
+  asm volatile(
+      "lea         0x00(,%1,4),%1                \n"
+      "lea         0x00(%1,%1,2),%4              \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movd        (%0),%%xmm0                   \n"
+      "movd        0x00(%0,%1,1),%%xmm1          \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"
+      "movd        0x00(%0,%1,2),%%xmm2          \n"
+      "movd        0x00(%0,%4,1),%%xmm3          \n"
+      "lea         0x00(%0,%1,4),%0              \n"
+      "punpckldq   %%xmm3,%%xmm2                 \n"
+      "punpcklqdq  %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),       // %0
+        "+r"(src_stepx_x4),   // %1
+        "+r"(dst_argb),       // %2
+        "+r"(dst_width),      // %3
+        "=&r"(src_stepx_x12)  // %4
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm2", "xmm3");
 }
 
 // Blends four 2x2 to 4x1.
 // Alignment requirement: dst_argb 16 byte aligned.
-void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
-                                  ptrdiff_t src_stride, int src_stepx,
-                                  uint8* dst_argb, int dst_width) {
+void ScaleARGBRowDownEvenBox_SSE2(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8_t* dst_argb,
+                                  int dst_width) {
   intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
   intptr_t src_stepx_x12;
   intptr_t row1 = (intptr_t)(src_stride);
-  asm volatile (
-    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
-    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
-    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
-
-    LABELALIGN
-  "1:                                          \n"
-    "movq      " MEMACCESS(0) ",%%xmm0         \n"
-    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
-    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
-    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
-    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
-    "movq      " MEMACCESS(5) ",%%xmm2         \n"
-    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
-    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
-    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
-    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "pavgb     %%xmm3,%%xmm1                   \n"
-    "movdqa    %%xmm0,%%xmm2                   \n"
-    "shufps    $0x88,%%xmm1,%%xmm0             \n"
-    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
-    "pavgb     %%xmm2,%%xmm0                   \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%3                         \n"
-    "jg        1b                              \n"
-  : "+r"(src_argb),        // %0
-    "+r"(src_stepx_x4),    // %1
-    "+r"(dst_argb),        // %2
-    "+rm"(dst_width),      // %3
-    "=&r"(src_stepx_x12),  // %4
-    "+r"(row1)             // %5
-  :: "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3"
-  );
+  asm volatile(
+      "lea         0x00(,%1,4),%1                \n"
+      "lea         0x00(%1,%1,2),%4              \n"
+      "lea         0x00(%0,%5,1),%5              \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movhps      0x00(%0,%1,1),%%xmm0          \n"
+      "movq        0x00(%0,%1,2),%%xmm1          \n"
+      "movhps      0x00(%0,%4,1),%%xmm1          \n"
+      "lea         0x00(%0,%1,4),%0              \n"
+      "movq        (%5),%%xmm2                   \n"
+      "movhps      0x00(%5,%1,1),%%xmm2          \n"
+      "movq        0x00(%5,%1,2),%%xmm3          \n"
+      "movhps      0x00(%5,%4,1),%%xmm3          \n"
+      "lea         0x00(%5,%1,4),%5              \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm2           \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),        // %0
+        "+r"(src_stepx_x4),    // %1
+        "+r"(dst_argb),        // %2
+        "+rm"(dst_width),      // %3
+        "=&r"(src_stepx_x12),  // %4
+        "+r"(row1)             // %5
+        ::"memory",
+        "cc", "xmm0", "xmm1", "xmm2", "xmm3");
 }
 
-void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx) {
+void ScaleARGBCols_SSE2(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx) {
   intptr_t x0, x1;
-  asm volatile (
-    "movd      %5,%%xmm2                       \n"
-    "movd      %6,%%xmm3                       \n"
-    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
-    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
-    "paddd     %%xmm0,%%xmm2                   \n"
-    "paddd     %%xmm3,%%xmm3                   \n"
-    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
-    "paddd     %%xmm0,%%xmm2                   \n"
-    "paddd     %%xmm3,%%xmm3                   \n"
-    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-    "pextrw    $0x1,%%xmm2,%k0                 \n"
-    "pextrw    $0x3,%%xmm2,%k1                 \n"
-    "cmp       $0x0,%4                         \n"
-    "jl        99f                             \n"
-    "sub       $0x4,%4                         \n"
-    "jl        49f                             \n"
-
-    LABELALIGN
-  "40:                                         \n"
-    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
-    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
-    "pextrw    $0x5,%%xmm2,%k0                 \n"
-    "pextrw    $0x7,%%xmm2,%k1                 \n"
-    "paddd     %%xmm3,%%xmm2                   \n"
-    "punpckldq %%xmm1,%%xmm0                   \n"
-    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
-    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
-    "pextrw    $0x1,%%xmm2,%k0                 \n"
-    "pextrw    $0x3,%%xmm2,%k1                 \n"
-    "punpckldq %%xmm4,%%xmm1                   \n"
-    "punpcklqdq %%xmm1,%%xmm0                  \n"
-    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x10,2) ",%2           \n"
-    "sub       $0x4,%4                         \n"
-    "jge       40b                             \n"
-
-  "49:                                         \n"
-    "test      $0x2,%4                         \n"
-    "je        29f                             \n"
-    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
-    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
-    "pextrw    $0x5,%%xmm2,%k0                 \n"
-    "punpckldq %%xmm1,%%xmm0                   \n"
-    "movq      %%xmm0," MEMACCESS(2) "         \n"
-    "lea       " MEMLEA(0x8,2) ",%2            \n"
-  "29:                                         \n"
-    "test      $0x1,%4                         \n"
-    "je        99f                             \n"
-    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
-    "movd      %%xmm0," MEMACCESS(2) "         \n"
-  "99:                                         \n"
-  : "=&a"(x0),         // %0
-    "=&d"(x1),         // %1
-    "+r"(dst_argb),    // %2
-    "+r"(src_argb),    // %3
-    "+r"(dst_width)    // %4
-  : "rm"(x),           // %5
-    "rm"(dx)           // %6
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
-  );
+  asm volatile(
+      "movd        %5,%%xmm2                     \n"
+      "movd        %6,%%xmm3                     \n"
+      "pshufd      $0x0,%%xmm2,%%xmm2            \n"
+      "pshufd      $0x11,%%xmm3,%%xmm0           \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pshufd      $0x5,%%xmm3,%%xmm0            \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pextrw      $0x1,%%xmm2,%k0               \n"
+      "pextrw      $0x3,%%xmm2,%k1               \n"
+      "cmp         $0x0,%4                       \n"
+      "jl          99f                           \n"
+      "sub         $0x4,%4                       \n"
+      "jl          49f                           \n"
+
+      LABELALIGN
+      "40:                                       \n"
+      "movd        0x00(%3,%0,4),%%xmm0          \n"
+      "movd        0x00(%3,%1,4),%%xmm1          \n"
+      "pextrw      $0x5,%%xmm2,%k0               \n"
+      "pextrw      $0x7,%%xmm2,%k1               \n"
+      "paddd       %%xmm3,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"
+      "movd        0x00(%3,%0,4),%%xmm1          \n"
+      "movd        0x00(%3,%1,4),%%xmm4          \n"
+      "pextrw      $0x1,%%xmm2,%k0               \n"
+      "pextrw      $0x3,%%xmm2,%k1               \n"
+      "punpckldq   %%xmm4,%%xmm1                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%4                       \n"
+      "jge         40b                           \n"
+
+      "49:                                       \n"
+      "test        $0x2,%4                       \n"
+      "je          29f                           \n"
+      "movd        0x00(%3,%0,4),%%xmm0          \n"
+      "movd        0x00(%3,%1,4),%%xmm1          \n"
+      "pextrw      $0x5,%%xmm2,%k0               \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"
+      "movq        %%xmm0,(%2)                   \n"
+      "lea         0x8(%2),%2                    \n"
+      "29:                                       \n"
+      "test        $0x1,%4                       \n"
+      "je          99f                           \n"
+      "movd        0x00(%3,%0,4),%%xmm0          \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "99:                                       \n"
+      : "=&a"(x0),       // %0
+        "=&d"(x1),       // %1
+        "+r"(dst_argb),  // %2
+        "+r"(src_argb),  // %3
+        "+r"(dst_width)  // %4
+      : "rm"(x),         // %5
+        "rm"(dx)         // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
 }
 
 // Reads 4 pixels, duplicates them and writes 8 pixels.
 // Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
-void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
-                           int dst_width, int x, int dx) {
-  asm volatile (
-    LABELALIGN
-  "1:                                          \n"
-    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
-    "lea       " MEMLEA(0x10,1) ",%1           \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpckldq %%xmm0,%%xmm0                   \n"
-    "punpckhdq %%xmm1,%%xmm1                   \n"
-    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
-    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
-    "lea       " MEMLEA(0x20,0) ",%0           \n"
-    "sub       $0x8,%2                         \n"
-    "jg        1b                              \n"
-
-  : "+r"(dst_argb),    // %0
-    "+r"(src_argb),    // %1
-    "+r"(dst_width)    // %2
-  :: "memory", "cc", NACL_R14
-    "xmm0", "xmm1"
-  );
+void ScaleARGBColsUp2_SSE2(uint8_t* dst_argb,
+                           const uint8_t* src_argb,
+                           int dst_width,
+                           int x,
+                           int dx) {
+  (void)x;
+  (void)dx;
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpckldq   %%xmm0,%%xmm0                 \n"
+      "punpckhdq   %%xmm1,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "movdqu      %%xmm1,0x10(%0)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+
+               : "+r"(dst_argb),  // %0
+                 "+r"(src_argb),  // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
 // Shuffle table for arranging 2 pixels into pairs for pmaddubsw
-static uvec8 kShuffleColARGB = {
-  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
-  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+static const uvec8 kShuffleColARGB = {
+    0u, 4u,  1u, 5u,  2u,  6u,  3u,  7u,  // bbggrraa 1st pixel
+    8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
 };
 
 // Shuffle table for duplicating 2 fractions into 8 bytes each
-static uvec8 kShuffleFractions = {
-  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+static const uvec8 kShuffleFractions = {
+    0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
 };
 
 // Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
-void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
-                               int dst_width, int x, int dx) {
+void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
+                               const uint8_t* src_argb,
+                               int dst_width,
+                               int x,
+                               int dx) {
   intptr_t x0, x1;
-  asm volatile (
-    "movdqa    %0,%%xmm4                       \n"
-    "movdqa    %1,%%xmm5                       \n"
-  :
-  : "m"(kShuffleColARGB),  // %0
-    "m"(kShuffleFractions)  // %1
+  asm volatile(
+      "movdqa      %0,%%xmm4                     \n"
+      "movdqa      %1,%%xmm5                     \n"
+      :
+      : "m"(kShuffleColARGB),   // %0
+        "m"(kShuffleFractions)  // %1
   );
 
-  asm volatile (
-    "movd      %5,%%xmm2                       \n"
-    "movd      %6,%%xmm3                       \n"
-    "pcmpeqb   %%xmm6,%%xmm6                   \n"
-    "psrlw     $0x9,%%xmm6                     \n"
-    "pextrw    $0x1,%%xmm2,%k3                 \n"
-    "sub       $0x2,%2                         \n"
-    "jl        29f                             \n"
-    "movdqa    %%xmm2,%%xmm0                   \n"
-    "paddd     %%xmm3,%%xmm0                   \n"
-    "punpckldq %%xmm0,%%xmm2                   \n"
-    "punpckldq %%xmm3,%%xmm3                   \n"
-    "paddd     %%xmm3,%%xmm3                   \n"
-    "pextrw    $0x3,%%xmm2,%k4                 \n"
-
-    LABELALIGN
-  "2:                                          \n"
-    "movdqa    %%xmm2,%%xmm1                   \n"
-    "paddd     %%xmm3,%%xmm2                   \n"
-    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
-    "psrlw     $0x9,%%xmm1                     \n"
-    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
-    "pshufb    %%xmm5,%%xmm1                   \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "pxor      %%xmm6,%%xmm1                   \n"
-    "pmaddubsw %%xmm1,%%xmm0                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "pextrw    $0x1,%%xmm2,%k3                 \n"
-    "pextrw    $0x3,%%xmm2,%k4                 \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movq      %%xmm0," MEMACCESS(0) "         \n"
-    "lea       " MEMLEA(0x8,0) ",%0            \n"
-    "sub       $0x2,%2                         \n"
-    "jge       2b                              \n"
-
-    LABELALIGN
-  "29:                                         \n"
-    "add       $0x1,%2                         \n"
-    "jl        99f                             \n"
-    "psrlw     $0x9,%%xmm2                     \n"
-    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
-    "pshufb    %%xmm5,%%xmm2                   \n"
-    "pshufb    %%xmm4,%%xmm0                   \n"
-    "pxor      %%xmm6,%%xmm2                   \n"
-    "pmaddubsw %%xmm2,%%xmm0                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movd      %%xmm0," MEMACCESS(0) "         \n"
-
-    LABELALIGN
-  "99:                                         \n"
-  : "+r"(dst_argb),    // %0
-    "+r"(src_argb),    // %1
-    "+rm"(dst_width),  // %2
-    "=&r"(x0),         // %3
-    "=&r"(x1)          // %4
-  : "rm"(x),           // %5
-    "rm"(dx)           // %6
-  : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
-  );
+  asm volatile(
+      "movd        %5,%%xmm2                     \n"
+      "movd        %6,%%xmm3                     \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrlw       $0x9,%%xmm6                   \n"
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "sub         $0x2,%2                       \n"
+      "jl          29f                           \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "paddd       %%xmm3,%%xmm0                 \n"
+      "punpckldq   %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm3,%%xmm3                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
+
+      LABELALIGN
+      "2:                                        \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"
+      "movq        0x00(%1,%3,4),%%xmm0          \n"
+      "psrlw       $0x9,%%xmm1                   \n"
+      "movhps      0x00(%1,%4,4),%%xmm0          \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pxor        %%xmm6,%%xmm1                 \n"
+      "pmaddubsw   %%xmm1,%%xmm0                 \n"
+      "psrlw       $0x7,%%xmm0                   \n"
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%0)                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "sub         $0x2,%2                       \n"
+      "jge         2b                            \n"
+
+      LABELALIGN
+      "29:                                       \n"
+      "add         $0x1,%2                       \n"
+      "jl          99f                           \n"
+      "psrlw       $0x9,%%xmm2                   \n"
+      "movq        0x00(%1,%3,4),%%xmm0          \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pxor        %%xmm6,%%xmm2                 \n"
+      "pmaddubsw   %%xmm2,%%xmm0                 \n"
+      "psrlw       $0x7,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movd        %%xmm0,(%0)                   \n"
+
+      LABELALIGN
+      "99:                                       \n"  // clang-format error.
+
+      : "+r"(dst_argb),    // %0
+        "+r"(src_argb),    // %1
+        "+rm"(dst_width),  // %2
+        "=&r"(x0),         // %3
+        "=&r"(x1)          // %4
+      : "rm"(x),           // %5
+        "rm"(dx)           // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 
 // Divide num by div and return as 16.16 fixed point result.
 int FixedDiv_X86(int num, int div) {
-  asm volatile (
-    "cdq                                       \n"
-    "shld      $0x10,%%eax,%%edx               \n"
-    "shl       $0x10,%%eax                     \n"
-    "idiv      %1                              \n"
-    "mov       %0, %%eax                       \n"
-    : "+a"(num)  // %0
-    : "c"(div)   // %1
-    : "memory", "cc", "edx"
-  );
+  asm volatile(
+      "cdq                                       \n"
+      "shld        $0x10,%%eax,%%edx             \n"
+      "shl         $0x10,%%eax                   \n"
+      "idiv        %1                            \n"
+      "mov         %0, %%eax                     \n"
+      : "+a"(num)  // %0
+      : "c"(div)   // %1
+      : "memory", "cc", "edx");
   return num;
 }
 
 // Divide num - 1 by div - 1 and return as 16.16 fixed point result.
 int FixedDiv1_X86(int num, int div) {
-  asm volatile (
-    "cdq                                       \n"
-    "shld      $0x10,%%eax,%%edx               \n"
-    "shl       $0x10,%%eax                     \n"
-    "sub       $0x10001,%%eax                  \n"
-    "sbb       $0x0,%%edx                      \n"
-    "sub       $0x1,%1                         \n"
-    "idiv      %1                              \n"
-    "mov       %0, %%eax                       \n"
-    : "+a"(num)  // %0
-    : "c"(div)   // %1
-    : "memory", "cc", "edx"
-  );
+  asm volatile(
+      "cdq                                       \n"
+      "shld        $0x10,%%eax,%%edx             \n"
+      "shl         $0x10,%%eax                   \n"
+      "sub         $0x10001,%%eax                \n"
+      "sbb         $0x0,%%edx                    \n"
+      "sub         $0x1,%1                       \n"
+      "idiv        %1                            \n"
+      "mov         %0, %%eax                     \n"
+      : "+a"(num)  // %0
+      : "c"(div)   // %1
+      : "memory", "cc", "edx");
   return num;
 }
 
+#if defined(HAS_SCALEUVROWDOWN2BOX_SSSE3) || \
+    defined(HAS_SCALEUVROWDOWN2BOX_AVX2)
+
+// Shuffle table for splitting UV into upper and lower part of register.
+static const uvec8 kShuffleSplitUV = {0u, 2u, 4u, 6u, 8u, 10u, 12u, 14u,
+                                      1u, 3u, 5u, 7u, 9u, 11u, 13u, 15u};
+static const uvec8 kShuffleMergeUV = {0u,   8u,   2u,   10u,  4u,   12u,
+                                      6u,   14u,  0x80, 0x80, 0x80, 0x80,
+                                      0x80, 0x80, 0x80, 0x80};
+#endif
+
+#ifdef HAS_SCALEUVROWDOWN2BOX_SSSE3
+
+void ScaleUVRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"  // 01010101
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5, %%xmm5                \n"  // zero
+      "movdqa      %4,%%xmm1                     \n"  // split shuffler
+      "movdqa      %5,%%xmm3                     \n"  // merge shuffler
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // 8 UV row 0
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"  // 8 UV row 1
+      "lea         0x10(%0),%0                   \n"
+      "pshufb      %%xmm1,%%xmm0                 \n"  // uuuuvvvv
+      "pshufb      %%xmm1,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"  // horizontal add
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"  // vertical add
+      "psrlw       $0x1,%%xmm0                   \n"  // round
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm3,%%xmm0                 \n"  // merge uv
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"  // 4 UV
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "m"(kShuffleSplitUV),         // %4
+        "m"(kShuffleMergeUV)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_SCALEUVROWDOWN2BOX_SSSE3
+
+#ifdef HAS_SCALEUVROWDOWN2BOX_AVX2
+void ScaleUVRowDown2Box_AVX2(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"  // 01010101
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"  // zero
+      "vbroadcastf128 %4,%%ymm1                  \n"  // split shuffler
+      "vbroadcastf128 %5,%%ymm3                  \n"  // merge shuffler
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // 16 UV row 0
+      "vmovdqu     0x00(%0,%3,1),%%ymm2          \n"  // 16 UV row 1
+      "lea         0x20(%0),%0                   \n"
+      "vpshufb     %%ymm1,%%ymm0,%%ymm0          \n"  // uuuuvvvv
+      "vpshufb     %%ymm1,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"  // horizontal add
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"  // vertical add
+      "vpsrlw      $0x1,%%ymm0,%%ymm0            \n"  // round
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm3,%%ymm0,%%ymm0          \n"  // merge uv
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"  // combine qwords
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"  // 8 UV
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "m"(kShuffleSplitUV),         // %4
+        "m"(kShuffleMergeUV)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_SCALEUVROWDOWN2BOX_AVX2
+
+static const uvec8 kUVLinearMadd31 = {3, 1, 3, 1, 1, 3, 1, 3,
+                                      3, 1, 3, 1, 1, 3, 1, 3};
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_SSSE3
+void ScaleUVRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm4,%%xmm4                 \n"
+      "psrlw       $15,%%xmm4                    \n"
+      "psllw       $1,%%xmm4                     \n"  // all 2
+      "movdqa      %3,%%xmm3                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 00112233 (1u1v)
+      "movq        2(%0),%%xmm1                  \n"  // 11223344 (1u1v)
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // 0101121223233434 (2u2v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm0,%%xmm2                 \n"  // 2323232334343434 (2u2v)
+      "punpckldq   %%xmm0,%%xmm0                 \n"  // 0101010112121212 (2u2v)
+      "pmaddubsw   %%xmm3,%%xmm2                 \n"  // 3*near+far (1u1v16, hi)
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"  // 3*near+far (1u1v16, lo)
+      "paddw       %%xmm4,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddw       %%xmm4,%%xmm2                 \n"  // 3*near+far+2 (hi)
+      "psrlw       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrlw       $2,%%xmm2                     \n"  // 3/4*near+1/4*far (hi)
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),        // %0
+        "+r"(dst_ptr),        // %1
+        "+r"(dst_width)       // %2
+      : "m"(kUVLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_SSSE3
+void ScaleUVRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm6,%%xmm6                 \n"
+      "psrlw       $15,%%xmm6                    \n"
+      "psllw       $3,%%xmm6                     \n"  // all 8
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 00112233 (1u1v)
+      "movq        2(%0),%%xmm1                  \n"  // 11223344 (1u1v)
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // 0101121223233434 (2u2v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm0,%%xmm2                 \n"  // 2323232334343434 (2u2v)
+      "punpckldq   %%xmm0,%%xmm0                 \n"  // 0101010112121212 (2u2v)
+      "pmaddubsw   %%xmm7,%%xmm2                 \n"  // 3*near+far (1u1v16, hi)
+      "pmaddubsw   %%xmm7,%%xmm0                 \n"  // 3*near+far (1u1v16, lo)
+
+      "movq        (%0,%3),%%xmm1                \n"
+      "movq        2(%0,%3),%%xmm4               \n"
+      "punpcklbw   %%xmm4,%%xmm1                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpckhdq   %%xmm1,%%xmm3                 \n"
+      "punpckldq   %%xmm1,%%xmm1                 \n"
+      "pmaddubsw   %%xmm7,%%xmm3                 \n"  // 3*near+far (2, hi)
+      "pmaddubsw   %%xmm7,%%xmm1                 \n"  // 3*near+far (2, lo)
+
+      // xmm0 xmm2
+      // xmm1 xmm3
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm1,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm1,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrlw       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm2,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm1,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm3,%%xmm1                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm2,%%xmm1                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrlw       $4,%%xmm1                     \n"  // ^ div by 16 (2, hi)
+
+      "packuswb    %%xmm0,%%xmm4                 \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packuswb    %%xmm1,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4)                \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kUVLinearMadd31)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_AVX2
+
+void ScaleUVRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $1,%%ymm4,%%ymm4              \n"  // all 2
+      "vbroadcastf128 %3,%%ymm3                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"
+      "vmovdqu     2(%0),%%xmm1                  \n"
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpunpckhdq  %%ymm0,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm3,%%ymm2,%%ymm1          \n"  // 3*near+far (hi)
+      "vpmaddubsw  %%ymm3,%%ymm0,%%ymm0          \n"  // 3*near+far (lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"  // 3*near+far+2 (hi)
+      "vpsrlw      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrlw      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 uv to 16 uv
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),        // %0
+        "+r"(dst_ptr),        // %1
+        "+r"(dst_width)       // %2
+      : "m"(kUVLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_AVX2
+void ScaleUVRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $15,%%ymm6,%%ymm6             \n"
+      "vpsllw      $3,%%ymm6,%%ymm6              \n"  // all 8
+      "vbroadcastf128 %5,%%ymm7                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"
+      "vmovdqu     2(%0),%%xmm1                  \n"
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpunpckhdq  %%ymm0,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm1          \n"  // 3*near+far (1, hi)
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm0          \n"  // 3*near+far (1, lo)
+
+      "vmovdqu     (%0,%3),%%xmm2                \n"  // 0123456789ABCDEF
+      "vmovdqu     2(%0,%3),%%xmm3               \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm2,%%ymm2     \n"
+      "vpermq      $0b11011000,%%ymm3,%%ymm3     \n"
+      "vpunpcklbw  %%ymm3,%%ymm2,%%ymm2          \n"
+      "vpunpckhdq  %%ymm2,%%ymm2,%%ymm4          \n"
+      "vpunpckldq  %%ymm2,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm7,%%ymm4,%%ymm3          \n"  // 3*near+far (2, hi)
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm2          \n"  // 3*near+far (2, lo)
+
+      // ymm0 ymm1
+      // ymm2 ymm3
+
+      "vpaddw      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddw      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddw      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrlw      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddw      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddw      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddw      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddw      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrlw      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddw      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddw      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddw      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddw      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddw      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddw      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddw      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrlw      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackuswb   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackuswb   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vmovdqu     %%ymm5,(%1,%4)                \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 uv to 16 uv
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kUVLinearMadd31)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+void ScaleUVRowUp2_Linear_16_SSE41(const uint16_t* src_ptr,
+                                   uint16_t* dst_ptr,
+                                   int dst_width) {
+  asm volatile(
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "pcmpeqd     %%xmm4,%%xmm4                 \n"
+      "psrld       $31,%%xmm4                    \n"
+      "pslld       $1,%%xmm4                     \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0011 (16b, 1u1v)
+      "movq        4(%0),%%xmm1                  \n"  // 1122 (16b, 1u1v)
+
+      "punpcklwd   %%xmm5,%%xmm0                 \n"  // 0011 (32b, 1u1v)
+      "punpcklwd   %%xmm5,%%xmm1                 \n"  // 1122 (32b, 1u1v)
+
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+
+      "pshufd      $0b01001110,%%xmm2,%%xmm2     \n"  // 1100 (lo, far)
+      "pshufd      $0b01001110,%%xmm3,%%xmm3     \n"  // 2211 (hi, far)
+
+      "paddd       %%xmm4,%%xmm2                 \n"  // far+2 (lo)
+      "paddd       %%xmm4,%%xmm3                 \n"  // far+2 (hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far+2 (lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far+2 (hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far+2 (hi)
+
+      "psrld       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrld       $2,%%xmm1                     \n"  // 3/4*near+1/4*far (hi)
+      "packusdw    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 2 uv to 4 uv
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+void ScaleUVRowUp2_Bilinear_16_SSE41(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint16_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width) {
+  asm volatile(
+      "pxor        %%xmm7,%%xmm7                 \n"
+      "pcmpeqd     %%xmm6,%%xmm6                 \n"
+      "psrld       $31,%%xmm6                    \n"
+      "pslld       $3,%%xmm6                     \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0011 (16b, 1u1v)
+      "movq        4(%0),%%xmm1                  \n"  // 1122 (16b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm0                 \n"  // 0011 (near) (32b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm1                 \n"  // 1122 (near) (32b, 1u1v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pshufd      $0b01001110,%%xmm2,%%xmm2     \n"  // 1100 (far) (1, lo)
+      "pshufd      $0b01001110,%%xmm3,%%xmm3     \n"  // 2211 (far) (1, hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far (1, lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far (1, hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (1, lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      "movq        (%0,%3,2),%%xmm2              \n"
+      "movq        4(%0,%3,2),%%xmm3             \n"
+      "punpcklwd   %%xmm7,%%xmm2                 \n"
+      "punpcklwd   %%xmm7,%%xmm3                 \n"
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "pshufd      $0b01001110,%%xmm4,%%xmm4     \n"  // 1100 (far) (2, lo)
+      "pshufd      $0b01001110,%%xmm5,%%xmm5     \n"  // 2211 (far) (2, hi)
+      "paddd       %%xmm2,%%xmm4                 \n"  // near+far (2, lo)
+      "paddd       %%xmm3,%%xmm5                 \n"  // near+far (2, hi)
+      "paddd       %%xmm2,%%xmm2                 \n"  // 2*near (2, lo)
+      "paddd       %%xmm3,%%xmm3                 \n"  // 2*near (2, hi)
+      "paddd       %%xmm4,%%xmm2                 \n"  // 3*near+far (2, lo)
+      "paddd       %%xmm5,%%xmm3                 \n"  // 3*near+far (2, hi)
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddd       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddd       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddd       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrld       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm2,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddd       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddd       %%xmm2,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddd       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrld       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddd       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (2, hi)
+      "paddd       %%xmm1,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrld       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"  // 6*near+2*far (2, hi)
+      "paddd       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (1, hi)
+      "paddd       %%xmm3,%%xmm2                 \n"  // 9*near+3*far (2, hi)
+      "paddd       %%xmm1,%%xmm2                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrld       $4,%%xmm2                     \n"  // ^ div by 16 (2, hi)
+
+      "packusdw    %%xmm0,%%xmm4                 \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packusdw    %%xmm2,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 2 uv to 4 uv
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+void ScaleUVRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrld      $31,%%ymm4,%%ymm4             \n"
+      "vpslld      $1,%%ymm4,%%ymm4              \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00112233 (16b, 1u1v)
+      "vmovdqu     4(%0),%%xmm1                  \n"  // 11223344 (16b, 1u1v)
+
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+
+      "vpshufd     $0b01001110,%%ymm0,%%ymm2     \n"  // 11003322 (lo, far)
+      "vpshufd     $0b01001110,%%ymm1,%%ymm3     \n"  // 22114433 (hi, far)
+
+      "vpaddd      %%ymm4,%%ymm2,%%ymm2          \n"  // far+2 (lo)
+      "vpaddd      %%ymm4,%%ymm3,%%ymm3          \n"  // far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far+2 (hi)
+
+      "vpsrld      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrld      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackusdw   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+void ScaleUVRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrld      $31,%%ymm6,%%ymm6             \n"
+      "vpslld      $3,%%ymm6,%%ymm6              \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00112233 (16b, 1u1v)
+      "vmovdqu     4(%0),%%xmm1                  \n"  // 11223344 (16b, 1u1v)
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b01001110,%%ymm0,%%ymm2     \n"  // 11003322 (lo, far)
+      "vpshufd     $0b01001110,%%ymm1,%%ymm3     \n"  // 22114433 (hi, far)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far (hi)
+
+      "vmovdqu     (%0,%3,2),%%xmm2              \n"  // 00112233 (16b, 1u1v)
+      "vmovdqu     4(%0,%3,2),%%xmm3             \n"  // 11223344 (16b, 1u1v)
+      "vpmovzxwd   %%xmm2,%%ymm2                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm3,%%ymm3                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b01001110,%%ymm2,%%ymm4     \n"  // 11003322 (lo, far)
+      "vpshufd     $0b01001110,%%ymm3,%%ymm5     \n"  // 22114433 (hi, far)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm4          \n"  // near+far (lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm5          \n"  // near+far (hi)
+      "vpaddd      %%ymm2,%%ymm2,%%ymm2          \n"  // 2*near (lo)
+      "vpaddd      %%ymm3,%%ymm3,%%ymm3          \n"  // 2*near (hi)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm2          \n"  // 3*near+far (lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm3          \n"  // 3*near+far (hi)
+
+      "vpaddd      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddd      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddd      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddd      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrld      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddd      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddd      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddd      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddd      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrld      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddd      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddd      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddd      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrld      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddd      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddd      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddd      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddd      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrld      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackusdw   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackusdw   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vmovdqu     %%ymm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
 #endif  // defined(__x86_64__) || defined(__i386__)
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/scale_lsx.cc b/media/libyuv/source/scale_lsx.cc
new file mode 100644
index 0000000000..bfe5e9fbba
--- /dev/null
+++ b/media/libyuv/source/scale_lsx.cc
@@ -0,0 +1,739 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "libyuv/scale_row.h"
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define LOAD_DATA(_src, _in, _out)                                       \
+  {                                                                      \
+    int _tmp1, _tmp2, _tmp3, _tmp4;                                      \
+    DUP4_ARG2(__lsx_vpickve2gr_w, _in, 0, _in, 1, _in, 2, _in, 3, _tmp1, \
+              _tmp2, _tmp3, _tmp4);                                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp1], 0);                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp2], 1);                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp3], 2);                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp4], 3);                      \
+  }
+
+void ScaleARGBRowDown2_LSX(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  (void)src_stride;
+  __m128i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src0, src1);
+    dst0 = __lsx_vpickod_w(src1, src0);
+    __lsx_vst(dst0, dst_argb, 0);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Linear_LSX(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  (void)src_stride;
+  __m128i src0, src1, tmp0, tmp1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src0, src1);
+    tmp0 = __lsx_vpickev_w(src1, src0);
+    tmp1 = __lsx_vpickod_w(src1, src0);
+    dst0 = __lsx_vavgr_bu(tmp1, tmp0);
+    __lsx_vst(dst0, dst_argb, 0);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Box_LSX(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  const uint8_t* s = src_argb;
+  const uint8_t* t = src_argb + src_stride;
+  __m128i src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i shuff = {0x0703060205010400, 0x0F0B0E0A0D090C08};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, s, 0, s, 16, src0, src1);
+    DUP2_ARG2(__lsx_vld, t, 0, t, 16, src2, src3);
+    DUP4_ARG3(__lsx_vshuf_b, src0, src0, shuff, src1, src1, shuff, src2, src2,
+              shuff, src3, src3, shuff, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vhaddw_hu_bu, tmp0, tmp0, tmp1, tmp1, tmp2, tmp2, tmp3,
+              tmp3, reg0, reg1, reg2, reg3);
+    DUP2_ARG2(__lsx_vsadd_hu, reg0, reg2, reg1, reg3, reg0, reg1);
+    dst0 = __lsx_vsrarni_b_h(reg1, reg0, 2);
+    __lsx_vst(dst0, dst_argb, 0);
+    s += 32;
+    t += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEven_LSX(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int32_t src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  int32_t stepx = src_stepx << 2;
+  (void)src_stride;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    dst0 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    dst1 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    dst2 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    dst3 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    __lsx_vstelm_w(dst0, dst_argb, 0, 0);
+    __lsx_vstelm_w(dst1, dst_argb, 4, 0);
+    __lsx_vstelm_w(dst2, dst_argb, 8, 0);
+    __lsx_vstelm_w(dst3, dst_argb, 12, 0);
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEvenBox_LSX(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  int32_t stepx = src_stepx * 4;
+  const uint8_t* next_argb = src_argb + src_stride;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, dst0;
+
+  for (x = 0; x < len; x++) {
+    tmp0 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp1 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp2 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp3 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp4 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    tmp5 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    tmp6 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    tmp7 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    DUP4_ARG2(__lsx_vilvl_d, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vaddwev_h_bu, src0, src2, src1, src3, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vaddwod_h_bu, src0, src2, src1, src3, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vpackev_w, tmp1, tmp0, tmp3, tmp2, reg0, reg1);
+    DUP2_ARG2(__lsx_vpackod_w, tmp1, tmp0, tmp3, tmp2, tmp4, tmp5);
+    DUP2_ARG2(__lsx_vadd_h, reg0, tmp4, reg1, tmp5, reg0, reg1);
+    dst0 = __lsx_vsrarni_b_h(reg1, reg0, 2);
+    dst0 = __lsx_vshuf4i_b(dst0, 0xD8);
+    __lsx_vst(dst0, dst_argb, 0);
+    dst_argb += 16;
+  }
+}
+
+void ScaleRowDown2_LSX(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  int len = dst_width / 32;
+  __m128i src0, src1, src2, src3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, dst0, dst1);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Linear_LSX(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  int x;
+  int len = dst_width / 32;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vavgr_bu, tmp0, tmp1, tmp2, tmp3, dst0, dst1);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Box_LSX(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  int len = dst_width / 32;
+  const uint8_t* src_nex = src_ptr + src_stride;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, src_nex, 0, src_nex, 16, src_nex, 32, src_nex, 48,
+              src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vaddwev_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vaddwod_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG3(__lsx_vsrarni_b_h, tmp1, tmp0, 2, tmp3, tmp2, 2, dst0, dst1);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    src_ptr += 64;
+    src_nex += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown4_LSX(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  int len = dst_width / 16;
+  __m128i src0, src1, src2, src3, tmp0, tmp1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp1);
+    dst0 = __lsx_vpickod_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst, 0);
+    src_ptr += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown4Box_LSX(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  int len = dst_width / 16;
+  const uint8_t* ptr1 = src_ptr + src_stride;
+  const uint8_t* ptr2 = ptr1 + src_stride;
+  const uint8_t* ptr3 = ptr2 + src_stride;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, ptr1, 0, ptr1, 16, ptr1, 32, ptr1, 48, src4, src5,
+              src6, src7);
+    DUP4_ARG2(__lsx_vaddwev_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vaddwod_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+              reg0, reg1, reg2, reg3);
+    DUP4_ARG2(__lsx_vld, ptr2, 0, ptr2, 16, ptr2, 32, ptr2, 48, src0, src1,
+              src2, src3);
+    DUP4_ARG2(__lsx_vld, ptr3, 0, ptr3, 16, ptr3, 32, ptr3, 48, src4, src5,
+              src6, src7);
+    DUP4_ARG2(__lsx_vaddwev_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vaddwod_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+              reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vadd_h, reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7,
+              reg0, reg1, reg2, reg3);
+    DUP4_ARG2(__lsx_vhaddw_wu_hu, reg0, reg0, reg1, reg1, reg2, reg2, reg3,
+              reg3, reg0, reg1, reg2, reg3);
+    DUP2_ARG3(__lsx_vsrarni_h_w, reg1, reg0, 4, reg3, reg2, 4, tmp0, tmp1);
+    dst0 = __lsx_vpickev_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst, 0);
+    src_ptr += 64;
+    ptr1 += 64;
+    ptr2 += 64;
+    ptr3 += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown38_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x, len;
+  __m128i src0, src1, tmp0;
+  __m128i shuff = {0x13100E0B08060300, 0x000000001E1B1816};
+
+  assert(dst_width % 3 == 0);
+  len = dst_width / 12;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+    tmp0 = __lsx_vshuf_b(src1, src0, shuff);
+    __lsx_vstelm_d(tmp0, dst, 0, 0);
+    __lsx_vstelm_w(tmp0, dst, 8, 2);
+    src_ptr += 32;
+    dst += 12;
+  }
+}
+
+void ScaleRowDown38_2_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, len;
+  const uint8_t* src_nex = src_ptr + src_stride;
+  __m128i src0, src1, src2, src3, dst0;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i shuff = {0x0A08160604120200, 0x000000001E0E0C1A};
+  __m128i const_0x2AAA = __lsx_vreplgr2vr_h(0x2AAA);
+  __m128i const_0x4000 = __lsx_vreplgr2vr_w(0x4000);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  len = dst_width / 12;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_nex, 0, src_nex, 16, src0,
+              src1, src2, src3);
+    DUP2_ARG2(__lsx_vaddwev_h_bu, src0, src2, src1, src3, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vaddwod_h_bu, src0, src2, src1, src3, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vpickev_h, tmp2, tmp0, tmp3, tmp1, reg0, reg1);
+    DUP2_ARG2(__lsx_vpackod_h, tmp1, tmp0, tmp3, tmp2, reg2, reg3);
+    tmp4 = __lsx_vpickev_w(reg3, reg2);
+    tmp5 = __lsx_vadd_h(reg0, reg1);
+    tmp6 = __lsx_vadd_h(tmp5, tmp4);
+    tmp7 = __lsx_vmuh_h(tmp6, const_0x2AAA);
+    tmp0 = __lsx_vpickod_w(reg3, reg2);
+    tmp1 = __lsx_vhaddw_wu_hu(tmp0, tmp0);
+    tmp2 = __lsx_vmul_w(tmp1, const_0x4000);
+    dst0 = __lsx_vshuf_b(tmp2, tmp7, shuff);
+    __lsx_vstelm_d(dst0, dst_ptr, 0, 0);
+    __lsx_vstelm_w(dst0, dst_ptr, 8, 2);
+    src_ptr += 32;
+    src_nex += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleRowDown38_3_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, len;
+  const uint8_t* ptr1 = src_ptr + src_stride;
+  const uint8_t* ptr2 = ptr1 + src_stride;
+  __m128i src0, src1, src2, src3, src4, src5;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, dst0;
+  __m128i zero = __lsx_vldi(0);
+  __m128i shuff = {0x0A08160604120200, 0x000000001E0E0C1A};
+  __m128i const_0x1C71 = __lsx_vreplgr2vr_h(0x1C71);
+  __m128i const_0x2AAA = __lsx_vreplgr2vr_w(0x2AAA);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  len = dst_width / 12;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, ptr1, 0, ptr1, 16, src0, src1,
+              src2, src3);
+    DUP2_ARG2(__lsx_vld, ptr2, 0, ptr2, 16, src4, src5);
+    DUP2_ARG2(__lsx_vaddwev_h_bu, src0, src2, src1, src3, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vaddwod_h_bu, src0, src2, src1, src3, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vpackev_b, zero, src4, zero, src5, tmp4, tmp6);
+    DUP2_ARG2(__lsx_vpackod_b, zero, src4, zero, src5, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp4, tmp1, tmp5, tmp2, tmp6, tmp3, tmp7,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG2(__lsx_vpickev_h, tmp2, tmp0, tmp3, tmp1, reg0, reg1);
+    DUP2_ARG2(__lsx_vpackod_h, tmp1, tmp0, tmp3, tmp2, reg2, reg3);
+    tmp4 = __lsx_vpickev_w(reg3, reg2);
+    tmp5 = __lsx_vadd_h(reg0, reg1);
+    tmp6 = __lsx_vadd_h(tmp5, tmp4);
+    tmp7 = __lsx_vmuh_h(tmp6, const_0x1C71);
+    tmp0 = __lsx_vpickod_w(reg3, reg2);
+    tmp1 = __lsx_vhaddw_wu_hu(tmp0, tmp0);
+    tmp2 = __lsx_vmul_w(tmp1, const_0x2AAA);
+    dst0 = __lsx_vshuf_b(tmp2, tmp7, shuff);
+    __lsx_vstelm_d(dst0, dst_ptr, 0, 0);
+    __lsx_vstelm_w(dst0, dst_ptr, 8, 2);
+    src_ptr += 32;
+    ptr1 += 32;
+    ptr2 += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleAddRow_LSX(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
+  int x;
+  int len = src_width / 16;
+  __m128i src0, tmp0, tmp1, dst0, dst1;
+  __m128i zero = __lsx_vldi(0);
+
+  assert(src_width > 0);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_ptr, 0);
+    DUP2_ARG2(__lsx_vld, dst_ptr, 0, dst_ptr, 16, dst0, dst1);
+    tmp0 = __lsx_vilvl_b(zero, src0);
+    tmp1 = __lsx_vilvh_b(zero, src0);
+    DUP2_ARG2(__lsx_vadd_h, dst0, tmp0, dst1, tmp1, dst0, dst1);
+    __lsx_vst(dst0, dst_ptr, 0);
+    __lsx_vst(dst1, dst_ptr, 16);
+    src_ptr += 16;
+    dst_ptr += 16;
+  }
+}
+
+void ScaleFilterCols_LSX(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  int j;
+  int len = dst_width / 16;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i vec0, vec1, dst0;
+  __m128i vec_x = __lsx_vreplgr2vr_w(x);
+  __m128i vec_dx = __lsx_vreplgr2vr_w(dx);
+  __m128i const1 = __lsx_vreplgr2vr_w(0xFFFF);
+  __m128i const2 = __lsx_vreplgr2vr_w(0x40);
+  __m128i const_tmp = {0x0000000100000000, 0x0000000300000002};
+
+  vec0 = __lsx_vmul_w(vec_dx, const_tmp);
+  vec1 = __lsx_vslli_w(vec_dx, 2);
+  vec_x = __lsx_vadd_w(vec_x, vec0);
+
+  for (j = 0; j < len; j++) {
+    tmp0 = __lsx_vsrai_w(vec_x, 16);
+    tmp4 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp1 = __lsx_vsrai_w(vec_x, 16);
+    tmp5 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp2 = __lsx_vsrai_w(vec_x, 16);
+    tmp6 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp3 = __lsx_vsrai_w(vec_x, 16);
+    tmp7 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    DUP4_ARG2(__lsx_vsrai_w, tmp4, 9, tmp5, 9, tmp6, 9, tmp7, 9, tmp4, tmp5,
+              tmp6, tmp7);
+    LOAD_DATA(src_ptr, tmp0, reg0);
+    LOAD_DATA(src_ptr, tmp1, reg1);
+    LOAD_DATA(src_ptr, tmp2, reg2);
+    LOAD_DATA(src_ptr, tmp3, reg3);
+    DUP4_ARG2(__lsx_vaddi_wu, tmp0, 1, tmp1, 1, tmp2, 1, tmp3, 1, tmp0, tmp1,
+              tmp2, tmp3);
+    LOAD_DATA(src_ptr, tmp0, reg4);
+    LOAD_DATA(src_ptr, tmp1, reg5);
+    LOAD_DATA(src_ptr, tmp2, reg6);
+    LOAD_DATA(src_ptr, tmp3, reg7);
+    DUP4_ARG2(__lsx_vsub_w, reg4, reg0, reg5, reg1, reg6, reg2, reg7, reg3,
+              reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vmul_w, reg4, tmp4, reg5, tmp5, reg6, tmp6, reg7, tmp7,
+              reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vadd_w, reg4, const2, reg5, const2, reg6, const2, reg7,
+              const2, reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vsrai_w, reg4, 7, reg5, 7, reg6, 7, reg7, 7, reg4, reg5,
+              reg6, reg7);
+    DUP4_ARG2(__lsx_vadd_w, reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7,
+              reg0, reg1, reg2, reg3);
+    DUP2_ARG2(__lsx_vpickev_h, reg1, reg0, reg3, reg2, tmp0, tmp1);
+    dst0 = __lsx_vpickev_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst_ptr, 0);
+    dst_ptr += 16;
+  }
+}
+
+void ScaleARGBCols_LSX(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  const uint32_t* src = (const uint32_t*)src_argb;
+  uint32_t* dst = (uint32_t*)dst_argb;
+  int j;
+  int len = dst_width / 4;
+  __m128i tmp0, tmp1, tmp2, dst0;
+  __m128i vec_x = __lsx_vreplgr2vr_w(x);
+  __m128i vec_dx = __lsx_vreplgr2vr_w(dx);
+  __m128i const_tmp = {0x0000000100000000, 0x0000000300000002};
+
+  tmp0 = __lsx_vmul_w(vec_dx, const_tmp);
+  tmp1 = __lsx_vslli_w(vec_dx, 2);
+  vec_x = __lsx_vadd_w(vec_x, tmp0);
+
+  for (j = 0; j < len; j++) {
+    tmp2 = __lsx_vsrai_w(vec_x, 16);
+    vec_x = __lsx_vadd_w(vec_x, tmp1);
+    LOAD_DATA(src, tmp2, dst0);
+    __lsx_vst(dst0, dst, 0);
+    dst += 4;
+  }
+}
+
+void ScaleARGBFilterCols_LSX(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx) {
+  const uint32_t* src = (const uint32_t*)src_argb;
+  int j;
+  int len = dst_width / 8;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i vec0, vec1, dst0, dst1;
+  __m128i vec_x = __lsx_vreplgr2vr_w(x);
+  __m128i vec_dx = __lsx_vreplgr2vr_w(dx);
+  __m128i const_tmp = {0x0000000100000000, 0x0000000300000002};
+  __m128i const_7f = __lsx_vldi(0x7F);
+
+  vec0 = __lsx_vmul_w(vec_dx, const_tmp);
+  vec1 = __lsx_vslli_w(vec_dx, 2);
+  vec_x = __lsx_vadd_w(vec_x, vec0);
+
+  for (j = 0; j < len; j++) {
+    tmp0 = __lsx_vsrai_w(vec_x, 16);
+    reg0 = __lsx_vsrai_w(vec_x, 9);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp1 = __lsx_vsrai_w(vec_x, 16);
+    reg1 = __lsx_vsrai_w(vec_x, 9);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    DUP2_ARG2(__lsx_vand_v, reg0, const_7f, reg1, const_7f, reg0, reg1);
+    DUP2_ARG2(__lsx_vshuf4i_b, reg0, 0, reg1, 0, reg0, reg1);
+    DUP2_ARG2(__lsx_vxor_v, reg0, const_7f, reg1, const_7f, reg2, reg3);
+    DUP2_ARG2(__lsx_vilvl_b, reg0, reg2, reg1, reg3, reg4, reg6);
+    DUP2_ARG2(__lsx_vilvh_b, reg0, reg2, reg1, reg3, reg5, reg7);
+    LOAD_DATA(src, tmp0, src0);
+    LOAD_DATA(src, tmp1, src1);
+    DUP2_ARG2(__lsx_vaddi_wu, tmp0, 1, tmp1, 1, tmp0, tmp1);
+    LOAD_DATA(src, tmp0, src2);
+    LOAD_DATA(src, tmp1, src3);
+    DUP2_ARG2(__lsx_vilvl_b, src2, src0, src3, src1, tmp4, tmp6);
+    DUP2_ARG2(__lsx_vilvh_b, src2, src0, src3, src1, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp4, reg4, tmp5, reg5, tmp6, reg6, tmp7, reg7,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG3(__lsx_vsrani_b_h, tmp1, tmp0, 7, tmp3, tmp2, 7, dst0, dst1);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    dst_argb += 32;
+  }
+}
+
+void ScaleRowDown34_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x;
+  (void)src_stride;
+  __m128i src0, src1, src2, src3;
+  __m128i dst0, dst1, dst2;
+  __m128i shuff0 = {0x0908070504030100, 0x141311100F0D0C0B};
+  __m128i shuff1 = {0x0F0D0C0B09080705, 0x1918171514131110};
+  __m128i shuff2 = {0x141311100F0D0C0B, 0x1F1D1C1B19181715};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0, src2, src1, shuff1, dst0,
+              dst1);
+    dst2 = __lsx_vshuf_b(src3, src2, shuff2);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    __lsx_vst(dst2, dst, 32);
+    src_ptr += 64;
+    dst += 48;
+  }
+}
+
+void ScaleRowDown34_0_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width) {
+  const uint8_t* src_nex = src_ptr + src_stride;
+  int x;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
+  __m128i tmp10, tmp11, dst0, dst1, dst2;
+  __m128i const0 = {0x0103030101010103, 0x0101010303010101};
+  __m128i const1 = {0x0301010101030301, 0x0103030101010103};
+  __m128i const2 = {0x0101010303010101, 0x0301010101030301};
+  __m128i shuff0 = {0x0504030202010100, 0x0A09090807060605};
+  __m128i shuff1 = {0x0F0E0E0D0D0C0B0A, 0x1514131212111110};
+  __m128i shuff2 = {0x0A09090807060605, 0x0F0E0E0D0D0C0B0A};
+  __m128i shift0 = {0x0002000200010002, 0x0001000200020001};
+  __m128i shift1 = {0x0002000100020002, 0x0002000200010002};
+  __m128i shift2 = {0x0001000200020001, 0x0002000100020002};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, src_nex, 0, src_nex, 16, src_nex, 32, src_nex, 48,
+              src4, src5, src6, src7);
+    DUP4_ARG3(__lsx_vshuf_b, src0, src0, shuff0, src1, src0, shuff1, src1, src1,
+              shuff2, src2, src2, shuff0, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG3(__lsx_vshuf_b, src3, src2, shuff1, src3, src3, shuff2, src4, src4,
+              shuff0, src5, src4, shuff1, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vshuf_b, src5, src5, shuff2, src6, src6, shuff0, src7, src6,
+              shuff1, src7, src7, shuff2, tmp8, tmp9, tmp10, tmp11);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp0, const0, tmp1, const1, tmp2, const2, tmp3,
+              const0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp4, const1, tmp5, const2, tmp6, const0, tmp7,
+              const1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp8, const2, tmp9, const0, tmp10, const1, tmp11,
+              const2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vsrar_h, src0, shift0, src1, shift1, src2, shift2, src3,
+              shift0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vsrar_h, src4, shift1, src5, shift2, src6, shift0, src7,
+              shift1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vsrar_h, tmp0, shift2, tmp1, shift0, tmp2, shift1, tmp3,
+              shift2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vslli_h, src0, 1, src1, 1, src2, 1, src3, 1, tmp5, tmp6,
+              tmp7, tmp8);
+    DUP2_ARG2(__lsx_vslli_h, src4, 1, src5, 1, tmp9, tmp10);
+    DUP4_ARG2(__lsx_vadd_h, src0, tmp5, src1, tmp6, src2, tmp7, src3, tmp8,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vadd_h, src4, tmp9, src5, tmp10, src4, src5);
+    DUP4_ARG2(__lsx_vadd_h, src0, src6, src1, src7, src2, tmp0, src3, tmp1,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vadd_h, src4, tmp2, src5, tmp3, src4, src5);
+    DUP2_ARG3(__lsx_vsrarni_b_h, src1, src0, 2, src3, src2, 2, dst0, dst1);
+    dst2 = __lsx_vsrarni_b_h(src5, src4, 2);
+    __lsx_vst(dst0, d, 0);
+    __lsx_vst(dst1, d, 16);
+    __lsx_vst(dst2, d, 32);
+    src_ptr += 64;
+    src_nex += 64;
+    d += 48;
+  }
+}
+
+void ScaleRowDown34_1_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width) {
+  const uint8_t* src_nex = src_ptr + src_stride;
+  int x;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
+  __m128i tmp10, tmp11, dst0, dst1, dst2;
+  __m128i const0 = {0x0103030101010103, 0x0101010303010101};
+  __m128i const1 = {0x0301010101030301, 0x0103030101010103};
+  __m128i const2 = {0x0101010303010101, 0x0301010101030301};
+  __m128i shuff0 = {0x0504030202010100, 0x0A09090807060605};
+  __m128i shuff1 = {0x0F0E0E0D0D0C0B0A, 0x1514131212111110};
+  __m128i shuff2 = {0x0A09090807060605, 0x0F0E0E0D0D0C0B0A};
+  __m128i shift0 = {0x0002000200010002, 0x0001000200020001};
+  __m128i shift1 = {0x0002000100020002, 0x0002000200010002};
+  __m128i shift2 = {0x0001000200020001, 0x0002000100020002};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, src_nex, 0, src_nex, 16, src_nex, 32, src_nex, 48,
+              src4, src5, src6, src7);
+    DUP4_ARG3(__lsx_vshuf_b, src0, src0, shuff0, src1, src0, shuff1, src1, src1,
+              shuff2, src2, src2, shuff0, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG3(__lsx_vshuf_b, src3, src2, shuff1, src3, src3, shuff2, src4, src4,
+              shuff0, src5, src4, shuff1, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vshuf_b, src5, src5, shuff2, src6, src6, shuff0, src7, src6,
+              shuff1, src7, src7, shuff2, tmp8, tmp9, tmp10, tmp11);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp0, const0, tmp1, const1, tmp2, const2, tmp3,
+              const0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp4, const1, tmp5, const2, tmp6, const0, tmp7,
+              const1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp8, const2, tmp9, const0, tmp10, const1, tmp11,
+              const2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vsrar_h, src0, shift0, src1, shift1, src2, shift2, src3,
+              shift0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vsrar_h, src4, shift1, src5, shift2, src6, shift0, src7,
+              shift1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vsrar_h, tmp0, shift2, tmp1, shift0, tmp2, shift1, tmp3,
+              shift2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vadd_h, src0, src6, src1, src7, src2, tmp0, src3, tmp1,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vadd_h, src4, tmp2, src5, tmp3, src4, src5);
+    DUP2_ARG3(__lsx_vsrarni_b_h, src1, src0, 1, src3, src2, 1, dst0, dst1);
+    dst2 = __lsx_vsrarni_b_h(src5, src4, 1);
+    __lsx_vst(dst0, d, 0);
+    __lsx_vst(dst1, d, 16);
+    __lsx_vst(dst2, d, 32);
+    src_ptr += 64;
+    src_nex += 64;
+    d += 48;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
diff --git a/media/libyuv/source/scale_mips.cc b/media/libyuv/source/scale_mips.cc
deleted file mode 100644
index ae953073fa..0000000000
--- a/media/libyuv/source/scale_mips.cc
+++ /dev/null
@@ -1,644 +0,0 @@
-/*
- *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/basic_types.h"
-#include "libyuv/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// This module is for GCC MIPS DSPR2
-#if !defined(LIBYUV_DISABLE_MIPS) && \
-    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32)
-
-void ScaleRowDown2_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst, int dst_width) {
-  __asm__ __volatile__(
-    ".set push                                     \n"
-    ".set noreorder                                \n"
-
-    "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
-    "beqz           $t9, 2f                        \n"
-    " nop                                          \n"
-
-  "1:                                              \n"
-    "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
-    "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
-    "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
-    "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
-    "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
-    "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
-    "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
-    "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
-    // TODO(fbarchard): Use odd pixels instead of even.
-    "precr.qb.ph    $t8, $t1, $t0                  \n"  // |6|4|2|0|
-    "precr.qb.ph    $t0, $t3, $t2                  \n"  // |14|12|10|8|
-    "precr.qb.ph    $t1, $t5, $t4                  \n"  // |22|20|18|16|
-    "precr.qb.ph    $t2, $t7, $t6                  \n"  // |30|28|26|24|
-    "addiu          %[src_ptr], %[src_ptr], 32     \n"
-    "addiu          $t9, $t9, -1                   \n"
-    "sw             $t8, 0(%[dst])                 \n"
-    "sw             $t0, 4(%[dst])                 \n"
-    "sw             $t1, 8(%[dst])                 \n"
-    "sw             $t2, 12(%[dst])                \n"
-    "bgtz           $t9, 1b                        \n"
-    " addiu         %[dst], %[dst], 16             \n"
-
-  "2:                                              \n"
-    "andi           $t9, %[dst_width], 0xf         \n"  // residue
-    "beqz           $t9, 3f                        \n"
-    " nop                                          \n"
-
-  "21:                                             \n"
-    "lbu            $t0, 0(%[src_ptr])             \n"
-    "addiu          %[src_ptr], %[src_ptr], 2      \n"
-    "addiu          $t9, $t9, -1                   \n"
-    "sb             $t0, 0(%[dst])                 \n"
-    "bgtz           $t9, 21b                       \n"
-    " addiu         %[dst], %[dst], 1              \n"
-
-  "3:                                              \n"
-    ".set pop                                      \n"
-  : [src_ptr] "+r" (src_ptr),
-    [dst] "+r" (dst)
-  : [dst_width] "r" (dst_width)
-  : "t0", "t1", "t2", "t3", "t4", "t5",
-    "t6", "t7", "t8", "t9"
-  );
-}
-
-void ScaleRowDown2Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width) {
-  const uint8* t = src_ptr + src_stride;
-
-  __asm__ __volatile__ (
-    ".set push                                    \n"
-    ".set noreorder                               \n"
-
-    "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
-    "bltz           $t9, 2f                       \n"
-    " nop                                         \n"
-
-  "1:                                             \n"
-    "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
-    "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
-    "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
-    "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
-    "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
-    "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
-    "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
-    "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
-    "addiu          $t9, $t9, -1                  \n"
-    "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
-    "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
-    "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
-    "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
-    "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
-    "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
-    "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
-    "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
-    "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
-    "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
-    "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
-    "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
-    "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
-    "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
-    "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
-    "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
-    "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
-    "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
-    "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
-    "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
-    "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
-    "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
-    "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
-    "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
-    "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
-    "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
-    "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
-    "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
-    "addiu          %[src_ptr], %[src_ptr], 16    \n"
-    "addiu          %[t], %[t], 16                \n"
-    "sb             $t0, 0(%[dst])                \n"
-    "sb             $t4, 1(%[dst])                \n"
-    "sb             $t1, 2(%[dst])                \n"
-    "sb             $t5, 3(%[dst])                \n"
-    "sb             $t2, 4(%[dst])                \n"
-    "sb             $t6, 5(%[dst])                \n"
-    "sb             $t3, 6(%[dst])                \n"
-    "sb             $t7, 7(%[dst])                \n"
-    "bgtz           $t9, 1b                       \n"
-    " addiu         %[dst], %[dst], 8             \n"
-
-  "2:                                             \n"
-    "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
-    "beqz           $t9, 3f                       \n"
-    " nop                                         \n"
-
-    "21:                                          \n"
-    "lwr            $t1, 0(%[src_ptr])            \n"
-    "lwl            $t1, 3(%[src_ptr])            \n"
-    "lwr            $t2, 0(%[t])                  \n"
-    "lwl            $t2, 3(%[t])                  \n"
-    "srl            $t8, $t1, 16                  \n"
-    "ins            $t1, $t2, 16, 16              \n"
-    "ins            $t2, $t8, 0, 16               \n"
-    "raddu.w.qb     $t1, $t1                      \n"
-    "raddu.w.qb     $t2, $t2                      \n"
-    "shra_r.w       $t1, $t1, 2                   \n"
-    "shra_r.w       $t2, $t2, 2                   \n"
-    "sb             $t1, 0(%[dst])                \n"
-    "sb             $t2, 1(%[dst])                \n"
-    "addiu          %[src_ptr], %[src_ptr], 4     \n"
-    "addiu          $t9, $t9, -2                  \n"
-    "addiu          %[t], %[t], 4                 \n"
-    "bgtz           $t9, 21b                      \n"
-    " addiu         %[dst], %[dst], 2             \n"
-
-  "3:                                             \n"
-    ".set pop                                     \n"
-
-  : [src_ptr] "+r" (src_ptr),
-    [dst] "+r" (dst), [t] "+r" (t)
-  : [dst_width] "r" (dst_width)
-  : "t0", "t1", "t2", "t3", "t4", "t5",
-    "t6", "t7", "t8", "t9"
-  );
-}
-
-void ScaleRowDown4_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst, int dst_width) {
-  __asm__ __volatile__ (
-      ".set push                                    \n"
-      ".set noreorder                               \n"
-
-      "srl            $t9, %[dst_width], 3          \n"
-      "beqz           $t9, 2f                       \n"
-      " nop                                         \n"
-
-     "1:                                            \n"
-      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
-      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
-      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
-      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
-      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
-      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
-      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
-      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
-      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
-      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
-      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
-      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
-      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |12|8|4|0|
-      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |28|24|20|16|
-      "addiu          %[src_ptr], %[src_ptr], 32    \n"
-      "addiu          $t9, $t9, -1                  \n"
-      "sw             $t1, 0(%[dst])                \n"
-      "sw             $t5, 4(%[dst])                \n"
-      "bgtz           $t9, 1b                       \n"
-      " addiu         %[dst], %[dst], 8             \n"
-
-    "2:                                             \n"
-      "andi           $t9, %[dst_width], 7          \n"  // residue
-      "beqz           $t9, 3f                       \n"
-      " nop                                         \n"
-
-    "21:                                            \n"
-      "lbu            $t1, 0(%[src_ptr])            \n"
-      "addiu          %[src_ptr], %[src_ptr], 4     \n"
-      "addiu          $t9, $t9, -1                  \n"
-      "sb             $t1, 0(%[dst])                \n"
-      "bgtz           $t9, 21b                      \n"
-      " addiu         %[dst], %[dst], 1             \n"
-
-    "3:                                             \n"
-      ".set pop                                     \n"
-      : [src_ptr] "+r" (src_ptr),
-        [dst] "+r" (dst)
-      : [dst_width] "r" (dst_width)
-      : "t1", "t2", "t3", "t4", "t5",
-        "t6", "t7", "t8", "t9"
-  );
-}
-
-void ScaleRowDown4Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width) {
-  intptr_t stride = src_stride;
-  const uint8* s1 = src_ptr + stride;
-  const uint8* s2 = s1 + stride;
-  const uint8* s3 = s2 + stride;
-
-  __asm__ __volatile__ (
-      ".set push                                  \n"
-      ".set noreorder                             \n"
-
-      "srl           $t9, %[dst_width], 1         \n"
-      "andi          $t8, %[dst_width], 1         \n"
-
-     "1:                                          \n"
-      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
-      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
-      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
-      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
-      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
-      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
-      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
-      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
-      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
-      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
-      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
-      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
-      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
-      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
-      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
-      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
-      "add           $t0, $t0, $t1                \n"
-      "add           $t1, $t2, $t3                \n"
-      "add           $t0, $t0, $t1                \n"
-      "add           $t4, $t4, $t5                \n"
-      "add           $t6, $t6, $t7                \n"
-      "add           $t4, $t4, $t6                \n"
-      "shra_r.w      $t0, $t0, 4                  \n"
-      "shra_r.w      $t4, $t4, 4                  \n"
-      "sb            $t0, 0(%[dst])               \n"
-      "sb            $t4, 1(%[dst])               \n"
-      "addiu         %[src_ptr], %[src_ptr], 8    \n"
-      "addiu         %[s1], %[s1], 8              \n"
-      "addiu         %[s2], %[s2], 8              \n"
-      "addiu         %[s3], %[s3], 8              \n"
-      "addiu         $t9, $t9, -1                 \n"
-      "bgtz          $t9, 1b                      \n"
-      " addiu        %[dst], %[dst], 2            \n"
-      "beqz          $t8, 2f                      \n"
-      " nop                                       \n"
-
-      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
-      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
-      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
-      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
-      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
-      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
-      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
-      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
-      "add           $t0, $t0, $t1                \n"
-      "add           $t1, $t2, $t3                \n"
-      "add           $t0, $t0, $t1                \n"
-      "shra_r.w      $t0, $t0, 4                  \n"
-      "sb            $t0, 0(%[dst])               \n"
-
-      "2:                                         \n"
-      ".set pop                                   \n"
-
-      : [src_ptr] "+r" (src_ptr),
-        [dst] "+r" (dst),
-        [s1] "+r" (s1),
-        [s2] "+r" (s2),
-        [s3] "+r" (s3)
-      : [dst_width] "r" (dst_width)
-      : "t0", "t1", "t2", "t3", "t4", "t5",
-        "t6","t7", "t8", "t9"
-  );
-}
-
-void ScaleRowDown34_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst, int dst_width) {
-  __asm__ __volatile__ (
-      ".set push                                          \n"
-      ".set noreorder                                     \n"
-    "1:                                                   \n"
-      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
-      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
-      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
-      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
-      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
-      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
-      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
-      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
-      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
-      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
-      "addiu           %[dst_width], %[dst_width], -24    \n"
-      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
-      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
-      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
-      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
-      "addiu           %[src_ptr], %[src_ptr], 32         \n"
-      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
-      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
-      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
-      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
-      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
-      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
-      "sw              $t1, 0(%[dst])                     \n"
-      "sw              $t0, 4(%[dst])                     \n"
-      "sw              $t3, 8(%[dst])                     \n"
-      "sw              $t5, 12(%[dst])                    \n"
-      "sw              $t9, 16(%[dst])                    \n"
-      "sw              $t7, 20(%[dst])                    \n"
-      "bnez            %[dst_width], 1b                   \n"
-      " addiu          %[dst], %[dst], 24                 \n"
-      ".set pop                                           \n"
-      : [src_ptr] "+r" (src_ptr),
-        [dst] "+r" (dst),
-        [dst_width] "+r" (dst_width)
-      :
-      : "t0", "t1", "t2", "t3", "t4", "t5",
-        "t6","t7", "t8", "t9"
-  );
-}
-
-void ScaleRowDown34_0_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* d, int dst_width) {
-  __asm__ __volatile__ (
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "repl.ph           $t3, 3                          \n"  // 0x00030003
-
-    "1:                                                  \n"
-      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
-      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
-      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
-      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
-      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
-      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
-      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
-      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
-      "raddu.w.qb        $t0, $t0                        \n"
-      "raddu.w.qb        $t1, $t1                        \n"
-      "shra_r.w          $t0, $t0, 1                     \n"
-      "shra_r.w          $t1, $t1, 1                     \n"
-      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
-      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
-      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
-      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
-      "addu.ph           $t2, $t2, $t4                   \n"
-      "addu.ph           $t6, $t6, $t5                   \n"
-      "sll               $t5, $t0, 1                     \n"
-      "add               $t0, $t5, $t0                   \n"
-      "shra_r.ph         $t2, $t2, 2                     \n"
-      "shra_r.ph         $t6, $t6, 2                     \n"
-      "shll.ph           $t4, $t2, 1                     \n"
-      "addq.ph           $t4, $t4, $t2                   \n"
-      "addu              $t0, $t0, $t1                   \n"
-      "addiu             %[src_ptr], %[src_ptr], 4       \n"
-      "shra_r.w          $t0, $t0, 2                     \n"
-      "addu.ph           $t6, $t6, $t4                   \n"
-      "shra_r.ph         $t6, $t6, 2                     \n"
-      "srl               $t1, $t6, 16                    \n"
-      "addiu             %[dst_width], %[dst_width], -3  \n"
-      "sb                $t1, 0(%[d])                    \n"
-      "sb                $t0, 1(%[d])                    \n"
-      "sb                $t6, 2(%[d])                    \n"
-      "bgtz              %[dst_width], 1b                \n"
-      " addiu            %[d], %[d], 3                   \n"
-    "3:                                                  \n"
-      ".set pop                                          \n"
-      : [src_ptr] "+r" (src_ptr),
-        [src_stride] "+r" (src_stride),
-        [d] "+r" (d),
-        [dst_width] "+r" (dst_width)
-      :
-      : "t0", "t1", "t2", "t3",
-        "t4", "t5", "t6"
-  );
-}
-
-void ScaleRowDown34_1_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* d, int dst_width) {
-  __asm__ __volatile__ (
-      ".set push                                           \n"
-      ".set noreorder                                      \n"
-      "repl.ph           $t2, 3                            \n"  // 0x00030003
-
-    "1:                                                    \n"
-      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
-      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
-      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
-      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
-      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
-      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
-      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
-      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
-      "raddu.w.qb        $t0, $t0                          \n"
-      "raddu.w.qb        $t1, $t1                          \n"
-      "shra_r.w          $t0, $t0, 1                       \n"
-      "shra_r.w          $t1, $t1, 1                       \n"
-      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
-      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
-      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
-      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
-      "addu.ph           $t4, $t4, $t3                     \n"
-      "addu.ph           $t6, $t6, $t5                     \n"
-      "shra_r.ph         $t6, $t6, 2                       \n"
-      "shra_r.ph         $t4, $t4, 2                       \n"
-      "addu.ph           $t6, $t6, $t4                     \n"
-      "addiu             %[src_ptr], %[src_ptr], 4         \n"
-      "shra_r.ph         $t6, $t6, 1                       \n"
-      "addu              $t0, $t0, $t1                     \n"
-      "addiu             %[dst_width], %[dst_width], -3    \n"
-      "shra_r.w          $t0, $t0, 1                       \n"
-      "srl               $t1, $t6, 16                      \n"
-      "sb                $t1, 0(%[d])                      \n"
-      "sb                $t0, 1(%[d])                      \n"
-      "sb                $t6, 2(%[d])                      \n"
-      "bgtz              %[dst_width], 1b                  \n"
-      " addiu            %[d], %[d], 3                     \n"
-    "3:                                                    \n"
-      ".set pop                                            \n"
-      : [src_ptr] "+r" (src_ptr),
-        [src_stride] "+r" (src_stride),
-        [d] "+r" (d),
-        [dst_width] "+r" (dst_width)
-      :
-      : "t0", "t1", "t2", "t3",
-        "t4", "t5", "t6"
-  );
-}
-
-void ScaleRowDown38_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst, int dst_width) {
-  __asm__ __volatile__ (
-      ".set push                                     \n"
-      ".set noreorder                                \n"
-
-    "1:                                              \n"
-      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
-      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
-      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
-      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
-      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
-      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
-      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
-      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
-      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
-      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
-      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
-      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
-      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
-      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
-      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
-      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
-      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
-      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
-      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
-      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
-      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
-      "addiu      %[src_ptr], %[src_ptr], 32         \n"
-      "addiu      %[dst_width], %[dst_width], -12    \n"
-      "addiu      $t8,%[dst_width], -12              \n"
-      "sw         $t1, 0(%[dst])                     \n"
-      "sw         $t4, 4(%[dst])                     \n"
-      "sw         $t6, 8(%[dst])                     \n"
-      "bgez       $t8, 1b                            \n"
-      " addiu     %[dst], %[dst], 12                 \n"
-      ".set pop                                      \n"
-      : [src_ptr] "+r" (src_ptr),
-        [dst] "+r" (dst),
-        [dst_width] "+r" (dst_width)
-      :
-      : "t0", "t1", "t2", "t3", "t4",
-        "t5", "t6", "t7", "t8"
-  );
-}
-
-void ScaleRowDown38_2_Box_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  intptr_t stride = src_stride;
-  const uint8* t = src_ptr + stride;
-  const int c = 0x2AAA;
-
-  __asm__ __volatile__ (
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-
-    "1:                                                  \n"
-      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
-      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
-      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
-      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
-      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
-      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
-      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
-      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
-      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
-      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
-      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
-      "srl             $t4, $t4, 2                       \n"  // t4 / 4
-      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
-      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
-      "addu            $t6, $t5, $t6                     \n"
-      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
-      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
-      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
-      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
-      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
-      "addu            $t0, $t0, $t2                     \n"
-      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
-      "addiu           %[src_ptr], %[src_ptr], 8         \n"
-      "addiu           %[t], %[t], 8                     \n"
-      "addiu           %[dst_width], %[dst_width], -3    \n"
-      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
-      "srl             $t6, $t6, 16                      \n"
-      "srl             $t0, $t0, 16                      \n"
-      "sb              $t4, -1(%[dst_ptr])               \n"
-      "sb              $t6, -2(%[dst_ptr])               \n"
-      "bgtz            %[dst_width], 1b                  \n"
-      " sb             $t0, -3(%[dst_ptr])               \n"
-      ".set pop                                          \n"
-      : [src_ptr] "+r" (src_ptr),
-        [dst_ptr] "+r" (dst_ptr),
-        [t] "+r" (t),
-        [dst_width] "+r" (dst_width)
-      : [c] "r" (c)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6"
-  );
-}
-
-void ScaleRowDown38_3_Box_DSPR2(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  intptr_t stride = src_stride;
-  const uint8* s1 = src_ptr + stride;
-  stride += stride;
-  const uint8* s2 = src_ptr + stride;
-  const int c1 = 0x1C71;
-  const int c2 = 0x2AAA;
-
-  __asm__ __volatile__ (
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-
-    "1:                                                  \n"
-      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
-      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
-      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
-      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
-      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
-      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
-      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
-      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
-      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
-      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
-      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
-      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
-      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
-      "addu            $t7, $t7, $t8                     \n"
-      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
-      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
-      "addu            $t6, $t6, $t8                     \n"
-      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
-      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
-      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
-      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
-      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
-      "addu            $t7, $t7, $t8                     \n"
-      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
-      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
-      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
-      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
-      "raddu.w.qb      $t0, $t0                          \n"
-      "raddu.w.qb      $t2, $t2                          \n"
-      "raddu.w.qb      $t4, $t4                          \n"
-      "addu            $t0, $t0, $t2                     \n"
-      "addu            $t0, $t0, $t4                     \n"
-      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
-      "addiu           %[src_ptr], %[src_ptr], 8         \n"
-      "addiu           %[s1], %[s1], 8                   \n"
-      "addiu           %[s2], %[s2], 8                   \n"
-      "addiu           %[dst_width], %[dst_width], -3    \n"
-      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
-      "srl             $t6, $t6, 16                      \n"
-      "srl             $t7, $t7, 16                      \n"
-      "srl             $t0, $t0, 16                      \n"
-      "sb              $t6, -1(%[dst_ptr])               \n"
-      "sb              $t7, -2(%[dst_ptr])               \n"
-      "bgtz            %[dst_width], 1b                  \n"
-      " sb             $t0, -3(%[dst_ptr])               \n"
-      ".set pop                                          \n"
-      : [src_ptr] "+r" (src_ptr),
-        [dst_ptr] "+r" (dst_ptr),
-        [s1] "+r" (s1),
-        [s2] "+r" (s2),
-        [dst_width] "+r" (dst_width)
-      : [c1] "r" (c1), [c2] "r" (c2)
-      : "t0", "t1", "t2", "t3", "t4",
-        "t5", "t6", "t7", "t8"
-  );
-}
-
-#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
-
diff --git a/media/libyuv/source/scale_mmi.cc b/media/libyuv/source/scale_mmi.cc
new file mode 100644
index 0000000000..1226ef3eaf
--- /dev/null
+++ b/media/libyuv/source/scale_mmi.cc
@@ -0,0 +1,1168 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Mips MMI.
+#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+// clang-format off
+
+// CPU agnostic row functions
+void ScaleRowDown2_MMI(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1, dest;
+  const uint64_t shift = 0x8ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+      "psrlh      %[src0],         %[src0],           %[shift]      \n\t"
+
+      "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+      "psrlh      %[src1],         %[src1],           %[shift]      \n\t"
+
+      "packushb   %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [shift] "f"(shift)
+      : "memory");
+}
+
+void ScaleRowDown2Linear_MMI(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1;
+  uint64_t dest, dest0, dest1;
+
+  const uint64_t mask = 0x00ff00ff00ff00ffULL;
+  const uint64_t shift = 0x8ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],          0x00(%[src_ptr])                \n\t"
+      "gsldlc1    %[src0],          0x07(%[src_ptr])                \n\t"
+      "and        %[dest0],         %[src0],          %[mask]       \n\t"
+      "gsldrc1    %[src1],          0x08(%[src_ptr])                \n\t"
+      "gsldlc1    %[src1],          0x0f(%[src_ptr])                \n\t"
+      "and        %[dest1],         %[src1],          %[mask]       \n\t"
+      "packushb   %[dest0],         %[dest0],         %[dest1]      \n\t"
+
+      "psrlh      %[src0],          %[src0],          %[shift]      \n\t"
+      "psrlh      %[src1],          %[src1],          %[shift]      \n\t"
+      "packushb   %[dest1],         %[src0],          %[src1]       \n\t"
+
+      "pavgb      %[dest],          %[dest0],         %[dest1]      \n\t"
+      "gssdlc1    %[dest],          0x07(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest],          0x00(%[dst_ptr])                \n\t"
+
+      "daddiu     %[src_ptr],       %[src_ptr],        0x10         \n\t"
+      "daddiu     %[dst_ptr],       %[dst_ptr],        0x08         \n\t"
+      "daddi      %[width],         %[width],         -0x08         \n\t"
+      "bnez       %[width],         1b                              \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest0] "=&f"(dest0),
+        [dest1] "=&f"(dest1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst), [mask] "f"(mask),
+        [shift] "f"(shift), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleRowDown2Box_MMI(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+
+  uint64_t s0, s1, t0, t1;
+  uint64_t dest, dest0, dest1;
+
+  const uint64_t ph = 0x0002000200020002ULL;
+  const uint64_t mask = 0x00ff00ff00ff00ffULL;
+  const uint64_t shift0 = 0x2ULL;
+  const uint64_t shift1 = 0x8ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[s0],            0x00(%[s])                      \n\t"
+      "gsldlc1    %[s0],            0x07(%[s])                      \n\t"
+      "psrlh      %[s1],            %[s0],            %[shift1]     \n\t"
+      "and        %[s0],            %[s0],            %[mask]       \n\t"
+
+      "gsldrc1    %[t0],            0x00(%[t])                      \n\t"
+      "gsldlc1    %[t0],            0x07(%[t])                      \n\t"
+      "psrlh      %[t1],            %[t0],            %[shift1]     \n\t"
+      "and        %[t0],            %[t0],            %[mask]       \n\t"
+
+      "paddh      %[dest0],         %[s0],            %[s1]         \n\t"
+      "paddh      %[dest0],         %[dest0],         %[t0]         \n\t"
+      "paddh      %[dest0],         %[dest0],         %[t1]         \n\t"
+      "paddh      %[dest0],         %[dest0],         %[ph]         \n\t"
+      "psrlh      %[dest0],         %[dest0],         %[shift0]     \n\t"
+
+      "gsldrc1    %[s0],            0x08(%[s])                      \n\t"
+      "gsldlc1    %[s0],            0x0f(%[s])                      \n\t"
+      "psrlh      %[s1],            %[s0],            %[shift1]     \n\t"
+      "and        %[s0],            %[s0],            %[mask]       \n\t"
+
+      "gsldrc1    %[t0],            0x08(%[t])                      \n\t"
+      "gsldlc1    %[t0],            0x0f(%[t])                      \n\t"
+      "psrlh      %[t1],            %[t0],            %[shift1]     \n\t"
+      "and        %[t0],            %[t0],            %[mask]       \n\t"
+
+      "paddh      %[dest1],         %[s0],            %[s1]         \n\t"
+      "paddh      %[dest1],         %[dest1],         %[t0]         \n\t"
+      "paddh      %[dest1],         %[dest1],         %[t1]         \n\t"
+      "paddh      %[dest1],         %[dest1],         %[ph]         \n\t"
+      "psrlh      %[dest1],         %[dest1],         %[shift0]     \n\t"
+
+      "packushb   %[dest],          %[dest0],         %[dest1]      \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[s],            %[s],              0x10          \n\t"
+      "daddiu     %[t],            %[t],              0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [s0] "=&f"(s0), [s1] "=&f"(s1), [t0] "=&f"(t0), [t1] "=&f"(t1),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest] "=&f"(dest)
+      : [s] "r"(s), [t] "r"(t), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [shift0] "f"(shift0), [shift1] "f"(shift1), [ph] "f"(ph),
+        [mask] "f"(mask)
+      : "memory");
+}
+
+void ScaleARGBRowDown2_MMI(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width) {
+  (void)src_stride;
+
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
+
+  uint64_t src0, src1, dest;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+      "punpckhwd  %[dest],         %[src0],           %[src1]       \n\t"
+
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src), [dst_ptr] "r"(dst), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleARGBRowDown2Linear_MMI(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1;
+  uint64_t dest, dest_hi, dest_lo;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "lwc1       %[src0],         0x00(%[src_ptr])                 \n\t"
+      "lwc1       %[src1],         0x08(%[src_ptr])                 \n\t"
+      "punpcklwd  %[dest_lo],      %[src0],           %[src1]       \n\t"
+      "lwc1       %[src0],         0x04(%[src_ptr])                 \n\t"
+      "lwc1       %[src1],         0x0c(%[src_ptr])                 \n\t"
+      "punpcklwd  %[dest_hi],      %[src0],           %[src1]       \n\t"
+
+      "pavgb      %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleARGBRowDown2Box_MMI(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  const uint8_t* s = src_argb;
+  const uint8_t* t = src_argb + src_stride;
+
+  uint64_t s0, s_hi, s_lo;
+  uint64_t t0, t_hi, t_lo;
+  uint64_t dest, dest_hi, dest_lo;
+
+  const uint64_t mask = 0x0ULL;
+  const uint64_t ph = 0x0002000200020002ULL;
+  const uint64_t shfit = 0x2ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[s0],            0x00(%[s])                      \n\t"
+      "gsldlc1    %[s0],            0x07(%[s])                      \n\t"
+      "punpcklbh  %[s_lo],          %[s0],           %[mask]        \n\t"
+      "punpckhbh  %[s_hi],          %[s0],           %[mask]        \n\t"
+      "paddh      %[dest_lo],       %[s_lo],         %[s_hi]        \n\t"
+
+      "gsldrc1    %[t0],            0x00(%[t])                      \n\t"
+      "gsldlc1    %[t0],            0x07(%[t])                      \n\t"
+      "punpcklbh  %[t_lo],          %[t0],           %[mask]        \n\t"
+      "punpckhbh  %[t_hi],          %[t0],           %[mask]        \n\t"
+      "paddh      %[dest_lo],       %[dest_lo],      %[t_lo]        \n\t"
+      "paddh      %[dest_lo],       %[dest_lo],      %[t_hi]        \n\t"
+
+      "paddh      %[dest_lo],      %[dest_lo],       %[ph]          \n\t"
+      "psrlh      %[dest_lo],      %[dest_lo],       %[shfit]       \n\t"
+
+      "gsldrc1    %[s0],            0x08(%[s])                      \n\t"
+      "gsldlc1    %[s0],            0x0f(%[s])                      \n\t"
+      "punpcklbh  %[s_lo],          %[s0],           %[mask]        \n\t"
+      "punpckhbh  %[s_hi],          %[s0],           %[mask]        \n\t"
+      "paddh      %[dest_hi],       %[s_lo],         %[s_hi]        \n\t"
+
+      "gsldrc1    %[t0],            0x08(%[t])                      \n\t"
+      "gsldlc1    %[t0],            0x0f(%[t])                      \n\t"
+      "punpcklbh  %[t_lo],          %[t0],           %[mask]        \n\t"
+      "punpckhbh  %[t_hi],          %[t0],           %[mask]        \n\t"
+      "paddh      %[dest_hi],       %[dest_hi],      %[t_lo]        \n\t"
+      "paddh      %[dest_hi],       %[dest_hi],      %[t_hi]        \n\t"
+
+      "paddh      %[dest_hi],      %[dest_hi],       %[ph]          \n\t"
+      "psrlh      %[dest_hi],      %[dest_hi],       %[shfit]       \n\t"
+
+      "packushb   %[dest],         %[dest_lo],       %[dest_hi]     \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[s],            %[s],              0x10          \n\t"
+      "daddiu     %[t],            %[t],              0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [s0] "=&f"(s0), [t0] "=&f"(t0), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [s_hi] "=&f"(s_hi), [s_lo] "=&f"(s_lo),
+        [t_hi] "=&f"(t_hi), [t_lo] "=&f"(t_lo), [dest] "=&f"(dest)
+      : [s] "r"(s), [t] "r"(t), [dst_ptr] "r"(dst_argb), [width] "r"(dst_width),
+        [mask] "f"(mask), [ph] "f"(ph), [shfit] "f"(shfit)
+      : "memory");
+}
+
+void ScaleRowDown2_16_MMI(const uint16_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint16_t* dst,
+                          int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1, dest;
+  const uint64_t shift = 0x10ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+      "psrlw      %[src0],         %[src0],           %[shift]      \n\t"
+
+      "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+      "psrlw      %[src1],         %[src1],           %[shift]      \n\t"
+
+      "packsswh   %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [shift] "f"(shift)
+      : "memory");
+}
+
+void ScaleRowDown2Linear_16_MMI(const uint16_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint16_t* dst,
+                                int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1;
+  uint64_t dest, dest_hi, dest_lo;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+      "punpcklhw  %[dest_lo],      %[src0],           %[src1]       \n\t"
+      "punpckhhw  %[dest_hi],      %[src0],           %[src1]       \n\t"
+
+      "punpcklhw  %[src0],         %[dest_lo],        %[dest_hi]    \n\t"
+      "punpckhhw  %[src1],         %[dest_lo],        %[dest_hi]    \n\t"
+
+      "pavgh      %[dest],         %[src0],           %[src1]       \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x10          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleRowDown2Box_16_MMI(const uint16_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint16_t* dst,
+                             int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+
+  uint64_t s0, s1, s_hi, s_lo;
+  uint64_t t0, t1, t_hi, t_lo;
+  uint64_t dest, dest0, dest1;
+
+  const uint64_t ph = 0x0000000200000002ULL;
+  const uint64_t mask = 0x0000ffff0000ffffULL;
+  const uint64_t shift0 = 0x10ULL;
+  const uint64_t shift1 = 0x2ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[s0],            0x00(%[s])                      \n\t"
+      "gsldlc1    %[s0],            0x07(%[s])                      \n\t"
+      "psrlw      %[s1],            %[s0],            %[shift0]     \n\t"
+      "and        %[s0],            %[s0],            %[mask]       \n\t"
+
+      "gsldrc1    %[t0],            0x00(%[t])                      \n\t"
+      "gsldlc1    %[t0],            0x07(%[t])                      \n\t"
+      "psrlw      %[t1],            %[t0],            %[shift0]     \n\t"
+      "and        %[t0],            %[t0],            %[mask]       \n\t"
+
+      "paddw      %[dest0],         %[s0],            %[s1]         \n\t"
+      "paddw      %[dest0],         %[dest0],         %[t0]         \n\t"
+      "paddw      %[dest0],         %[dest0],         %[t1]         \n\t"
+      "paddw      %[dest0],         %[dest0],         %[ph]         \n\t"
+      "psrlw      %[dest0],         %[dest0],         %[shift1]     \n\t"
+
+      "gsldrc1    %[s0],            0x08(%[s])                      \n\t"
+      "gsldlc1    %[s0],            0x0f(%[s])                      \n\t"
+      "psrlw      %[s1],            %[s0],            %[shift0]     \n\t"
+      "and        %[s0],            %[s0],            %[mask]       \n\t"
+
+      "gsldrc1    %[t0],            0x08(%[t])                      \n\t"
+      "gsldlc1    %[t0],            0x0f(%[t])                      \n\t"
+      "psrlw      %[t1],            %[t0],            %[shift0]     \n\t"
+      "and        %[t0],            %[t0],            %[mask]       \n\t"
+
+      "paddw      %[dest1],         %[s0],            %[s1]         \n\t"
+      "paddw      %[dest1],         %[dest1],         %[t0]         \n\t"
+      "paddw      %[dest1],         %[dest1],         %[t1]         \n\t"
+      "paddw      %[dest1],         %[dest1],         %[ph]         \n\t"
+      "psrlw      %[dest1],         %[dest1],         %[shift1]     \n\t"
+
+      "packsswh   %[dest],          %[dest0],         %[dest1]      \n\t"
+      "gssdlc1    %[dest],          0x07(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest],          0x00(%[dst_ptr])                \n\t"
+
+      "daddiu     %[s],             %[s],              0x10         \n\t"
+      "daddiu     %[t],             %[t],              0x10         \n\t"
+      "daddiu     %[dst_ptr],       %[dst_ptr],        0x08         \n\t"
+      "daddi      %[width],         %[width],         -0x04         \n\t"
+      "bnez       %[width],         1b                              \n\t"
+      : [s0] "=&f"(s0), [s1] "=&f"(s1), [t0] "=&f"(t0), [t1] "=&f"(t1),
+        [s_hi] "=&f"(s_hi), [s_lo] "=&f"(s_lo), [t_hi] "=&f"(t_hi),
+        [t_lo] "=&f"(t_lo), [dest0] "=&f"(dest0), [dest1] "=&f"(dest1),
+        [dest] "=&f"(dest)
+      : [s] "r"(s), [t] "r"(t), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [shift0] "f"(shift0), [shift1] "f"(shift1), [ph] "f"(ph),
+        [mask] "f"(mask)
+      : "memory");
+}
+
+void ScaleRowDown4_MMI(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1;
+  uint64_t dest, dest_hi, dest_lo;
+
+  const uint64_t shift = 0x10ULL;
+  const uint64_t mask = 0x000000ff000000ffULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+      "psrlw      %[src0],         %[src0],           %[shift]      \n\t"
+      "and        %[src0],         %[src0],           %[mask]       \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+      "psrlw      %[src1],         %[src1],           %[shift]      \n\t"
+      "and        %[src1],         %[src1],           %[mask]       \n\t"
+      "packsswh   %[dest_lo],      %[src0],           %[src1]       \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_ptr])                 \n\t"
+      "psrlw      %[src0],         %[src0],           %[shift]      \n\t"
+      "and        %[src0],         %[src0],           %[mask]       \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_ptr])                 \n\t"
+      "psrlw      %[src1],         %[src1],           %[shift]      \n\t"
+      "and        %[src1],         %[src1],           %[mask]       \n\t"
+      "packsswh   %[dest_hi],      %[src0],           %[src1]       \n\t"
+
+      "packushb   %[dest],         %[dest_lo],         %[dest_hi]   \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x20          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [shift] "f"(shift), [mask] "f"(mask)
+      : "memory");
+}
+
+void ScaleRowDown4_16_MMI(const uint16_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint16_t* dst,
+                          int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1;
+  uint64_t dest, dest_hi, dest_lo;
+
+  const uint64_t mask = 0x0ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+      "punpckhhw  %[dest_lo],      %[src0],           %[src1]       \n\t"
+      "punpcklhw  %[dest_lo],      %[dest_lo],        %[mask]       \n\t"
+
+      "gsldrc1    %[src0],         0x10(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src1],         0x18(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src1],         0x1f(%[src_ptr])                 \n\t"
+      "punpckhhw  %[dest_hi],      %[src0],           %[src1]       \n\t"
+      "punpcklhw  %[dest_hi],      %[dest_hi],        %[mask]       \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x20          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest_hi] "=&f"(dest_hi),
+        [dest_lo] "=&f"(dest_lo), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [mask] "f"(mask)
+      : "memory");
+}
+
+#define DO_SCALEROWDOWN4BOX_PUNPCKADD()                              \
+  "punpcklbh  %[src_lo],       %[src],           %[mask0]      \n\t" \
+  "punpckhbh  %[src_hi],       %[src],           %[mask0]      \n\t" \
+  "paddh      %[dest_lo],      %[dest_lo],       %[src_lo]     \n\t" \
+  "paddh      %[dest_hi],      %[dest_hi],       %[src_hi]     \n\t"
+
+#define DO_SCALEROWDOWN4BOX_LOOP(reg)                                \
+  "ldc1       %[src],          0x00(%[src0_ptr])               \n\t" \
+  "punpcklbh  %[dest_lo],      %[src],           %[mask0]      \n\t" \
+  "punpckhbh  %[dest_hi],      %[src],           %[mask0]      \n\t" \
+                                                                     \
+  "ldc1       %[src],          0x00(%[src1_ptr])               \n\t" \
+  DO_SCALEROWDOWN4BOX_PUNPCKADD()                                    \
+                                                                     \
+  "ldc1       %[src],          0x00(%[src2_ptr])               \n\t" \
+  DO_SCALEROWDOWN4BOX_PUNPCKADD()                                    \
+                                                                     \
+  "ldc1       %[src],          0x00(%[src3_ptr])               \n\t" \
+  DO_SCALEROWDOWN4BOX_PUNPCKADD()                                    \
+                                                                     \
+  "pmaddhw    %[dest_lo],      %[dest_lo],       %[mask1]      \n\t" \
+  "pmaddhw    %[dest_hi],      %[dest_hi],       %[mask1]      \n\t" \
+  "packsswh   " #reg   ",      %[dest_lo],       %[dest_hi]    \n\t" \
+  "pmaddhw    " #reg   ",      " #reg   ",       %[mask1]      \n\t" \
+  "paddh      " #reg   ",      " #reg   ",       %[ph]         \n\t" \
+  "psrlh      " #reg   ",      " #reg   ",       %[shift]      \n\t" \
+                                                                     \
+  "daddiu     %[src0_ptr],     %[src0_ptr],      0x08          \n\t" \
+  "daddiu     %[src1_ptr],     %[src1_ptr],      0x08          \n\t" \
+  "daddiu     %[src2_ptr],     %[src2_ptr],      0x08          \n\t" \
+  "daddiu     %[src3_ptr],     %[src3_ptr],      0x08          \n\t"
+
+/* LibYUVScaleTest.ScaleDownBy4_Box */
+void ScaleRowDown4Box_MMI(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  const uint8_t* src0_ptr = src_ptr;
+  const uint8_t* src1_ptr = src_ptr + src_stride;
+  const uint8_t* src2_ptr = src_ptr + src_stride * 2;
+  const uint8_t* src3_ptr = src_ptr + src_stride * 3;
+
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest, dest_hi, dest_lo, dest0, dest1, dest2, dest3;
+
+  const uint64_t mask0 = 0x0ULL;
+  const uint64_t mask1 = 0x0001000100010001ULL;
+  const uint64_t ph = 0x0008000800080008ULL;
+  const uint64_t shift = 0x4ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+
+      DO_SCALEROWDOWN4BOX_LOOP(%[dest0])
+      DO_SCALEROWDOWN4BOX_LOOP(%[dest1])
+      DO_SCALEROWDOWN4BOX_LOOP(%[dest2])
+      DO_SCALEROWDOWN4BOX_LOOP(%[dest3])
+
+      "packsswh   %[dest_lo],      %[dest0],          %[dest1]      \n\t"
+      "packsswh   %[dest_hi],      %[dest2],          %[dest3]      \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi]    \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3), [src] "=&f"(src), [dest] "=&f"(dest)
+      : [src0_ptr] "r"(src0_ptr), [src1_ptr] "r"(src1_ptr),
+        [src2_ptr] "r"(src2_ptr), [src3_ptr] "r"(src3_ptr), [dst_ptr] "r"(dst),
+        [width] "r"(dst_width), [shift] "f"(shift), [mask0] "f"(mask0),
+        [ph] "f"(ph), [mask1] "f"(mask1)
+      : "memory");
+}
+
+#define DO_SCALEROWDOWN4BOX_16_PUNPCKADD()                            \
+  "punpcklbh  %[src_lo],       %[src],            %[mask0]      \n\t" \
+  "punpckhbh  %[src_hi],       %[src],            %[mask0]      \n\t" \
+  "paddh      %[dest_lo],      %[dest_lo],        %[src_lo]     \n\t" \
+  "paddh      %[dest_hi],      %[dest_hi],        %[src_hi]     \n\t"
+
+#define DO_SCALEROWDOWN4BOX_16_LOOP(reg)                              \
+  "ldc1       %[src],          0x00(%[src0_ptr])                \n\t" \
+  "punpcklbh  %[dest_lo],      %[src],            %[mask0]      \n\t" \
+  "punpckhbh  %[dest_hi],      %[src],            %[mask0]      \n\t" \
+                                                                      \
+  "ldc1       %[src],          0x00(%[src1_ptr])                \n\t" \
+  DO_SCALEROWDOWN4BOX_16_PUNPCKADD()                                  \
+                                                                      \
+  "ldc1       %[src],          0x00(%[src2_ptr])                \n\t" \
+  DO_SCALEROWDOWN4BOX_16_PUNPCKADD()                                  \
+                                                                      \
+  "ldc1       %[src],          0x00(%[src3_ptr])                \n\t" \
+  DO_SCALEROWDOWN4BOX_16_PUNPCKADD()                                  \
+                                                                      \
+  "paddw      %[dest],         %[dest_lo],        %[dest_hi]    \n\t" \
+  "punpckhwd  %[dest_hi],      %[dest],           %[dest]       \n\t" \
+  "paddw      %[dest],         %[dest_hi],        %[dest]       \n\t" \
+  "paddw      %[dest],         %[dest],           %[ph]         \n\t" \
+  "psraw      %[dest],         %[dest],           %[shift]      \n\t" \
+  "and        " #reg ",        %[dest],           %[mask1]      \n\t" \
+                                                                      \
+  "daddiu     %[src0_ptr],     %[src0_ptr],       0x08          \n\t" \
+  "daddiu     %[src1_ptr],     %[src1_ptr],       0x08          \n\t" \
+  "daddiu     %[src2_ptr],     %[src2_ptr],       0x08          \n\t" \
+  "daddiu     %[src3_ptr],     %[src3_ptr],       0x08          \n\t"
+
+/* LibYUVScaleTest.ScaleDownBy4_Box_16 */
+void ScaleRowDown4Box_16_MMI(const uint16_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint16_t* dst,
+                             int dst_width) {
+  const uint16_t* src0_ptr = src_ptr;
+  const uint16_t* src1_ptr = src_ptr + src_stride;
+  const uint16_t* src2_ptr = src_ptr + src_stride * 2;
+  const uint16_t* src3_ptr = src_ptr + src_stride * 3;
+
+  uint64_t src, src_hi, src_lo;
+  uint64_t dest, dest_hi, dest_lo, dest0, dest1, dest2, dest3;
+
+  const uint64_t mask0 = 0x0ULL;
+  const uint64_t mask1 = 0x00000000ffffffffULL;
+  const uint64_t ph = 0x0000000800000008ULL;
+  const uint64_t shift = 0x04ULL;
+
+  __asm__ volatile(
+      "1:                                                        \n\t"
+
+      DO_SCALEROWDOWN4BOX_16_LOOP(%[dest0])
+      DO_SCALEROWDOWN4BOX_16_LOOP(%[dest1])
+      DO_SCALEROWDOWN4BOX_16_LOOP(%[dest2])
+      DO_SCALEROWDOWN4BOX_16_LOOP(%[dest3])
+      "punpcklwd  %[dest_lo],      %[dest0],          %[dest1]   \n\t"
+      "punpcklwd  %[dest_hi],      %[dest2],          %[dest3]   \n\t"
+
+      "packushb   %[dest],         %[dest_lo],        %[dest_hi] \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])              \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])              \n\t"
+
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08       \n\t"
+      "daddi      %[width],        %[width],         -0x04       \n\t"
+      "bnez       %[width],        1b                            \n\t"
+      : [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
+        [dest3] "=&f"(dest3), [src] "=&f"(src), [dest] "=&f"(dest)
+      : [src0_ptr] "r"(src0_ptr), [src1_ptr] "r"(src1_ptr),
+        [src2_ptr] "r"(src2_ptr), [src3_ptr] "r"(src3_ptr), [dst_ptr] "r"(dst),
+        [width] "r"(dst_width), [shift] "f"(shift), [mask0] "f"(mask0),
+        [ph] "f"(ph), [mask1] "f"(mask1)
+      : "memory");
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleColsUp2_MMI(uint8_t* dst_ptr,
+                      const uint8_t* src_ptr,
+                      int dst_width,
+                      int x,
+                      int dx) {
+  uint64_t src, dest;
+
+  (void)x;
+  (void)dx;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "lwc1       %[src],          0x00(%[src_ptr])                 \n\t"
+
+      "punpcklbh  %[dest],         %[src],            %[src]        \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x04          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst_ptr), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleColsUp2_16_MMI(uint16_t* dst_ptr,
+                         const uint16_t* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  uint64_t src, dest;
+
+  (void)x;
+  (void)dx;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+
+      "punpcklhw  %[dest],         %[src],            %[src]        \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "punpckhhw  %[dest],         %[src],            %[src]        \n\t"
+      "gssdlc1    %[dest],         0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x08(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src] "=&f"(src), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst_ptr), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleAddRow_MMI(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
+  uint64_t src, src_hi, src_lo, dest0, dest1;
+  const uint64_t mask = 0x0ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "punpcklbh  %[src_lo],       %[src],            %[mask]       \n\t"
+      "punpckhbh  %[src_hi],       %[src],            %[mask]       \n\t"
+
+      "gsldrc1    %[dest0],        0x00(%[dst_ptr])                 \n\t"
+      "gsldlc1    %[dest0],        0x07(%[dst_ptr])                 \n\t"
+      "paddush    %[dest0],        %[dest0],          %[src_lo]     \n\t"
+      "gsldrc1    %[dest1],        0x08(%[dst_ptr])                 \n\t"
+      "gsldlc1    %[dest1],        0x0f(%[dst_ptr])                 \n\t"
+      "paddush    %[dest1],        %[dest1],          %[src_hi]     \n\t"
+
+      "gssdlc1    %[dest0],        0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_ptr])                 \n\t"
+      "gssdlc1    %[dest1],        0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest1],        0x08(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x08          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [src_hi] "=&f"(src_hi),
+        [src_lo] "=&f"(src_lo), [src] "=&f"(src)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst_ptr), [width] "r"(src_width),
+        [mask] "f"(mask)
+      : "memory");
+}
+
+void ScaleAddRow_16_MMI(const uint16_t* src_ptr,
+                        uint32_t* dst_ptr,
+                        int src_width) {
+  uint64_t src, src_hi, src_lo, dest0, dest1;
+  const uint64_t mask = 0x0ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src],          0x00(%[src_ptr])                 \n\t"
+      "gsldlc1    %[src],          0x07(%[src_ptr])                 \n\t"
+      "punpcklhw  %[src_lo],       %[src],            %[mask]       \n\t"
+      "punpckhhw  %[src_hi],       %[src],            %[mask]       \n\t"
+
+      "gsldrc1    %[dest0],        0x00(%[dst_ptr])                 \n\t"
+      "gsldlc1    %[dest0],        0x07(%[dst_ptr])                 \n\t"
+      "paddw      %[dest0],        %[dest0],          %[src_lo]     \n\t"
+      "gssdlc1    %[dest0],        0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest0],        0x00(%[dst_ptr])                 \n\t"
+
+      "gsldrc1    %[dest1],        0x08(%[dst_ptr])                 \n\t"
+      "gsldlc1    %[dest1],        0x0f(%[dst_ptr])                 \n\t"
+      "paddw      %[dest1],        %[dest1],          %[src_hi]     \n\t"
+      "gssdlc1    %[dest1],        0x0f(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest1],        0x08(%[dst_ptr])                 \n\t"
+
+      "daddiu     %[src_ptr],      %[src_ptr],        0x08          \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+      "daddi      %[width],        %[width],         -0x04          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [src_hi] "=&f"(src_hi),
+        [src_lo] "=&f"(src_lo), [src] "=&f"(src)
+      : [src_ptr] "r"(src_ptr), [dst_ptr] "r"(dst_ptr), [width] "r"(src_width),
+        [mask] "f"(mask)
+      : "memory");
+}
+
+void ScaleARGBRowDownEven_MMI(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  (void)src_stride;
+
+  uint64_t src0, src1, dest;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "lwc1       %[src0],          0x00(%[src_ptr])                \n\t"
+      "dadd       %[src_ptr],       %[src_ptr],       %[src_stepx_4]\n\t"
+      "lwc1       %[src1],          0x00(%[src_ptr])                \n\t"
+      "punpcklwd  %[dest],          %[src0],          %[src1]       \n\t"
+
+      "gssdlc1    %[dest],          0x07(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest],          0x00(%[dst_ptr])                \n\t"
+
+      "dadd       %[src_ptr],       %[src_ptr],       %[src_stepx_4]\n\t"
+      "daddiu     %[dst_ptr],       %[dst_ptr],       0x08          \n\t"
+      "daddi      %[width],         %[width],        -0x02          \n\t"
+      "bnez       %[width],         1b                              \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb),
+        [src_stepx_4] "r"(src_stepx << 2), [width] "r"(dst_width)
+      : "memory");
+}
+
+void ScaleARGBRowDownEvenBox_MMI(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  const uint8_t* src0_ptr = src_argb;
+  const uint8_t* src1_ptr = src_argb + src_stride;
+
+  uint64_t src0, src1, src_hi, src_lo;
+  uint64_t dest, dest_hi, dest_lo, dest0, dest1;
+
+  const uint64_t mask = 0x0ULL;
+  const uint64_t ph = 0x0002000200020002ULL;
+  const uint64_t shift = 0x2ULL;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+
+      "lwc1       %[src0],         0x00(%[src0_ptr])                \n\t"
+      "punpcklbh  %[dest_lo],      %[src0],          %[mask]        \n\t"
+      "lwc1       %[src0],         0x04(%[src0_ptr])                \n\t"
+      "punpcklbh  %[dest_hi],      %[src0],          %[mask]        \n\t"
+
+      "lwc1       %[src1],         0x00(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src_lo],       %[src1],          %[mask]        \n\t"
+      "lwc1       %[src1],         0x04(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src_hi],       %[src1],          %[mask]        \n\t"
+      "paddh      %[dest_lo],      %[dest_lo],       %[src_lo]      \n\t"
+      "paddh      %[dest_hi],      %[dest_hi],       %[src_hi]      \n\t"
+      "paddh      %[dest0],        %[dest_hi],       %[dest_lo]     \n\t"
+      "paddh      %[dest0],        %[dest0],         %[ph]          \n\t"
+      "psrlh      %[dest0],        %[dest0],         %[shift]       \n\t"
+
+      "dadd       %[src0_ptr],     %[src0_ptr],      %[src_stepx_4] \n\t"
+      "dadd       %[src1_ptr],     %[src1_ptr],      %[src_stepx_4] \n\t"
+
+      "lwc1       %[src0],         0x00(%[src0_ptr])                \n\t"
+      "punpcklbh  %[dest_lo],      %[src0],          %[mask]        \n\t"
+      "lwc1       %[src0],         0x04(%[src0_ptr])                \n\t"
+      "punpcklbh  %[dest_hi],      %[src0],          %[mask]        \n\t"
+
+      "lwc1       %[src1],         0x00(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src_lo],       %[src1],          %[mask]        \n\t"
+      "lwc1       %[src1],         0x04(%[src1_ptr])                \n\t"
+      "punpcklbh  %[src_hi],       %[src1],          %[mask]        \n\t"
+      "paddh      %[dest_lo],      %[dest_lo],       %[src_lo]      \n\t"
+      "paddh      %[dest_hi],      %[dest_hi],       %[src_hi]      \n\t"
+      "paddh      %[dest1],        %[dest_hi],       %[dest_lo]     \n\t"
+      "paddh      %[dest1],        %[dest1],         %[ph]          \n\t"
+      "psrlh      %[dest1],        %[dest1],         %[shift]       \n\t"
+
+      "packushb   %[dest],         %[dest0],          %[dest1]      \n\t"
+      "gssdlc1    %[dest],         0x07(%[dst_ptr])                 \n\t"
+      "gssdrc1    %[dest],         0x00(%[dst_ptr])                 \n\t"
+
+      "dadd       %[src0_ptr],     %[src0_ptr],      %[src_stepx_4] \n\t"
+      "dadd       %[src1_ptr],     %[src1_ptr],      %[src_stepx_4] \n\t"
+      "daddiu     %[dst_ptr],      %[dst_ptr],        0x08          \n\t"
+      "daddi      %[width],        %[width],         -0x02          \n\t"
+      "bnez       %[width],        1b                               \n\t"
+      : [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
+        [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo),
+        [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [src0] "=&f"(src0),
+        [src1] "=&f"(src1), [dest] "=&f"(dest)
+      : [src0_ptr] "r"(src0_ptr), [src1_ptr] "r"(src1_ptr),
+        [dst_ptr] "r"(dst_argb), [width] "r"(dst_width),
+        [src_stepx_4] "r"(src_stepx << 2), [shift] "f"(shift), [mask] "f"(mask),
+        [ph] "f"(ph)
+      : "memory");
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleARGBCols_MMI(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
+
+  const uint32_t* src_tmp;
+
+  uint64_t dest, offset;
+
+  const uint64_t shift0 = 16;
+  const uint64_t shift1 = 2;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "srav       %[offset],        %[x],             %[shift0]     \n\t"
+      "sllv       %[offset],        %[offset],        %[shift1]     \n\t"
+      "dadd       %[src_tmp],       %[src_ptr],       %[offset]     \n\t"
+      "lwc1       %[dest],          0x00(%[src_tmp])                \n\t"
+      "swc1       %[dest],          0x00(%[dst_ptr])                \n\t"
+
+      "dadd       %[x],             %[x],             %[dx]         \n\t"
+
+      "daddiu     %[dst_ptr],       %[dst_ptr],       0x04          \n\t"
+      "daddi      %[width],         %[width],        -0x01          \n\t"
+      "bnez       %[width],         1b                              \n\t"
+      : [dest] "=&f"(dest), [offset] "=&r"(offset), [src_tmp] "=&r"(src_tmp)
+      : [src_ptr] "r"(src), [dst_ptr] "r"(dst), [width] "r"(dst_width),
+        [dx] "r"(dx), [x] "r"(x), [shift0] "r"(shift0), [shift1] "r"(shift1)
+      : "memory");
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleARGBColsUp2_MMI(uint8_t* dst_argb,
+                          const uint8_t* src_argb,
+                          int dst_width,
+                          int x,
+                          int dx) {
+  uint64_t src, dest0, dest1;
+  (void)x;
+  (void)dx;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src],           0x00(%[src_ptr])                \n\t"
+      "gsldlc1    %[src],           0x07(%[src_ptr])                \n\t"
+      "punpcklwd  %[dest0],         %[src],           %[src]        \n\t"
+      "gssdlc1    %[dest0],         0x07(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest0],         0x00(%[dst_ptr])                \n\t"
+      "punpckhwd  %[dest1],         %[src],           %[src]        \n\t"
+      "gssdlc1    %[dest1],         0x0f(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest1],         0x08(%[dst_ptr])                \n\t"
+
+      "daddiu     %[src_ptr],       %[src_ptr],       0x08          \n\t"
+      "daddiu     %[dst_ptr],       %[dst_ptr],       0x10          \n\t"
+      "daddi      %[width],         %[width],        -0x04          \n\t"
+      "bnez       %[width],         1b                              \n\t"
+      : [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [src] "=&f"(src)
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_argb), [width] "r"(dst_width)
+      : "memory");
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+/* LibYUVBaseTest.TestFixedDiv */
+int FixedDiv_MIPS(int num, int div) {
+  int quotient = 0;
+  const int shift = 16;
+
+  asm(
+      "dsll    %[num],     %[num],     %[shift]    \n\t"
+      "ddiv    %[num],     %[div]                  \t\n"
+      "mflo    %[quo]                              \t\n"
+      : [quo] "+&r"(quotient)
+      : [num] "r"(num), [div] "r"(div), [shift] "r"(shift));
+
+  return quotient;
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+/* LibYUVScaleTest.ARGBScaleTo320x240_Linear */
+int FixedDiv1_MIPS(int num, int div) {
+  int quotient = 0;
+  const int shift = 16;
+  const int val1 = 1;
+  const int64_t val11 = 0x00010001ULL;
+
+  asm(
+      "dsll    %[num],     %[num],     %[shift]    \n\t"
+      "dsub    %[num],     %[num],     %[val11]    \n\t"
+      "dsub    %[div],     %[div],     %[val1]     \n\t"
+      "ddiv    %[num],     %[div]                  \t\n"
+      "mflo    %[quo]                              \t\n"
+      : [quo] "+&r"(quotient)
+      : [num] "r"(num), [div] "r"(div), [val1] "r"(val1), [val11] "r"(val11),
+        [shift] "r"(shift));
+
+  return quotient;
+}
+
+// Read 8x2 upsample with filtering and write 16x1.
+// actually reads an extra pixel, so 9x2.
+void ScaleRowUp2_16_MMI(const uint16_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint16_t* dst,
+                        int dst_width) {
+  const uint16_t* src2_ptr = src_ptr + src_stride;
+
+  uint64_t src0, src1;
+  uint64_t dest, dest04, dest15, dest26, dest37;
+  uint64_t tmp0, tmp1, tmp2, tmp3;
+
+  const uint64_t mask0 = 0x0003000900030009ULL;
+  const uint64_t mask1 = 0x0001000300010003ULL;
+  const uint64_t mask2 = 0x0009000300090003ULL;
+  const uint64_t mask3 = 0x0003000100030001ULL;
+  const uint64_t ph = 0x0000000800000008ULL;
+  const uint64_t shift = 4;
+
+  __asm__ volatile(
+      "1:                                                           \n\t"
+      "gsldrc1    %[src0],          0x00(%[src1_ptr])               \n\t"
+      "gsldlc1    %[src0],          0x07(%[src1_ptr])               \n\t"
+      "pmaddhw    %[dest04],        %[src0],          %[mask0]      \n\t"
+      "gsldrc1    %[src1],          0x00(%[src2_ptr])               \n\t"
+      "gsldlc1    %[src1],          0x07(%[src2_ptr])               \n\t"
+      "pmaddhw    %[dest],          %[src1],          %[mask1]      \n\t"
+      "paddw      %[dest04],        %[dest04],        %[dest]       \n\t"
+      "paddw      %[dest04],        %[dest04],        %[ph]         \n\t"
+      "psrlw      %[dest04],        %[dest04],        %[shift]      \n\t"
+
+      "pmaddhw    %[dest15],        %[src0],          %[mask2]      \n\t"
+      "pmaddhw    %[dest],          %[src1],          %[mask3]      \n\t"
+      "paddw      %[dest15],        %[dest15],        %[dest]       \n\t"
+      "paddw      %[dest15],        %[dest15],        %[ph]         \n\t"
+      "psrlw      %[dest15],        %[dest15],        %[shift]      \n\t"
+
+      "gsldrc1    %[src0],          0x02(%[src1_ptr])               \n\t"
+      "gsldlc1    %[src0],          0x09(%[src1_ptr])               \n\t"
+      "pmaddhw    %[dest26],        %[src0],          %[mask0]      \n\t"
+      "gsldrc1    %[src1],          0x02(%[src2_ptr])               \n\t"
+      "gsldlc1    %[src1],          0x09(%[src2_ptr])               \n\t"
+      "pmaddhw    %[dest],          %[src1],          %[mask1]      \n\t"
+      "paddw      %[dest26],        %[dest26],        %[dest]       \n\t"
+      "paddw      %[dest26],        %[dest26],        %[ph]         \n\t"
+      "psrlw      %[dest26],        %[dest26],        %[shift]      \n\t"
+
+      "pmaddhw    %[dest37],        %[src0],          %[mask2]      \n\t"
+      "pmaddhw    %[dest],          %[src1],          %[mask3]      \n\t"
+      "paddw      %[dest37],        %[dest37],        %[dest]       \n\t"
+      "paddw      %[dest37],        %[dest37],        %[ph]         \n\t"
+      "psrlw      %[dest37],        %[dest37],        %[shift]      \n\t"
+
+      /* tmp0 = ( 00 04 02 06 ) */
+      "packsswh   %[tmp0],          %[dest04],        %[dest26]     \n\t"
+      /* tmp1 = ( 01 05 03 07 ) */
+      "packsswh   %[tmp1],          %[dest15],        %[dest37]     \n\t"
+
+      /* tmp2 = ( 00 01 04 05 )*/
+      "punpcklhw  %[tmp2],          %[tmp0],          %[tmp1]       \n\t"
+      /* tmp3 = ( 02 03 06 07 )*/
+      "punpckhhw  %[tmp3],          %[tmp0],          %[tmp1]       \n\t"
+
+      /* ( 00 01 02 03 ) */
+      "punpcklwd  %[dest],          %[tmp2],          %[tmp3]       \n\t"
+      "gssdlc1    %[dest],          0x07(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest],          0x00(%[dst_ptr])                \n\t"
+
+      /* ( 04 05 06 07 ) */
+      "punpckhwd  %[dest],          %[tmp2],          %[tmp3]       \n\t"
+      "gssdlc1    %[dest],          0x0f(%[dst_ptr])                \n\t"
+      "gssdrc1    %[dest],          0x08(%[dst_ptr])                \n\t"
+
+      "daddiu     %[src1_ptr],      %[src1_ptr],      0x08          \n\t"
+      "daddiu     %[src2_ptr],      %[src2_ptr],      0x08          \n\t"
+      "daddiu     %[dst_ptr],       %[dst_ptr],       0x10          \n\t"
+      "daddi      %[width],         %[width],        -0x08          \n\t"
+      "bnez       %[width],         1b                              \n\t"
+      : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest04] "=&f"(dest04),
+        [dest15] "=&f"(dest15), [dest26] "=&f"(dest26), [dest37] "=&f"(dest37),
+        [tmp0] "=&f"(tmp0), [tmp1] "=&f"(tmp1), [tmp2] "=&f"(tmp2),
+        [tmp3] "=&f"(tmp3), [dest] "=&f"(dest)
+      : [src1_ptr] "r"(src_ptr), [src2_ptr] "r"(src2_ptr), [dst_ptr] "r"(dst),
+        [width] "r"(dst_width), [mask0] "f"(mask0), [mask1] "f"(mask1),
+        [mask2] "f"(mask2), [mask3] "f"(mask3), [shift] "f"(shift), [ph] "f"(ph)
+      : "memory");
+}
+
+void ScaleRowDown34_MMI(const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint8_t* dst,
+                      int dst_width) {
+  (void)src_stride;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  uint64_t src[2];
+  uint64_t tmp[2];
+  __asm__ volatile (
+    "1:                                                           \n\t"
+    "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+    "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+    "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+    "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+    "and        %[tmp1],         %[src0],        %[mask1]         \n\t"
+    "psrlw      %[tmp0],         %[src0],        %[rmov]          \n\t"
+    "psllw      %[tmp0],         %[tmp0],        %[lmov1]         \n\t"
+    "or         %[src0],         %[tmp0],        %[tmp1]          \n\t"
+    "punpckhwd  %[tmp0],         %[src0],        %[src0]          \n\t"
+    "psllw      %[tmp1],         %[tmp0],        %[rmov]          \n\t"
+    "or         %[src0],         %[src0],        %[tmp1]          \n\t"
+    "psrlw      %[tmp0],         %[tmp0],        %[rmov8]         \n\t"
+    "pextrh     %[tmp0],         %[tmp0],        %[zero]          \n\t"
+    "pinsrh_2   %[src0],         %[src0],        %[tmp0]          \n\t"
+    "pextrh     %[tmp0],         %[src1],        %[zero]          \n\t"
+    "pinsrh_3   %[src0],         %[src0],        %[tmp0]          \n\t"
+
+    "punpckhwd  %[tmp0],         %[src1],        %[src1]          \n\t"
+    "pextrh     %[tmp1],         %[tmp0],        %[zero]          \n\t"
+    "psrlw      %[src1],         %[src1],        %[rmov]          \n\t"
+    "psllw      %[tmp1],         %[tmp1],        %[rmov8]         \n\t"
+    "or         %[src1],         %[src1],        %[tmp1]          \n\t"
+    "and        %[tmp0],         %[tmp0],        %[mask2]         \n\t"
+    "or         %[src1],         %[src1],        %[tmp0]          \n\t"
+
+    "gssdlc1    %[src0],         0x07(%[dst_ptr])                 \n\t"
+    "gssdrc1    %[src0],         0x00(%[dst_ptr])                 \n\t"
+    "gsswlc1    %[src1],         0x0b(%[dst_ptr])                 \n\t"
+    "gsswrc1    %[src1],         0x08(%[dst_ptr])                 \n\t"
+
+    "daddiu     %[src_ptr],      %[src_ptr],     0x10             \n\t"
+    "daddi      %[width],        %[width],      -0x0c             \n\t"
+    "daddiu     %[dst_ptr],      %[dst_ptr],     0x0c             \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [src0]"=&f"(src[0]),              [src1]"=&f"(src[1]),
+      [tmp0]"=&f"(tmp[0]),              [tmp1]"=&f"(tmp[1])
+    : [src_ptr]"r"(src_ptr),            [dst_ptr]"r"(dst),
+      [lmov]"f"(0xc),                   [rmov]"f"(0x18),
+      [mask1]"f"(0xffff0000ffff),       [rmov8]"f"(0x8),
+      [zero]"f"(0x0),                   [mask2]"f"(0xff000000),
+      [width]"r"(dst_width),            [lmov1]"f"(0x10)
+    : "memory"
+  );
+}
+// clang-format on
+
+#endif  // !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/scale_msa.cc b/media/libyuv/source/scale_msa.cc
new file mode 100644
index 0000000000..482a521f0d
--- /dev/null
+++ b/media/libyuv/source/scale_msa.cc
@@ -0,0 +1,949 @@
+/*
+ *  Copyright 2016 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "libyuv/scale_row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define LOAD_INDEXED_DATA(srcp, indx0, out0) \
+  {                                          \
+    out0[0] = srcp[indx0[0]];                \
+    out0[1] = srcp[indx0[1]];                \
+    out0[2] = srcp[indx0[2]];                \
+    out0[3] = srcp[indx0[3]];                \
+  }
+
+void ScaleARGBRowDown2_MSA(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    dst0 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0);
+    ST_UB(dst0, dst_argb);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Linear_MSA(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  v16u8 src0, src1, vec0, vec1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    vec0 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0);
+    vec1 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0);
+    dst0 = (v16u8)__msa_aver_u_b((v16u8)vec0, (v16u8)vec1);
+    ST_UB(dst0, dst_argb);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Box_MSA(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  const uint8_t* s = src_argb;
+  const uint8_t* t = src_argb + src_stride;
+  v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0;
+  v8u16 reg0, reg1, reg2, reg3;
+  v16i8 shuffler = {0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15};
+
+  for (x = 0; x < dst_width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    vec0 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src1, (v16i8)src1);
+    vec2 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src2, (v16i8)src2);
+    vec3 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src3, (v16i8)src3);
+    reg0 = __msa_hadd_u_h(vec0, vec0);
+    reg1 = __msa_hadd_u_h(vec1, vec1);
+    reg2 = __msa_hadd_u_h(vec2, vec2);
+    reg3 = __msa_hadd_u_h(vec3, vec3);
+    reg0 += reg2;
+    reg1 += reg3;
+    reg0 = (v8u16)__msa_srari_h((v8i16)reg0, 2);
+    reg1 = (v8u16)__msa_srari_h((v8i16)reg1, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    ST_UB(dst0, dst_argb);
+    s += 32;
+    t += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEven_MSA(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int32_t src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  int32_t stepx = src_stepx * 4;
+  int32_t data0, data1, data2, data3;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 4) {
+    data0 = LW(src_argb);
+    data1 = LW(src_argb + stepx);
+    data2 = LW(src_argb + stepx * 2);
+    data3 = LW(src_argb + stepx * 3);
+    SW(data0, dst_argb);
+    SW(data1, dst_argb + 4);
+    SW(data2, dst_argb + 8);
+    SW(data3, dst_argb + 12);
+    src_argb += stepx * 4;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEvenBox_MSA(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  const uint8_t* nxt_argb = src_argb + src_stride;
+  int32_t stepx = src_stepx * 4;
+  int64_t data0, data1, data2, data3;
+  v16u8 src0 = {0}, src1 = {0}, src2 = {0}, src3 = {0};
+  v16u8 vec0, vec1, vec2, vec3;
+  v8u16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 dst0;
+
+  for (x = 0; x < dst_width; x += 4) {
+    data0 = LD(src_argb);
+    data1 = LD(src_argb + stepx);
+    data2 = LD(src_argb + stepx * 2);
+    data3 = LD(src_argb + stepx * 3);
+    src0 = (v16u8)__msa_insert_d((v2i64)src0, 0, data0);
+    src0 = (v16u8)__msa_insert_d((v2i64)src0, 1, data1);
+    src1 = (v16u8)__msa_insert_d((v2i64)src1, 0, data2);
+    src1 = (v16u8)__msa_insert_d((v2i64)src1, 1, data3);
+    data0 = LD(nxt_argb);
+    data1 = LD(nxt_argb + stepx);
+    data2 = LD(nxt_argb + stepx * 2);
+    data3 = LD(nxt_argb + stepx * 3);
+    src2 = (v16u8)__msa_insert_d((v2i64)src2, 0, data0);
+    src2 = (v16u8)__msa_insert_d((v2i64)src2, 1, data1);
+    src3 = (v16u8)__msa_insert_d((v2i64)src3, 0, data2);
+    src3 = (v16u8)__msa_insert_d((v2i64)src3, 1, data3);
+    vec0 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec2 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec3 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    reg0 = __msa_hadd_u_h(vec0, vec0);
+    reg1 = __msa_hadd_u_h(vec1, vec1);
+    reg2 = __msa_hadd_u_h(vec2, vec2);
+    reg3 = __msa_hadd_u_h(vec3, vec3);
+    reg4 = (v8u16)__msa_pckev_d((v2i64)reg2, (v2i64)reg0);
+    reg5 = (v8u16)__msa_pckev_d((v2i64)reg3, (v2i64)reg1);
+    reg6 = (v8u16)__msa_pckod_d((v2i64)reg2, (v2i64)reg0);
+    reg7 = (v8u16)__msa_pckod_d((v2i64)reg3, (v2i64)reg1);
+    reg4 += reg6;
+    reg5 += reg7;
+    reg4 = (v8u16)__msa_srari_h((v8i16)reg4, 2);
+    reg5 = (v8u16)__msa_srari_h((v8i16)reg5, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    ST_UB(dst0, dst_argb);
+    src_argb += stepx * 4;
+    nxt_argb += stepx * 4;
+    dst_argb += 16;
+  }
+}
+
+void ScaleRowDown2_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Linear_MSA(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    dst0 = __msa_aver_u_b(vec1, vec0);
+    dst1 = __msa_aver_u_b(vec3, vec2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = __msa_hadd_u_h(src0, src0);
+    vec1 = __msa_hadd_u_h(src1, src1);
+    vec2 = __msa_hadd_u_h(src2, src2);
+    vec3 = __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    vec0 = (v8u16)__msa_srari_h((v8i16)vec0, 2);
+    vec1 = (v8u16)__msa_srari_h((v8i16)vec1, 2);
+    vec2 = (v8u16)__msa_srari_h((v8i16)vec2, 2);
+    vec3 = (v8u16)__msa_srari_h((v8i16)vec3, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    ST_UB2(dst0, dst1, dst, 16);
+    s += 64;
+    t += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown4_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst);
+    src_ptr += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown4Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t0 = s + src_stride;
+  const uint8_t* t1 = s + src_stride * 2;
+  const uint8_t* t2 = s + src_stride * 3;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v4u32 reg0, reg1, reg2, reg3;
+
+  for (x = 0; x < dst_width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t0, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t0, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t0, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t0, 48);
+    vec0 = __msa_hadd_u_h(src0, src0);
+    vec1 = __msa_hadd_u_h(src1, src1);
+    vec2 = __msa_hadd_u_h(src2, src2);
+    vec3 = __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)t1, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)t1, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t1, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t1, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t2, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t2, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t2, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t2, 48);
+    vec0 += __msa_hadd_u_h(src0, src0);
+    vec1 += __msa_hadd_u_h(src1, src1);
+    vec2 += __msa_hadd_u_h(src2, src2);
+    vec3 += __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    reg0 = __msa_hadd_u_w(vec0, vec0);
+    reg1 = __msa_hadd_u_w(vec1, vec1);
+    reg2 = __msa_hadd_u_w(vec2, vec2);
+    reg3 = __msa_hadd_u_w(vec3, vec3);
+    reg0 = (v4u32)__msa_srari_w((v4i32)reg0, 4);
+    reg1 = (v4u32)__msa_srari_w((v4i32)reg1, 4);
+    reg2 = (v4u32)__msa_srari_w((v4i32)reg2, 4);
+    reg3 = (v4u32)__msa_srari_w((v4i32)reg3, 4);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst);
+    s += 64;
+    t0 += 64;
+    t1 += 64;
+    t2 += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown38_MSA(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x, width;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, vec0;
+  v16i8 mask = {0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0};
+  (void)src_stride;
+
+  assert(dst_width % 3 == 0);
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    vec0 = (v16u8)__msa_vshf_b(mask, (v16i8)src1, (v16i8)src0);
+    dst0 = __msa_copy_u_d((v2i64)vec0, 0);
+    dst1 = __msa_copy_u_w((v4i32)vec0, 2);
+    SD(dst0, dst);
+    SW(dst1, dst + 8);
+    src_ptr += 32;
+    dst += 12;
+  }
+}
+
+void ScaleRowDown38_2_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, width;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, src2, src3, out;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4u32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  v8i16 zero = {0};
+  v8i16 mask = {0, 1, 2, 8, 3, 4, 5, 9};
+  v16i8 dst_mask = {0, 2, 16, 4, 6, 18, 8, 10, 20, 12, 14, 22, 0, 0, 0, 0};
+  v4u32 const_0x2AAA = (v4u32)__msa_fill_w(0x2AAA);
+  v4u32 const_0x4000 = (v4u32)__msa_fill_w(0x4000);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec0);
+    vec5 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec1);
+    vec6 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec2);
+    vec7 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec3);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    vec1 = (v8u16)__msa_pckod_w((v4i32)vec3, (v4i32)vec2);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    tmp0 = __msa_hadd_u_w(vec4, vec4);
+    tmp1 = __msa_hadd_u_w(vec5, vec5);
+    tmp2 = __msa_hadd_u_w(vec6, vec6);
+    tmp3 = __msa_hadd_u_w(vec7, vec7);
+    tmp4 = __msa_hadd_u_w(vec0, vec0);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    tmp0 = __msa_hadd_u_w(vec0, vec0);
+    tmp1 = __msa_hadd_u_w(vec1, vec1);
+    tmp0 *= const_0x2AAA;
+    tmp1 *= const_0x2AAA;
+    tmp4 *= const_0x4000;
+    tmp0 = (v4u32)__msa_srai_w((v4i32)tmp0, 16);
+    tmp1 = (v4u32)__msa_srai_w((v4i32)tmp1, 16);
+    tmp4 = (v4u32)__msa_srai_w((v4i32)tmp4, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp4, (v8i16)tmp4);
+    out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0);
+    dst0 = __msa_copy_u_d((v2i64)out, 0);
+    dst1 = __msa_copy_u_w((v4i32)out, 2);
+    SD(dst0, dst_ptr);
+    SW(dst1, dst_ptr + 8);
+    s += 32;
+    t += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleRowDown38_3_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, width;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t0 = s + src_stride;
+  const uint8_t* t1 = s + src_stride * 2;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, src2, src3, src4, src5, out;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4u32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  v8u16 zero = {0};
+  v8i16 mask = {0, 1, 2, 8, 3, 4, 5, 9};
+  v16i8 dst_mask = {0, 2, 16, 4, 6, 18, 8, 10, 20, 12, 14, 22, 0, 0, 0, 0};
+  v4u32 const_0x1C71 = (v4u32)__msa_fill_w(0x1C71);
+  v4u32 const_0x2AAA = (v4u32)__msa_fill_w(0x2AAA);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t0, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t0, 16);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t1, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t1, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src4);
+    vec5 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src4);
+    vec6 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src5);
+    vec7 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src5);
+    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec0 += __msa_hadd_u_h((v16u8)vec4, (v16u8)vec4);
+    vec1 += __msa_hadd_u_h((v16u8)vec5, (v16u8)vec5);
+    vec2 += __msa_hadd_u_h((v16u8)vec6, (v16u8)vec6);
+    vec3 += __msa_hadd_u_h((v16u8)vec7, (v16u8)vec7);
+    vec4 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec0);
+    vec5 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec1);
+    vec6 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec2);
+    vec7 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec3);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    vec1 = (v8u16)__msa_pckod_w((v4i32)vec3, (v4i32)vec2);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    tmp0 = __msa_hadd_u_w(vec4, vec4);
+    tmp1 = __msa_hadd_u_w(vec5, vec5);
+    tmp2 = __msa_hadd_u_w(vec6, vec6);
+    tmp3 = __msa_hadd_u_w(vec7, vec7);
+    tmp4 = __msa_hadd_u_w(vec0, vec0);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    tmp0 = __msa_hadd_u_w(vec0, vec0);
+    tmp1 = __msa_hadd_u_w(vec1, vec1);
+    tmp0 *= const_0x1C71;
+    tmp1 *= const_0x1C71;
+    tmp4 *= const_0x2AAA;
+    tmp0 = (v4u32)__msa_srai_w((v4i32)tmp0, 16);
+    tmp1 = (v4u32)__msa_srai_w((v4i32)tmp1, 16);
+    tmp4 = (v4u32)__msa_srai_w((v4i32)tmp4, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp4, (v8i16)tmp4);
+    out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0);
+    dst0 = __msa_copy_u_d((v2i64)out, 0);
+    dst1 = __msa_copy_u_w((v4i32)out, 2);
+    SD(dst0, dst_ptr);
+    SW(dst1, dst_ptr + 8);
+    s += 32;
+    t0 += 32;
+    t1 += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleAddRow_MSA(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
+  int x;
+  v16u8 src0;
+  v8u16 dst0, dst1;
+  v16i8 zero = {0};
+
+  assert(src_width > 0);
+
+  for (x = 0; x < src_width; x += 16) {
+    src0 = LD_UB(src_ptr);
+    dst0 = (v8u16)__msa_ld_h((v8i16*)dst_ptr, 0);
+    dst1 = (v8u16)__msa_ld_h((v8i16*)dst_ptr, 16);
+    dst0 += (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
+    dst1 += (v8u16)__msa_ilvl_b(zero, (v16i8)src0);
+    ST_UH2(dst0, dst1, dst_ptr, 8);
+    src_ptr += 16;
+    dst_ptr += 16;
+  }
+}
+
+void ScaleFilterCols_MSA(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  int j;
+  v4i32 vec_x = __msa_fill_w(x);
+  v4i32 vec_dx = __msa_fill_w(dx);
+  v4i32 vec_const = {0, 1, 2, 3};
+  v4i32 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8u16 reg0, reg1;
+  v16u8 dst0;
+  v4i32 const_0xFFFF = __msa_fill_w(0xFFFF);
+  v4i32 const_0x40 = __msa_fill_w(0x40);
+
+  vec0 = vec_dx * vec_const;
+  vec1 = vec_dx * 4;
+  vec_x += vec0;
+
+  for (j = 0; j < dst_width - 1; j += 16) {
+    vec2 = vec_x >> 16;
+    vec6 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec3 = vec_x >> 16;
+    vec7 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec4 = vec_x >> 16;
+    vec8 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec5 = vec_x >> 16;
+    vec9 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec6 >>= 9;
+    vec7 >>= 9;
+    vec8 >>= 9;
+    vec9 >>= 9;
+    LOAD_INDEXED_DATA(src_ptr, vec2, tmp0);
+    LOAD_INDEXED_DATA(src_ptr, vec3, tmp1);
+    LOAD_INDEXED_DATA(src_ptr, vec4, tmp2);
+    LOAD_INDEXED_DATA(src_ptr, vec5, tmp3);
+    vec2 += 1;
+    vec3 += 1;
+    vec4 += 1;
+    vec5 += 1;
+    LOAD_INDEXED_DATA(src_ptr, vec2, tmp4);
+    LOAD_INDEXED_DATA(src_ptr, vec3, tmp5);
+    LOAD_INDEXED_DATA(src_ptr, vec4, tmp6);
+    LOAD_INDEXED_DATA(src_ptr, vec5, tmp7);
+    tmp4 -= tmp0;
+    tmp5 -= tmp1;
+    tmp6 -= tmp2;
+    tmp7 -= tmp3;
+    tmp4 *= vec6;
+    tmp5 *= vec7;
+    tmp6 *= vec8;
+    tmp7 *= vec9;
+    tmp4 += const_0x40;
+    tmp5 += const_0x40;
+    tmp6 += const_0x40;
+    tmp7 += const_0x40;
+    tmp4 >>= 7;
+    tmp5 >>= 7;
+    tmp6 >>= 7;
+    tmp7 >>= 7;
+    tmp0 += tmp4;
+    tmp1 += tmp5;
+    tmp2 += tmp6;
+    tmp3 += tmp7;
+    reg0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    reg1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    __msa_st_b(dst0, dst_ptr, 0);
+    dst_ptr += 16;
+  }
+}
+
+void ScaleARGBCols_MSA(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  uint32_t* dst = (uint32_t*)(dst_argb);
+  int j;
+  v4i32 x_vec = __msa_fill_w(x);
+  v4i32 dx_vec = __msa_fill_w(dx);
+  v4i32 const_vec = {0, 1, 2, 3};
+  v4i32 vec0, vec1, vec2;
+  v4i32 dst0;
+
+  vec0 = dx_vec * const_vec;
+  vec1 = dx_vec * 4;
+  x_vec += vec0;
+
+  for (j = 0; j < dst_width; j += 4) {
+    vec2 = x_vec >> 16;
+    x_vec += vec1;
+    LOAD_INDEXED_DATA(src, vec2, dst0);
+    __msa_st_w(dst0, dst, 0);
+    dst += 4;
+  }
+}
+
+void ScaleARGBFilterCols_MSA(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx) {
+  const uint32_t* src = (const uint32_t*)(src_argb);
+  int j;
+  v4u32 src0, src1, src2, src3;
+  v4u32 vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 mult0, mult1, mult2, mult3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 dst0, dst1;
+  v4u32 vec_x = (v4u32)__msa_fill_w(x);
+  v4u32 vec_dx = (v4u32)__msa_fill_w(dx);
+  v4u32 vec_const = {0, 1, 2, 3};
+  v16u8 const_0x7f = (v16u8)__msa_fill_b(0x7f);
+
+  vec0 = vec_dx * vec_const;
+  vec1 = vec_dx * 4;
+  vec_x += vec0;
+
+  for (j = 0; j < dst_width - 1; j += 8) {
+    vec2 = vec_x >> 16;
+    reg0 = (v16u8)(vec_x >> 9);
+    vec_x += vec1;
+    vec3 = vec_x >> 16;
+    reg1 = (v16u8)(vec_x >> 9);
+    vec_x += vec1;
+    reg0 = reg0 & const_0x7f;
+    reg1 = reg1 & const_0x7f;
+    reg0 = (v16u8)__msa_shf_b((v16i8)reg0, 0);
+    reg1 = (v16u8)__msa_shf_b((v16i8)reg1, 0);
+    reg2 = reg0 ^ const_0x7f;
+    reg3 = reg1 ^ const_0x7f;
+    mult0 = (v16u8)__msa_ilvr_b((v16i8)reg0, (v16i8)reg2);
+    mult1 = (v16u8)__msa_ilvl_b((v16i8)reg0, (v16i8)reg2);
+    mult2 = (v16u8)__msa_ilvr_b((v16i8)reg1, (v16i8)reg3);
+    mult3 = (v16u8)__msa_ilvl_b((v16i8)reg1, (v16i8)reg3);
+    LOAD_INDEXED_DATA(src, vec2, src0);
+    LOAD_INDEXED_DATA(src, vec3, src1);
+    vec2 += 1;
+    vec3 += 1;
+    LOAD_INDEXED_DATA(src, vec2, src2);
+    LOAD_INDEXED_DATA(src, vec3, src3);
+    reg4 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    reg5 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    reg6 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    reg7 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    tmp0 = __msa_dotp_u_h(reg4, mult0);
+    tmp1 = __msa_dotp_u_h(reg5, mult1);
+    tmp2 = __msa_dotp_u_h(reg6, mult2);
+    tmp3 = __msa_dotp_u_h(reg7, mult3);
+    tmp0 >>= 7;
+    tmp1 >>= 7;
+    tmp2 >>= 7;
+    tmp3 >>= 7;
+    dst0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+    __msa_st_b(dst0, dst_argb, 0);
+    __msa_st_b(dst1, dst_argb, 16);
+    dst_argb += 32;
+  }
+}
+
+void ScaleRowDown34_MSA(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x;
+  (void)src_stride;
+  v16u8 src0, src1, src2, src3;
+  v16u8 vec0, vec1, vec2;
+  v16i8 mask0 = {0, 1, 3, 4, 5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20};
+  v16i8 mask1 = {5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20, 21, 23, 24, 25};
+  v16i8 mask2 = {11, 12, 13, 15, 16, 17, 19, 20,
+                 21, 23, 24, 25, 27, 28, 29, 31};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src2, (v16i8)src1);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src2);
+    __msa_st_b((v16i8)vec0, dst, 0);
+    __msa_st_b((v16i8)vec1, dst, 16);
+    __msa_st_b((v16i8)vec2, dst, 32);
+    src_ptr += 64;
+    dst += 48;
+  }
+}
+
+void ScaleRowDown34_0_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1, dst2;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 vec6, vec7, vec8, vec9, vec10, vec11;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v8i16 reg6, reg7, reg8, reg9, reg10, reg11;
+  v16u8 const0 = {3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1};
+  v16u8 const1 = {1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1};
+  v16u8 const2 = {1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3};
+  v16i8 mask0 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
+  v16i8 mask1 = {10, 11, 12, 13, 13, 14, 14, 15,
+                 16, 17, 17, 18, 18, 19, 20, 21};
+  v16i8 mask2 = {5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15};
+  v8i16 shft0 = {2, 1, 2, 2, 1, 2, 2, 1};
+  v8i16 shft1 = {2, 2, 1, 2, 2, 1, 2, 2};
+  v8i16 shft2 = {1, 2, 2, 1, 2, 2, 1, 2};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src1, (v16i8)src1);
+    vec3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src2, (v16i8)src2);
+    vec4 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec5 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src3);
+    vec6 = (v16u8)__msa_vshf_b(mask0, (v16i8)src4, (v16i8)src4);
+    vec7 = (v16u8)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec8 = (v16u8)__msa_vshf_b(mask2, (v16i8)src5, (v16i8)src5);
+    vec9 = (v16u8)__msa_vshf_b(mask0, (v16i8)src6, (v16i8)src6);
+    vec10 = (v16u8)__msa_vshf_b(mask1, (v16i8)src7, (v16i8)src6);
+    vec11 = (v16u8)__msa_vshf_b(mask2, (v16i8)src7, (v16i8)src7);
+    reg0 = (v8i16)__msa_dotp_u_h(vec0, const0);
+    reg1 = (v8i16)__msa_dotp_u_h(vec1, const1);
+    reg2 = (v8i16)__msa_dotp_u_h(vec2, const2);
+    reg3 = (v8i16)__msa_dotp_u_h(vec3, const0);
+    reg4 = (v8i16)__msa_dotp_u_h(vec4, const1);
+    reg5 = (v8i16)__msa_dotp_u_h(vec5, const2);
+    reg6 = (v8i16)__msa_dotp_u_h(vec6, const0);
+    reg7 = (v8i16)__msa_dotp_u_h(vec7, const1);
+    reg8 = (v8i16)__msa_dotp_u_h(vec8, const2);
+    reg9 = (v8i16)__msa_dotp_u_h(vec9, const0);
+    reg10 = (v8i16)__msa_dotp_u_h(vec10, const1);
+    reg11 = (v8i16)__msa_dotp_u_h(vec11, const2);
+    reg0 = __msa_srar_h(reg0, shft0);
+    reg1 = __msa_srar_h(reg1, shft1);
+    reg2 = __msa_srar_h(reg2, shft2);
+    reg3 = __msa_srar_h(reg3, shft0);
+    reg4 = __msa_srar_h(reg4, shft1);
+    reg5 = __msa_srar_h(reg5, shft2);
+    reg6 = __msa_srar_h(reg6, shft0);
+    reg7 = __msa_srar_h(reg7, shft1);
+    reg8 = __msa_srar_h(reg8, shft2);
+    reg9 = __msa_srar_h(reg9, shft0);
+    reg10 = __msa_srar_h(reg10, shft1);
+    reg11 = __msa_srar_h(reg11, shft2);
+    reg0 = reg0 * 3 + reg6;
+    reg1 = reg1 * 3 + reg7;
+    reg2 = reg2 * 3 + reg8;
+    reg3 = reg3 * 3 + reg9;
+    reg4 = reg4 * 3 + reg10;
+    reg5 = reg5 * 3 + reg11;
+    reg0 = __msa_srari_h(reg0, 2);
+    reg1 = __msa_srari_h(reg1, 2);
+    reg2 = __msa_srari_h(reg2, 2);
+    reg3 = __msa_srari_h(reg3, 2);
+    reg4 = __msa_srari_h(reg4, 2);
+    reg5 = __msa_srari_h(reg5, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    __msa_st_b((v16i8)dst0, d, 0);
+    __msa_st_b((v16i8)dst1, d, 16);
+    __msa_st_b((v16i8)dst2, d, 32);
+    s += 64;
+    t += 64;
+    d += 48;
+  }
+}
+
+void ScaleRowDown34_1_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1, dst2;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 vec6, vec7, vec8, vec9, vec10, vec11;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v8i16 reg6, reg7, reg8, reg9, reg10, reg11;
+  v16u8 const0 = {3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1};
+  v16u8 const1 = {1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1};
+  v16u8 const2 = {1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3};
+  v16i8 mask0 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
+  v16i8 mask1 = {10, 11, 12, 13, 13, 14, 14, 15,
+                 16, 17, 17, 18, 18, 19, 20, 21};
+  v16i8 mask2 = {5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15};
+  v8i16 shft0 = {2, 1, 2, 2, 1, 2, 2, 1};
+  v8i16 shft1 = {2, 2, 1, 2, 2, 1, 2, 2};
+  v8i16 shft2 = {1, 2, 2, 1, 2, 2, 1, 2};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src1, (v16i8)src1);
+    vec3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src2, (v16i8)src2);
+    vec4 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec5 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src3);
+    vec6 = (v16u8)__msa_vshf_b(mask0, (v16i8)src4, (v16i8)src4);
+    vec7 = (v16u8)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec8 = (v16u8)__msa_vshf_b(mask2, (v16i8)src5, (v16i8)src5);
+    vec9 = (v16u8)__msa_vshf_b(mask0, (v16i8)src6, (v16i8)src6);
+    vec10 = (v16u8)__msa_vshf_b(mask1, (v16i8)src7, (v16i8)src6);
+    vec11 = (v16u8)__msa_vshf_b(mask2, (v16i8)src7, (v16i8)src7);
+    reg0 = (v8i16)__msa_dotp_u_h(vec0, const0);
+    reg1 = (v8i16)__msa_dotp_u_h(vec1, const1);
+    reg2 = (v8i16)__msa_dotp_u_h(vec2, const2);
+    reg3 = (v8i16)__msa_dotp_u_h(vec3, const0);
+    reg4 = (v8i16)__msa_dotp_u_h(vec4, const1);
+    reg5 = (v8i16)__msa_dotp_u_h(vec5, const2);
+    reg6 = (v8i16)__msa_dotp_u_h(vec6, const0);
+    reg7 = (v8i16)__msa_dotp_u_h(vec7, const1);
+    reg8 = (v8i16)__msa_dotp_u_h(vec8, const2);
+    reg9 = (v8i16)__msa_dotp_u_h(vec9, const0);
+    reg10 = (v8i16)__msa_dotp_u_h(vec10, const1);
+    reg11 = (v8i16)__msa_dotp_u_h(vec11, const2);
+    reg0 = __msa_srar_h(reg0, shft0);
+    reg1 = __msa_srar_h(reg1, shft1);
+    reg2 = __msa_srar_h(reg2, shft2);
+    reg3 = __msa_srar_h(reg3, shft0);
+    reg4 = __msa_srar_h(reg4, shft1);
+    reg5 = __msa_srar_h(reg5, shft2);
+    reg6 = __msa_srar_h(reg6, shft0);
+    reg7 = __msa_srar_h(reg7, shft1);
+    reg8 = __msa_srar_h(reg8, shft2);
+    reg9 = __msa_srar_h(reg9, shft0);
+    reg10 = __msa_srar_h(reg10, shft1);
+    reg11 = __msa_srar_h(reg11, shft2);
+    reg0 += reg6;
+    reg1 += reg7;
+    reg2 += reg8;
+    reg3 += reg9;
+    reg4 += reg10;
+    reg5 += reg11;
+    reg0 = __msa_srari_h(reg0, 1);
+    reg1 = __msa_srari_h(reg1, 1);
+    reg2 = __msa_srari_h(reg2, 1);
+    reg3 = __msa_srari_h(reg3, 1);
+    reg4 = __msa_srari_h(reg4, 1);
+    reg5 = __msa_srari_h(reg5, 1);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    __msa_st_b((v16i8)dst0, d, 0);
+    __msa_st_b((v16i8)dst1, d, 16);
+    __msa_st_b((v16i8)dst2, d, 32);
+    s += 64;
+    t += 64;
+    d += 48;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
diff --git a/media/libyuv/source/scale_neon.cc b/media/libyuv/source/scale_neon.cc
index 44b0c8080d..6a0d6e1b49 100644
--- a/media/libyuv/source/scale_neon.cc
+++ b/media/libyuv/source/scale_neon.cc
@@ -23,576 +23,1019 @@ extern "C" {
 // Provided by Fritz Koenig
 
 // Read 32x1 throw away even pixels, and write 16x1.
-void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    // load even pixels into q0, odd into q1
-    MEMACCESS(0)
-    "vld2.8     {q0, q1}, [%0]!                \n"
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst),              // %1
-    "+r"(dst_width)         // %2
-  :
-  : "q0", "q1"              // Clobber List
+void ScaleRowDown2_NEON(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      // load even pixels into q0, odd into q1
+      "vld2.8      {q0, q1}, [%0]!               \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q1}, [%1]!                   \n"  // store odd pixels
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "q0", "q1"  // Clobber List
   );
 }
 
 // Read 32x1 average down and write 16x1.
-void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0, q1}, [%0]!                \n"  // load pixels and post inc
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop
-    "vpaddl.u8  q0, q0                         \n"  // add adjacent
-    "vpaddl.u8  q1, q1                         \n"
-    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #1                     \n"
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst),              // %1
-    "+r"(dst_width)         // %2
-  :
-  : "q0", "q1"     // Clobber List
+void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst,
+                              int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 32 pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vrhadd.u8   q0, q0, q1                    \n"  // rounding half add
+      "vst1.8      {q0}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "q0", "q1"  // Clobber List
   );
 }
 
 // Read 32x2 average down and write 16x1.
-void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width) {
-  asm volatile (
-    // change the stride to row 2 pointer
-    "add        %1, %0                         \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
-    MEMACCESS(1)
-    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
-    "vpaddl.u8  q1, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
-    "vpadal.u8  q1, q3                         \n"
-    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #2                     \n"
-    MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(src_stride),       // %1
-    "+r"(dst),              // %2
-    "+r"(dst_width)         // %3
-  :
-  : "q0", "q1", "q2", "q3"     // Clobber List
+void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %0                        \n"
+      "1:                                        \n"
+      "vld1.8      {q0, q1}, [%0]!               \n"  // load row 1 and post inc
+      "vld1.8      {q2, q3}, [%1]!               \n"  // load row 2 and post inc
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop
+      "vpaddl.u8   q0, q0                        \n"  // row 1 add adjacent
+      "vpaddl.u8   q1, q1                        \n"
+      "vpadal.u8   q0, q2                        \n"  // row 2 add adjacent +
+                                                      // row1
+      "vpadal.u8   q1, q3                        \n"
+      "vrshrn.u16  d0, q0, #2                    \n"  // downshift, round and
+                                                      // pack
+      "vrshrn.u16  d1, q1, #2                    \n"
+      "vst1.8      {q0}, [%2]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "q0", "q1", "q2", "q3"  // Clobber List
   );
 }
 
-void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
-    "subs       %2, %2, #8                     \n" // 8 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {d2}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width)         // %2
-  :
-  : "q0", "q1", "memory", "cc"
-  );
+void ScaleRowDown4_NEON(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "vst1.8      {d2}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "q0", "q1", "memory", "cc");
 }
 
-void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  const uint8* src_ptr1 = src_ptr + src_stride;
-  const uint8* src_ptr2 = src_ptr + src_stride * 2;
-  const uint8* src_ptr3 = src_ptr + src_stride * 3;
-asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
-    MEMACCESS(3)
-    "vld1.8     {q1}, [%3]!                    \n"
-    MEMACCESS(4)
-    "vld1.8     {q2}, [%4]!                    \n"
-    MEMACCESS(5)
-    "vld1.8     {q3}, [%5]!                    \n"
-    "subs       %2, %2, #4                     \n"
-    "vpaddl.u8  q0, q0                         \n"
-    "vpadal.u8  q0, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"
-    "vpadal.u8  q0, q3                         \n"
-    "vpaddl.u16 q0, q0                         \n"
-    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
-    "vmovn.u16  d0, q0                         \n"
-    MEMACCESS(1)
-    "vst1.32    {d0[0]}, [%1]!                 \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),   // %0
-    "+r"(dst_ptr),   // %1
-    "+r"(dst_width), // %2
-    "+r"(src_ptr1),  // %3
-    "+r"(src_ptr2),  // %4
-    "+r"(src_ptr3)   // %5
-  :
-  : "q0", "q1", "q2", "q3", "memory", "cc"
-  );
+void ScaleRowDown4Box_NEON(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  const uint8_t* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8_t* src_ptr3 = src_ptr + src_stride * 3;
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load up 16x4
+      "vld1.8      {q1}, [%3]!                   \n"
+      "vld1.8      {q2}, [%4]!                   \n"
+      "vld1.8      {q3}, [%5]!                   \n"
+      "subs        %2, %2, #4                    \n"
+      "vpaddl.u8   q0, q0                        \n"
+      "vpadal.u8   q0, q1                        \n"
+      "vpadal.u8   q0, q2                        \n"
+      "vpadal.u8   q0, q3                        \n"
+      "vpaddl.u16  q0, q0                        \n"
+      "vrshrn.u32  d0, q0, #4                    \n"  // divide by 16 w/rounding
+      "vmovn.u16   d0, q0                        \n"
+      "vst1.32     {d0[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_ptr1),   // %3
+        "+r"(src_ptr2),   // %4
+        "+r"(src_ptr3)    // %5
+      :
+      : "q0", "q1", "q2", "q3", "memory", "cc");
 }
 
 // Down scale from 4 to 3 pixels. Use the neon multilane read/write
 // to load up the every 4th pixel into a 4 different registers.
 // Point samples 32 pixels to 24 pixels.
-void ScaleRowDown34_NEON(const uint8* src_ptr,
+void ScaleRowDown34_NEON(const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "subs       %2, %2, #24                  \n"
-    "vmov       d2, d3                       \n" // order d0, d1, d2
-    MEMACCESS(1)
-    "vst3.8     {d0, d1, d2}, [%1]!          \n"
-    "bgt        1b                           \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width)         // %2
-  :
-  : "d0", "d1", "d2", "d3", "memory", "cc"
-  );
+                         uint8_t* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "subs        %2, %2, #24                   \n"
+      "vmov        d2, d3                        \n"  // order d0, d1, d2
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "d0", "d1", "d2", "d3", "memory", "cc");
 }
 
-void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown34_0_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vmov.u8    d24, #3                        \n"
-    "add        %3, %0                         \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-    "subs         %2, %2, #24                  \n"
-
-    // filter src line 0 with src line 1
-    // expand chars to shorts to allow for room
-    // when adding lines together
-    "vmovl.u8     q8, d4                       \n"
-    "vmovl.u8     q9, d5                       \n"
-    "vmovl.u8     q10, d6                      \n"
-    "vmovl.u8     q11, d7                      \n"
-
-    // 3 * line_0 + line_1
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vmlal.u8     q9, d1, d24                  \n"
-    "vmlal.u8     q10, d2, d24                 \n"
-    "vmlal.u8     q11, d3, d24                 \n"
-
-    // (3 * line_0 + line_1) >> 2
-    "vqrshrn.u16  d0, q8, #2                   \n"
-    "vqrshrn.u16  d1, q9, #2                   \n"
-    "vqrshrn.u16  d2, q10, #2                  \n"
-    "vqrshrn.u16  d3, q11, #2                  \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q8, d1                       \n"
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vqrshrn.u16  d0, q8, #2                   \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q8, d2                       \n"
-    "vmlal.u8     q8, d3, d24                  \n"
-    "vqrshrn.u16  d2, q8, #2                   \n"
-
-    MEMACCESS(1)
-    "vst3.8       {d0, d1, d2}, [%1]!          \n"
-
-    "bgt          1b                           \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width),        // %2
-    "+r"(src_stride)        // %3
-  :
-  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
-  );
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vmov.u8     d24, #3                       \n"
+      "add         %3, %0                        \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"  // src line 1
+      "subs        %2, %2, #24                   \n"
+
+      // filter src line 0 with src line 1
+      // expand chars to shorts to allow for room
+      // when adding lines together
+      "vmovl.u8    q8, d4                        \n"
+      "vmovl.u8    q9, d5                        \n"
+      "vmovl.u8    q10, d6                       \n"
+      "vmovl.u8    q11, d7                       \n"
+
+      // 3 * line_0 + line_1
+      "vmlal.u8    q8, d0, d24                   \n"
+      "vmlal.u8    q9, d1, d24                   \n"
+      "vmlal.u8    q10, d2, d24                  \n"
+      "vmlal.u8    q11, d3, d24                  \n"
+
+      // (3 * line_0 + line_1 + 2) >> 2
+      "vqrshrn.u16 d0, q8, #2                    \n"
+      "vqrshrn.u16 d1, q9, #2                    \n"
+      "vqrshrn.u16 d2, q10, #2                   \n"
+      "vqrshrn.u16 d3, q11, #2                   \n"
+
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "vmovl.u8    q8, d1                        \n"
+      "vmlal.u8    q8, d0, d24                   \n"
+      "vqrshrn.u16 d0, q8, #2                    \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "vrhadd.u8   d1, d1, d2                    \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "vmovl.u8    q8, d2                        \n"
+      "vmlal.u8    q8, d3, d24                   \n"
+      "vqrshrn.u16 d2, q8, #2                    \n"
+
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"
+
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_stride)  // %3
+      :
+      : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory",
+        "cc");
 }
 
-void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown34_1_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vmov.u8    d24, #3                        \n"
-    "add        %3, %0                         \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-    "subs         %2, %2, #24                  \n"
-    // average src line 0 with src line 1
-    "vrhadd.u8    q0, q0, q2                   \n"
-    "vrhadd.u8    q1, q1, q3                   \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q3, d1                       \n"
-    "vmlal.u8     q3, d0, d24                  \n"
-    "vqrshrn.u16  d0, q3, #2                   \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q3, d2                       \n"
-    "vmlal.u8     q3, d3, d24                  \n"
-    "vqrshrn.u16  d2, q3, #2                   \n"
-
-    MEMACCESS(1)
-    "vst3.8       {d0, d1, d2}, [%1]!          \n"
-    "bgt          1b                           \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width),        // %2
-    "+r"(src_stride)        // %3
-  :
-  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
-  );
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vmov.u8     d24, #3                       \n"
+      "add         %3, %0                        \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"  // src line 1
+      "subs        %2, %2, #24                   \n"
+      // average src line 0 with src line 1
+      "vrhadd.u8   q0, q0, q2                    \n"
+      "vrhadd.u8   q1, q1, q3                    \n"
+
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "vmovl.u8    q3, d1                        \n"
+      "vmlal.u8    q3, d0, d24                   \n"
+      "vqrshrn.u16 d0, q3, #2                    \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "vrhadd.u8   d1, d1, d2                    \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "vmovl.u8    q3, d2                        \n"
+      "vmlal.u8    q3, d3, d24                   \n"
+      "vqrshrn.u16 d2, q3, #2                    \n"
+
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_stride)  // %3
+      :
+      : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc");
 }
 
 #define HAS_SCALEROWDOWN38_NEON
-static uvec8 kShuf38 =
-  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
-static uvec8 kShuf38_2 =
-  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
-static vec16 kMult38_Div6 =
-  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
-    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
-static vec16 kMult38_Div9 =
-  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
-    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+static const uvec8 kShuf38 = {0,  3,  6,  8,  11, 14, 16, 19,
+                              22, 24, 27, 30, 0,  0,  0,  0};
+static const uvec8 kShuf38_2 = {0,  8, 16, 2,  10, 17, 4, 12,
+                                18, 6, 14, 19, 0,  0,  0, 0};
+static const vec16 kMult38_Div6 = {65536 / 12, 65536 / 12, 65536 / 12,
+                                   65536 / 12, 65536 / 12, 65536 / 12,
+                                   65536 / 12, 65536 / 12};
+static const vec16 kMult38_Div9 = {65536 / 18, 65536 / 18, 65536 / 18,
+                                   65536 / 18, 65536 / 18, 65536 / 18,
+                                   65536 / 18, 65536 / 18};
 
 // 32 -> 12
-void ScaleRowDown38_NEON(const uint8* src_ptr,
+void ScaleRowDown38_NEON(const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    MEMACCESS(3)
-    "vld1.8     {q3}, [%3]                     \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
-    "subs       %2, %2, #12                    \n"
-    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
-    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
-    MEMACCESS(1)
-    "vst1.8     {d4}, [%1]!                    \n"
-    MEMACCESS(1)
-    "vst1.32    {d5[0]}, [%1]!                 \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width)         // %2
-  : "r"(&kShuf38)           // %3
-  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
-  );
+                         uint8_t* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "vld1.8      {q3}, [%3]                    \n"
+      "1:                                        \n"
+      "vld1.8      {d0, d1, d2, d3}, [%0]!       \n"
+      "subs        %2, %2, #12                   \n"
+      "vtbl.u8     d4, {d0, d1, d2, d3}, d6      \n"
+      "vtbl.u8     d5, {d0, d1, d2, d3}, d7      \n"
+      "vst1.8      {d4}, [%1]!                   \n"
+      "vst1.32     {d5[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      : "r"(&kShuf38)    // %3
+      : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc");
 }
 
 // 32x3 -> 12x1
-void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
                                       ptrdiff_t src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+                                      uint8_t* dst_ptr,
+                                      int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride * 2;
+
+  asm volatile(
+      "vld1.16     {q13}, [%5]                   \n"
+      "vld1.8      {q14}, [%6]                   \n"
+      "vld1.8      {q15}, [%7]                   \n"
+      "add         %3, %0                        \n"
+      "1:                                        \n"
+
+      // d0 = 00 40 01 41 02 42 03 43
+      // d1 = 10 50 11 51 12 52 13 53
+      // d2 = 20 60 21 61 22 62 23 63
+      // d3 = 30 70 31 71 32 72 33 73
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"
+      "vld4.8      {d16, d17, d18, d19}, [%4]!   \n"
+      "subs        %2, %2, #12                   \n"
+
+      // Shuffle the input data around to get align the data
+      //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+      // d0 = 00 10 01 11 02 12 03 13
+      // d1 = 40 50 41 51 42 52 43 53
+      "vtrn.u8     d0, d1                        \n"
+      "vtrn.u8     d4, d5                        \n"
+      "vtrn.u8     d16, d17                      \n"
+
+      // d2 = 20 30 21 31 22 32 23 33
+      // d3 = 60 70 61 71 62 72 63 73
+      "vtrn.u8     d2, d3                        \n"
+      "vtrn.u8     d6, d7                        \n"
+      "vtrn.u8     d18, d19                      \n"
+
+      // d0 = 00+10 01+11 02+12 03+13
+      // d2 = 40+50 41+51 42+52 43+53
+      "vpaddl.u8   q0, q0                        \n"
+      "vpaddl.u8   q2, q2                        \n"
+      "vpaddl.u8   q8, q8                        \n"
+
+      // d3 = 60+70 61+71 62+72 63+73
+      "vpaddl.u8   d3, d3                        \n"
+      "vpaddl.u8   d7, d7                        \n"
+      "vpaddl.u8   d19, d19                      \n"
+
+      // combine source lines
+      "vadd.u16    q0, q2                        \n"
+      "vadd.u16    q0, q8                        \n"
+      "vadd.u16    d4, d3, d7                    \n"
+      "vadd.u16    d4, d19                       \n"
+
+      // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+      //             + s[6 + st * 1] + s[7 + st * 1]
+      //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+      "vqrdmulh.s16 q2, q2, q13                  \n"
+      "vmovn.u16   d4, q2                        \n"
+
+      // Shuffle 2,3 reg around so that 2 can be added to the
+      //  0,1 reg and 3 can be added to the 4,5 reg. This
+      //  requires expanding from u8 to u16 as the 0,1 and 4,5
+      //  registers are already expanded. Then do transposes
+      //  to get aligned.
+      // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+      "vmovl.u8    q1, d2                        \n"
+      "vmovl.u8    q3, d6                        \n"
+      "vmovl.u8    q9, d18                       \n"
+
+      // combine source lines
+      "vadd.u16    q1, q3                        \n"
+      "vadd.u16    q1, q9                        \n"
+
+      // d4 = xx 20 xx 30 xx 22 xx 32
+      // d5 = xx 21 xx 31 xx 23 xx 33
+      "vtrn.u32    d2, d3                        \n"
+
+      // d4 = xx 20 xx 21 xx 22 xx 23
+      // d5 = xx 30 xx 31 xx 32 xx 33
+      "vtrn.u16    d2, d3                        \n"
+
+      // 0+1+2, 3+4+5
+      "vadd.u16    q0, q1                        \n"
+
+      // Need to divide, but can't downshift as the the value
+      //  isn't a power of 2. So multiply by 65536 / n
+      //  and take the upper 16 bits.
+      "vqrdmulh.s16 q0, q0, q15                  \n"
+
+      // Align for table lookup, vtbl requires registers to
+      //  be adjacent
+      "vmov.u8     d2, d4                        \n"
+
+      "vtbl.u8     d3, {d0, d1, d2}, d28         \n"
+      "vtbl.u8     d4, {d0, d1, d2}, d29         \n"
+
+      "vst1.8      {d3}, [%1]!                   \n"
+      "vst1.32     {d4[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),       // %0
+        "+r"(dst_ptr),       // %1
+        "+r"(dst_width),     // %2
+        "+r"(src_stride),    // %3
+        "+r"(src_ptr1)       // %4
+      : "r"(&kMult38_Div6),  // %5
+        "r"(&kShuf38_2),     // %6
+        "r"(&kMult38_Div9)   // %7
+      : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory",
+        "cc");
+}
 
-  asm volatile (
-    MEMACCESS(5)
-    "vld1.16    {q13}, [%5]                    \n"
-    MEMACCESS(6)
-    "vld1.8     {q14}, [%6]                    \n"
-    MEMACCESS(7)
-    "vld1.8     {q15}, [%7]                    \n"
-    "add        %3, %0                         \n"
-  "1:                                          \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
-    MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
-    MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
-    MEMACCESS(4)
-    "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
-    "subs         %2, %2, #12                  \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-    "vtrn.u8      d16, d17                     \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-    "vtrn.u8      d18, d19                     \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-    "vpaddl.u8    q8, q8                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-    "vpaddl.u8    d19, d19                     \n"
-
-    // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     q0, q8                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-    "vadd.u16     d4, d19                      \n"
-
-    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
-    //             + s[6 + st * 1] + s[7 + st * 1]
-    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
-    "vqrdmulh.s16 q2, q2, q13                  \n"
-    "vmovn.u16    d4, q2                       \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg. This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded. Then do transposes
-    //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-    "vmovl.u8     q9, d18                      \n"
-
-    // combine source lines
-    "vadd.u16     q1, q3                       \n"
-    "vadd.u16     q1, q9                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
-
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
-
-    // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2. So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q0, q15                  \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
-
-    MEMACCESS(1)
-    "vst1.8       {d3}, [%1]!                  \n"
-    MEMACCESS(1)
-    "vst1.32      {d4[0]}, [%1]!               \n"
-    "bgt          1b                           \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width),        // %2
-    "+r"(src_stride),       // %3
-    "+r"(src_ptr1)          // %4
-  : "r"(&kMult38_Div6),     // %5
-    "r"(&kShuf38_2),        // %6
-    "r"(&kMult38_Div9)      // %7
-  : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory", "cc"
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vld1.16     {q13}, [%4]                   \n"
+      "vld1.8      {q14}, [%5]                   \n"
+      "add         %3, %0                        \n"
+      "1:                                        \n"
+
+      // d0 = 00 40 01 41 02 42 03 43
+      // d1 = 10 50 11 51 12 52 13 53
+      // d2 = 20 60 21 61 22 62 23 63
+      // d3 = 30 70 31 71 32 72 33 73
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"
+      "subs        %2, %2, #12                   \n"
+
+      // Shuffle the input data around to get align the data
+      //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+      // d0 = 00 10 01 11 02 12 03 13
+      // d1 = 40 50 41 51 42 52 43 53
+      "vtrn.u8     d0, d1                        \n"
+      "vtrn.u8     d4, d5                        \n"
+
+      // d2 = 20 30 21 31 22 32 23 33
+      // d3 = 60 70 61 71 62 72 63 73
+      "vtrn.u8     d2, d3                        \n"
+      "vtrn.u8     d6, d7                        \n"
+
+      // d0 = 00+10 01+11 02+12 03+13
+      // d2 = 40+50 41+51 42+52 43+53
+      "vpaddl.u8   q0, q0                        \n"
+      "vpaddl.u8   q2, q2                        \n"
+
+      // d3 = 60+70 61+71 62+72 63+73
+      "vpaddl.u8   d3, d3                        \n"
+      "vpaddl.u8   d7, d7                        \n"
+
+      // combine source lines
+      "vadd.u16    q0, q2                        \n"
+      "vadd.u16    d4, d3, d7                    \n"
+
+      // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+      "vqrshrn.u16 d4, q2, #2                    \n"
+
+      // Shuffle 2,3 reg around so that 2 can be added to the
+      //  0,1 reg and 3 can be added to the 4,5 reg. This
+      //  requires expanding from u8 to u16 as the 0,1 and 4,5
+      //  registers are already expanded. Then do transposes
+      //  to get aligned.
+      // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+      "vmovl.u8    q1, d2                        \n"
+      "vmovl.u8    q3, d6                        \n"
+
+      // combine source lines
+      "vadd.u16    q1, q3                        \n"
+
+      // d4 = xx 20 xx 30 xx 22 xx 32
+      // d5 = xx 21 xx 31 xx 23 xx 33
+      "vtrn.u32    d2, d3                        \n"
+
+      // d4 = xx 20 xx 21 xx 22 xx 23
+      // d5 = xx 30 xx 31 xx 32 xx 33
+      "vtrn.u16    d2, d3                        \n"
+
+      // 0+1+2, 3+4+5
+      "vadd.u16    q0, q1                        \n"
+
+      // Need to divide, but can't downshift as the the value
+      //  isn't a power of 2. So multiply by 65536 / n
+      //  and take the upper 16 bits.
+      "vqrdmulh.s16 q0, q0, q13                  \n"
+
+      // Align for table lookup, vtbl requires registers to
+      //  be adjacent
+      "vmov.u8     d2, d4                        \n"
+
+      "vtbl.u8     d3, {d0, d1, d2}, d28         \n"
+      "vtbl.u8     d4, {d0, d1, d2}, d29         \n"
+
+      "vst1.8      {d3}, [%1]!                   \n"
+      "vst1.32     {d4[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),       // %0
+        "+r"(dst_ptr),       // %1
+        "+r"(dst_width),     // %2
+        "+r"(src_stride)     // %3
+      : "r"(&kMult38_Div6),  // %4
+        "r"(&kShuf38_2)      // %5
+      : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc");
+}
+
+void ScaleRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "vmov.u8     d30, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 01234567
+      "vld1.8      {d5}, [%3]!                   \n"  // 12345678
+
+      "vmovl.u8    q0, d4                        \n"  // 01234567 (16b)
+      "vmovl.u8    q1, d5                        \n"  // 12345678 (16b)
+      "vmlal.u8    q0, d5, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u8    q1, d4, d30                   \n"  // 3*near+far (even)
+
+      "vrshrn.u16  d1, q0, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u16  d0, q1, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.8      {d0, d1}, [%1]!               \n"  // store
+      "subs        %2, %2, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q15"  // Clobber List
   );
 }
 
-// 32x2 -> 12x1
-void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+void ScaleRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    MEMACCESS(4)
-    "vld1.16    {q13}, [%4]                    \n"
-    MEMACCESS(5)
-    "vld1.8     {q14}, [%5]                    \n"
-    "add        %3, %0                         \n"
-  "1:                                          \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
-    MEMACCESS(0)
-    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
-    MEMACCESS(3)
-    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
-    "subs         %2, %2, #12                  \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-
-    // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-
-    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-    "vqrshrn.u16  d4, q2, #2                   \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg. This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded. Then do transposes
-    //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-
-    // combine source lines
-    "vadd.u16     q1, q3                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
-
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
-
-    // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2. So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q0, q13                  \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
-
-    MEMACCESS(1)
-    "vst1.8       {d3}, [%1]!                  \n"
-    MEMACCESS(1)
-    "vst1.32      {d4[0]}, [%1]!               \n"
-    "bgt          1b                           \n"
-  : "+r"(src_ptr),       // %0
-    "+r"(dst_ptr),       // %1
-    "+r"(dst_width),     // %2
-    "+r"(src_stride)     // %3
-  : "r"(&kMult38_Div6),  // %4
-    "r"(&kShuf38_2)      // %5
-  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 1;
+  const uint8_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+      "vmov.u8     d28, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 01234567
+      "vld1.8      {d5}, [%5]!                   \n"  // 12345678
+
+      "vmovl.u8    q0, d4                        \n"  // 01234567 (16b)
+      "vmovl.u8    q1, d5                        \n"  // 12345678 (16b)
+      "vmlal.u8    q0, d5, d28                   \n"  // 3*near+far (1, odd)
+      "vmlal.u8    q1, d4, d28                   \n"  // 3*near+far (1, even)
+
+      "vld1.8      {d8}, [%1]!                   \n"
+      "vld1.8      {d9}, [%6]!                   \n"
+
+      "vmovl.u8    q2, d8                        \n"
+      "vmovl.u8    q3, d9                        \n"
+      "vmlal.u8    q2, d9, d28                   \n"  // 3*near+far (2, odd)
+      "vmlal.u8    q3, d8, d28                   \n"  // 3*near+far (2, even)
+
+      // e  o
+      // q1 q0
+      // q3 q2
+
+      "vmovq       q4, q2                        \n"
+      "vmovq       q5, q3                        \n"
+      "vmla.u16    q4, q0, q15                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u16    q5, q1, q15                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u16    q0, q2, q15                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u16    q1, q3, q15                   \n"  // 9 3 3 1 (2, even)
+
+      // e  o
+      // q5 q4
+      // q1 q0
+
+      "vrshrn.u16  d2, q1, #4                    \n"  // 2, even
+      "vrshrn.u16  d3, q0, #4                    \n"  // 2, odd
+      "vrshrn.u16  d0, q5, #4                    \n"  // 1, even
+      "vrshrn.u16  d1, q4, #4                    \n"  // 1, odd
+
+      "vst2.8      {d0, d1}, [%2]!               \n"  // store
+      "vst2.8      {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "d28",
+        "q15"  // Clobber List
   );
 }
 
-void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                    uint16* dst_ptr, int src_width, int src_height) {
-  const uint8* src_tmp;
-  asm volatile (
-  "1:                                          \n"
-    "mov       %0, %1                          \n"
-    "mov       r12, %5                         \n"
-    "veor      q2, q2, q2                      \n"
-    "veor      q3, q3, q3                      \n"
-  "2:                                          \n"
-    // load 16 pixels into q0
-    MEMACCESS(0)
-    "vld1.8     {q0}, [%0], %3                 \n"
-    "vaddw.u8   q3, q3, d1                     \n"
-    "vaddw.u8   q2, q2, d0                     \n"
-    "subs       r12, r12, #1                   \n"
-    "bgt        2b                             \n"
-    MEMACCESS(2)
-    "vst1.16    {q2, q3}, [%2]!                \n"  // store pixels
-    "add        %1, %1, #16                    \n"
-    "subs       %4, %4, #16                    \n"  // 16 processed per loop
-    "bgt        1b                             \n"
-  : "=&r"(src_tmp),    // %0
-    "+r"(src_ptr),     // %1
-    "+r"(dst_ptr),     // %2
-    "+r"(src_stride),  // %3
-    "+r"(src_width),   // %4
-    "+r"(src_height)   // %5
-  :
-  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"  // Clobber List
+void ScaleRowUp2_Linear_12_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q1}, [%0]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q0}, [%3]!                   \n"  // 12345678 (16b)
+
+      "vmovq       q2, q0                        \n"
+      "vmla.u16    q0, q1, q15                   \n"  // 3*near+far (odd)
+      "vmla.u16    q1, q2, q15                   \n"  // 3*near+far (even)
+
+      "vrshr.u16   q0, q0, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshr.u16   q1, q1, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.16     {d0, d1, d2, d3}, [%1]!       \n"  // store
+      "subs        %2, %2, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_12_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q1}, [%5]!                   \n"  // 12345678 (16b)
+
+      "vmovq       q2, q0                        \n"
+      "vmla.u16    q0, q1, q15                   \n"  // 3*near+far (odd)
+      "vmla.u16    q1, q2, q15                   \n"  // 3*near+far (even)
+
+      "vld1.16     {q2}, [%1]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q3}, [%6]!                   \n"  // 12345678 (16b)
+
+      "vmovq       q4, q2                        \n"
+      "vmla.u16    q2, q3, q15                   \n"  // 3*near+far (odd)
+      "vmla.u16    q3, q4, q15                   \n"  // 3*near+far (even)
+
+      "vmovq       q4, q2                        \n"
+      "vmovq       q5, q3                        \n"
+      "vmla.u16    q4, q0, q15                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u16    q5, q1, q15                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u16    q0, q2, q15                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u16    q1, q3, q15                   \n"  // 9 3 3 1 (2, even)
+
+      "vrshr.u16   q2, q1, #4                    \n"  // 2, even
+      "vrshr.u16   q3, q0, #4                    \n"  // 2, odd
+      "vrshr.u16   q0, q5, #4                    \n"  // 1, even
+      "vrshr.u16   q1, q4, #4                    \n"  // 1, odd
+
+      "vst2.16     {d0, d1, d2, d3}, [%2]!       \n"  // store
+      "vst2.16     {d4, d5, d6, d7}, [%3]!       \n"  // store
+      "subs        %4, %4, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5",
+        "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "vmov.u16    d31, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q1}, [%3]!                   \n"  // 12345678 (16b)
+
+      "vmovl.u16   q2, d0                        \n"  // 0123 (32b)
+      "vmovl.u16   q3, d1                        \n"  // 4567 (32b)
+      "vmovl.u16   q4, d2                        \n"  // 1234 (32b)
+      "vmovl.u16   q5, d3                        \n"  // 5678 (32b)
+
+      "vmlal.u16   q2, d2, d31                   \n"
+      "vmlal.u16   q3, d3, d31                   \n"
+      "vmlal.u16   q4, d0, d31                   \n"
+      "vmlal.u16   q5, d1, d31                   \n"
+
+      "vrshrn.u32  d0, q4, #2                    \n"
+      "vrshrn.u32  d1, q5, #2                    \n"
+      "vrshrn.u32  d2, q2, #2                    \n"
+      "vrshrn.u32  d3, q3, #2                    \n"
+
+      "vst2.16     {q0, q1}, [%1]!               \n"  // store
+      "subs        %2, %2, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "vmov.u16    d31, #3                       \n"
+      "vmov.u32    q14, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {d0}, [%0]!                   \n"  // 0123 (16b)
+      "vld1.16     {d1}, [%5]!                   \n"  // 1234 (16b)
+      "vmovl.u16   q2, d0                        \n"  // 0123 (32b)
+      "vmovl.u16   q3, d1                        \n"  // 1234 (32b)
+      "vmlal.u16   q2, d1, d31                   \n"
+      "vmlal.u16   q3, d0, d31                   \n"
+
+      "vld1.16     {d0}, [%1]!                   \n"  // 0123 (16b)
+      "vld1.16     {d1}, [%6]!                   \n"  // 1234 (16b)
+      "vmovl.u16   q4, d0                        \n"  // 0123 (32b)
+      "vmovl.u16   q5, d1                        \n"  // 1234 (32b)
+      "vmlal.u16   q4, d1, d31                   \n"
+      "vmlal.u16   q5, d0, d31                   \n"
+
+      "vmovq       q0, q4                        \n"
+      "vmovq       q1, q5                        \n"
+      "vmla.u32    q4, q2, q14                   \n"
+      "vmla.u32    q5, q3, q14                   \n"
+      "vmla.u32    q2, q0, q14                   \n"
+      "vmla.u32    q3, q1, q14                   \n"
+
+      "vrshrn.u32  d1, q4, #4                    \n"
+      "vrshrn.u32  d0, q5, #4                    \n"
+      "vrshrn.u32  d3, q2, #4                    \n"
+      "vrshrn.u32  d2, q3, #4                    \n"
+
+      "vst2.16     {d0, d1}, [%2]!               \n"  // store
+      "vst2.16     {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #8                    \n"  // 4 sample -> 8 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q14",
+        "d31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  const uint8_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "vmov.u8     d30, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 00112233 (1u1v)
+      "vld1.8      {d5}, [%3]!                   \n"  // 11223344 (1u1v)
+
+      "vmovl.u8    q0, d4                        \n"  // 00112233 (1u1v, 16b)
+      "vmovl.u8    q1, d5                        \n"  // 11223344 (1u1v, 16b)
+      "vmlal.u8    q0, d5, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u8    q1, d4, d30                   \n"  // 3*near+far (even)
+
+      "vrshrn.u16  d1, q0, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u16  d0, q1, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.16     {d0, d1}, [%1]!               \n"  // store
+      "subs        %2, %2, #8                    \n"  // 4 uv -> 8 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "d30"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 2;
+  const uint8_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+      "vmov.u8     d28, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 00112233 (1u1v)
+      "vld1.8      {d5}, [%5]!                   \n"  // 11223344 (1u1v)
+
+      "vmovl.u8    q0, d4                        \n"  // 00112233 (1u1v, 16b)
+      "vmovl.u8    q1, d5                        \n"  // 11223344 (1u1v, 16b)
+      "vmlal.u8    q0, d5, d28                   \n"  // 3*near+far (1, odd)
+      "vmlal.u8    q1, d4, d28                   \n"  // 3*near+far (1, even)
+
+      "vld1.8      {d8}, [%1]!                   \n"  // 00112233 (1u1v)
+      "vld1.8      {d9}, [%6]!                   \n"  // 11223344 (1u1v)
+
+      "vmovl.u8    q2, d8                        \n"  // 00112233 (1u1v, 16b)
+      "vmovl.u8    q3, d9                        \n"  // 11223344 (1u1v, 16b)
+      "vmlal.u8    q2, d9, d28                   \n"  // 3*near+far (2, odd)
+      "vmlal.u8    q3, d8, d28                   \n"  // 3*near+far (2, even)
+
+      // e  o
+      // q1 q0
+      // q3 q2
+
+      "vmovq       q4, q2                        \n"
+      "vmovq       q5, q3                        \n"
+      "vmla.u16    q4, q0, q15                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u16    q5, q1, q15                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u16    q0, q2, q15                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u16    q1, q3, q15                   \n"  // 9 3 3 1 (2, even)
+
+      // e  o
+      // q5 q4
+      // q1 q0
+
+      "vrshrn.u16  d2, q1, #4                    \n"  // 2, even
+      "vrshrn.u16  d3, q0, #4                    \n"  // 2, odd
+      "vrshrn.u16  d0, q5, #4                    \n"  // 1, even
+      "vrshrn.u16  d1, q4, #4                    \n"  // 1, odd
+
+      "vst2.16     {d0, d1}, [%2]!               \n"  // store
+      "vst2.16     {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #8                    \n"  // 4 uv -> 8 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "d28",
+        "q15"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width) {
+  const uint16_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "vmov.u16    d30, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // 00112233 (1u1v, 16)
+      "vld1.16     {q1}, [%3]!                   \n"  // 11223344 (1u1v, 16)
+
+      "vmovl.u16   q2, d0                        \n"  // 0011 (1u1v, 32b)
+      "vmovl.u16   q3, d2                        \n"  // 1122 (1u1v, 32b)
+      "vmovl.u16   q4, d1                        \n"  // 2233 (1u1v, 32b)
+      "vmovl.u16   q5, d3                        \n"  // 3344 (1u1v, 32b)
+      "vmlal.u16   q2, d2, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u16   q3, d0, d30                   \n"  // 3*near+far (even)
+      "vmlal.u16   q4, d3, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u16   q5, d1, d30                   \n"  // 3*near+far (even)
+
+      "vrshrn.u32  d1, q2, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u32  d0, q3, #2                    \n"  // 3/4*near+1/4*far (even)
+      "vrshrn.u32  d3, q4, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u32  d2, q5, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.32     {d0, d1}, [%1]!               \n"  // store
+      "vst2.32     {d2, d3}, [%1]!               \n"  // store
+      "subs        %2, %2, #8                    \n"  // 4 uv -> 8 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5",
+        "d30"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 2;
+  const uint16_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "vmov.u16    d30, #3                       \n"
+      "vmov.u32    q14, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0]!                   \n"  // 0011 (1u1v)
+      "vld1.8      {d1}, [%5]!                   \n"  // 1122 (1u1v)
+      "vmovl.u16   q2, d0                        \n"  // 0011 (1u1v, 32b)
+      "vmovl.u16   q3, d1                        \n"  // 1122 (1u1v, 32b)
+      "vmlal.u16   q2, d1, d30                   \n"  // 3*near+far (1, odd)
+      "vmlal.u16   q3, d0, d30                   \n"  // 3*near+far (1, even)
+
+      "vld1.8      {d0}, [%1]!                   \n"  // 0011 (1u1v)
+      "vld1.8      {d1}, [%6]!                   \n"  // 1122 (1u1v)
+      "vmovl.u16   q4, d0                        \n"  // 0011 (1u1v, 32b)
+      "vmovl.u16   q5, d1                        \n"  // 1122 (1u1v, 32b)
+      "vmlal.u16   q4, d1, d30                   \n"  // 3*near+far (2, odd)
+      "vmlal.u16   q5, d0, d30                   \n"  // 3*near+far (2, even)
+
+      "vmovq       q0, q4                        \n"
+      "vmovq       q1, q5                        \n"
+      "vmla.u32    q4, q2, q14                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u32    q5, q3, q14                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u32    q2, q0, q14                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u32    q3, q1, q14                   \n"  // 9 3 3 1 (2, even)
+
+      "vrshrn.u32  d1, q4, #4                    \n"  // 1, odd
+      "vrshrn.u32  d0, q5, #4                    \n"  // 1, even
+      "vrshrn.u32  d3, q2, #4                    \n"  // 2, odd
+      "vrshrn.u32  d2, q3, #4                    \n"  // 2, even
+
+      "vst2.32     {d0, d1}, [%2]!               \n"  // store
+      "vst2.32     {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #4                    \n"  // 2 uv -> 4 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q14",
+        "d30"  // Clobber List
+  );
+}
+
+// Add a row of bytes to a row of shorts.  Used for box filter.
+// Reads 16 bytes and accumulates to 16 shorts at a time.
+void ScaleAddRow_NEON(const uint8_t* src_ptr,
+                      uint16_t* dst_ptr,
+                      int src_width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {q1, q2}, [%1]                \n"  // load accumulator
+      "vld1.8      {q0}, [%0]!                   \n"  // load 16 bytes
+      "vaddw.u8    q2, q2, d1                    \n"  // add
+      "vaddw.u8    q1, q1, d0                    \n"
+      "vst1.16     {q1, q2}, [%1]!               \n"  // store accumulator
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(src_width)  // %2
+      :
+      : "memory", "cc", "q0", "q1", "q2"  // Clobber List
   );
 }
 
 // TODO(Yang Zhang): Investigate less load instructions for
 // the x/dx stepping
-#define LOAD2_DATA8_LANE(n)                                    \
-    "lsr        %5, %3, #16                    \n"             \
-    "add        %6, %1, %5                     \n"             \
-    "add        %3, %3, %4                     \n"             \
-    MEMACCESS(6)                                               \
-    "vld2.8     {d6["#n"], d7["#n"]}, [%6]     \n"
-
-// The NEON version mimics this formula:
-// #define BLENDER(a, b, f) (uint8)((int)(a) +
-//    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
-
-void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
-                          int dst_width, int x, int dx) {
+#define LOAD2_DATA8_LANE(n)                      \
+  "lsr        %5, %3, #16                    \n" \
+  "add        %6, %1, %5                     \n" \
+  "add        %3, %3, %4                     \n" \
+  "vld2.8     {d6[" #n "], d7[" #n "]}, [%6] \n"
+
+// The NEON version mimics this formula (from row_common.cc):
+// #define BLENDER(a, b, f) (uint8_t)((int)(a) +
+//    ((((int)((f)) * ((int)(b) - (int)(a))) + 0x8000) >> 16))
+
+void ScaleFilterCols_NEON(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          int dst_width,
+                          int x,
+                          int dx) {
   int dx_offset[4] = {0, 1, 2, 3};
   int* tmp = dx_offset;
-  const uint8* src_tmp = src_ptr;
+  const uint8_t* src_tmp = src_ptr;
   asm volatile (
-    "vdup.32    q0, %3                         \n"  // x
-    "vdup.32    q1, %4                         \n"  // dx
-    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
-    "vshl.i32   q3, q1, #2                     \n"  // 4 * dx
-    "vmul.s32   q1, q1, q2                     \n"
+      "vdup.32     q0, %3                        \n"  // x
+      "vdup.32     q1, %4                        \n"  // dx
+      "vld1.32     {q2}, [%5]                    \n"  // 0 1 2 3
+      "vshl.i32    q3, q1, #2                    \n"  // 4 * dx
+      "vmul.s32    q1, q1, q2                    \n"
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "vadd.s32   q1, q1, q0                     \n"
+      "vadd.s32    q1, q1, q0                    \n"
     // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
-    "vadd.s32   q2, q1, q3                     \n"
-    "vshl.i32   q0, q3, #1                     \n"  // 8 * dx
-  "1:                                          \n"
+      "vadd.s32    q2, q1, q3                    \n"
+      "vshl.i32    q0, q3, #1                    \n"  // 8 * dx
+      "1:                                        \n"
     LOAD2_DATA8_LANE(0)
     LOAD2_DATA8_LANE(1)
     LOAD2_DATA8_LANE(2)
@@ -601,28 +1044,27 @@ void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
     LOAD2_DATA8_LANE(5)
     LOAD2_DATA8_LANE(6)
     LOAD2_DATA8_LANE(7)
-    "vmov       q10, q1                        \n"
-    "vmov       q11, q2                        \n"
-    "vuzp.16    q10, q11                       \n"
-    "vmovl.u8   q8, d6                         \n"
-    "vmovl.u8   q9, d7                         \n"
-    "vsubl.s16  q11, d18, d16                  \n"
-    "vsubl.s16  q12, d19, d17                  \n"
-    "vmovl.u16  q13, d20                       \n"
-    "vmovl.u16  q10, d21                       \n"
-    "vmul.s32   q11, q11, q13                  \n"
-    "vmul.s32   q12, q12, q10                  \n"
-    "vrshrn.s32  d18, q11, #16                 \n"
-    "vrshrn.s32  d19, q12, #16                 \n"
-    "vadd.s16   q8, q8, q9                     \n"
-    "vmovn.s16  d6, q8                         \n"
-
-    MEMACCESS(0)
-    "vst1.8     {d6}, [%0]!                    \n"  // store pixels
-    "vadd.s32   q1, q1, q0                     \n"
-    "vadd.s32   q2, q2, q0                     \n"
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    "bgt        1b                             \n"
+      "vmov        q10, q1                       \n"
+      "vmov        q11, q2                       \n"
+      "vuzp.16     q10, q11                      \n"
+      "vmovl.u8    q8, d6                        \n"
+      "vmovl.u8    q9, d7                        \n"
+      "vsubl.s16   q11, d18, d16                 \n"
+      "vsubl.s16   q12, d19, d17                 \n"
+      "vmovl.u16   q13, d20                      \n"
+      "vmovl.u16   q10, d21                      \n"
+      "vmul.s32    q11, q11, q13                 \n"
+      "vmul.s32    q12, q12, q10                 \n"
+      "vrshrn.s32  d18, q11, #16                 \n"
+      "vrshrn.s32  d19, q12, #16                 \n"
+      "vadd.s16    q8, q8, q9                    \n"
+      "vmovn.s16   d6, q8                        \n"
+
+      "vst1.8      {d6}, [%0]!                   \n"  // store pixels
+      "vadd.s32    q1, q1, q0                    \n"
+      "vadd.s32    q2, q2, q0                    \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "bgt         1b                            \n"
   : "+r"(dst_ptr),          // %0
     "+r"(src_ptr),          // %1
     "+r"(dst_width),        // %2
@@ -639,336 +1081,310 @@ void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
 #undef LOAD2_DATA8_LANE
 
 // 16x2 -> 16x1
-void ScaleFilterRows_NEON(uint8* dst_ptr,
-                          const uint8* src_ptr, ptrdiff_t src_stride,
-                          int dst_width, int source_y_fraction) {
-  asm volatile (
-    "cmp          %4, #0                       \n"
-    "beq          100f                         \n"
-    "add          %2, %1                       \n"
-    "cmp          %4, #64                      \n"
-    "beq          75f                          \n"
-    "cmp          %4, #128                     \n"
-    "beq          50f                          \n"
-    "cmp          %4, #192                     \n"
-    "beq          25f                          \n"
-
-    "vdup.8       d5, %4                       \n"
-    "rsb          %4, #256                     \n"
-    "vdup.8       d4, %4                       \n"
-    // General purpose row blend.
-  "1:                                          \n"
-    MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
-    MEMACCESS(2)
-    "vld1.8       {q1}, [%2]!                  \n"
-    "subs         %3, %3, #16                  \n"
-    "vmull.u8     q13, d0, d4                  \n"
-    "vmull.u8     q14, d1, d4                  \n"
-    "vmlal.u8     q13, d2, d5                  \n"
-    "vmlal.u8     q14, d3, d5                  \n"
-    "vrshrn.u16   d0, q13, #8                  \n"
-    "vrshrn.u16   d1, q14, #8                  \n"
-    MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          1b                           \n"
-    "b            99f                          \n"
-
-    // Blend 25 / 75.
-  "25:                                         \n"
-    MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
-    MEMACCESS(2)
-    "vld1.8       {q1}, [%2]!                  \n"
-    "subs         %3, %3, #16                  \n"
-    "vrhadd.u8    q0, q1                       \n"
-    "vrhadd.u8    q0, q1                       \n"
-    MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          25b                          \n"
-    "b            99f                          \n"
-
-    // Blend 50 / 50.
-  "50:                                         \n"
-    MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
-    MEMACCESS(2)
-    "vld1.8       {q1}, [%2]!                  \n"
-    "subs         %3, %3, #16                  \n"
-    "vrhadd.u8    q0, q1                       \n"
-    MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          50b                          \n"
-    "b            99f                          \n"
-
-    // Blend 75 / 25.
-  "75:                                         \n"
-    MEMACCESS(1)
-    "vld1.8       {q1}, [%1]!                  \n"
-    MEMACCESS(2)
-    "vld1.8       {q0}, [%2]!                  \n"
-    "subs         %3, %3, #16                  \n"
-    "vrhadd.u8    q0, q1                       \n"
-    "vrhadd.u8    q0, q1                       \n"
-    MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          75b                          \n"
-    "b            99f                          \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-  "100:                                        \n"
-    MEMACCESS(1)
-    "vld1.8       {q0}, [%1]!                  \n"
-    "subs         %3, %3, #16                  \n"
-    MEMACCESS(0)
-    "vst1.8       {q0}, [%0]!                  \n"
-    "bgt          100b                         \n"
-
-  "99:                                         \n"
-    MEMACCESS(0)
-    "vst1.8       {d1[7]}, [%0]                \n"
-  : "+r"(dst_ptr),          // %0
-    "+r"(src_ptr),          // %1
-    "+r"(src_stride),       // %2
-    "+r"(dst_width),        // %3
-    "+r"(source_y_fraction) // %4
-  :
-  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
-  );
+void ScaleFilterRows_NEON(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          int dst_width,
+                          int source_y_fraction) {
+  asm volatile(
+      "cmp         %4, #0                        \n"
+      "beq         100f                          \n"
+      "add         %2, %1                        \n"
+      "cmp         %4, #64                       \n"
+      "beq         75f                           \n"
+      "cmp         %4, #128                      \n"
+      "beq         50f                           \n"
+      "cmp         %4, #192                      \n"
+      "beq         25f                           \n"
+
+      "vdup.8      d5, %4                        \n"
+      "rsb         %4, #256                      \n"
+      "vdup.8      d4, %4                        \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vmull.u8    q13, d0, d4                   \n"
+      "vmull.u8    q14, d1, d4                   \n"
+      "vmlal.u8    q13, d2, d5                   \n"
+      "vmlal.u8    q14, d3, d5                   \n"
+      "vrshrn.u16  d0, q13, #8                   \n"
+      "vrshrn.u16  d1, q14, #8                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 25 / 75.
+      "25:                                       \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         25b                           \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 75 / 25.
+      "75:                                       \n"
+      "vld1.8      {q1}, [%1]!                   \n"
+      "vld1.8      {q0}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         75b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         100b                          \n"
+
+      "99:                                       \n"
+      "vst1.8      {d1[7]}, [%0]                 \n"
+      : "+r"(dst_ptr),           // %0
+        "+r"(src_ptr),           // %1
+        "+r"(src_stride),        // %2
+        "+r"(dst_width),         // %3
+        "+r"(source_y_fraction)  // %4
+      :
+      : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc");
 }
 
-void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    // load even pixels into q0, odd into q1
-    MEMACCESS(0)
-    "vld2.32    {q0, q1}, [%0]!                \n"
-    MEMACCESS(0)
-    "vld2.32    {q2, q3}, [%0]!                \n"
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    MEMACCESS(1)
-    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
-    MEMACCESS(1)
-    "vst1.8     {q3}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst),              // %1
-    "+r"(dst_width)         // %2
-  :
-  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+void ScaleARGBRowDown2_NEON(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst,
+                            int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.32     {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.32     {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "vmov        q2, q1                        \n"  // load next 8 ARGB
+      "vst2.32     {q2, q3}, [%1]!               \n"  // store odd pixels
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
   );
 }
 
-void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
-                                  uint8* dst_argb, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
-    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #1                     \n"
-    "vrshrn.u16 d2, q2, #1                     \n"
-    "vrshrn.u16 d3, q3, #1                     \n"
-    MEMACCESS(1)
-    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"
-    "bgt       1b                              \n"
-  : "+r"(src_argb),         // %0
-    "+r"(dst_argb),         // %1
-    "+r"(dst_width)         // %2
-  :
-  : "memory", "cc", "q0", "q1", "q2", "q3"     // Clobber List
+//  46:  f964 018d   vld4.32  {d16,d18,d20,d22}, [r4]!
+//  4a:  3e04        subs  r6, #4
+//  4c:  f964 118d   vld4.32  {d17,d19,d21,d23}, [r4]!
+//  50:  ef64 21f4   vorr  q9, q10, q10
+//  54:  f942 038d   vst2.32  {d16-d19}, [r2]!
+//  58:  d1f5        bne.n  46 <ScaleARGBRowDown2_C+0x46>
+
+void ScaleARGBRowDown2Linear_NEON(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_argb,
+                                  int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.32     {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.32     {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "vrhadd.u8   q0, q0, q1                    \n"  // rounding half add
+      "vrhadd.u8   q1, q2, q3                    \n"  // rounding half add
+      "vst2.32     {q0, q1}, [%1]!               \n"
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
   );
 }
 
-void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst, int dst_width) {
-  asm volatile (
-    // change the stride to row 2 pointer
-    "add        %1, %1, %0                     \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    MEMACCESS(0)
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
-    MEMACCESS(1)
-    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
-    MEMACCESS(1)
-    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
-    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
-    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #2                     \n"
-    "vrshrn.u16 d2, q2, #2                     \n"
-    "vrshrn.u16 d3, q3, #2                     \n"
-    MEMACCESS(2)
-    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
-    "bgt        1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(src_stride),       // %1
-    "+r"(dst),              // %2
-    "+r"(dst_width)         // %3
-  :
-  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
-  );
+void ScaleARGBRowDown2Box_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst,
+                               int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vpaddl.u8   q3, q3                        \n"  // A 16 bytes -> 8 shorts.
+      "vld4.8      {d16, d18, d20, d22}, [%1]!   \n"  // load 8 more ARGB
+      "vld4.8      {d17, d19, d21, d23}, [%1]!   \n"  // load last 8 ARGB
+      "vpadal.u8   q0, q8                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q9                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q10                       \n"  // R 16 bytes -> 8 shorts.
+      "vpadal.u8   q3, q11                       \n"  // A 16 bytes -> 8 shorts.
+      "vrshrn.u16  d0, q0, #2                    \n"  // round and pack to bytes
+      "vrshrn.u16  d1, q1, #2                    \n"
+      "vrshrn.u16  d2, q2, #2                    \n"
+      "vrshrn.u16  d3, q3, #2                    \n"
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
 }
 
 // Reads 4 pixels at a time.
 // Alignment requirement: src_argb 4 byte aligned.
-void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
-                               int src_stepx, uint8* dst_argb, int dst_width) {
-  asm volatile (
-    "mov        r12, %3, lsl #2                \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.32    {d0[0]}, [%0], r12             \n"
-    MEMACCESS(0)
-    "vld1.32    {d0[1]}, [%0], r12             \n"
-    MEMACCESS(0)
-    "vld1.32    {d1[0]}, [%0], r12             \n"
-    MEMACCESS(0)
-    "vld1.32    {d1[1]}, [%0], r12             \n"
-    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
-    MEMACCESS(1)
-    "vst1.8     {q0}, [%1]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_argb),    // %0
-    "+r"(dst_argb),    // %1
-    "+r"(dst_width)    // %2
-  : "r"(src_stepx)     // %3
-  : "memory", "cc", "r12", "q0"
-  );
+void ScaleARGBRowDownEven_NEON(const uint8_t* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8_t* dst_argb,
+                               int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "mov         r12, %3, lsl #2               \n"
+      "1:                                        \n"
+      "vld1.32     {d0[0]}, [%0], r12            \n"
+      "vld1.32     {d0[1]}, [%0], r12            \n"
+      "vld1.32     {d1[0]}, [%0], r12            \n"
+      "vld1.32     {d1[1]}, [%0], r12            \n"
+      "subs        %2, %2, #4                    \n"  // 4 pixels per loop.
+      "vst1.8      {q0}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(dst_width)  // %2
+      : "r"(src_stepx)   // %3
+      : "memory", "cc", "r12", "q0");
 }
 
 // Reads 4 pixels at a time.
 // Alignment requirement: src_argb 4 byte aligned.
-void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+void ScaleARGBRowDownEvenBox_NEON(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
                                   int src_stepx,
-                                  uint8* dst_argb, int dst_width) {
-  asm volatile (
-    "mov        r12, %4, lsl #2                \n"
-    "add        %1, %1, %0                     \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
-    MEMACCESS(1)
-    "vld1.8     {d1}, [%1], r12                \n"
-    MEMACCESS(0)
-    "vld1.8     {d2}, [%0], r12                \n"
-    MEMACCESS(1)
-    "vld1.8     {d3}, [%1], r12                \n"
-    MEMACCESS(0)
-    "vld1.8     {d4}, [%0], r12                \n"
-    MEMACCESS(1)
-    "vld1.8     {d5}, [%1], r12                \n"
-    MEMACCESS(0)
-    "vld1.8     {d6}, [%0], r12                \n"
-    MEMACCESS(1)
-    "vld1.8     {d7}, [%1], r12                \n"
-    "vaddl.u8   q0, d0, d1                     \n"
-    "vaddl.u8   q1, d2, d3                     \n"
-    "vaddl.u8   q2, d4, d5                     \n"
-    "vaddl.u8   q3, d6, d7                     \n"
-    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
-    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
-    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
-    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
-    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
-    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
-    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
-    MEMACCESS(2)
-    "vst1.8     {q0}, [%2]!                    \n"
-    "bgt        1b                             \n"
-  : "+r"(src_argb),    // %0
-    "+r"(src_stride),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(dst_width)    // %3
-  : "r"(src_stepx)     // %4
-  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
-  );
+                                  uint8_t* dst_argb,
+                                  int dst_width) {
+  asm volatile(
+      "mov         r12, %4, lsl #2               \n"
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0], r12               \n"  // 4 2x2 blocks -> 2x1
+      "vld1.8      {d1}, [%1], r12               \n"
+      "vld1.8      {d2}, [%0], r12               \n"
+      "vld1.8      {d3}, [%1], r12               \n"
+      "vld1.8      {d4}, [%0], r12               \n"
+      "vld1.8      {d5}, [%1], r12               \n"
+      "vld1.8      {d6}, [%0], r12               \n"
+      "vld1.8      {d7}, [%1], r12               \n"
+      "vaddl.u8    q0, d0, d1                    \n"
+      "vaddl.u8    q1, d2, d3                    \n"
+      "vaddl.u8    q2, d4, d5                    \n"
+      "vaddl.u8    q3, d6, d7                    \n"
+      "vswp.8      d1, d2                        \n"  // ab_cd -> ac_bd
+      "vswp.8      d5, d6                        \n"  // ef_gh -> eg_fh
+      "vadd.u16    q0, q0, q1                    \n"  // (a+b)_(c+d)
+      "vadd.u16    q2, q2, q3                    \n"  // (e+f)_(g+h)
+      "vrshrn.u16  d0, q0, #2                    \n"  // first 2 pixels.
+      "vrshrn.u16  d1, q2, #2                    \n"  // next 2 pixels.
+      "subs        %3, %3, #4                    \n"  // 4 pixels per loop.
+      "vst1.8      {q0}, [%2]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(src_stepx)     // %4
+      : "memory", "cc", "r12", "q0", "q1", "q2", "q3");
 }
 
 // TODO(Yang Zhang): Investigate less load instructions for
 // the x/dx stepping
-#define LOAD1_DATA32_LANE(dn, n)                               \
-    "lsr        %5, %3, #16                    \n"             \
-    "add        %6, %1, %5, lsl #2             \n"             \
-    "add        %3, %3, %4                     \n"             \
-    MEMACCESS(6)                                               \
-    "vld1.32    {"#dn"["#n"]}, [%6]            \n"
-
-void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx) {
+#define LOAD1_DATA32_LANE(dn, n)                 \
+  "lsr        %5, %3, #16                    \n" \
+  "add        %6, %1, %5, lsl #2             \n" \
+  "add        %3, %3, %4                     \n" \
+  "vld1.32    {" #dn "[" #n "]}, [%6]        \n"
+
+void ScaleARGBCols_NEON(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx) {
   int tmp;
-  const uint8* src_tmp = src_argb;
-  asm volatile (
-  "1:                                          \n"
-    LOAD1_DATA32_LANE(d0, 0)
-    LOAD1_DATA32_LANE(d0, 1)
-    LOAD1_DATA32_LANE(d1, 0)
-    LOAD1_DATA32_LANE(d1, 1)
-    LOAD1_DATA32_LANE(d2, 0)
-    LOAD1_DATA32_LANE(d2, 1)
-    LOAD1_DATA32_LANE(d3, 0)
-    LOAD1_DATA32_LANE(d3, 1)
-
-    MEMACCESS(0)
-    "vst1.32     {q0, q1}, [%0]!               \n"  // store pixels
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    "bgt        1b                             \n"
-  : "+r"(dst_argb),   // %0
-    "+r"(src_argb),   // %1
-    "+r"(dst_width),  // %2
-    "+r"(x),          // %3
-    "+r"(dx),         // %4
-    "=&r"(tmp),       // %5
-    "+r"(src_tmp)     // %6
-  :
-  : "memory", "cc", "q0", "q1"
-  );
+  const uint8_t* src_tmp = src_argb;
+  asm volatile(
+      "1:                                        \n"
+      // clang-format off
+      LOAD1_DATA32_LANE(d0, 0)
+      LOAD1_DATA32_LANE(d0, 1)
+      LOAD1_DATA32_LANE(d1, 0)
+      LOAD1_DATA32_LANE(d1, 1)
+      LOAD1_DATA32_LANE(d2, 0)
+      LOAD1_DATA32_LANE(d2, 1)
+      LOAD1_DATA32_LANE(d3, 0)
+      LOAD1_DATA32_LANE(d3, 1)
+      // clang-format on
+      "vst1.32     {q0, q1}, [%0]!               \n"  // store pixels
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "bgt         1b                            \n"
+      : "+r"(dst_argb),   // %0
+        "+r"(src_argb),   // %1
+        "+r"(dst_width),  // %2
+        "+r"(x),          // %3
+        "+r"(dx),         // %4
+        "=&r"(tmp),       // %5
+        "+r"(src_tmp)     // %6
+      :
+      : "memory", "cc", "q0", "q1");
 }
 
 #undef LOAD1_DATA32_LANE
 
 // TODO(Yang Zhang): Investigate less load instructions for
 // the x/dx stepping
-#define LOAD2_DATA32_LANE(dn1, dn2, n)                         \
-    "lsr        %5, %3, #16                           \n"      \
-    "add        %6, %1, %5, lsl #2                    \n"      \
-    "add        %3, %3, %4                            \n"      \
-    MEMACCESS(6)                                               \
-    "vld2.32    {"#dn1"["#n"], "#dn2"["#n"]}, [%6]    \n"
-
-void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
-                              int dst_width, int x, int dx) {
+#define LOAD2_DATA32_LANE(dn1, dn2, n)                       \
+  "lsr        %5, %3, #16                                \n" \
+  "add        %6, %1, %5, lsl #2                         \n" \
+  "add        %3, %3, %4                                 \n" \
+  "vld2.32    {" #dn1 "[" #n "], " #dn2 "[" #n "]}, [%6] \n"
+
+void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
+                              const uint8_t* src_argb,
+                              int dst_width,
+                              int x,
+                              int dx) {
   int dx_offset[4] = {0, 1, 2, 3};
   int* tmp = dx_offset;
-  const uint8* src_tmp = src_argb;
+  const uint8_t* src_tmp = src_argb;
   asm volatile (
-    "vdup.32    q0, %3                         \n"  // x
-    "vdup.32    q1, %4                         \n"  // dx
-    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
-    "vshl.i32   q9, q1, #2                     \n"  // 4 * dx
-    "vmul.s32   q1, q1, q2                     \n"
-    "vmov.i8    q3, #0x7f                      \n"  // 0x7F
-    "vmov.i16   q15, #0x7f                     \n"  // 0x7F
+      "vdup.32     q0, %3                        \n"  // x
+      "vdup.32     q1, %4                        \n"  // dx
+      "vld1.32     {q2}, [%5]                    \n"  // 0 1 2 3
+      "vshl.i32    q9, q1, #2                    \n"  // 4 * dx
+      "vmul.s32    q1, q1, q2                    \n"
+      "vmov.i8     q3, #0x7f                     \n"  // 0x7F
+      "vmov.i16    q15, #0x7f                    \n"  // 0x7F
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "vadd.s32   q8, q1, q0                     \n"
-  "1:                                          \n"
+      "vadd.s32    q8, q1, q0                    \n"
+      "1:                                        \n"
     // d0, d1: a
     // d2, d3: b
     LOAD2_DATA32_LANE(d0, d2, 0)
@@ -993,7 +1409,6 @@ void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
     "vshrn.i16   d0, q11, #7                   \n"
     "vshrn.i16   d1, q12, #7                   \n"
 
-    MEMACCESS(0)
     "vst1.32     {d0, d1}, [%0]!               \n"  // store pixels
     "vadd.s32    q8, q8, q9                    \n"
     "subs        %2, %2, #4                    \n"  // 4 processed per loop
@@ -1013,6 +1428,64 @@ void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
 
 #undef LOAD2_DATA32_LANE
 
+void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "vld2.8      {d0, d2}, [%0]!               \n"  // load 8 UV pixels.
+      "vld2.8      {d1, d3}, [%0]!               \n"  // load next 8 UV
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vpaddl.u8   q0, q0                        \n"  // U 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // V 16 bytes -> 8 shorts.
+      "vld2.8      {d16, d18}, [%1]!             \n"  // load 8 more UV
+      "vld2.8      {d17, d19}, [%1]!             \n"  // load last 8 UV
+      "vpadal.u8   q0, q8                        \n"  // U 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q9                        \n"  // V 16 bytes -> 8 shorts.
+      "vrshrn.u16  d0, q0, #2                    \n"  // round and pack to bytes
+      "vrshrn.u16  d1, q1, #2                    \n"
+      "vst2.8      {d0, d1}, [%2]!               \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q8", "q9");
+}
+
+// Reads 4 pixels at a time.
+void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             int src_stepx,  // pixel step
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src1_ptr = src_ptr + src_stepx * 2;
+  const uint8_t* src2_ptr = src_ptr + src_stepx * 4;
+  const uint8_t* src3_ptr = src_ptr + src_stepx * 6;
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {d0[0]}, [%0], %6             \n"
+      "vld1.16     {d0[1]}, [%1], %6             \n"
+      "vld1.16     {d0[2]}, [%2], %6             \n"
+      "vld1.16     {d0[3]}, [%3], %6             \n"
+      "subs        %5, %5, #4                    \n"  // 4 pixels per loop.
+      "vst1.8      {d0}, [%4]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),      // %0
+        "+r"(src1_ptr),     // %1
+        "+r"(src2_ptr),     // %2
+        "+r"(src3_ptr),     // %3
+        "+r"(dst_ptr),      // %4
+        "+r"(dst_width)     // %5
+      : "r"(src_stepx * 8)  // %6
+      : "memory", "cc", "d0");
+}
+
 #endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/scale_neon64.cc b/media/libyuv/source/scale_neon64.cc
index ff277f26ff..9f9636e646 100644
--- a/media/libyuv/source/scale_neon64.cc
+++ b/media/libyuv/source/scale_neon64.cc
@@ -8,8 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "libyuv/scale.h"
 #include "libyuv/row.h"
+#include "libyuv/scale.h"
 #include "libyuv/scale_row.h"
 
 #ifdef __cplusplus
@@ -21,592 +21,1059 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 // Read 32x1 throw away even pixels, and write 16x1.
-void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    // load even pixels into v0, odd into v1
-    MEMACCESS(0)
-    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"
-    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
-    MEMACCESS(1)
-    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
-    "b.gt       1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst),              // %1
-    "+r"(dst_width)         // %2
-  :
-  : "v0", "v1"              // Clobber List
+void ScaleRowDown2_NEON(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      // load even pixels into v0, odd into v1
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v1.16b}, [%1], #16           \n"  // store odd pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "v0", "v1"  // Clobber List
   );
 }
 
 // Read 32x1 average down and write 16x1.
-void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b,v1.16b}, [%0], #32     \n"  // load pixels and post inc
-    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
-    "uaddlp     v0.8h, v0.16b                  \n"  // add adjacent
-    "uaddlp     v1.8h, v1.16b                  \n"
-    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
-    "rshrn2     v0.16b, v1.8h, #1              \n"
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst),              // %1
-    "+r"(dst_width)         // %2
-  :
-  : "v0", "v1"     // Clobber List
+void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst,
+                              int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      // load even pixels into v0, odd into v1
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"  // rounding half add
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v0.16b}, [%1], #16           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "v0", "v1"  // Clobber List
   );
 }
 
 // Read 32x2 average down and write 16x1.
-void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst, int dst_width) {
-  asm volatile (
-    // change the stride to row 2 pointer
-    "add        %1, %1, %0                     \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.16b,v1.16b}, [%0], #32    \n"  // load row 1 and post inc
-    MEMACCESS(1)
-    "ld1        {v2.16b, v3.16b}, [%1], #32    \n"  // load row 2 and post inc
-    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
-    "uaddlp     v0.8h, v0.16b                  \n"  // row 1 add adjacent
-    "uaddlp     v1.8h, v1.16b                  \n"
-    "uadalp     v0.8h, v2.16b                  \n"  // row 2 add adjacent + row1
-    "uadalp     v1.8h, v3.16b                  \n"
-    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
-    "rshrn2     v0.16b, v1.8h, #2              \n"
-    MEMACCESS(2)
-    "st1        {v0.16b}, [%2], #16            \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(src_stride),       // %1
-    "+r"(dst),              // %2
-    "+r"(dst_width)         // %3
-  :
-  : "v0", "v1", "v2", "v3"     // Clobber List
+void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst,
+                           int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "ld1         {v0.16b, v1.16b}, [%0], #32   \n"  // load row 1 and post inc
+      "ld1         {v2.16b, v3.16b}, [%1], #32   \n"  // load row 2 and post inc
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "uaddlp      v0.8h, v0.16b                 \n"  // row 1 add adjacent
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uaddlp      v1.8h, v1.16b                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v0.8h, v2.16b                 \n"  // += row 2 add adjacent
+      "uadalp      v1.8h, v3.16b                 \n"
+      "rshrn       v0.8b, v0.8h, #2              \n"  // round and pack
+      "rshrn2      v0.16b, v1.8h, #2             \n"
+      "st1         {v0.16b}, [%2], #16           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld4     {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32          \n"  // src line 0
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-    MEMACCESS(1)
-    "st1     {v2.8b}, [%1], #8                 \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width)         // %2
-  :
-  : "v0", "v1", "v2", "v3", "memory", "cc"
-  );
+void ScaleRowDown4_NEON(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst_ptr,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v2.8b}, [%1], #8             \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "v0", "v1", "v2", "v3", "memory", "cc");
 }
 
-void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  const uint8* src_ptr1 = src_ptr + src_stride;
-  const uint8* src_ptr2 = src_ptr + src_stride * 2;
-  const uint8* src_ptr3 = src_ptr + src_stride * 3;
-asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1     {v0.16b}, [%0], #16               \n"   // load up 16x4
-    MEMACCESS(3)
-    "ld1     {v1.16b}, [%2], #16               \n"
-    MEMACCESS(4)
-    "ld1     {v2.16b}, [%3], #16               \n"
-    MEMACCESS(5)
-    "ld1     {v3.16b}, [%4], #16               \n"
-    "subs    %w5, %w5, #4                      \n"
-    "uaddlp  v0.8h, v0.16b                     \n"
-    "uadalp  v0.8h, v1.16b                     \n"
-    "uadalp  v0.8h, v2.16b                     \n"
-    "uadalp  v0.8h, v3.16b                     \n"
-    "addp    v0.8h, v0.8h, v0.8h               \n"
-    "rshrn   v0.8b, v0.8h, #4                  \n"   // divide by 16 w/rounding
-    MEMACCESS(1)
-    "st1    {v0.s}[0], [%1], #4                \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_ptr),   // %0
-    "+r"(dst_ptr),   // %1
-    "+r"(src_ptr1),  // %2
-    "+r"(src_ptr2),  // %3
-    "+r"(src_ptr3),  // %4
-    "+r"(dst_width)  // %5
-  :
-  : "v0", "v1", "v2", "v3", "memory", "cc"
-  );
+void ScaleRowDown4Box_NEON(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_ptr,
+                           int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  const uint8_t* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8_t* src_ptr3 = src_ptr + src_stride * 3;
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load up 16x4
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "ld1         {v2.16b}, [%3], #16           \n"
+      "ld1         {v3.16b}, [%4], #16           \n"
+      "subs        %w5, %w5, #4                  \n"
+      "uaddlp      v0.8h, v0.16b                 \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uadalp      v0.8h, v1.16b                 \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "uadalp      v0.8h, v2.16b                 \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "uadalp      v0.8h, v3.16b                 \n"
+      "prfm        pldl1keep, [%4, 448]          \n"
+      "addp        v0.8h, v0.8h, v0.8h           \n"
+      "rshrn       v0.8b, v0.8h, #4              \n"  // divide by 16 w/rounding
+      "st1         {v0.s}[0], [%1], #4           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(src_ptr1),  // %2
+        "+r"(src_ptr2),  // %3
+        "+r"(src_ptr3),  // %4
+        "+r"(dst_width)  // %5
+      :
+      : "v0", "v1", "v2", "v3", "memory", "cc");
 }
 
 // Down scale from 4 to 3 pixels. Use the neon multilane read/write
 // to load up the every 4th pixel into a 4 different registers.
 // Point samples 32 pixels to 24 pixels.
-void ScaleRowDown34_NEON(const uint8* src_ptr,
+void ScaleRowDown34_NEON(const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "1:                                                  \n"
-    MEMACCESS(0)
-    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
-    "subs      %w2, %w2, #24                           \n"
-    "orr       v2.16b, v3.16b, v3.16b                  \n"  // order v0, v1, v2
-    MEMACCESS(1)
-    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
-    "b.gt      1b                                      \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width)         // %2
-  :
-  : "v0", "v1", "v2", "v3", "memory", "cc"
-  );
+                         uint8_t* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "subs        %w2, %w2, #24                 \n"
+      "orr         v2.16b, v3.16b, v3.16b        \n"  // order v0,v1,v2
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st3         {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "v0", "v1", "v2", "v3", "memory", "cc");
 }
 
-void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown34_0_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movi      v20.8b, #3                              \n"
-    "add       %3, %3, %0                              \n"
-  "1:                                                  \n"
-    MEMACCESS(0)
-    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
-    MEMACCESS(3)
-    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
-    "subs         %w2, %w2, #24                        \n"
-
-    // filter src line 0 with src line 1
-    // expand chars to shorts to allow for room
-    // when adding lines together
-    "ushll     v16.8h, v4.8b, #0                       \n"
-    "ushll     v17.8h, v5.8b, #0                       \n"
-    "ushll     v18.8h, v6.8b, #0                       \n"
-    "ushll     v19.8h, v7.8b, #0                       \n"
-
-    // 3 * line_0 + line_1
-    "umlal     v16.8h, v0.8b, v20.8b                   \n"
-    "umlal     v17.8h, v1.8b, v20.8b                   \n"
-    "umlal     v18.8h, v2.8b, v20.8b                   \n"
-    "umlal     v19.8h, v3.8b, v20.8b                   \n"
-
-    // (3 * line_0 + line_1) >> 2
-    "uqrshrn   v0.8b, v16.8h, #2                       \n"
-    "uqrshrn   v1.8b, v17.8h, #2                       \n"
-    "uqrshrn   v2.8b, v18.8h, #2                       \n"
-    "uqrshrn   v3.8b, v19.8h, #2                       \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "ushll     v16.8h, v1.8b, #0                       \n"
-    "umlal     v16.8h, v0.8b, v20.8b                   \n"
-    "uqrshrn   v0.8b, v16.8h, #2                       \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "ushll     v16.8h, v2.8b, #0                       \n"
-    "umlal     v16.8h, v3.8b, v20.8b                   \n"
-    "uqrshrn   v2.8b, v16.8h, #2                       \n"
-
-    MEMACCESS(1)
-    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
-
-    "b.gt      1b                                      \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width),        // %2
-    "+r"(src_stride)        // %3
-  :
-  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19",
-    "v20", "memory", "cc"
-  );
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "movi        v20.8b, #3                    \n"
+      "add         %3, %3, %0                    \n"
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n"  // src line 1
+      "subs        %w2, %w2, #24                 \n"
+
+      // filter src line 0 with src line 1
+      // expand chars to shorts to allow for room
+      // when adding lines together
+      "ushll       v16.8h, v4.8b, #0             \n"
+      "ushll       v17.8h, v5.8b, #0             \n"
+      "ushll       v18.8h, v6.8b, #0             \n"
+      "ushll       v19.8h, v7.8b, #0             \n"
+
+      // 3 * line_0 + line_1
+      "umlal       v16.8h, v0.8b, v20.8b         \n"
+      "umlal       v17.8h, v1.8b, v20.8b         \n"
+      "umlal       v18.8h, v2.8b, v20.8b         \n"
+      "umlal       v19.8h, v3.8b, v20.8b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      // (3 * line_0 + line_1 + 2) >> 2
+      "uqrshrn     v0.8b, v16.8h, #2             \n"
+      "uqrshrn     v1.8b, v17.8h, #2             \n"
+      "uqrshrn     v2.8b, v18.8h, #2             \n"
+      "uqrshrn     v3.8b, v19.8h, #2             \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "ushll       v16.8h, v1.8b, #0             \n"
+      "umlal       v16.8h, v0.8b, v20.8b         \n"
+      "uqrshrn     v0.8b, v16.8h, #2             \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "urhadd      v1.8b, v1.8b, v2.8b           \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "ushll       v16.8h, v2.8b, #0             \n"
+      "umlal       v16.8h, v3.8b, v20.8b         \n"
+      "uqrshrn     v2.8b, v16.8h, #2             \n"
+
+      "st3         {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
+
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_stride)  // %3
+      :
+      : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
+        "v19", "v20", "memory", "cc");
 }
 
-void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown34_1_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "movi      v20.8b, #3                              \n"
-    "add       %3, %3, %0                              \n"
-  "1:                                                  \n"
-    MEMACCESS(0)
-    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
-    MEMACCESS(3)
-    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
-    "subs         %w2, %w2, #24                        \n"
-    // average src line 0 with src line 1
-    "urhadd    v0.8b, v0.8b, v4.8b                     \n"
-    "urhadd    v1.8b, v1.8b, v5.8b                     \n"
-    "urhadd    v2.8b, v2.8b, v6.8b                     \n"
-    "urhadd    v3.8b, v3.8b, v7.8b                     \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "ushll     v4.8h, v1.8b, #0                        \n"
-    "umlal     v4.8h, v0.8b, v20.8b                    \n"
-    "uqrshrn   v0.8b, v4.8h, #2                        \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "ushll     v4.8h, v2.8b, #0                        \n"
-    "umlal     v4.8h, v3.8b, v20.8b                    \n"
-    "uqrshrn   v2.8b, v4.8h, #2                        \n"
-
-    MEMACCESS(1)
-    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
-    "b.gt      1b                                      \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width),        // %2
-    "+r"(src_stride)        // %3
-  :
-  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc"
-  );
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "movi        v20.8b, #3                    \n"
+      "add         %3, %3, %0                    \n"
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n"  // src line 1
+      "subs        %w2, %w2, #24                 \n"
+      // average src line 0 with src line 1
+      "urhadd      v0.8b, v0.8b, v4.8b           \n"
+      "urhadd      v1.8b, v1.8b, v5.8b           \n"
+      "urhadd      v2.8b, v2.8b, v6.8b           \n"
+      "urhadd      v3.8b, v3.8b, v7.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "ushll       v4.8h, v1.8b, #0              \n"
+      "umlal       v4.8h, v0.8b, v20.8b          \n"
+      "uqrshrn     v0.8b, v4.8h, #2              \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "urhadd      v1.8b, v1.8b, v2.8b           \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "ushll       v4.8h, v2.8b, #0              \n"
+      "umlal       v4.8h, v3.8b, v20.8b          \n"
+      "uqrshrn     v2.8b, v4.8h, #2              \n"
+
+      "st3         {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_stride)  // %3
+      :
+      : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc");
 }
 
-static uvec8 kShuf38 =
-  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
-static uvec8 kShuf38_2 =
-  { 0, 16, 32, 2, 18, 33, 4, 20, 34, 6, 22, 35, 0, 0, 0, 0 };
-static vec16 kMult38_Div6 =
-  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
-    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
-static vec16 kMult38_Div9 =
-  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
-    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+static const uvec8 kShuf38 = {0,  3,  6,  8,  11, 14, 16, 19,
+                              22, 24, 27, 30, 0,  0,  0,  0};
+static const uvec8 kShuf38_2 = {0,  16, 32, 2,  18, 33, 4, 20,
+                                34, 6,  22, 35, 0,  0,  0, 0};
+static const vec16 kMult38_Div6 = {65536 / 12, 65536 / 12, 65536 / 12,
+                                   65536 / 12, 65536 / 12, 65536 / 12,
+                                   65536 / 12, 65536 / 12};
+static const vec16 kMult38_Div9 = {65536 / 18, 65536 / 18, 65536 / 18,
+                                   65536 / 18, 65536 / 18, 65536 / 18,
+                                   65536 / 18, 65536 / 18};
 
 // 32 -> 12
-void ScaleRowDown38_NEON(const uint8* src_ptr,
+void ScaleRowDown38_NEON(const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    MEMACCESS(3)
-    "ld1       {v3.16b}, [%3]                          \n"
-  "1:                                                  \n"
-    MEMACCESS(0)
-    "ld1       {v0.16b,v1.16b}, [%0], #32             \n"
-    "subs      %w2, %w2, #12                           \n"
-    "tbl       v2.16b, {v0.16b,v1.16b}, v3.16b        \n"
-    MEMACCESS(1)
-    "st1       {v2.8b}, [%1], #8                       \n"
-    MEMACCESS(1)
-    "st1       {v2.s}[2], [%1], #4                     \n"
-    "b.gt      1b                                      \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(dst_width)         // %2
-  : "r"(&kShuf38)           // %3
-  : "v0", "v1", "v2", "v3", "memory", "cc"
-  );
+                         uint8_t* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "ld1         {v3.16b}, [%3]                \n"
+      "1:                                        \n"
+      "ld1         {v0.16b,v1.16b}, [%0], #32    \n"
+      "subs        %w2, %w2, #12                 \n"
+      "tbl         v2.16b, {v0.16b,v1.16b}, v3.16b \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v2.8b}, [%1], #8             \n"
+      "st1         {v2.s}[2], [%1], #4           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      : "r"(&kShuf38)    // %3
+      : "v0", "v1", "v2", "v3", "memory", "cc");
 }
 
 // 32x3 -> 12x1
-void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
                                       ptrdiff_t src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+                                      uint8_t* dst_ptr,
+                                      int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride * 2;
   ptrdiff_t tmp_src_stride = src_stride;
 
-  asm volatile (
-    MEMACCESS(5)
-    "ld1       {v29.8h}, [%5]                          \n"
-    MEMACCESS(6)
-    "ld1       {v30.16b}, [%6]                         \n"
-    MEMACCESS(7)
-    "ld1       {v31.8h}, [%7]                          \n"
-    "add       %2, %2, %0                              \n"
-  "1:                                                  \n"
-
-    // 00 40 01 41 02 42 03 43
-    // 10 50 11 51 12 52 13 53
-    // 20 60 21 61 22 62 23 63
-    // 30 70 31 71 32 72 33 73
-    MEMACCESS(0)
-    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
-    MEMACCESS(3)
-    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
-    MEMACCESS(4)
-    "ld4       {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32              \n"
-    "subs      %w4, %w4, #12                           \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
-    // 00 10 01 11 02 12 03 13
-    // 40 50 41 51 42 52 43 53
-    "trn1      v20.8b, v0.8b, v1.8b                    \n"
-    "trn2      v21.8b, v0.8b, v1.8b                    \n"
-    "trn1      v22.8b, v4.8b, v5.8b                    \n"
-    "trn2      v23.8b, v4.8b, v5.8b                    \n"
-    "trn1      v24.8b, v16.8b, v17.8b                  \n"
-    "trn2      v25.8b, v16.8b, v17.8b                  \n"
-
-    // 20 30 21 31 22 32 23 33
-    // 60 70 61 71 62 72 63 73
-    "trn1      v0.8b, v2.8b, v3.8b                     \n"
-    "trn2      v1.8b, v2.8b, v3.8b                     \n"
-    "trn1      v4.8b, v6.8b, v7.8b                     \n"
-    "trn2      v5.8b, v6.8b, v7.8b                     \n"
-    "trn1      v16.8b, v18.8b, v19.8b                  \n"
-    "trn2      v17.8b, v18.8b, v19.8b                  \n"
-
-    // 00+10 01+11 02+12 03+13
-    // 40+50 41+51 42+52 43+53
-    "uaddlp    v20.4h, v20.8b                          \n"
-    "uaddlp    v21.4h, v21.8b                          \n"
-    "uaddlp    v22.4h, v22.8b                          \n"
-    "uaddlp    v23.4h, v23.8b                          \n"
-    "uaddlp    v24.4h, v24.8b                          \n"
-    "uaddlp    v25.4h, v25.8b                          \n"
-
-    // 60+70 61+71 62+72 63+73
-    "uaddlp    v1.4h, v1.8b                            \n"
-    "uaddlp    v5.4h, v5.8b                            \n"
-    "uaddlp    v17.4h, v17.8b                          \n"
-
-    // combine source lines
-    "add       v20.4h, v20.4h, v22.4h                  \n"
-    "add       v21.4h, v21.4h, v23.4h                  \n"
-    "add       v20.4h, v20.4h, v24.4h                  \n"
-    "add       v21.4h, v21.4h, v25.4h                  \n"
-    "add       v2.4h, v1.4h, v5.4h                     \n"
-    "add       v2.4h, v2.4h, v17.4h                    \n"
-
-    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
-    //             + s[6 + st * 1] + s[7 + st * 1]
-    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
-    "sqrdmulh  v2.8h, v2.8h, v29.8h                    \n"
-    "xtn       v2.8b,  v2.8h                           \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg. This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded. Then do transposes
-    //  to get aligned.
-    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "ushll     v16.8h, v16.8b, #0                      \n"
-    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
-
-    // combine source lines
-    "add       v0.8h, v0.8h, v16.8h                    \n"
-
-    // xx 20 xx 21 xx 22 xx 23
-    // xx 30 xx 31 xx 32 xx 33
-    "trn1      v1.8h, v0.8h, v0.8h                     \n"
-    "trn2      v4.8h, v0.8h, v0.8h                     \n"
-    "xtn       v0.4h, v1.4s                            \n"
-    "xtn       v4.4h, v4.4s                            \n"
-
-    // 0+1+2, 3+4+5
-    "add       v20.8h, v20.8h, v0.8h                   \n"
-    "add       v21.8h, v21.8h, v4.8h                   \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2. So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "sqrdmulh  v0.8h, v20.8h, v31.8h                   \n"
-    "sqrdmulh  v1.8h, v21.8h, v31.8h                   \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
-
-    MEMACCESS(1)
-    "st1       {v3.8b}, [%1], #8                       \n"
-    MEMACCESS(1)
-    "st1       {v3.s}[2], [%1], #4                     \n"
-    "b.gt      1b                                      \n"
-  : "+r"(src_ptr),          // %0
-    "+r"(dst_ptr),          // %1
-    "+r"(tmp_src_stride),   // %2
-    "+r"(src_ptr1),         // %3
-    "+r"(dst_width)         // %4
-  : "r"(&kMult38_Div6),     // %5
-    "r"(&kShuf38_2),        // %6
-    "r"(&kMult38_Div9)      // %7
-  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
-    "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29",
-    "v30", "v31", "memory", "cc"
-  );
+  asm volatile(
+      "ld1         {v29.8h}, [%5]                \n"
+      "ld1         {v30.16b}, [%6]               \n"
+      "ld1         {v31.8h}, [%7]                \n"
+      "add         %2, %2, %0                    \n"
+      "1:                                        \n"
+
+      // 00 40 01 41 02 42 03 43
+      // 10 50 11 51 12 52 13 53
+      // 20 60 21 61 22 62 23 63
+      // 30 70 31 71 32 72 33 73
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32 \n"
+      "subs        %w4, %w4, #12                 \n"
+
+      // Shuffle the input data around to get align the data
+      //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+      // 00 10 01 11 02 12 03 13
+      // 40 50 41 51 42 52 43 53
+      "trn1        v20.8b, v0.8b, v1.8b          \n"
+      "trn2        v21.8b, v0.8b, v1.8b          \n"
+      "trn1        v22.8b, v4.8b, v5.8b          \n"
+      "trn2        v23.8b, v4.8b, v5.8b          \n"
+      "trn1        v24.8b, v16.8b, v17.8b        \n"
+      "trn2        v25.8b, v16.8b, v17.8b        \n"
+
+      // 20 30 21 31 22 32 23 33
+      // 60 70 61 71 62 72 63 73
+      "trn1        v0.8b, v2.8b, v3.8b           \n"
+      "trn2        v1.8b, v2.8b, v3.8b           \n"
+      "trn1        v4.8b, v6.8b, v7.8b           \n"
+      "trn2        v5.8b, v6.8b, v7.8b           \n"
+      "trn1        v16.8b, v18.8b, v19.8b        \n"
+      "trn2        v17.8b, v18.8b, v19.8b        \n"
+
+      // 00+10 01+11 02+12 03+13
+      // 40+50 41+51 42+52 43+53
+      "uaddlp      v20.4h, v20.8b                \n"
+      "uaddlp      v21.4h, v21.8b                \n"
+      "uaddlp      v22.4h, v22.8b                \n"
+      "uaddlp      v23.4h, v23.8b                \n"
+      "uaddlp      v24.4h, v24.8b                \n"
+      "uaddlp      v25.4h, v25.8b                \n"
+
+      // 60+70 61+71 62+72 63+73
+      "uaddlp      v1.4h, v1.8b                  \n"
+      "uaddlp      v5.4h, v5.8b                  \n"
+      "uaddlp      v17.4h, v17.8b                \n"
+
+      // combine source lines
+      "add         v20.4h, v20.4h, v22.4h        \n"
+      "add         v21.4h, v21.4h, v23.4h        \n"
+      "add         v20.4h, v20.4h, v24.4h        \n"
+      "add         v21.4h, v21.4h, v25.4h        \n"
+      "add         v2.4h, v1.4h, v5.4h           \n"
+      "add         v2.4h, v2.4h, v17.4h          \n"
+
+      // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+      //             + s[6 + st * 1] + s[7 + st * 1]
+      //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+      "sqrdmulh    v2.8h, v2.8h, v29.8h          \n"
+      "xtn         v2.8b,  v2.8h                 \n"
+
+      // Shuffle 2,3 reg around so that 2 can be added to the
+      //  0,1 reg and 3 can be added to the 4,5 reg. This
+      //  requires expanding from u8 to u16 as the 0,1 and 4,5
+      //  registers are already expanded. Then do transposes
+      //  to get aligned.
+      // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+      "ushll       v16.8h, v16.8b, #0            \n"
+      "uaddl       v0.8h, v0.8b, v4.8b           \n"
+
+      // combine source lines
+      "add         v0.8h, v0.8h, v16.8h          \n"
+
+      // xx 20 xx 21 xx 22 xx 23
+      // xx 30 xx 31 xx 32 xx 33
+      "trn1        v1.8h, v0.8h, v0.8h           \n"
+      "trn2        v4.8h, v0.8h, v0.8h           \n"
+      "xtn         v0.4h, v1.4s                  \n"
+      "xtn         v4.4h, v4.4s                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      // 0+1+2, 3+4+5
+      "add         v20.8h, v20.8h, v0.8h         \n"
+      "add         v21.8h, v21.8h, v4.8h         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+
+      // Need to divide, but can't downshift as the the value
+      //  isn't a power of 2. So multiply by 65536 / n
+      //  and take the upper 16 bits.
+      "sqrdmulh    v0.8h, v20.8h, v31.8h         \n"
+      "sqrdmulh    v1.8h, v21.8h, v31.8h         \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+
+      // Align for table lookup, vtbl requires registers to be adjacent
+      "tbl         v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
+
+      "st1         {v3.8b}, [%1], #8             \n"
+      "st1         {v3.s}[2], [%1], #4           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),         // %0
+        "+r"(dst_ptr),         // %1
+        "+r"(tmp_src_stride),  // %2
+        "+r"(src_ptr1),        // %3
+        "+r"(dst_width)        // %4
+      : "r"(&kMult38_Div6),    // %5
+        "r"(&kShuf38_2),       // %6
+        "r"(&kMult38_Div9)     // %7
+      : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
+        "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29", "v30", "v31",
+        "memory", "cc");
 }
 
 // 32x2 -> 12x1
-void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
+                               uint8_t* dst_ptr,
+                               int dst_width) {
   // TODO(fbarchard): use src_stride directly for clang 3.5+.
   ptrdiff_t tmp_src_stride = src_stride;
-  asm volatile (
-    MEMACCESS(4)
-    "ld1       {v30.8h}, [%4]                          \n"
-    MEMACCESS(5)
-    "ld1       {v31.16b}, [%5]                         \n"
-    "add       %2, %2, %0                              \n"
-  "1:                                                  \n"
-
-    // 00 40 01 41 02 42 03 43
-    // 10 50 11 51 12 52 13 53
-    // 20 60 21 61 22 62 23 63
-    // 30 70 31 71 32 72 33 73
-    MEMACCESS(0)
-    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
-    MEMACCESS(3)
-    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
-    "subs      %w3, %w3, #12                           \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
-    // 00 10 01 11 02 12 03 13
-    // 40 50 41 51 42 52 43 53
-    "trn1      v16.8b, v0.8b, v1.8b                    \n"
-    "trn2      v17.8b, v0.8b, v1.8b                    \n"
-    "trn1      v18.8b, v4.8b, v5.8b                    \n"
-    "trn2      v19.8b, v4.8b, v5.8b                    \n"
-
-    // 20 30 21 31 22 32 23 33
-    // 60 70 61 71 62 72 63 73
-    "trn1      v0.8b, v2.8b, v3.8b                     \n"
-    "trn2      v1.8b, v2.8b, v3.8b                     \n"
-    "trn1      v4.8b, v6.8b, v7.8b                     \n"
-    "trn2      v5.8b, v6.8b, v7.8b                     \n"
-
-    // 00+10 01+11 02+12 03+13
-    // 40+50 41+51 42+52 43+53
-    "uaddlp    v16.4h, v16.8b                          \n"
-    "uaddlp    v17.4h, v17.8b                          \n"
-    "uaddlp    v18.4h, v18.8b                          \n"
-    "uaddlp    v19.4h, v19.8b                          \n"
-
-    // 60+70 61+71 62+72 63+73
-    "uaddlp    v1.4h, v1.8b                            \n"
-    "uaddlp    v5.4h, v5.8b                            \n"
-
-    // combine source lines
-    "add       v16.4h, v16.4h, v18.4h                  \n"
-    "add       v17.4h, v17.4h, v19.4h                  \n"
-    "add       v2.4h, v1.4h, v5.4h                     \n"
-
-    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-    "uqrshrn   v2.8b, v2.8h, #2                        \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg. This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded. Then do transposes
-    //  to get aligned.
-    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-
-    // combine source lines
-    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
-
-    // xx 20 xx 21 xx 22 xx 23
-    // xx 30 xx 31 xx 32 xx 33
-    "trn1      v1.8h, v0.8h, v0.8h                     \n"
-    "trn2      v4.8h, v0.8h, v0.8h                     \n"
-    "xtn       v0.4h, v1.4s                            \n"
-    "xtn       v4.4h, v4.4s                            \n"
-
-    // 0+1+2, 3+4+5
-    "add       v16.8h, v16.8h, v0.8h                   \n"
-    "add       v17.8h, v17.8h, v4.8h                   \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2. So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "sqrdmulh  v0.8h, v16.8h, v30.8h                   \n"
-    "sqrdmulh  v1.8h, v17.8h, v30.8h                   \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-
-    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
-
-    MEMACCESS(1)
-    "st1       {v3.8b}, [%1], #8                       \n"
-    MEMACCESS(1)
-    "st1       {v3.s}[2], [%1], #4                     \n"
-    "b.gt      1b                                      \n"
-  : "+r"(src_ptr),         // %0
-    "+r"(dst_ptr),         // %1
-    "+r"(tmp_src_stride),  // %2
-    "+r"(dst_width)        // %3
-  : "r"(&kMult38_Div6),    // %4
-    "r"(&kShuf38_2)        // %5
-  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
-    "v18", "v19", "v30", "v31", "memory", "cc"
+  asm volatile(
+      "ld1         {v30.8h}, [%4]                \n"
+      "ld1         {v31.16b}, [%5]               \n"
+      "add         %2, %2, %0                    \n"
+      "1:                                        \n"
+
+      // 00 40 01 41 02 42 03 43
+      // 10 50 11 51 12 52 13 53
+      // 20 60 21 61 22 62 23 63
+      // 30 70 31 71 32 72 33 73
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n"
+      "subs        %w3, %w3, #12                 \n"
+
+      // Shuffle the input data around to get align the data
+      //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+      // 00 10 01 11 02 12 03 13
+      // 40 50 41 51 42 52 43 53
+      "trn1        v16.8b, v0.8b, v1.8b          \n"
+      "trn2        v17.8b, v0.8b, v1.8b          \n"
+      "trn1        v18.8b, v4.8b, v5.8b          \n"
+      "trn2        v19.8b, v4.8b, v5.8b          \n"
+
+      // 20 30 21 31 22 32 23 33
+      // 60 70 61 71 62 72 63 73
+      "trn1        v0.8b, v2.8b, v3.8b           \n"
+      "trn2        v1.8b, v2.8b, v3.8b           \n"
+      "trn1        v4.8b, v6.8b, v7.8b           \n"
+      "trn2        v5.8b, v6.8b, v7.8b           \n"
+
+      // 00+10 01+11 02+12 03+13
+      // 40+50 41+51 42+52 43+53
+      "uaddlp      v16.4h, v16.8b                \n"
+      "uaddlp      v17.4h, v17.8b                \n"
+      "uaddlp      v18.4h, v18.8b                \n"
+      "uaddlp      v19.4h, v19.8b                \n"
+
+      // 60+70 61+71 62+72 63+73
+      "uaddlp      v1.4h, v1.8b                  \n"
+      "uaddlp      v5.4h, v5.8b                  \n"
+
+      // combine source lines
+      "add         v16.4h, v16.4h, v18.4h        \n"
+      "add         v17.4h, v17.4h, v19.4h        \n"
+      "add         v2.4h, v1.4h, v5.4h           \n"
+
+      // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+      "uqrshrn     v2.8b, v2.8h, #2              \n"
+
+      // Shuffle 2,3 reg around so that 2 can be added to the
+      //  0,1 reg and 3 can be added to the 4,5 reg. This
+      //  requires expanding from u8 to u16 as the 0,1 and 4,5
+      //  registers are already expanded. Then do transposes
+      //  to get aligned.
+      // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+
+      // combine source lines
+      "uaddl       v0.8h, v0.8b, v4.8b           \n"
+
+      // xx 20 xx 21 xx 22 xx 23
+      // xx 30 xx 31 xx 32 xx 33
+      "trn1        v1.8h, v0.8h, v0.8h           \n"
+      "trn2        v4.8h, v0.8h, v0.8h           \n"
+      "xtn         v0.4h, v1.4s                  \n"
+      "xtn         v4.4h, v4.4s                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      // 0+1+2, 3+4+5
+      "add         v16.8h, v16.8h, v0.8h         \n"
+      "add         v17.8h, v17.8h, v4.8h         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+
+      // Need to divide, but can't downshift as the the value
+      //  isn't a power of 2. So multiply by 65536 / n
+      //  and take the upper 16 bits.
+      "sqrdmulh    v0.8h, v16.8h, v30.8h         \n"
+      "sqrdmulh    v1.8h, v17.8h, v30.8h         \n"
+
+      // Align for table lookup, vtbl requires registers to
+      //  be adjacent
+
+      "tbl         v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
+
+      "st1         {v3.8b}, [%1], #8             \n"
+      "st1         {v3.s}[2], [%1], #4           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),         // %0
+        "+r"(dst_ptr),         // %1
+        "+r"(tmp_src_stride),  // %2
+        "+r"(dst_width)        // %3
+      : "r"(&kMult38_Div6),    // %4
+        "r"(&kShuf38_2)        // %5
+      : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
+        "v19", "v30", "v31", "memory", "cc");
+}
+
+void ScaleRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 01234567
+      "ldr         d1, [%1], #8                  \n"  // 12345678
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"  // 01234567 (16b)
+      "ushll       v3.8h, v1.8b, #0              \n"  // 12345678 (16b)
+
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (even)
+
+      "rshrn       v2.8b, v2.8h, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v1.8b, v3.8h, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.8b, v2.8b}, [%2], #16     \n"  // store
+      "subs        %w3, %w3, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31"  // Clobber List
   );
 }
 
-void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                    uint16* dst_ptr, int src_width, int src_height) {
-  const uint8* src_tmp;
-  asm volatile (
-  "1:                                          \n"
-    "mov       %0, %1                          \n"
-    "mov       w12, %w5                        \n"
-    "eor       v2.16b, v2.16b, v2.16b          \n"
-    "eor       v3.16b, v3.16b, v3.16b          \n"
-  "2:                                          \n"
-    // load 16 pixels into q0
-    MEMACCESS(0)
-    "ld1       {v0.16b}, [%0], %3              \n"
-    "uaddw2    v3.8h, v3.8h, v0.16b            \n"
-    "uaddw     v2.8h, v2.8h, v0.8b             \n"
-    "subs      w12, w12, #1                    \n"
-    "b.gt      2b                              \n"
-    MEMACCESS(2)
-    "st1      {v2.8h, v3.8h}, [%2], #32        \n"  // store pixels
-    "add      %1, %1, #16                      \n"
-    "subs     %w4, %w4, #16                    \n"  // 16 processed per loop
-    "b.gt     1b                               \n"
-  : "=&r"(src_tmp),    // %0
-    "+r"(src_ptr),     // %1
-    "+r"(dst_ptr),     // %2
-    "+r"(src_stride),  // %3
-    "+r"(src_width),   // %4
-    "+r"(src_height)   // %5
-  :
-  : "memory", "cc", "w12", "v0", "v1", "v2", "v3"  // Clobber List
+void ScaleRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 1;
+  const uint8_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+      "movi        v30.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 01234567
+      "ldr         d1, [%2], #8                  \n"  // 12345678
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"  // 01234567 (16b)
+      "ushll       v3.8h, v1.8b, #0              \n"  // 12345678 (16b)
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (1, odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"
+      "ldr         d1, [%3], #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v4.8h, v0.8b, #0              \n"  // 01234567 (16b)
+      "ushll       v5.8h, v1.8b, #0              \n"  // 12345678 (16b)
+      "umlal       v4.8h, v1.8b, v31.8b          \n"  // 3*near+far (2, odd)
+      "umlal       v5.8h, v0.8b, v31.8b          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.8h, v2.8h, v30.8h          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.8h, v3.8h, v30.8h          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.8h, v0.8h, v30.8h          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.8h, v1.8h, v30.8h          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v2.8b, v2.8h, #4              \n"  // 2, odd
+      "rshrn       v1.8b, v3.8h, #4              \n"  // 2, even
+      "rshrn       v4.8b, v4.8h, #4              \n"  // 1, odd
+      "rshrn       v3.8b, v5.8h, #4              \n"  // 1, even
+
+      "st2         {v1.8b, v2.8b}, [%5], #16     \n"  // store 1
+      "st2         {v3.8b, v4.8b}, [%4], #16     \n"  // store 2
+      "subs        %w6, %w6, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_12_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v1.8h}, [%1], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "mov         v2.16b, v0.16b                \n"
+      "mla         v0.8h, v1.8h, v31.8h          \n"  // 3*near+far (odd)
+      "mla         v1.8h, v2.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "urshr       v2.8h, v0.8h, #2              \n"  // 3/4*near+1/4*far (odd)
+      "urshr       v1.8h, v1.8h, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.8h, v2.8h}, [%2], #32     \n"  // store
+      "subs        %w3, %w3, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_12_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v2.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v3.8h}, [%2], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "mov         v0.16b, v2.16b                \n"
+      "mla         v2.8h, v3.8h, v31.8h          \n"  // 3*near+far (odd)
+      "mla         v3.8h, v0.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "ld1         {v4.8h}, [%1], #16            \n"  // 01234567 (16b)
+      "ld1         {v5.8h}, [%3], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+
+      "mov         v0.16b, v4.16b                \n"
+      "mla         v4.8h, v5.8h, v31.8h          \n"  // 3*near+far (odd)
+      "mla         v5.8h, v0.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.8h, v2.8h, v31.8h          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.8h, v3.8h, v31.8h          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.8h, v0.8h, v31.8h          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.8h, v1.8h, v31.8h          \n"  // 9 3 3 1 (2, even)
+
+      "urshr       v2.8h, v2.8h, #4              \n"  // 2, odd
+      "urshr       v1.8h, v3.8h, #4              \n"  // 2, even
+      "urshr       v4.8h, v4.8h, #4              \n"  // 1, odd
+      "urshr       v3.8h, v5.8h, #4              \n"  // 1, even
+
+      "st2         {v3.8h, v4.8h}, [%4], #32     \n"  // store 1
+      "st2         {v1.8h, v2.8h}, [%5], #32     \n"  // store 2
+
+      "subs        %w6, %w6, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v1.8h}, [%1], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0123 (32b)
+      "ushll2      v3.4s, v0.8h, #0              \n"  // 4567 (32b)
+      "ushll       v4.4s, v1.4h, #0              \n"  // 1234 (32b)
+      "ushll2      v5.4s, v1.8h, #0              \n"  // 5678 (32b)
+
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (1, odd)
+      "umlal2      v3.4s, v1.8h, v31.8h          \n"  // 3*near+far (2, odd)
+      "umlal       v4.4s, v0.4h, v31.4h          \n"  // 3*near+far (1, even)
+      "umlal2      v5.4s, v0.8h, v31.8h          \n"  // 3*near+far (2, even)
+
+      "rshrn       v0.4h, v4.4s, #2              \n"  // 3/4*near+1/4*far
+      "rshrn2      v0.8h, v5.4s, #2              \n"  // 3/4*near+1/4*far (even)
+      "rshrn       v1.4h, v2.4s, #2              \n"  // 3/4*near+1/4*far
+      "rshrn2      v1.8h, v3.4s, #2              \n"  // 3/4*near+1/4*far (odd)
+
+      "st2         {v0.8h, v1.8h}, [%2], #32     \n"  // store
+      "subs        %w3, %w3, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "movi        v31.4h, #3                    \n"
+      "movi        v30.4s, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 0123 (16b)
+      "ldr         d1, [%2], #8                  \n"  // 1234 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0123 (32b)
+      "ushll       v3.4s, v1.4h, #0              \n"  // 1234 (32b)
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (1, odd)
+      "umlal       v3.4s, v0.4h, v31.4h          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"  // 0123 (16b)
+      "ldr         d1, [%3], #8                  \n"  // 1234 (16b)
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v4.4s, v0.4h, #0              \n"  // 0123 (32b)
+      "ushll       v5.4s, v1.4h, #0              \n"  // 1234 (32b)
+      "umlal       v4.4s, v1.4h, v31.4h          \n"  // 3*near+far (2, odd)
+      "umlal       v5.4s, v0.4h, v31.4h          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.4s, v2.4s, v30.4s          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.4s, v3.4s, v30.4s          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.4s, v0.4s, v30.4s          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.4s, v1.4s, v30.4s          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v1.4h, v4.4s, #4              \n"  // 3/4*near+1/4*far
+      "rshrn       v0.4h, v5.4s, #4              \n"  // 3/4*near+1/4*far
+      "rshrn       v5.4h, v2.4s, #4              \n"  // 3/4*near+1/4*far
+      "rshrn       v4.4h, v3.4s, #4              \n"  // 3/4*near+1/4*far
+
+      "st2         {v0.4h, v1.4h}, [%4], #16     \n"  // store 1
+      "st2         {v4.4h, v5.4h}, [%5], #16     \n"  // store 2
+
+      "subs        %w6, %w6, #8                  \n"  // 4 sample -> 8 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  const uint8_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 00112233 (1u1v)
+      "ldr         d1, [%1], #8                  \n"  // 11223344 (1u1v)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"  // 00112233 (1u1v, 16b)
+      "ushll       v3.8h, v1.8b, #0              \n"  // 11223344 (1u1v, 16b)
+
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (even)
+
+      "rshrn       v2.8b, v2.8h, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v1.8b, v3.8h, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.4h, v2.4h}, [%2], #16     \n"  // store
+      "subs        %w3, %w3, #8                  \n"  // 4 uv -> 8 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 2;
+  const uint8_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+      "movi        v30.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"
+      "ldr         d1, [%2], #8                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"
+      "ushll       v3.8h, v1.8b, #0              \n"
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (1, odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"
+      "ldr         d1, [%3], #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v4.8h, v0.8b, #0              \n"
+      "ushll       v5.8h, v1.8b, #0              \n"
+      "umlal       v4.8h, v1.8b, v31.8b          \n"  // 3*near+far (2, odd)
+      "umlal       v5.8h, v0.8b, v31.8b          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.8h, v2.8h, v30.8h          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.8h, v3.8h, v30.8h          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.8h, v0.8h, v30.8h          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.8h, v1.8h, v30.8h          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v2.8b, v2.8h, #4              \n"  // 2, odd
+      "rshrn       v1.8b, v3.8h, #4              \n"  // 2, even
+      "rshrn       v4.8b, v4.8h, #4              \n"  // 1, odd
+      "rshrn       v3.8b, v5.8h, #4              \n"  // 1, even
+
+      "st2         {v1.4h, v2.4h}, [%5], #16     \n"  // store 2
+      "st2         {v3.4h, v4.4h}, [%4], #16     \n"  // store 1
+      "subs        %w6, %w6, #8                  \n"  // 4 uv -> 8 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width) {
+  const uint16_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v1.8h}, [%1], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0011 (1u1v, 32b)
+      "ushll       v3.4s, v1.4h, #0              \n"  // 1122 (1u1v, 32b)
+      "ushll2      v4.4s, v0.8h, #0              \n"  // 2233 (1u1v, 32b)
+      "ushll2      v5.4s, v1.8h, #0              \n"  // 3344 (1u1v, 32b)
+
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (odd)
+      "umlal       v3.4s, v0.4h, v31.4h          \n"  // 3*near+far (even)
+      "umlal2      v4.4s, v1.8h, v31.8h          \n"  // 3*near+far (odd)
+      "umlal2      v5.4s, v0.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "rshrn       v2.4h, v2.4s, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v1.4h, v3.4s, #2              \n"  // 3/4*near+1/4*far (even)
+      "rshrn       v4.4h, v4.4s, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v3.4h, v5.4s, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.2s, v2.2s}, [%2], #16     \n"  // store
+      "st2         {v3.2s, v4.2s}, [%2], #16     \n"  // store
+      "subs        %w3, %w3, #8                  \n"  // 4 uv -> 8 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 2;
+  const uint16_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "movi        v31.4h, #3                    \n"
+      "movi        v30.4s, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"
+      "ldr         d1, [%2], #8                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0011 (1u1v, 32b)
+      "ushll       v3.4s, v1.4h, #0              \n"  // 1122 (1u1v, 32b)
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (1, odd)
+      "umlal       v3.4s, v0.4h, v31.4h          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"
+      "ldr         d1, [%3], #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v4.4s, v0.4h, #0              \n"  // 0011 (1u1v, 32b)
+      "ushll       v5.4s, v1.4h, #0              \n"  // 1122 (1u1v, 32b)
+      "umlal       v4.4s, v1.4h, v31.4h          \n"  // 3*near+far (2, odd)
+      "umlal       v5.4s, v0.4h, v31.4h          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.4s, v2.4s, v30.4s          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.4s, v3.4s, v30.4s          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.4s, v0.4s, v30.4s          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.4s, v1.4s, v30.4s          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v1.4h, v2.4s, #4              \n"  // 2, odd
+      "rshrn       v0.4h, v3.4s, #4              \n"  // 2, even
+      "rshrn       v3.4h, v4.4s, #4              \n"  // 1, odd
+      "rshrn       v2.4h, v5.4s, #4              \n"  // 1, even
+
+      "st2         {v0.2s, v1.2s}, [%5], #16     \n"  // store 2
+      "st2         {v2.2s, v3.2s}, [%4], #16     \n"  // store 1
+      "subs        %w6, %w6, #4                  \n"  // 2 uv -> 4 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+// Add a row of bytes to a row of shorts.  Used for box filter.
+// Reads 16 bytes and accumulates to 16 shorts at a time.
+void ScaleAddRow_NEON(const uint8_t* src_ptr,
+                      uint16_t* dst_ptr,
+                      int src_width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v1.8h, v2.8h}, [%1]          \n"  // load accumulator
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 16 bytes
+      "uaddw2      v2.8h, v2.8h, v0.16b          \n"  // add
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uaddw       v1.8h, v1.8h, v0.8b           \n"
+      "st1         {v1.8h, v2.8h}, [%1], #32     \n"  // store accumulator
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(src_width)  // %2
+      :
+      : "memory", "cc", "v0", "v1", "v2"  // Clobber List
   );
 }
 
 // TODO(Yang Zhang): Investigate less load instructions for
 // the x/dx stepping
-#define LOAD2_DATA8_LANE(n)                                    \
-    "lsr        %5, %3, #16                    \n"             \
-    "add        %6, %1, %5                    \n"              \
-    "add        %3, %3, %4                     \n"             \
-    MEMACCESS(6)                                               \
-    "ld2        {v4.b, v5.b}["#n"], [%6]      \n"
-
-void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
-                          int dst_width, int x, int dx) {
+#define LOAD2_DATA8_LANE(n)                      \
+  "lsr        %5, %3, #16                    \n" \
+  "add        %6, %1, %5                     \n" \
+  "add        %3, %3, %4                     \n" \
+  "ld2        {v4.b, v5.b}[" #n "], [%6]     \n"
+
+// The NEON version mimics this formula (from row_common.cc):
+// #define BLENDER(a, b, f) (uint8_t)((int)(a) +
+//    ((((int)((f)) * ((int)(b) - (int)(a))) + 0x8000) >> 16))
+
+void ScaleFilterCols_NEON(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          int dst_width,
+                          int x,
+                          int dx) {
   int dx_offset[4] = {0, 1, 2, 3};
   int* tmp = dx_offset;
-  const uint8* src_tmp = src_ptr;
-  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
-  int64 x64 = (int64) x;
-  int64 dx64 = (int64) dx;
+  const uint8_t* src_tmp = src_ptr;
+  int64_t x64 = (int64_t)x;    // NOLINT
+  int64_t dx64 = (int64_t)dx;  // NOLINT
   asm volatile (
-    "dup        v0.4s, %w3                     \n"  // x
-    "dup        v1.4s, %w4                     \n"  // dx
-    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
-    "shl        v3.4s, v1.4s, #2               \n"  // 4 * dx
-    "mul        v1.4s, v1.4s, v2.4s            \n"
+      "dup         v0.4s, %w3                    \n"  // x
+      "dup         v1.4s, %w4                    \n"  // dx
+      "ld1         {v2.4s}, [%5]                 \n"  // 0 1 2 3
+      "shl         v3.4s, v1.4s, #2              \n"  // 4 * dx
+      "mul         v1.4s, v1.4s, v2.4s           \n"
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "add        v1.4s, v1.4s, v0.4s            \n"
+      "add         v1.4s, v1.4s, v0.4s           \n"
     // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
-    "add        v2.4s, v1.4s, v3.4s            \n"
-    "shl        v0.4s, v3.4s, #1               \n"  // 8 * dx
-  "1:                                          \n"
+      "add         v2.4s, v1.4s, v3.4s           \n"
+      "shl         v0.4s, v3.4s, #1              \n"  // 8 * dx
+      "1:                                        \n"
     LOAD2_DATA8_LANE(0)
     LOAD2_DATA8_LANE(1)
     LOAD2_DATA8_LANE(2)
@@ -615,31 +1082,30 @@ void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
     LOAD2_DATA8_LANE(5)
     LOAD2_DATA8_LANE(6)
     LOAD2_DATA8_LANE(7)
-    "mov       v6.16b, v1.16b                  \n"
-    "mov       v7.16b, v2.16b                  \n"
-    "uzp1      v6.8h, v6.8h, v7.8h             \n"
-    "ushll     v4.8h, v4.8b, #0                \n"
-    "ushll     v5.8h, v5.8b, #0                \n"
-    "ssubl     v16.4s, v5.4h, v4.4h            \n"
-    "ssubl2    v17.4s, v5.8h, v4.8h            \n"
-    "ushll     v7.4s, v6.4h, #0                \n"
-    "ushll2    v6.4s, v6.8h, #0                \n"
-    "mul       v16.4s, v16.4s, v7.4s           \n"
-    "mul       v17.4s, v17.4s, v6.4s           \n"
-    "rshrn      v6.4h, v16.4s, #16             \n"
-    "rshrn2     v6.8h, v17.4s, #16             \n"
-    "add       v4.8h, v4.8h, v6.8h             \n"
-    "xtn       v4.8b, v4.8h                    \n"
-
-    MEMACCESS(0)
-    "st1       {v4.8b}, [%0], #8               \n"  // store pixels
-    "add       v1.4s, v1.4s, v0.4s             \n"
-    "add       v2.4s, v2.4s, v0.4s             \n"
-    "subs      %w2, %w2, #8                    \n"  // 8 processed per loop
-    "b.gt      1b                              \n"
+      "mov         v6.16b, v1.16b                \n"
+      "mov         v7.16b, v2.16b                \n"
+      "uzp1        v6.8h, v6.8h, v7.8h           \n"
+      "ushll       v4.8h, v4.8b, #0              \n"
+      "ushll       v5.8h, v5.8b, #0              \n"
+      "ssubl       v16.4s, v5.4h, v4.4h          \n"
+      "ssubl2      v17.4s, v5.8h, v4.8h          \n"
+      "ushll       v7.4s, v6.4h, #0              \n"
+      "ushll2      v6.4s, v6.8h, #0              \n"
+      "mul         v16.4s, v16.4s, v7.4s         \n"
+      "mul         v17.4s, v17.4s, v6.4s         \n"
+      "rshrn       v6.4h, v16.4s, #16            \n"
+      "rshrn2      v6.8h, v17.4s, #16            \n"
+      "add         v4.8h, v4.8h, v6.8h           \n"
+      "xtn         v4.8b, v4.8h                  \n"
+
+      "st1         {v4.8b}, [%0], #8             \n"  // store pixels
+      "add         v1.4s, v1.4s, v0.4s           \n"
+      "add         v2.4s, v2.4s, v0.4s           \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "b.gt        1b                            \n"
   : "+r"(dst_ptr),          // %0
     "+r"(src_ptr),          // %1
-    "+r"(dst_width64),      // %2
+    "+r"(dst_width),        // %2
     "+r"(x64),              // %3
     "+r"(dx64),             // %4
     "+r"(tmp),              // %5
@@ -653,342 +1119,328 @@ void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
 #undef LOAD2_DATA8_LANE
 
 // 16x2 -> 16x1
-void ScaleFilterRows_NEON(uint8* dst_ptr,
-                          const uint8* src_ptr, ptrdiff_t src_stride,
-                          int dst_width, int source_y_fraction) {
-    int y_fraction = 256 - source_y_fraction;
-  asm volatile (
-    "cmp          %w4, #0                      \n"
-    "b.eq         100f                         \n"
-    "add          %2, %2, %1                   \n"
-    "cmp          %w4, #64                     \n"
-    "b.eq         75f                          \n"
-    "cmp          %w4, #128                    \n"
-    "b.eq         50f                          \n"
-    "cmp          %w4, #192                    \n"
-    "b.eq         25f                          \n"
-
-    "dup          v5.8b, %w4                   \n"
-    "dup          v4.8b, %w5                   \n"
-    // General purpose row blend.
-  "1:                                          \n"
-    MEMACCESS(1)
-    "ld1          {v0.16b}, [%1], #16          \n"
-    MEMACCESS(2)
-    "ld1          {v1.16b}, [%2], #16          \n"
-    "subs         %w3, %w3, #16                \n"
-    "umull        v6.8h, v0.8b, v4.8b          \n"
-    "umull2       v7.8h, v0.16b, v4.16b        \n"
-    "umlal        v6.8h, v1.8b, v5.8b          \n"
-    "umlal2       v7.8h, v1.16b, v5.16b        \n"
-    "rshrn        v0.8b, v6.8h, #8             \n"
-    "rshrn2       v0.16b, v7.8h, #8            \n"
-    MEMACCESS(0)
-    "st1          {v0.16b}, [%0], #16          \n"
-    "b.gt         1b                           \n"
-    "b            99f                          \n"
-
-    // Blend 25 / 75.
-  "25:                                         \n"
-    MEMACCESS(1)
-    "ld1          {v0.16b}, [%1], #16          \n"
-    MEMACCESS(2)
-    "ld1          {v1.16b}, [%2], #16          \n"
-    "subs         %w3, %w3, #16                \n"
-    "urhadd       v0.16b, v0.16b, v1.16b       \n"
-    "urhadd       v0.16b, v0.16b, v1.16b       \n"
-    MEMACCESS(0)
-    "st1          {v0.16b}, [%0], #16          \n"
-    "b.gt         25b                          \n"
-    "b            99f                          \n"
-
-    // Blend 50 / 50.
-  "50:                                         \n"
-    MEMACCESS(1)
-    "ld1          {v0.16b}, [%1], #16          \n"
-    MEMACCESS(2)
-    "ld1          {v1.16b}, [%2], #16          \n"
-    "subs         %w3, %w3, #16                \n"
-    "urhadd       v0.16b, v0.16b, v1.16b       \n"
-    MEMACCESS(0)
-    "st1          {v0.16b}, [%0], #16          \n"
-    "b.gt         50b                          \n"
-    "b            99f                          \n"
-
-    // Blend 75 / 25.
-  "75:                                         \n"
-    MEMACCESS(1)
-    "ld1          {v1.16b}, [%1], #16          \n"
-    MEMACCESS(2)
-    "ld1          {v0.16b}, [%2], #16          \n"
-    "subs         %w3, %w3, #16                \n"
-    "urhadd       v0.16b, v0.16b, v1.16b       \n"
-    "urhadd       v0.16b, v0.16b, v1.16b       \n"
-    MEMACCESS(0)
-    "st1          {v0.16b}, [%0], #16          \n"
-    "b.gt         75b                          \n"
-    "b            99f                          \n"
-
-    // Blend 100 / 0 - Copy row unchanged.
-  "100:                                        \n"
-    MEMACCESS(1)
-    "ld1          {v0.16b}, [%1], #16          \n"
-    "subs         %w3, %w3, #16                \n"
-    MEMACCESS(0)
-    "st1          {v0.16b}, [%0], #16          \n"
-    "b.gt         100b                         \n"
-
-  "99:                                         \n"
-    MEMACCESS(0)
-    "st1          {v0.b}[15], [%0]             \n"
-  : "+r"(dst_ptr),          // %0
-    "+r"(src_ptr),          // %1
-    "+r"(src_stride),       // %2
-    "+r"(dst_width),        // %3
-    "+r"(source_y_fraction),// %4
-    "+r"(y_fraction)        // %5
-  :
-  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc"
-  );
+void ScaleFilterRows_NEON(uint8_t* dst_ptr,
+                          const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          int dst_width,
+                          int source_y_fraction) {
+  int y_fraction = 256 - source_y_fraction;
+  asm volatile(
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "add         %2, %2, %1                    \n"
+      "cmp         %w4, #64                      \n"
+      "b.eq        75f                           \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+      "cmp         %w4, #192                     \n"
+      "b.eq        25f                           \n"
+
+      "dup         v5.8b, %w4                    \n"
+      "dup         v4.8b, %w5                    \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "umull       v6.8h, v0.8b, v4.8b           \n"
+      "umull2      v7.8h, v0.16b, v4.16b         \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "umlal       v6.8h, v1.8b, v5.8b           \n"
+      "umlal2      v7.8h, v1.16b, v5.16b         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "rshrn       v0.8b, v6.8h, #8              \n"
+      "rshrn2      v0.16b, v7.8h, #8             \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 25 / 75.
+      "25:                                       \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        25b                           \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 75 / 25.
+      "75:                                       \n"
+      "ld1         {v1.16b}, [%1], #16           \n"
+      "ld1         {v0.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        75b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        100b                          \n"
+
+      "99:                                       \n"
+      "st1         {v0.b}[15], [%0]              \n"
+      : "+r"(dst_ptr),            // %0
+        "+r"(src_ptr),            // %1
+        "+r"(src_stride),         // %2
+        "+r"(dst_width),          // %3
+        "+r"(source_y_fraction),  // %4
+        "+r"(y_fraction)          // %5
+      :
+      : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc");
 }
 
-void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    // load even pixels into q0, odd into q1
-    MEMACCESS (0)
-    "ld2        {v0.4s, v1.4s}, [%0], #32      \n"
-    MEMACCESS (0)
-    "ld2        {v2.4s, v3.4s}, [%0], #32      \n"
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-    MEMACCESS (1)
-    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
-    MEMACCESS (1)
-    "st1        {v3.16b}, [%1], #16            \n"
-    "b.gt       1b                             \n"
-  : "+r" (src_ptr),          // %0
-    "+r" (dst),              // %1
-    "+r" (dst_width)         // %2
-  :
-  : "memory", "cc", "v0", "v1", "v2", "v3"  // Clobber List
+void ScaleARGBRowDown2_NEON(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst,
+                            int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      // load 16 ARGB pixels with even pixels into q0/q2, odd into q1/q3
+      "ld4         {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "mov         v2.16b, v3.16b                \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st2         {v1.4s,v2.4s}, [%1], #32      \n"  // store 8 odd pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
-                                  uint8* dst_argb, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS (0)
-    // load 8 ARGB pixels.
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
-    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
-    "rshrn      v1.8b, v1.8h, #1               \n"
-    "rshrn      v2.8b, v2.8h, #1               \n"
-    "rshrn      v3.8b, v3.8h, #1               \n"
-    MEMACCESS (1)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32     \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),         // %0
-    "+r"(dst_argb),         // %1
-    "+r"(dst_width)         // %2
-  :
-  : "memory", "cc", "v0", "v1", "v2", "v3"    // Clobber List
+void ScaleARGBRowDown2Linear_NEON(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_argb,
+                                  int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      // load 16 ARGB pixels with even pixels into q0/q2, odd into q1/q3
+      "ld4         {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"  // rounding half add
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v1.16b, v2.16b, v3.16b        \n"
+      "st2         {v0.4s,v1.4s}, [%1], #32      \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
-void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst, int dst_width) {
-  asm volatile (
-    // change the stride to row 2 pointer
-    "add        %1, %1, %0                     \n"
-  "1:                                          \n"
-    MEMACCESS (0)
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"  // load 8 ARGB pixels.
-    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
-    MEMACCESS (1)
-    "ld4        {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8 more ARGB pixels.
-    "uadalp     v0.8h, v16.16b                 \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v17.16b                 \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v18.16b                 \n"  // R 16 bytes -> 8 shorts.
-    "uadalp     v3.8h, v19.16b                 \n"  // A 16 bytes -> 8 shorts.
-    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
-    "rshrn      v1.8b, v1.8h, #2               \n"
-    "rshrn      v2.8b, v2.8h, #2               \n"
-    "rshrn      v3.8b, v3.8h, #2               \n"
-    MEMACCESS (2)
-    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32     \n"
-    "b.gt       1b                             \n"
-  : "+r" (src_ptr),          // %0
-    "+r" (src_stride),       // %1
-    "+r" (dst),              // %2
-    "+r" (dst_width)         // %3
-  :
-  : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19"
-  );
+void ScaleARGBRowDown2Box_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst,
+                               int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "uaddlp      v3.8h, v3.16b                 \n"  // A 16 bytes -> 8 shorts.
+      "ld4         {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8
+      "uadalp      v0.8h, v16.16b                \n"  // B 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v17.16b                \n"  // G 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uadalp      v2.8h, v18.16b                \n"  // R 16 bytes -> 8 shorts.
+      "uadalp      v3.8h, v19.16b                \n"  // A 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "rshrn       v0.8b, v0.8h, #2              \n"  // round and pack
+      "rshrn       v1.8b, v1.8h, #2              \n"
+      "rshrn       v2.8b, v2.8h, #2              \n"
+      "rshrn       v3.8b, v3.8h, #2              \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
 }
 
 // Reads 4 pixels at a time.
 // Alignment requirement: src_argb 4 byte aligned.
-void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
-                               int src_stepx, uint8* dst_argb, int dst_width) {
-  asm volatile (
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.s}[0], [%0], %3            \n"
-    MEMACCESS(0)
-    "ld1        {v0.s}[1], [%0], %3            \n"
-    MEMACCESS(0)
-    "ld1        {v0.s}[2], [%0], %3            \n"
-    MEMACCESS(0)
-    "ld1        {v0.s}[3], [%0], %3            \n"
-    "subs       %w2, %w2, #4                   \n"  // 4 pixels per loop.
-    MEMACCESS(1)
-    "st1        {v0.16b}, [%1], #16            \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),    // %0
-    "+r"(dst_argb),    // %1
-    "+r"(dst_width)    // %2
-  : "r"((int64)(src_stepx * 4)) // %3
-  : "memory", "cc", "v0"
-  );
+void ScaleARGBRowDownEven_NEON(const uint8_t* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8_t* dst_argb,
+                               int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.s}[0], [%0], %3           \n"
+      "ld1         {v0.s}[1], [%0], %3           \n"
+      "ld1         {v0.s}[2], [%0], %3           \n"
+      "ld1         {v0.s}[3], [%0], %3           \n"
+      "subs        %w2, %w2, #4                  \n"  // 4 pixels per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v0.16b}, [%1], #16           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),                // %0
+        "+r"(dst_argb),                // %1
+        "+r"(dst_width)                // %2
+      : "r"((int64_t)(src_stepx * 4))  // %3
+      : "memory", "cc", "v0");
 }
 
 // Reads 4 pixels at a time.
 // Alignment requirement: src_argb 4 byte aligned.
 // TODO(Yang Zhang): Might be worth another optimization pass in future.
 // It could be upgraded to 8 pixels at a time to start with.
-void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+void ScaleARGBRowDownEvenBox_NEON(const uint8_t* src_argb,
+                                  ptrdiff_t src_stride,
                                   int src_stepx,
-                                  uint8* dst_argb, int dst_width) {
-  asm volatile (
-    "add        %1, %1, %0                     \n"
-  "1:                                          \n"
-    MEMACCESS(0)
-    "ld1        {v0.8b}, [%0], %4              \n"  // Read 4 2x2 blocks -> 2x1
-    MEMACCESS(1)
-    "ld1        {v1.8b}, [%1], %4              \n"
-    MEMACCESS(0)
-    "ld1        {v2.8b}, [%0], %4              \n"
-    MEMACCESS(1)
-    "ld1        {v3.8b}, [%1], %4              \n"
-    MEMACCESS(0)
-    "ld1        {v4.8b}, [%0], %4              \n"
-    MEMACCESS(1)
-    "ld1        {v5.8b}, [%1], %4              \n"
-    MEMACCESS(0)
-    "ld1        {v6.8b}, [%0], %4              \n"
-    MEMACCESS(1)
-    "ld1        {v7.8b}, [%1], %4              \n"
-    "uaddl      v0.8h, v0.8b, v1.8b            \n"
-    "uaddl      v2.8h, v2.8b, v3.8b            \n"
-    "uaddl      v4.8h, v4.8b, v5.8b            \n"
-    "uaddl      v6.8h, v6.8b, v7.8b            \n"
-    "mov        v16.d[1], v0.d[1]              \n"  // ab_cd -> ac_bd
-    "mov        v0.d[1], v2.d[0]               \n"
-    "mov        v2.d[0], v16.d[1]              \n"
-    "mov        v16.d[1], v4.d[1]              \n"  // ef_gh -> eg_fh
-    "mov        v4.d[1], v6.d[0]               \n"
-    "mov        v6.d[0], v16.d[1]              \n"
-    "add        v0.8h, v0.8h, v2.8h            \n"  // (a+b)_(c+d)
-    "add        v4.8h, v4.8h, v6.8h            \n"  // (e+f)_(g+h)
-    "rshrn      v0.8b, v0.8h, #2               \n"  // first 2 pixels.
-    "rshrn2     v0.16b, v4.8h, #2              \n"  // next 2 pixels.
-    "subs       %w3, %w3, #4                   \n"  // 4 pixels per loop.
-    MEMACCESS(2)
-    "st1     {v0.16b}, [%2], #16               \n"
-    "b.gt       1b                             \n"
-  : "+r"(src_argb),    // %0
-    "+r"(src_stride),  // %1
-    "+r"(dst_argb),    // %2
-    "+r"(dst_width)    // %3
-  : "r"((int64)(src_stepx * 4)) // %4
-  : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
-  );
+                                  uint8_t* dst_argb,
+                                  int dst_width) {
+  asm volatile(
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "ld1         {v0.8b}, [%0], %4             \n"  // Read 4 2x2 -> 2x1
+      "ld1         {v1.8b}, [%1], %4             \n"
+      "ld1         {v2.8b}, [%0], %4             \n"
+      "ld1         {v3.8b}, [%1], %4             \n"
+      "ld1         {v4.8b}, [%0], %4             \n"
+      "ld1         {v5.8b}, [%1], %4             \n"
+      "ld1         {v6.8b}, [%0], %4             \n"
+      "ld1         {v7.8b}, [%1], %4             \n"
+      "uaddl       v0.8h, v0.8b, v1.8b           \n"
+      "uaddl       v2.8h, v2.8b, v3.8b           \n"
+      "uaddl       v4.8h, v4.8b, v5.8b           \n"
+      "uaddl       v6.8h, v6.8b, v7.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "mov         v16.d[1], v0.d[1]             \n"  // ab_cd -> ac_bd
+      "mov         v0.d[1], v2.d[0]              \n"
+      "mov         v2.d[0], v16.d[1]             \n"
+      "mov         v16.d[1], v4.d[1]             \n"  // ef_gh -> eg_fh
+      "mov         v4.d[1], v6.d[0]              \n"
+      "mov         v6.d[0], v16.d[1]             \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "add         v0.8h, v0.8h, v2.8h           \n"  // (a+b)_(c+d)
+      "add         v4.8h, v4.8h, v6.8h           \n"  // (e+f)_(g+h)
+      "rshrn       v0.8b, v0.8h, #2              \n"  // first 2 pixels.
+      "rshrn2      v0.16b, v4.8h, #2             \n"  // next 2 pixels.
+      "subs        %w3, %w3, #4                  \n"  // 4 pixels per loop.
+      "st1         {v0.16b}, [%2], #16           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),                // %0
+        "+r"(src_stride),              // %1
+        "+r"(dst_argb),                // %2
+        "+r"(dst_width)                // %3
+      : "r"((int64_t)(src_stepx * 4))  // %4
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
 }
 
 // TODO(Yang Zhang): Investigate less load instructions for
 // the x/dx stepping
-#define LOAD1_DATA32_LANE(vn, n)                               \
-    "lsr        %5, %3, #16                    \n"             \
-    "add        %6, %1, %5, lsl #2             \n"             \
-    "add        %3, %3, %4                     \n"             \
-    MEMACCESS(6)                                               \
-    "ld1        {"#vn".s}["#n"], [%6]          \n"
-
-void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx) {
-  const uint8* src_tmp = src_argb;
-  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
-  int64 x64 = (int64) x;
-  int64 dx64 = (int64) dx;
-  int64 tmp64;
-  asm volatile (
-  "1:                                          \n"
-    LOAD1_DATA32_LANE(v0, 0)
-    LOAD1_DATA32_LANE(v0, 1)
-    LOAD1_DATA32_LANE(v0, 2)
-    LOAD1_DATA32_LANE(v0, 3)
-    LOAD1_DATA32_LANE(v1, 0)
-    LOAD1_DATA32_LANE(v1, 1)
-    LOAD1_DATA32_LANE(v1, 2)
-    LOAD1_DATA32_LANE(v1, 3)
-
-    MEMACCESS(0)
-    "st1        {v0.4s, v1.4s}, [%0], #32      \n"  // store pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-    "b.gt        1b                            \n"
-  : "+r"(dst_argb),     // %0
-    "+r"(src_argb),     // %1
-    "+r"(dst_width64),  // %2
-    "+r"(x64),          // %3
-    "+r"(dx64),         // %4
-    "=&r"(tmp64),       // %5
-    "+r"(src_tmp)       // %6
-  :
-  : "memory", "cc", "v0", "v1"
-  );
+#define LOAD1_DATA32_LANE(vn, n)                 \
+  "lsr        %5, %3, #16                    \n" \
+  "add        %6, %1, %5, lsl #2             \n" \
+  "add        %3, %3, %4                     \n" \
+  "ld1        {" #vn ".s}[" #n "], [%6]      \n"
+
+void ScaleARGBCols_NEON(uint8_t* dst_argb,
+                        const uint8_t* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx) {
+  const uint8_t* src_tmp = src_argb;
+  int64_t x64 = (int64_t)x;    // NOLINT
+  int64_t dx64 = (int64_t)dx;  // NOLINT
+  int64_t tmp64;
+  asm volatile(
+      "1:                                        \n"
+      // clang-format off
+      LOAD1_DATA32_LANE(v0, 0)
+      LOAD1_DATA32_LANE(v0, 1)
+      LOAD1_DATA32_LANE(v0, 2)
+      LOAD1_DATA32_LANE(v0, 3)
+      LOAD1_DATA32_LANE(v1, 0)
+      LOAD1_DATA32_LANE(v1, 1)
+      LOAD1_DATA32_LANE(v1, 2)
+      LOAD1_DATA32_LANE(v1, 3)
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      // clang-format on
+      "st1         {v0.4s, v1.4s}, [%0], #32     \n"  // store pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "b.gt        1b                            \n"
+      : "+r"(dst_argb),   // %0
+        "+r"(src_argb),   // %1
+        "+r"(dst_width),  // %2
+        "+r"(x64),        // %3
+        "+r"(dx64),       // %4
+        "=&r"(tmp64),     // %5
+        "+r"(src_tmp)     // %6
+      :
+      : "memory", "cc", "v0", "v1");
 }
 
 #undef LOAD1_DATA32_LANE
 
 // TODO(Yang Zhang): Investigate less load instructions for
 // the x/dx stepping
-#define LOAD2_DATA32_LANE(vn1, vn2, n)                         \
-    "lsr        %5, %3, #16                           \n"      \
-    "add        %6, %1, %5, lsl #2                    \n"      \
-    "add        %3, %3, %4                            \n"      \
-    MEMACCESS(6)                                               \
-    "ld2        {"#vn1".s, "#vn2".s}["#n"], [%6]      \n"
-
-void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
-                              int dst_width, int x, int dx) {
+#define LOAD2_DATA32_LANE(vn1, vn2, n)                  \
+  "lsr        %5, %3, #16                           \n" \
+  "add        %6, %1, %5, lsl #2                    \n" \
+  "add        %3, %3, %4                            \n" \
+  "ld2        {" #vn1 ".s, " #vn2 ".s}[" #n "], [%6]  \n"
+
+void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
+                              const uint8_t* src_argb,
+                              int dst_width,
+                              int x,
+                              int dx) {
   int dx_offset[4] = {0, 1, 2, 3};
   int* tmp = dx_offset;
-  const uint8* src_tmp = src_argb;
-  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
-  int64 x64 = (int64) x;
-  int64 dx64 = (int64) dx;
+  const uint8_t* src_tmp = src_argb;
+  int64_t x64 = (int64_t)x;    // NOLINT
+  int64_t dx64 = (int64_t)dx;  // NOLINT
   asm volatile (
-    "dup        v0.4s, %w3                     \n"  // x
-    "dup        v1.4s, %w4                     \n"  // dx
-    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
-    "shl        v6.4s, v1.4s, #2               \n"  // 4 * dx
-    "mul        v1.4s, v1.4s, v2.4s            \n"
-    "movi       v3.16b, #0x7f                  \n"  // 0x7F
-    "movi       v4.8h, #0x7f                   \n"  // 0x7F
+      "dup         v0.4s, %w3                    \n"  // x
+      "dup         v1.4s, %w4                    \n"  // dx
+      "ld1         {v2.4s}, [%5]                 \n"  // 0 1 2 3
+      "shl         v6.4s, v1.4s, #2              \n"  // 4 * dx
+      "mul         v1.4s, v1.4s, v2.4s           \n"
+      "movi        v3.16b, #0x7f                 \n"  // 0x7F
+      "movi        v4.8h, #0x7f                  \n"  // 0x7F
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "add        v5.4s, v1.4s, v0.4s            \n"
-  "1:                                          \n"
+      "add         v5.4s, v1.4s, v0.4s           \n"
+      "1:                                        \n"
     // d0, d1: a
     // d2, d3: b
     LOAD2_DATA32_LANE(v0, v1, 0)
@@ -1009,19 +1461,18 @@ void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
     "umull2     v17.8h, v0.16b, v7.16b         \n"
     "umull      v18.8h, v1.8b, v2.8b           \n"
     "umull2     v19.8h, v1.16b, v2.16b         \n"
+    "prfm       pldl1keep, [%1, 448]           \n"  // prefetch 7 lines ahead
     "add        v16.8h, v16.8h, v18.8h         \n"
     "add        v17.8h, v17.8h, v19.8h         \n"
     "shrn       v0.8b, v16.8h, #7              \n"
     "shrn2      v0.16b, v17.8h, #7             \n"
-
-    MEMACCESS(0)
     "st1     {v0.4s}, [%0], #16                \n"  // store pixels
     "add     v5.4s, v5.4s, v6.4s               \n"
     "subs    %w2, %w2, #4                      \n"  // 4 processed per loop
-    "b.gt    1b                                \n"
+    "b.gt       1b                             \n"
   : "+r"(dst_argb),         // %0
     "+r"(src_argb),         // %1
-    "+r"(dst_width64),      // %2
+    "+r"(dst_width),        // %2
     "+r"(x64),              // %3
     "+r"(dx64),             // %4
     "+r"(tmp),              // %5
@@ -1034,6 +1485,147 @@ void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
 
 #undef LOAD2_DATA32_LANE
 
+// Read 16x2 average down and write 8x1.
+void ScaleRowDown2Box_16_NEON(const uint16_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint16_t* dst,
+                              int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %0, %1, lsl #1            \n"  // ptr + stide * 2
+      "1:                                        \n"
+      "ld1         {v0.8h, v1.8h}, [%0], #32     \n"  // load row 1 and post inc
+      "ld1         {v2.8h, v3.8h}, [%1], #32     \n"  // load row 2 and post inc
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop
+      "uaddlp      v0.4s, v0.8h                  \n"  // row 1 add adjacent
+      "uaddlp      v1.4s, v1.8h                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uadalp      v0.4s, v2.8h                  \n"  // +row 2 add adjacent
+      "uadalp      v1.4s, v3.8h                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "rshrn       v0.4h, v0.4s, #2              \n"  // round and pack
+      "rshrn2      v0.8h, v1.4s, #2              \n"
+      "st1         {v0.8h}, [%2], #16            \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+// Read 8x2 upsample with filtering and write 16x1.
+// Actually reads an extra pixel, so 9x2.
+void ScaleRowUp2_16_NEON(const uint16_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint16_t* dst,
+                         int dst_width) {
+  asm volatile(
+      "add         %1, %0, %1, lsl #1            \n"  // ptr + stide * 2
+      "movi        v0.8h, #9                     \n"  // constants
+      "movi        v1.4s, #3                     \n"
+
+      "1:                                        \n"
+      "ld1         {v3.8h}, [%0], %4             \n"  // TL read first 8
+      "ld1         {v4.8h}, [%0], %5             \n"  // TR read 8 offset by 1
+      "ld1         {v5.8h}, [%1], %4             \n"  // BL read 8 from next row
+      "ld1         {v6.8h}, [%1], %5             \n"  // BR offset by 1
+      "subs        %w3, %w3, #16                 \n"  // 16 dst pixels per loop
+      "umull       v16.4s, v3.4h, v0.4h          \n"
+      "umull2      v7.4s, v3.8h, v0.8h           \n"
+      "umull       v18.4s, v4.4h, v0.4h          \n"
+      "umull2      v17.4s, v4.8h, v0.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uaddw       v16.4s, v16.4s, v6.4h         \n"
+      "uaddl2      v19.4s, v6.8h, v3.8h          \n"
+      "uaddl       v3.4s, v6.4h, v3.4h           \n"
+      "uaddw2      v6.4s, v7.4s, v6.8h           \n"
+      "uaddl2      v7.4s, v5.8h, v4.8h           \n"
+      "uaddl       v4.4s, v5.4h, v4.4h           \n"
+      "uaddw       v18.4s, v18.4s, v5.4h         \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "mla         v16.4s, v4.4s, v1.4s          \n"
+      "mla         v18.4s, v3.4s, v1.4s          \n"
+      "mla         v6.4s, v7.4s, v1.4s           \n"
+      "uaddw2      v4.4s, v17.4s, v5.8h          \n"
+      "uqrshrn     v16.4h,  v16.4s, #4           \n"
+      "mla         v4.4s, v19.4s, v1.4s          \n"
+      "uqrshrn2    v16.8h, v6.4s, #4             \n"
+      "uqrshrn     v17.4h, v18.4s, #4            \n"
+      "uqrshrn2    v17.8h, v4.4s, #4             \n"
+      "st2         {v16.8h-v17.8h}, [%2], #32    \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      : "r"(2LL),          // %4
+        "r"(14LL)          // %5
+      : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
+        "v19"  // Clobber List
+  );
+}
+
+void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 UV
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uaddlp      v0.8h, v0.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "uaddlp      v1.8h, v1.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "ld2         {v16.16b,v17.16b}, [%1], #32  \n"  // load 16
+      "uadalp      v0.8h, v16.16b                \n"  // U 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v17.16b                \n"  // V 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "rshrn       v0.8b, v0.8h, #2              \n"  // round and pack
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "rshrn       v1.8b, v1.8h, #2              \n"
+      "st2         {v0.8b,v1.8b}, [%2], #16      \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "v0", "v1", "v16", "v17");
+}
+
+// Reads 4 pixels at a time.
+void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             int src_stepx,  // pixel step
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src1_ptr = src_ptr + src_stepx * 2;
+  const uint8_t* src2_ptr = src_ptr + src_stepx * 4;
+  const uint8_t* src3_ptr = src_ptr + src_stepx * 6;
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.h}[0], [%0], %6           \n"
+      "ld1         {v1.h}[0], [%1], %6           \n"
+      "ld1         {v2.h}[0], [%2], %6           \n"
+      "ld1         {v3.h}[0], [%3], %6           \n"
+      "subs        %w5, %w5, #4                  \n"  // 4 pixels per loop.
+      "st4         {v0.h, v1.h, v2.h, v3.h}[0], [%4], #8 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),                 // %0
+        "+r"(src1_ptr),                // %1
+        "+r"(src2_ptr),                // %2
+        "+r"(src3_ptr),                // %3
+        "+r"(dst_ptr),                 // %4
+        "+r"(dst_width)                // %5
+      : "r"((int64_t)(src_stepx * 8))  // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3");
+}
+
 #endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 #ifdef __cplusplus
diff --git a/media/libyuv/source/scale_rgb.cc b/media/libyuv/source/scale_rgb.cc
new file mode 100644
index 0000000000..8db59b56fc
--- /dev/null
+++ b/media/libyuv/source/scale_rgb.cc
@@ -0,0 +1,66 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h" /* For FilterMode */
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/convert_argb.h"
+#include "libyuv/convert_from_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/scale_argb.h"
+#include "libyuv/scale_rgb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Scale a 24 bit image.
+// Converts to ARGB as intermediate step
+
+LIBYUV_API
+int RGBScale(const uint8_t* src_rgb,
+             int src_stride_rgb,
+             int src_width,
+             int src_height,
+             uint8_t* dst_rgb,
+             int dst_stride_rgb,
+             int dst_width,
+             int dst_height,
+             enum FilterMode filtering) {
+  int r;
+  uint8_t* src_argb =
+      (uint8_t*)malloc(src_width * src_height * 4 + dst_width * dst_height * 4);
+  uint8_t* dst_argb = src_argb + src_width * src_height * 4;
+
+  if (!src_argb) {
+    return 1;
+  }
+
+  r = RGB24ToARGB(src_rgb, src_stride_rgb, src_argb, src_width * 4, src_width,
+                  src_height);
+  if (!r) {
+    r = ARGBScale(src_argb, src_width * 4, src_width, src_height, dst_argb,
+                  dst_width * 4, dst_width, dst_height, filtering);
+    if (!r) {
+      r = ARGBToRGB24(dst_argb, dst_width * 4, dst_rgb, dst_stride_rgb,
+                      dst_width, dst_height);
+    }
+  }
+  free(src_argb);
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/scale_uv.cc b/media/libyuv/source/scale_uv.cc
new file mode 100644
index 0000000000..5b92d04321
--- /dev/null
+++ b/media/libyuv/source/scale_uv.cc
@@ -0,0 +1,1161 @@
+/*
+ *  Copyright 2020 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyUV
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Macros to enable specialized scalers
+
+#ifndef HAS_SCALEUVDOWN2
+#define HAS_SCALEUVDOWN2 1
+#endif
+#ifndef HAS_SCALEUVDOWN4BOX
+#define HAS_SCALEUVDOWN4BOX 1
+#endif
+#ifndef HAS_SCALEUVDOWNEVEN
+#define HAS_SCALEUVDOWNEVEN 1
+#endif
+#ifndef HAS_SCALEUVBILINEARDOWN
+#define HAS_SCALEUVBILINEARDOWN 1
+#endif
+#ifndef HAS_SCALEUVBILINEARUP
+#define HAS_SCALEUVBILINEARUP 1
+#endif
+#ifndef HAS_UVCOPY
+#define HAS_UVCOPY 1
+#endif
+#ifndef HAS_SCALEPLANEVERTICAL
+#define HAS_SCALEPLANEVERTICAL 1
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// ScaleUV, 1/2
+// This is an optimized version for scaling down a UV to 1/2 of
+// its original size.
+#if HAS_SCALEUVDOWN2
+static void ScaleUVDown2(int src_width,
+                         int src_height,
+                         int dst_width,
+                         int dst_height,
+                         int src_stride,
+                         int dst_stride,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_uv,
+                         int x,
+                         int dx,
+                         int y,
+                         int dy,
+                         enum FilterMode filtering) {
+  int j;
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleUVRowDown2)(const uint8_t* src_uv, ptrdiff_t src_stride,
+                          uint8_t* dst_uv, int dst_width) =
+      filtering == kFilterNone
+          ? ScaleUVRowDown2_C
+          : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_C
+                                        : ScaleUVRowDown2Box_C);
+  (void)src_width;
+  (void)src_height;
+  (void)dx;
+  assert(dx == 65536 * 2);      // Test scale factor of 2.
+  assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
+  // Advance to odd row, even column.
+  if (filtering == kFilterBilinear) {
+    src_uv += (y >> 16) * (intptr_t)src_stride + (x >> 16) * 2;
+  } else {
+    src_uv += (y >> 16) * (intptr_t)src_stride + ((x >> 16) - 1) * 2;
+  }
+
+#if defined(HAS_SCALEUVROWDOWN2BOX_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && filtering) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && filtering) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && filtering) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_NEON;
+    }
+  }
+#endif
+
+// This code is not enabled.  Only box filter is available at this time.
+#if defined(HAS_SCALEUVROWDOWN2_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleUVRowDown2 =
+        filtering == kFilterNone
+            ? ScaleUVRowDown2_Any_SSSE3
+            : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_Any_SSSE3
+                                          : ScaleUVRowDown2Box_Any_SSSE3);
+    if (IS_ALIGNED(dst_width, 2)) {
+      ScaleUVRowDown2 =
+          filtering == kFilterNone
+              ? ScaleUVRowDown2_SSSE3
+              : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_SSSE3
+                                            : ScaleUVRowDown2Box_SSSE3);
+    }
+  }
+#endif
+// This code is not enabled.  Only box filter is available at this time.
+#if defined(HAS_SCALEUVROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVRowDown2 =
+        filtering == kFilterNone
+            ? ScaleUVRowDown2_Any_NEON
+            : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_Any_NEON
+                                          : ScaleUVRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 =
+          filtering == kFilterNone
+              ? ScaleUVRowDown2_NEON
+              : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_NEON
+                                            : ScaleUVRowDown2Box_NEON);
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVRowDown2 =
+        filtering == kFilterNone
+            ? ScaleUVRowDown2_Any_MSA
+            : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_Any_MSA
+                                          : ScaleUVRowDown2Box_Any_MSA);
+    if (IS_ALIGNED(dst_width, 2)) {
+      ScaleUVRowDown2 =
+          filtering == kFilterNone
+              ? ScaleUVRowDown2_MSA
+              : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_MSA
+                                            : ScaleUVRowDown2Box_MSA);
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVRowDown2(src_uv, src_stride, dst_uv, dst_width);
+    src_uv += row_stride;
+    dst_uv += dst_stride;
+  }
+}
+#endif  // HAS_SCALEUVDOWN2
+
+// ScaleUV, 1/4
+// This is an optimized version for scaling down a UV to 1/4 of
+// its original size.
+#if HAS_SCALEUVDOWN4BOX
+static void ScaleUVDown4Box(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_uv,
+                            uint8_t* dst_uv,
+                            int x,
+                            int dx,
+                            int y,
+                            int dy) {
+  int j;
+  // Allocate 2 rows of UV.
+  const int row_size = (dst_width * 2 * 2 + 15) & ~15;
+  align_buffer_64(row, row_size * 2);
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleUVRowDown2)(const uint8_t* src_uv, ptrdiff_t src_stride,
+                          uint8_t* dst_uv, int dst_width) =
+      ScaleUVRowDown2Box_C;
+  // Advance to odd row, even column.
+  src_uv += (y >> 16) * (intptr_t)src_stride + (x >> 16) * 2;
+  (void)src_width;
+  (void)src_height;
+  (void)dx;
+  assert(dx == 65536 * 4);      // Test scale factor of 4.
+  assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
+
+#if defined(HAS_SCALEUVROWDOWN2BOX_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_NEON;
+    }
+  }
+#endif
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVRowDown2(src_uv, src_stride, row, dst_width * 2);
+    ScaleUVRowDown2(src_uv + src_stride * 2, src_stride, row + row_size,
+                    dst_width * 2);
+    ScaleUVRowDown2(row, row_size, dst_uv, dst_width);
+    src_uv += row_stride;
+    dst_uv += dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+#endif  // HAS_SCALEUVDOWN4BOX
+
+// ScaleUV Even
+// This is an optimized version for scaling down a UV to even
+// multiple of its original size.
+#if HAS_SCALEUVDOWNEVEN
+static void ScaleUVDownEven(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_uv,
+                            uint8_t* dst_uv,
+                            int x,
+                            int dx,
+                            int y,
+                            int dy,
+                            enum FilterMode filtering) {
+  int j;
+  int col_step = dx >> 16;
+  ptrdiff_t row_stride = (ptrdiff_t)((dy >> 16) * (intptr_t)src_stride);
+  void (*ScaleUVRowDownEven)(const uint8_t* src_uv, ptrdiff_t src_stride,
+                             int src_step, uint8_t* dst_uv, int dst_width) =
+      filtering ? ScaleUVRowDownEvenBox_C : ScaleUVRowDownEven_C;
+  (void)src_width;
+  (void)src_height;
+  assert(IS_ALIGNED(src_width, 2));
+  assert(IS_ALIGNED(src_height, 2));
+  src_uv += (y >> 16) * (intptr_t)src_stride + (x >> 16) * 2;
+#if defined(HAS_SCALEUVROWDOWNEVEN_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleUVRowDownEven = filtering ? ScaleUVRowDownEvenBox_Any_SSSE3
+                                   : ScaleUVRowDownEven_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven =
+          filtering ? ScaleUVRowDownEvenBox_SSE2 : ScaleUVRowDownEven_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && !filtering) {
+    ScaleUVRowDownEven = ScaleUVRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven = ScaleUVRowDownEven_NEON;
+    }
+  }
+#endif  // TODO(fbarchard): Enable Box filter
+#if defined(HAS_SCALEUVROWDOWNEVENBOX_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVRowDownEven = filtering ? ScaleUVRowDownEvenBox_Any_NEON
+                                   : ScaleUVRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven =
+          filtering ? ScaleUVRowDownEvenBox_NEON : ScaleUVRowDownEven_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWNEVEN_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVRowDownEven =
+        filtering ? ScaleUVRowDownEvenBox_Any_MSA : ScaleUVRowDownEven_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven =
+          filtering ? ScaleUVRowDownEvenBox_MSA : ScaleUVRowDownEven_MSA;
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVRowDownEven(src_uv, src_stride, col_step, dst_uv, dst_width);
+    src_uv += row_stride;
+    dst_uv += dst_stride;
+  }
+}
+#endif
+
+// Scale UV down with bilinear interpolation.
+#if HAS_SCALEUVBILINEARDOWN
+static void ScaleUVBilinearDown(int src_width,
+                                int src_height,
+                                int dst_width,
+                                int dst_height,
+                                int src_stride,
+                                int dst_stride,
+                                const uint8_t* src_uv,
+                                uint8_t* dst_uv,
+                                int x,
+                                int dx,
+                                int y,
+                                int dy,
+                                enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8_t* dst_uv, const uint8_t* src_uv,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleUVFilterCols)(uint8_t* dst_uv, const uint8_t* src_uv,
+                            int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleUVFilterCols64_C : ScaleUVFilterCols_C;
+  int64_t xlast = x + (int64_t)(dst_width - 1) * dx;
+  int64_t xl = (dx >= 0) ? x : xlast;
+  int64_t xr = (dx >= 0) ? xlast : x;
+  int clip_src_width;
+  xl = (xl >> 16) & ~3;    // Left edge aligned.
+  xr = (xr >> 16) + 1;     // Right most pixel used.  Bilinear uses 2 pixels.
+  xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
+  if (xr > src_width) {
+    xr = src_width;
+  }
+  clip_src_width = (int)(xr - xl) * 2;  // Width aligned to 2.
+  src_uv += xl * 2;
+  x -= (int)(xl << 16);
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVFilterCols = ScaleUVFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_MSA;
+    }
+  }
+#endif
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row of UV.
+  {
+    align_buffer_64(row, clip_src_width * 2);
+
+    const int max_y = (src_height - 1) << 16;
+    if (y > max_y) {
+      y = max_y;
+    }
+    for (j = 0; j < dst_height; ++j) {
+      int yi = y >> 16;
+      const uint8_t* src = src_uv + yi * (intptr_t)src_stride;
+      if (filtering == kFilterLinear) {
+        ScaleUVFilterCols(dst_uv, src, dst_width, x, dx);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(row, src, src_stride, clip_src_width, yf);
+        ScaleUVFilterCols(dst_uv, row, dst_width, x, dx);
+      }
+      dst_uv += dst_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+#endif
+
+// Scale UV up with bilinear interpolation.
+#if HAS_SCALEUVBILINEARUP
+static void ScaleUVBilinearUp(int src_width,
+                              int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint8_t* src_uv,
+                              uint8_t* dst_uv,
+                              int x,
+                              int dx,
+                              int y,
+                              int dy,
+                              enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8_t* dst_uv, const uint8_t* src_uv,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleUVFilterCols)(uint8_t* dst_uv, const uint8_t* src_uv,
+                            int dst_width, int x, int dx) =
+      filtering ? ScaleUVFilterCols_C : ScaleUVCols_C;
+  const int max_y = (src_height - 1) << 16;
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
+  if (src_width >= 32768) {
+    ScaleUVFilterCols = filtering ? ScaleUVFilterCols64_C : ScaleUVCols64_C;
+  }
+#if defined(HAS_SCALEUVFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVFilterCols = ScaleUVFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_SSSE3)
+  if (!filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVFilterCols = ScaleUVCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_NEON)
+  if (!filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVFilterCols = ScaleUVCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleUVFilterCols = ScaleUVCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_MSA)
+  if (!filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVFilterCols = ScaleUVCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVCols_MSA;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleUVFilterCols = ScaleUVColsUp2_C;
+#if defined(HAS_SCALEUVCOLSUP2_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVColsUp2_SSSE3;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  {
+    int yi = y >> 16;
+    const uint8_t* src = src_uv + yi * (intptr_t)src_stride;
+
+    // Allocate 2 rows of UV.
+    const int row_size = (dst_width * 2 + 15) & ~15;
+    align_buffer_64(row, row_size * 2);
+
+    uint8_t* rowptr = row;
+    int rowstride = row_size;
+    int lasty = yi;
+
+    ScaleUVFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleUVFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    if (src_height > 2) {
+      src += src_stride;
+    }
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_uv + yi * (intptr_t)src_stride;
+        }
+        if (yi != lasty) {
+          ScaleUVFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_uv, rowptr, 0, dst_width * 2, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_uv, rowptr, rowstride, dst_width * 2, yf);
+      }
+      dst_uv += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+#endif  // HAS_SCALEUVBILINEARUP
+
+// Scale UV, horizontally up by 2 times.
+// Uses linear filter horizontally, nearest vertically.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// This is used to scale U and V planes of NV16 to NV24.
+void ScaleUVLinearUp2(int src_width,
+                      int src_height,
+                      int dst_width,
+                      int dst_height,
+                      int src_stride,
+                      int dst_stride,
+                      const uint8_t* src_uv,
+                      uint8_t* dst_uv) {
+  void (*ScaleRowUp)(const uint8_t* src_uv, uint8_t* dst_uv, int dst_width) =
+      ScaleUVRowUp2_Linear_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_uv + ((src_height - 1) / 2) * (intptr_t)src_stride, dst_uv,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_uv + (y >> 16) * (intptr_t)src_stride, dst_uv, dst_width);
+      dst_uv += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale plane, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// This is used to scale U and V planes of NV12 to NV24.
+void ScaleUVBilinearUp2(int src_width,
+                        int src_height,
+                        int dst_width,
+                        int dst_height,
+                        int src_stride,
+                        int dst_stride,
+                        const uint8_t* src_ptr,
+                        uint8_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                      uint8_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleUVRowUp2_Bilinear_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    // TODO(fbarchard): Test performance of writing one row of destination at a
+    // time.
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+// Scale 16 bit UV, horizontally up by 2 times.
+// Uses linear filter horizontally, nearest vertically.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// This is used to scale U and V planes of P210 to P410.
+void ScaleUVLinearUp2_16(int src_width,
+                         int src_height,
+                         int dst_width,
+                         int dst_height,
+                         int src_stride,
+                         int dst_stride,
+                         const uint16_t* src_uv,
+                         uint16_t* dst_uv) {
+  void (*ScaleRowUp)(const uint16_t* src_uv, uint16_t* dst_uv, int dst_width) =
+      ScaleUVRowUp2_Linear_16_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_uv + ((src_height - 1) / 2) * (intptr_t)src_stride, dst_uv,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_uv + (y >> 16) * (intptr_t)src_stride, dst_uv, dst_width);
+      dst_uv += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale 16 bit UV, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// This is used to scale U and V planes of P010 to P410.
+void ScaleUVBilinearUp2_16(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint16_t* src_ptr,
+                           uint16_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleUVRowUp2_Bilinear_16_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    // TODO(fbarchard): Test performance of writing one row of destination at a
+    // time.
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+// Scale UV to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScaleUVSimple(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          int src_stride,
+                          int dst_stride,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_uv,
+                          int x,
+                          int dx,
+                          int y,
+                          int dy) {
+  int j;
+  void (*ScaleUVCols)(uint8_t* dst_uv, const uint8_t* src_uv, int dst_width,
+                      int x, int dx) =
+      (src_width >= 32768) ? ScaleUVCols64_C : ScaleUVCols_C;
+  (void)src_height;
+#if defined(HAS_SCALEUVCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVCols = ScaleUVCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVCols = ScaleUVCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVCols = ScaleUVCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVCols = ScaleUVCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVCols = ScaleUVCols_MSA;
+    }
+  }
+#endif
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleUVCols = ScaleUVColsUp2_C;
+#if defined(HAS_SCALEUVCOLSUP2_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(dst_width, 8)) {
+      ScaleUVCols = ScaleUVColsUp2_SSSE3;
+    }
+#endif
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVCols(dst_uv, src_uv + (y >> 16) * (intptr_t)src_stride, dst_width, x,
+                dx);
+    dst_uv += dst_stride;
+    y += dy;
+  }
+}
+
+// Copy UV with optional flipping
+#if HAS_UVCOPY
+static int UVCopy(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_uv,
+                  int dst_stride_uv,
+                  int width,
+                  int height) {
+  if (!src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * (intptr_t)src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  CopyPlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, width * 2, height);
+  return 0;
+}
+
+static int UVCopy_16(const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint16_t* dst_uv,
+                     int dst_stride_uv,
+                     int width,
+                     int height) {
+  if (!src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * (intptr_t)src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  CopyPlane_16(src_uv, src_stride_uv, dst_uv, dst_stride_uv, width * 2, height);
+  return 0;
+}
+#endif  // HAS_UVCOPY
+
+// Scale a UV plane (from NV12)
+// This function in turn calls a scaling function
+// suitable for handling the desired resolutions.
+static void ScaleUV(const uint8_t* src,
+                    int src_stride,
+                    int src_width,
+                    int src_height,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int dst_width,
+                    int dst_height,
+                    int clip_x,
+                    int clip_y,
+                    int clip_width,
+                    int clip_height,
+                    enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // UV does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * (intptr_t)src_stride;
+    src_stride = -src_stride;
+  }
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
+  src_width = Abs(src_width);
+  if (clip_x) {
+    int64_t clipf = (int64_t)(clip_x)*dx;
+    x += (clipf & 0xffff);
+    src += (clipf >> 16) * 2;
+    dst += clip_x * 2;
+  }
+  if (clip_y) {
+    int64_t clipf = (int64_t)(clip_y)*dy;
+    y += (clipf & 0xffff);
+    src += (clipf >> 16) * (intptr_t)src_stride;
+    dst += clip_y * dst_stride;
+  }
+
+  // Special case for integer step values.
+  if (((dx | dy) & 0xffff) == 0) {
+    if (!dx || !dy) {  // 1 pixel wide and/or tall.
+      filtering = kFilterNone;
+    } else {
+      // Optimized even scale down. ie 2, 4, 6, 8, 10x.
+      if (!(dx & 0x10000) && !(dy & 0x10000)) {
+#if HAS_SCALEUVDOWN2
+        if (dx == 0x20000) {
+          // Optimized 1/2 downsample.
+          ScaleUVDown2(src_width, src_height, clip_width, clip_height,
+                       src_stride, dst_stride, src, dst, x, dx, y, dy,
+                       filtering);
+          return;
+        }
+#endif
+#if HAS_SCALEUVDOWN4BOX
+        if (dx == 0x40000 && filtering == kFilterBox) {
+          // Optimized 1/4 box downsample.
+          ScaleUVDown4Box(src_width, src_height, clip_width, clip_height,
+                          src_stride, dst_stride, src, dst, x, dx, y, dy);
+          return;
+        }
+#endif
+#if HAS_SCALEUVDOWNEVEN
+        ScaleUVDownEven(src_width, src_height, clip_width, clip_height,
+                        src_stride, dst_stride, src, dst, x, dx, y, dy,
+                        filtering);
+        return;
+#endif
+      }
+      // Optimized odd scale down. ie 3, 5, 7, 9x.
+      if ((dx & 0x10000) && (dy & 0x10000)) {
+        filtering = kFilterNone;
+#ifdef HAS_UVCOPY
+        if (dx == 0x10000 && dy == 0x10000) {
+          // Straight copy.
+          UVCopy(src + (y >> 16) * (intptr_t)src_stride + (x >> 16) * 2,
+                 src_stride, dst, dst_stride, clip_width, clip_height);
+          return;
+        }
+#endif
+      }
+    }
+  }
+  // HAS_SCALEPLANEVERTICAL
+  if (dx == 0x10000 && (x & 0xffff) == 0) {
+    // Arbitrary scale vertically, but unscaled horizontally.
+    ScalePlaneVertical(src_height, clip_width, clip_height, src_stride,
+                       dst_stride, src, dst, x, y, dy, /*bpp=*/2, filtering);
+    return;
+  }
+  if (filtering && (dst_width + 1) / 2 == src_width) {
+    ScaleUVLinearUp2(src_width, src_height, clip_width, clip_height, src_stride,
+                     dst_stride, src, dst);
+    return;
+  }
+  if ((clip_height + 1) / 2 == src_height &&
+      (clip_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScaleUVBilinearUp2(src_width, src_height, clip_width, clip_height,
+                       src_stride, dst_stride, src, dst);
+    return;
+  }
+#if HAS_SCALEUVBILINEARUP
+  if (filtering && dy < 65536) {
+    ScaleUVBilinearUp(src_width, src_height, clip_width, clip_height,
+                      src_stride, dst_stride, src, dst, x, dx, y, dy,
+                      filtering);
+    return;
+  }
+#endif
+#if HAS_SCALEUVBILINEARDOWN
+  if (filtering) {
+    ScaleUVBilinearDown(src_width, src_height, clip_width, clip_height,
+                        src_stride, dst_stride, src, dst, x, dx, y, dy,
+                        filtering);
+    return;
+  }
+#endif
+  ScaleUVSimple(src_width, src_height, clip_width, clip_height, src_stride,
+                dst_stride, src, dst, x, dx, y, dy);
+}
+
+// Scale an UV image.
+LIBYUV_API
+int UVScale(const uint8_t* src_uv,
+            int src_stride_uv,
+            int src_width,
+            int src_height,
+            uint8_t* dst_uv,
+            int dst_stride_uv,
+            int dst_width,
+            int dst_height,
+            enum FilterMode filtering) {
+  if (!src_uv || src_width <= 0 || src_height == 0 || src_width > 32768 ||
+      src_height > 32768 || !dst_uv || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+  ScaleUV(src_uv, src_stride_uv, src_width, src_height, dst_uv, dst_stride_uv,
+          dst_width, dst_height, 0, 0, dst_width, dst_height, filtering);
+  return 0;
+}
+
+// Scale a 16 bit UV image.
+// This function is currently incomplete, it can't handle all cases.
+LIBYUV_API
+int UVScale_16(const uint16_t* src_uv,
+               int src_stride_uv,
+               int src_width,
+               int src_height,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int dst_width,
+               int dst_height,
+               enum FilterMode filtering) {
+  int dy = 0;
+
+  if (!src_uv || src_width <= 0 || src_height == 0 || src_width > 32768 ||
+      src_height > 32768 || !dst_uv || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  // UV does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src_uv = src_uv + (src_height - 1) * (intptr_t)src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+  src_width = Abs(src_width);
+
+#ifdef HAS_UVCOPY
+  if (!filtering && src_width == dst_width && (src_height % dst_height == 0)) {
+    if (dst_height == 1) {
+      UVCopy_16(src_uv + ((src_height - 1) / 2) * (intptr_t)src_stride_uv,
+                src_stride_uv, dst_uv, dst_stride_uv, dst_width, dst_height);
+    } else {
+      dy = src_height / dst_height;
+      UVCopy_16(src_uv + ((dy - 1) / 2) * (intptr_t)src_stride_uv,
+                (int)(dy * (intptr_t)src_stride_uv), dst_uv, dst_stride_uv,
+                dst_width, dst_height);
+    }
+
+    return 0;
+  }
+#endif
+
+  if (filtering && (dst_width + 1) / 2 == src_width) {
+    ScaleUVLinearUp2_16(src_width, src_height, dst_width, dst_height,
+                        src_stride_uv, dst_stride_uv, src_uv, dst_uv);
+    return 0;
+  }
+
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScaleUVBilinearUp2_16(src_width, src_height, dst_width, dst_height,
+                          src_stride_uv, dst_stride_uv, src_uv, dst_uv);
+    return 0;
+  }
+
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/media/libyuv/source/scale_win.cc b/media/libyuv/source/scale_win.cc
index f17097365c..ea1f95c6c3 100644
--- a/media/libyuv/source/scale_win.cc
+++ b/media/libyuv/source/scale_win.cc
@@ -16,98 +16,95 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for 32 bit Visual C x86 and clangcl
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+// This module is for 32 bit Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(_M_IX86)
 
 // Offsets for source bytes 0 to 9
-static uvec8 kShuf0 =
-  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf0 = {0,   1,   3,   4,   5,   7,   8,   9,
+                             128, 128, 128, 128, 128, 128, 128, 128};
 
 // Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
-static uvec8 kShuf1 =
-  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf1 = {3,   4,   5,   7,   8,   9,   11,  12,
+                             128, 128, 128, 128, 128, 128, 128, 128};
 
 // Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-static uvec8 kShuf2 =
-  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf2 = {5,   7,   8,   9,   11,  12,  13,  15,
+                             128, 128, 128, 128, 128, 128, 128, 128};
 
 // Offsets for source bytes 0 to 10
-static uvec8 kShuf01 =
-  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+static const uvec8 kShuf01 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
 
 // Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
-static uvec8 kShuf11 =
-  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+static const uvec8 kShuf11 = {2, 3, 4, 5,  5,  6,  6,  7,
+                              8, 9, 9, 10, 10, 11, 12, 13};
 
 // Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-static uvec8 kShuf21 =
-  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+static const uvec8 kShuf21 = {5,  6,  6,  7,  8,  9,  9,  10,
+                              10, 11, 12, 13, 13, 14, 14, 15};
 
 // Coefficients for source bytes 0 to 10
-static uvec8 kMadd01 =
-  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+static const uvec8 kMadd01 = {3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2};
 
 // Coefficients for source bytes 10 to 21
-static uvec8 kMadd11 =
-  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+static const uvec8 kMadd11 = {1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1};
 
 // Coefficients for source bytes 21 to 31
-static uvec8 kMadd21 =
-  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+static const uvec8 kMadd21 = {2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3};
 
 // Coefficients for source bytes 21 to 31
-static vec16 kRound34 =
-  { 2, 2, 2, 2, 2, 2, 2, 2 };
+static const vec16 kRound34 = {2, 2, 2, 2, 2, 2, 2, 2};
 
-static uvec8 kShuf38a =
-  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShuf38a = {0,   3,   6,   8,   11,  14,  128, 128,
+                               128, 128, 128, 128, 128, 128, 128, 128};
 
-static uvec8 kShuf38b =
-  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+static const uvec8 kShuf38b = {128, 128, 128, 128, 128, 128, 0,   3,
+                               6,   8,   11,  14,  128, 128, 128, 128};
 
 // Arrange words 0,3,6 into 0,1,2
-static uvec8 kShufAc =
-  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAc = {0,   1,   6,   7,   12,  13,  128, 128,
+                              128, 128, 128, 128, 128, 128, 128, 128};
 
 // Arrange words 0,3,6 into 3,4,5
-static uvec8 kShufAc3 =
-  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+static const uvec8 kShufAc3 = {128, 128, 128, 128, 128, 128, 0,   1,
+                               6,   7,   12,  13,  128, 128, 128, 128};
 
 // Scaling values for boxes of 3x3 and 2x3
-static uvec16 kScaleAc33 =
-  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+static const uvec16 kScaleAc33 = {65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9,
+                                  65536 / 9, 65536 / 6, 0,         0};
 
 // Arrange first value for pixels 0,1,2,3,4,5
-static uvec8 kShufAb0 =
-  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAb0 = {0,  128, 3,  128, 6,   128, 8,   128,
+                               11, 128, 14, 128, 128, 128, 128, 128};
 
 // Arrange second value for pixels 0,1,2,3,4,5
-static uvec8 kShufAb1 =
-  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAb1 = {1,  128, 4,  128, 7,   128, 9,   128,
+                               12, 128, 15, 128, 128, 128, 128, 128};
 
 // Arrange third value for pixels 0,1,2,3,4,5
-static uvec8 kShufAb2 =
-  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+static const uvec8 kShufAb2 = {2,  128, 5,   128, 128, 128, 10,  128,
+                               13, 128, 128, 128, 128, 128, 128, 128};
 
 // Scaling values for boxes of 3x2 and 2x2
-static uvec16 kScaleAb2 =
-  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+static const uvec16 kScaleAb2 = {65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3,
+                                 65536 / 3, 65536 / 2, 0,         0};
 
 // Reads 32 pixels, throws half away and writes 16 pixels.
-__declspec(naked)
-void ScaleRowDown2_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                         uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown2_SSSE3(const uint8_t* src_ptr,
+                                           ptrdiff_t src_stride,
+                                           uint8_t* dst_ptr,
+                                           int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]  // src_ptr
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_ptr
+    mov        ecx, [esp + 16]  // dst_width
 
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
-    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm0, 8          // isolate odd pixels.
     psrlw      xmm1, 8
     packuswb   xmm0, xmm1
     movdqu     [edx], xmm0
@@ -120,27 +117,28 @@ void ScaleRowDown2_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Blends 32x1 rectangle to 16x1.
-__declspec(naked)
-void ScaleRowDown2Linear_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                               uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown2Linear_SSSE3(const uint8_t* src_ptr,
+                                                 ptrdiff_t src_stride,
+                                                 uint8_t* dst_ptr,
+                                                 int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]  // src_ptr
+    // src_stride
+    mov        edx, [esp + 12]  // dst_ptr
+    mov        ecx, [esp + 16]  // dst_width
 
-    pcmpeqb    xmm4, xmm4            // constant 0x0101
+    pcmpeqb    xmm4, xmm4  // constant 0x0101
     psrlw      xmm4, 15
     packuswb   xmm4, xmm4
-    pxor       xmm5, xmm5            // constant 0
+    pxor       xmm5, xmm5  // constant 0
 
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
-    pmaddubsw  xmm0, xmm4      // horizontal add
+    pmaddubsw  xmm0, xmm4  // horizontal add
     pmaddubsw  xmm1, xmm4
-    pavgw      xmm0, xmm5      // (x + 1) / 2
+    pavgw      xmm0, xmm5       // (x + 1) / 2
     pavgw      xmm1, xmm5
     packuswb   xmm0, xmm1
     movdqu     [edx], xmm0
@@ -153,20 +151,21 @@ void ScaleRowDown2Linear_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Blends 32x2 rectangle to 16x1.
-__declspec(naked)
-void ScaleRowDown2Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                            uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                                              ptrdiff_t src_stride,
+                                              uint8_t* dst_ptr,
+                                              int dst_width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
+    mov        eax, [esp + 4 + 4]  // src_ptr
+    mov        esi, [esp + 4 + 8]  // src_stride
+    mov        edx, [esp + 4 + 12]  // dst_ptr
+    mov        ecx, [esp + 4 + 16]  // dst_width
 
-    pcmpeqb    xmm4, xmm4            // constant 0x0101
+    pcmpeqb    xmm4, xmm4  // constant 0x0101
     psrlw      xmm4, 15
     packuswb   xmm4, xmm4
-    pxor       xmm5, xmm5            // constant 0
+    pxor       xmm5, xmm5  // constant 0
 
   wloop:
     movdqu     xmm0, [eax]
@@ -174,15 +173,15 @@ void ScaleRowDown2Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
     movdqu     xmm2, [eax + esi]
     movdqu     xmm3, [eax + esi + 16]
     lea        eax,  [eax + 32]
-    pmaddubsw  xmm0, xmm4      // horizontal add
+    pmaddubsw  xmm0, xmm4  // horizontal add
     pmaddubsw  xmm1, xmm4
     pmaddubsw  xmm2, xmm4
     pmaddubsw  xmm3, xmm4
-    paddw      xmm0, xmm2      // vertical add
+    paddw      xmm0, xmm2  // vertical add
     paddw      xmm1, xmm3
     psrlw      xmm0, 1
     psrlw      xmm1, 1
-    pavgw      xmm0, xmm5      // (x + 1) / 2
+    pavgw      xmm0, xmm5  // (x + 1) / 2
     pavgw      xmm1, xmm5
     packuswb   xmm0, xmm1
     movdqu     [edx], xmm0
@@ -197,23 +196,24 @@ void ScaleRowDown2Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 
 #ifdef HAS_SCALEROWDOWN2_AVX2
 // Reads 64 pixels, throws half away and writes 32 pixels.
-__declspec(naked)
-void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown2_AVX2(const uint8_t* src_ptr,
+                                          ptrdiff_t src_stride,
+                                          uint8_t* dst_ptr,
+                                          int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]  // src_ptr
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_ptr
+    mov        ecx, [esp + 16]  // dst_width
 
   wloop:
     vmovdqu     ymm0, [eax]
     vmovdqu     ymm1, [eax + 32]
     lea         eax,  [eax + 64]
-    vpsrlw      ymm0, ymm0, 8        // isolate odd pixels.
+    vpsrlw      ymm0, ymm0, 8  // isolate odd pixels.
     vpsrlw      ymm1, ymm1, 8
     vpackuswb   ymm0, ymm0, ymm1
-    vpermq      ymm0, ymm0, 0xd8     // unmutate vpackuswb
+    vpermq      ymm0, ymm0, 0xd8       // unmutate vpackuswb
     vmovdqu     [edx], ymm0
     lea         edx, [edx + 32]
     sub         ecx, 32
@@ -225,30 +225,31 @@ void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Blends 64x1 rectangle to 32x1.
-__declspec(naked)
-void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                              uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown2Linear_AVX2(const uint8_t* src_ptr,
+                                                ptrdiff_t src_stride,
+                                                uint8_t* dst_ptr,
+                                                int dst_width) {
   __asm {
-    mov         eax, [esp + 4]        // src_ptr
-                                      // src_stride
-    mov         edx, [esp + 12]       // dst_ptr
-    mov         ecx, [esp + 16]       // dst_width
+    mov         eax, [esp + 4]  // src_ptr
+    // src_stride
+    mov         edx, [esp + 12]  // dst_ptr
+    mov         ecx, [esp + 16]  // dst_width
 
-    vpcmpeqb    ymm4, ymm4, ymm4      // '1' constant, 8b
+    vpcmpeqb    ymm4, ymm4, ymm4  // '1' constant, 8b
     vpsrlw      ymm4, ymm4, 15
     vpackuswb   ymm4, ymm4, ymm4
-    vpxor       ymm5, ymm5, ymm5      // constant 0
+    vpxor       ymm5, ymm5, ymm5  // constant 0
 
   wloop:
     vmovdqu     ymm0, [eax]
     vmovdqu     ymm1, [eax + 32]
     lea         eax,  [eax + 64]
-    vpmaddubsw  ymm0, ymm0, ymm4      // horizontal add
+    vpmaddubsw  ymm0, ymm0, ymm4  // horizontal add
     vpmaddubsw  ymm1, ymm1, ymm4
-    vpavgw      ymm0, ymm0, ymm5      // (x + 1) / 2
+    vpavgw      ymm0, ymm0, ymm5  // (x + 1) / 2
     vpavgw      ymm1, ymm1, ymm5
     vpackuswb   ymm0, ymm0, ymm1
-    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpermq      ymm0, ymm0, 0xd8       // unmutate vpackuswb
     vmovdqu     [edx], ymm0
     lea         edx, [edx + 32]
     sub         ecx, 32
@@ -262,20 +263,21 @@ void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
 // For rounding, average = (sum + 2) / 4
 // becomes average((sum >> 1), 0)
 // Blends 64x2 rectangle to 32x1.
-__declspec(naked)
-void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown2Box_AVX2(const uint8_t* src_ptr,
+                                             ptrdiff_t src_stride,
+                                             uint8_t* dst_ptr,
+                                             int dst_width) {
   __asm {
     push        esi
-    mov         eax, [esp + 4 + 4]    // src_ptr
-    mov         esi, [esp + 4 + 8]    // src_stride
-    mov         edx, [esp + 4 + 12]   // dst_ptr
-    mov         ecx, [esp + 4 + 16]   // dst_width
+    mov         eax, [esp + 4 + 4]  // src_ptr
+    mov         esi, [esp + 4 + 8]  // src_stride
+    mov         edx, [esp + 4 + 12]  // dst_ptr
+    mov         ecx, [esp + 4 + 16]  // dst_width
 
-    vpcmpeqb    ymm4, ymm4, ymm4      // '1' constant, 8b
+    vpcmpeqb    ymm4, ymm4, ymm4  // '1' constant, 8b
     vpsrlw      ymm4, ymm4, 15
     vpackuswb   ymm4, ymm4, ymm4
-    vpxor       ymm5, ymm5, ymm5      // constant 0
+    vpxor       ymm5, ymm5, ymm5  // constant 0
 
   wloop:
     vmovdqu     ymm0, [eax]
@@ -283,18 +285,18 @@ void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
     vmovdqu     ymm2, [eax + esi]
     vmovdqu     ymm3, [eax + esi + 32]
     lea         eax,  [eax + 64]
-    vpmaddubsw  ymm0, ymm0, ymm4      // horizontal add
+    vpmaddubsw  ymm0, ymm0, ymm4  // horizontal add
     vpmaddubsw  ymm1, ymm1, ymm4
     vpmaddubsw  ymm2, ymm2, ymm4
     vpmaddubsw  ymm3, ymm3, ymm4
-    vpaddw      ymm0, ymm0, ymm2      // vertical add
+    vpaddw      ymm0, ymm0, ymm2  // vertical add
     vpaddw      ymm1, ymm1, ymm3
-    vpsrlw      ymm0, ymm0, 1         // (x + 2) / 4 = (x / 2 + 1) / 2
+    vpsrlw      ymm0, ymm0, 1  // (x + 2) / 4 = (x / 2 + 1) / 2
     vpsrlw      ymm1, ymm1, 1
-    vpavgw      ymm0, ymm0, ymm5      // (x + 1) / 2
+    vpavgw      ymm0, ymm0, ymm5  // (x + 1) / 2
     vpavgw      ymm1, ymm1, ymm5
     vpackuswb   ymm0, ymm0, ymm1
-    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpermq      ymm0, ymm0, 0xd8  // unmutate vpackuswb
     vmovdqu     [edx], ymm0
     lea         edx, [edx + 32]
     sub         ecx, 32
@@ -308,15 +310,16 @@ void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
 #endif  // HAS_SCALEROWDOWN2_AVX2
 
 // Point samples 32 pixels to 8 pixels.
-__declspec(naked)
-void ScaleRowDown4_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown4_SSSE3(const uint8_t* src_ptr,
+                                           ptrdiff_t src_stride,
+                                           uint8_t* dst_ptr,
+                                           int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff0000
+    mov        eax, [esp + 4]  // src_ptr
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_ptr
+    mov        ecx, [esp + 16]  // dst_width
+    pcmpeqb    xmm5, xmm5       // generate mask 0x00ff0000
     psrld      xmm5, 24
     pslld      xmm5, 16
 
@@ -339,50 +342,51 @@ void ScaleRowDown4_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Blends 32x4 rectangle to 8x1.
-__declspec(naked)
-void ScaleRowDown4Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown4Box_SSSE3(const uint8_t* src_ptr,
+                                              ptrdiff_t src_stride,
+                                              uint8_t* dst_ptr,
+                                              int dst_width) {
   __asm {
     push       esi
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_ptr
-    mov        esi, [esp + 8 + 8]    // src_stride
-    mov        edx, [esp + 8 + 12]   // dst_ptr
-    mov        ecx, [esp + 8 + 16]   // dst_width
+    mov        eax, [esp + 8 + 4]  // src_ptr
+    mov        esi, [esp + 8 + 8]  // src_stride
+    mov        edx, [esp + 8 + 12]  // dst_ptr
+    mov        ecx, [esp + 8 + 16]  // dst_width
     lea        edi, [esi + esi * 2]  // src_stride * 3
-    pcmpeqb    xmm4, xmm4            // constant 0x0101
+    pcmpeqb    xmm4, xmm4  // constant 0x0101
     psrlw      xmm4, 15
     movdqa     xmm5, xmm4
     packuswb   xmm4, xmm4
-    psllw      xmm5, 3               // constant 0x0008
+    psllw      xmm5, 3  // constant 0x0008
 
   wloop:
-    movdqu     xmm0, [eax]           // average rows
+    movdqu     xmm0, [eax]  // average rows
     movdqu     xmm1, [eax + 16]
     movdqu     xmm2, [eax + esi]
     movdqu     xmm3, [eax + esi + 16]
-    pmaddubsw  xmm0, xmm4      // horizontal add
+    pmaddubsw  xmm0, xmm4  // horizontal add
     pmaddubsw  xmm1, xmm4
     pmaddubsw  xmm2, xmm4
     pmaddubsw  xmm3, xmm4
-    paddw      xmm0, xmm2      // vertical add rows 0, 1
+    paddw      xmm0, xmm2  // vertical add rows 0, 1
     paddw      xmm1, xmm3
     movdqu     xmm2, [eax + esi * 2]
     movdqu     xmm3, [eax + esi * 2 + 16]
     pmaddubsw  xmm2, xmm4
     pmaddubsw  xmm3, xmm4
-    paddw      xmm0, xmm2      // add row 2
+    paddw      xmm0, xmm2  // add row 2
     paddw      xmm1, xmm3
     movdqu     xmm2, [eax + edi]
     movdqu     xmm3, [eax + edi + 16]
     lea        eax, [eax + 32]
     pmaddubsw  xmm2, xmm4
     pmaddubsw  xmm3, xmm4
-    paddw      xmm0, xmm2      // add row 3
+    paddw      xmm0, xmm2  // add row 3
     paddw      xmm1, xmm3
     phaddw     xmm0, xmm1
-    paddw      xmm0, xmm5      // + 8 for round
-    psrlw      xmm0, 4         // /16 for average of 4 * 4
+    paddw      xmm0, xmm5  // + 8 for round
+    psrlw      xmm0, 4  // /16 for average of 4 * 4
     packuswb   xmm0, xmm0
     movq       qword ptr [edx], xmm0
     lea        edx, [edx + 8]
@@ -397,15 +401,16 @@ void ScaleRowDown4Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 
 #ifdef HAS_SCALEROWDOWN4_AVX2
 // Point samples 64 pixels to 16 pixels.
-__declspec(naked)
-void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                        uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown4_AVX2(const uint8_t* src_ptr,
+                                          ptrdiff_t src_stride,
+                                          uint8_t* dst_ptr,
+                                          int dst_width) {
   __asm {
-    mov         eax, [esp + 4]        // src_ptr
-                                      // src_stride ignored
-    mov         edx, [esp + 12]       // dst_ptr
-    mov         ecx, [esp + 16]       // dst_width
-    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0x00ff0000
+    mov         eax, [esp + 4]  // src_ptr
+    // src_stride ignored
+    mov         edx, [esp + 12]  // dst_ptr
+    mov         ecx, [esp + 16]  // dst_width
+    vpcmpeqb    ymm5, ymm5, ymm5  // generate mask 0x00ff0000
     vpsrld      ymm5, ymm5, 24
     vpslld      ymm5, ymm5, 16
 
@@ -416,10 +421,10 @@ void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
     vpand       ymm0, ymm0, ymm5
     vpand       ymm1, ymm1, ymm5
     vpackuswb   ymm0, ymm0, ymm1
-    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpermq      ymm0, ymm0, 0xd8  // unmutate vpackuswb
     vpsrlw      ymm0, ymm0, 8
     vpackuswb   ymm0, ymm0, ymm0
-    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpermq      ymm0, ymm0, 0xd8       // unmutate vpackuswb
     vmovdqu     [edx], xmm0
     lea         edx, [edx + 16]
     sub         ecx, 16
@@ -431,52 +436,53 @@ void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Blends 64x4 rectangle to 16x1.
-__declspec(naked)
-void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
-                           uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown4Box_AVX2(const uint8_t* src_ptr,
+                                             ptrdiff_t src_stride,
+                                             uint8_t* dst_ptr,
+                                             int dst_width) {
   __asm {
     push        esi
     push        edi
-    mov         eax, [esp + 8 + 4]    // src_ptr
-    mov         esi, [esp + 8 + 8]    // src_stride
-    mov         edx, [esp + 8 + 12]   // dst_ptr
-    mov         ecx, [esp + 8 + 16]   // dst_width
+    mov         eax, [esp + 8 + 4]  // src_ptr
+    mov         esi, [esp + 8 + 8]  // src_stride
+    mov         edx, [esp + 8 + 12]  // dst_ptr
+    mov         ecx, [esp + 8 + 16]  // dst_width
     lea         edi, [esi + esi * 2]  // src_stride * 3
-    vpcmpeqb    ymm4, ymm4, ymm4            // constant 0x0101
+    vpcmpeqb    ymm4, ymm4, ymm4  // constant 0x0101
     vpsrlw      ymm4, ymm4, 15
-    vpsllw      ymm5, ymm4, 3               // constant 0x0008
+    vpsllw      ymm5, ymm4, 3  // constant 0x0008
     vpackuswb   ymm4, ymm4, ymm4
 
   wloop:
-    vmovdqu     ymm0, [eax]           // average rows
+    vmovdqu     ymm0, [eax]  // average rows
     vmovdqu     ymm1, [eax + 32]
     vmovdqu     ymm2, [eax + esi]
     vmovdqu     ymm3, [eax + esi + 32]
-    vpmaddubsw  ymm0, ymm0, ymm4      // horizontal add
+    vpmaddubsw  ymm0, ymm0, ymm4  // horizontal add
     vpmaddubsw  ymm1, ymm1, ymm4
     vpmaddubsw  ymm2, ymm2, ymm4
     vpmaddubsw  ymm3, ymm3, ymm4
-    vpaddw      ymm0, ymm0, ymm2      // vertical add rows 0, 1
+    vpaddw      ymm0, ymm0, ymm2  // vertical add rows 0, 1
     vpaddw      ymm1, ymm1, ymm3
     vmovdqu     ymm2, [eax + esi * 2]
     vmovdqu     ymm3, [eax + esi * 2 + 32]
     vpmaddubsw  ymm2, ymm2, ymm4
     vpmaddubsw  ymm3, ymm3, ymm4
-    vpaddw      ymm0, ymm0, ymm2      // add row 2
+    vpaddw      ymm0, ymm0, ymm2  // add row 2
     vpaddw      ymm1, ymm1, ymm3
     vmovdqu     ymm2, [eax + edi]
     vmovdqu     ymm3, [eax + edi + 32]
     lea         eax,  [eax + 64]
     vpmaddubsw  ymm2, ymm2, ymm4
     vpmaddubsw  ymm3, ymm3, ymm4
-    vpaddw      ymm0, ymm0, ymm2      // add row 3
+    vpaddw      ymm0, ymm0, ymm2  // add row 3
     vpaddw      ymm1, ymm1, ymm3
-    vphaddw     ymm0, ymm0, ymm1      // mutates
-    vpermq      ymm0, ymm0, 0xd8      // unmutate vphaddw
-    vpaddw      ymm0, ymm0, ymm5      // + 8 for round
-    vpsrlw      ymm0, ymm0, 4         // /32 for average of 4 * 4
+    vphaddw     ymm0, ymm0, ymm1  // mutates
+    vpermq      ymm0, ymm0, 0xd8  // unmutate vphaddw
+    vpaddw      ymm0, ymm0, ymm5  // + 8 for round
+    vpsrlw      ymm0, ymm0, 4  // /32 for average of 4 * 4
     vpackuswb   ymm0, ymm0, ymm0
-    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpermq      ymm0, ymm0, 0xd8  // unmutate vpackuswb
     vmovdqu     [edx], xmm0
     lea         edx, [edx + 16]
     sub         ecx, 16
@@ -494,14 +500,15 @@ void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
 // Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
 // Then shuffled to do the scaling.
 
-__declspec(naked)
-void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown34_SSSE3(const uint8_t* src_ptr,
+                                            ptrdiff_t src_stride,
+                                            uint8_t* dst_ptr,
+                                            int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]   // src_ptr
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_ptr
+    mov        ecx, [esp + 16]  // dst_width
     movdqa     xmm3, xmmword ptr kShuf0
     movdqa     xmm4, xmmword ptr kShuf1
     movdqa     xmm5, xmmword ptr kShuf2
@@ -541,16 +548,16 @@ void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 // xmm7 kRound34
 
 // Note that movdqa+palign may be better than movdqu.
-__declspec(naked)
-void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown34_1_Box_SSSE3(const uint8_t* src_ptr,
+                                                  ptrdiff_t src_stride,
+                                                  uint8_t* dst_ptr,
+                                                  int dst_width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
+    mov        eax, [esp + 4 + 4]  // src_ptr
+    mov        esi, [esp + 4 + 8]  // src_stride
+    mov        edx, [esp + 4 + 12]  // dst_ptr
+    mov        ecx, [esp + 4 + 16]  // dst_width
     movdqa     xmm2, xmmword ptr kShuf01
     movdqa     xmm3, xmmword ptr kShuf11
     movdqa     xmm4, xmmword ptr kShuf21
@@ -559,7 +566,7 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     movdqa     xmm7, xmmword ptr kRound34
 
   wloop:
-    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm0, [eax]  // pixels 0..7
     movdqu     xmm1, [eax + esi]
     pavgb      xmm0, xmm1
     pshufb     xmm0, xmm2
@@ -568,7 +575,7 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
     movq       qword ptr [edx], xmm0
-    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm0, [eax + 8]  // pixels 8..15
     movdqu     xmm1, [eax + esi + 8]
     pavgb      xmm0, xmm1
     pshufb     xmm0, xmm3
@@ -577,7 +584,7 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
     movq       qword ptr [edx + 8], xmm0
-    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm0, [eax + 16]  // pixels 16..23
     movdqu     xmm1, [eax + esi + 16]
     lea        eax, [eax + 32]
     pavgb      xmm0, xmm1
@@ -598,16 +605,16 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
 }
 
 // Note that movdqa+palign may be better than movdqu.
-__declspec(naked)
-void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown34_0_Box_SSSE3(const uint8_t* src_ptr,
+                                                  ptrdiff_t src_stride,
+                                                  uint8_t* dst_ptr,
+                                                  int dst_width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
+    mov        eax, [esp + 4 + 4]  // src_ptr
+    mov        esi, [esp + 4 + 8]  // src_stride
+    mov        edx, [esp + 4 + 12]  // dst_ptr
+    mov        ecx, [esp + 4 + 16]  // dst_width
     movdqa     xmm2, xmmword ptr kShuf01
     movdqa     xmm3, xmmword ptr kShuf11
     movdqa     xmm4, xmmword ptr kShuf21
@@ -616,7 +623,7 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     movdqa     xmm7, xmmword ptr kRound34
 
   wloop:
-    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm0, [eax]  // pixels 0..7
     movdqu     xmm1, [eax + esi]
     pavgb      xmm1, xmm0
     pavgb      xmm0, xmm1
@@ -626,7 +633,7 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
     movq       qword ptr [edx], xmm0
-    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm0, [eax + 8]  // pixels 8..15
     movdqu     xmm1, [eax + esi + 8]
     pavgb      xmm1, xmm0
     pavgb      xmm0, xmm1
@@ -636,7 +643,7 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
     psrlw      xmm0, 2
     packuswb   xmm0, xmm0
     movq       qword ptr [edx + 8], xmm0
-    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm0, [eax + 16]  // pixels 16..23
     movdqu     xmm1, [eax + esi + 16]
     lea        eax, [eax + 32]
     pavgb      xmm1, xmm0
@@ -660,26 +667,27 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
 // 3/8 point sampler
 
 // Scale 32 pixels to 12
-__declspec(naked)
-void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
-                          uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown38_SSSE3(const uint8_t* src_ptr,
+                                            ptrdiff_t src_stride,
+                                            uint8_t* dst_ptr,
+                                            int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]  // src_ptr
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_ptr
+    mov        ecx, [esp + 16]  // dst_width
     movdqa     xmm4, xmmword ptr kShuf38a
     movdqa     xmm5, xmmword ptr kShuf38b
 
   xloop:
-    movdqu     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
-    movdqu     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    movdqu     xmm0, [eax]  // 16 pixels -> 0,1,2,3,4,5
+    movdqu     xmm1, [eax + 16]  // 16 pixels -> 6,7,8,9,10,11
     lea        eax, [eax + 32]
     pshufb     xmm0, xmm4
     pshufb     xmm1, xmm5
     paddusb    xmm0, xmm1
 
-    movq       qword ptr [edx], xmm0  // write 12 pixels
+    movq       qword ptr [edx], xmm0       // write 12 pixels
     movhlps    xmm1, xmm0
     movd       [edx + 8], xmm1
     lea        edx, [edx + 12]
@@ -691,23 +699,23 @@ void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Scale 16x3 pixels to 6x1 with interpolation
-__declspec(naked)
-void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown38_3_Box_SSSE3(const uint8_t* src_ptr,
+                                                  ptrdiff_t src_stride,
+                                                  uint8_t* dst_ptr,
+                                                  int dst_width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
+    mov        eax, [esp + 4 + 4]  // src_ptr
+    mov        esi, [esp + 4 + 8]  // src_stride
+    mov        edx, [esp + 4 + 12]  // dst_ptr
+    mov        ecx, [esp + 4 + 16]  // dst_width
     movdqa     xmm2, xmmword ptr kShufAc
     movdqa     xmm3, xmmword ptr kShufAc3
     movdqa     xmm4, xmmword ptr kScaleAc33
     pxor       xmm5, xmm5
 
   xloop:
-    movdqu     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqu     xmm0, [eax]  // sum up 3 rows into xmm0/1
     movdqu     xmm6, [eax + esi]
     movhlps    xmm1, xmm0
     movhlps    xmm7, xmm6
@@ -725,14 +733,14 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     paddusw    xmm0, xmm6
     paddusw    xmm1, xmm7
 
-    movdqa     xmm6, xmm0            // 8 pixels -> 0,1,2 of xmm6
+    movdqa     xmm6, xmm0  // 8 pixels -> 0,1,2 of xmm6
     psrldq     xmm0, 2
     paddusw    xmm6, xmm0
     psrldq     xmm0, 2
     paddusw    xmm6, xmm0
     pshufb     xmm6, xmm2
 
-    movdqa     xmm7, xmm1            // 8 pixels -> 3,4,5 of xmm6
+    movdqa     xmm7, xmm1  // 8 pixels -> 3,4,5 of xmm6
     psrldq     xmm1, 2
     paddusw    xmm7, xmm1
     psrldq     xmm1, 2
@@ -740,10 +748,10 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
     pshufb     xmm7, xmm3
     paddusw    xmm6, xmm7
 
-    pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
+    pmulhuw    xmm6, xmm4  // divide by 9,9,6, 9,9,6
     packuswb   xmm6, xmm6
 
-    movd       [edx], xmm6           // write 6 pixels
+    movd       [edx], xmm6  // write 6 pixels
     psrlq      xmm6, 16
     movd       [edx + 2], xmm6
     lea        edx, [edx + 6]
@@ -756,28 +764,28 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
 }
 
 // Scale 16x2 pixels to 6x1 with interpolation
-__declspec(naked)
-void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr, int dst_width) {
+__declspec(naked) void ScaleRowDown38_2_Box_SSSE3(const uint8_t* src_ptr,
+                                                  ptrdiff_t src_stride,
+                                                  uint8_t* dst_ptr,
+                                                  int dst_width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
+    mov        eax, [esp + 4 + 4]  // src_ptr
+    mov        esi, [esp + 4 + 8]  // src_stride
+    mov        edx, [esp + 4 + 12]  // dst_ptr
+    mov        ecx, [esp + 4 + 16]  // dst_width
     movdqa     xmm2, xmmword ptr kShufAb0
     movdqa     xmm3, xmmword ptr kShufAb1
     movdqa     xmm4, xmmword ptr kShufAb2
     movdqa     xmm5, xmmword ptr kScaleAb2
 
   xloop:
-    movdqu     xmm0, [eax]           // average 2 rows into xmm0
+    movdqu     xmm0, [eax]  // average 2 rows into xmm0
     movdqu     xmm1, [eax + esi]
     lea        eax, [eax + 16]
     pavgb      xmm0, xmm1
 
-    movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
+    movdqa     xmm1, xmm0  // 16 pixels -> 0,1,2,3,4,5 of xmm1
     pshufb     xmm1, xmm2
     movdqa     xmm6, xmm0
     pshufb     xmm6, xmm3
@@ -785,10 +793,10 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
     pshufb     xmm0, xmm4
     paddusw    xmm1, xmm0
 
-    pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
+    pmulhuw    xmm1, xmm5  // divide by 3,3,2, 3,3,2
     packuswb   xmm1, xmm1
 
-    movd       [edx], xmm1           // write 6 pixels
+    movd       [edx], xmm1  // write 6 pixels
     psrlq      xmm1, 16
     movd       [edx + 2], xmm1
     lea        edx, [edx + 6]
@@ -801,26 +809,27 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
 }
 
 // Reads 16 bytes and accumulates to 16 shorts at a time.
-__declspec(naked)
-void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+__declspec(naked) void ScaleAddRow_SSE2(const uint8_t* src_ptr,
+                                        uint16_t* dst_ptr,
+                                        int src_width) {
   __asm {
-    mov        eax, [esp + 4]   // src_ptr
-    mov        edx, [esp + 8]   // dst_ptr
+    mov        eax, [esp + 4]  // src_ptr
+    mov        edx, [esp + 8]  // dst_ptr
     mov        ecx, [esp + 12]  // src_width
     pxor       xmm5, xmm5
 
-  // sum rows
+        // sum rows
   xloop:
-    movdqu     xmm3, [eax]       // read 16 bytes
+    movdqu     xmm3, [eax]  // read 16 bytes
     lea        eax, [eax + 16]
-    movdqu     xmm0, [edx]       // read 16 words from destination
+    movdqu     xmm0, [edx]  // read 16 words from destination
     movdqu     xmm1, [edx + 16]
     movdqa     xmm2, xmm3
     punpcklbw  xmm2, xmm5
     punpckhbw  xmm3, xmm5
-    paddusw    xmm0, xmm2        // sum 16 words
+    paddusw    xmm0, xmm2  // sum 16 words
     paddusw    xmm1, xmm3
-    movdqu     [edx], xmm0       // write 16 words to destination
+    movdqu     [edx], xmm0  // write 16 words to destination
     movdqu     [edx + 16], xmm1
     lea        edx, [edx + 32]
     sub        ecx, 16
@@ -831,24 +840,25 @@ void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
 
 #ifdef HAS_SCALEADDROW_AVX2
 // Reads 32 bytes and accumulates to 32 shorts at a time.
-__declspec(naked)
-void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+__declspec(naked) void ScaleAddRow_AVX2(const uint8_t* src_ptr,
+                                        uint16_t* dst_ptr,
+                                        int src_width) {
   __asm {
-    mov         eax, [esp + 4]   // src_ptr
-    mov         edx, [esp + 8]   // dst_ptr
+    mov         eax, [esp + 4]  // src_ptr
+    mov         edx, [esp + 8]  // dst_ptr
     mov         ecx, [esp + 12]  // src_width
     vpxor       ymm5, ymm5, ymm5
 
-  // sum rows
+        // sum rows
   xloop:
-    vmovdqu     ymm3, [eax]       // read 32 bytes
+    vmovdqu     ymm3, [eax]  // read 32 bytes
     lea         eax, [eax + 32]
     vpermq      ymm3, ymm3, 0xd8  // unmutate for vpunpck
     vpunpcklbw  ymm2, ymm3, ymm5
     vpunpckhbw  ymm3, ymm3, ymm5
-    vpaddusw    ymm0, ymm2, [edx] // sum 16 words
+    vpaddusw    ymm0, ymm2, [edx]  // sum 16 words
     vpaddusw    ymm1, ymm3, [edx + 32]
-    vmovdqu     [edx], ymm0       // write 32 words to destination
+    vmovdqu     [edx], ymm0  // write 32 words to destination
     vmovdqu     [edx + 32], ymm1
     lea         edx, [edx + 64]
     sub         ecx, 32
@@ -862,86 +872,87 @@ void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
 
 // Constant for making pixels signed to avoid pmaddubsw
 // saturation.
-static uvec8 kFsub80 =
-  { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
+static const uvec8 kFsub80 = {0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+                              0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
 
 // Constant for making pixels unsigned and adding .5 for rounding.
-static uvec16 kFadd40 =
-  { 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040 };
+static const uvec16 kFadd40 = {0x4040, 0x4040, 0x4040, 0x4040,
+                               0x4040, 0x4040, 0x4040, 0x4040};
 
 // Bilinear column filtering. SSSE3 version.
-__declspec(naked)
-void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                           int dst_width, int x, int dx) {
+__declspec(naked) void ScaleFilterCols_SSSE3(uint8_t* dst_ptr,
+                                             const uint8_t* src_ptr,
+                                             int dst_width,
+                                             int x,
+                                             int dx) {
   __asm {
     push       ebx
     push       esi
     push       edi
-    mov        edi, [esp + 12 + 4]    // dst_ptr
-    mov        esi, [esp + 12 + 8]    // src_ptr
-    mov        ecx, [esp + 12 + 12]   // dst_width
+    mov        edi, [esp + 12 + 4]  // dst_ptr
+    mov        esi, [esp + 12 + 8]  // src_ptr
+    mov        ecx, [esp + 12 + 12]  // dst_width
     movd       xmm2, [esp + 12 + 16]  // x
     movd       xmm3, [esp + 12 + 20]  // dx
-    mov        eax, 0x04040000      // shuffle to line up fractions with pixel.
+    mov        eax, 0x04040000  // shuffle to line up fractions with pixel.
     movd       xmm5, eax
-    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    pcmpeqb    xmm6, xmm6  // generate 0x007f for inverting fraction.
     psrlw      xmm6, 9
-    pcmpeqb    xmm7, xmm7           // generate 0x0001
+    pcmpeqb    xmm7, xmm7  // generate 0x0001
     psrlw      xmm7, 15
-    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    pextrw     eax, xmm2, 1  // get x0 integer. preroll
     sub        ecx, 2
     jl         xloop29
 
-    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    movdqa     xmm0, xmm2  // x1 = x0 + dx
     paddd      xmm0, xmm3
-    punpckldq  xmm2, xmm0           // x0 x1
-    punpckldq  xmm3, xmm3           // dx dx
-    paddd      xmm3, xmm3           // dx * 2, dx * 2
-    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+    punpckldq  xmm2, xmm0  // x0 x1
+    punpckldq  xmm3, xmm3  // dx dx
+    paddd      xmm3, xmm3  // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3  // get x1 integer. preroll
 
     // 2 Pixel loop.
   xloop2:
-    movdqa     xmm1, xmm2           // x0, x1 fractions.
-    paddd      xmm2, xmm3           // x += dx
+    movdqa     xmm1, xmm2  // x0, x1 fractions.
+    paddd      xmm2, xmm3  // x += dx
     movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
     movd       xmm0, ebx
-    psrlw      xmm1, 9              // 7 bit fractions.
+    psrlw      xmm1, 9  // 7 bit fractions.
     movzx      ebx, word ptr [esi + edx]  // 2 source x1 pixels
     movd       xmm4, ebx
-    pshufb     xmm1, xmm5           // 0011
+    pshufb     xmm1, xmm5  // 0011
     punpcklwd  xmm0, xmm4
     psubb      xmm0, xmmword ptr kFsub80  // make pixels signed.
-    pxor       xmm1, xmm6           // 0..7f and 7f..0
-    paddusb    xmm1, xmm7           // +1 so 0..7f and 80..1
-    pmaddubsw  xmm1, xmm0           // 16 bit, 2 pixels.
-    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
-    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    pxor       xmm1, xmm6  // 0..7f and 7f..0
+    paddusb    xmm1, xmm7  // +1 so 0..7f and 80..1
+    pmaddubsw  xmm1, xmm0  // 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1  // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3  // get x1 integer. next iteration.
     paddw      xmm1, xmmword ptr kFadd40  // make pixels unsigned and round.
-    psrlw      xmm1, 7              // 8.7 fixed point to low 8 bits.
-    packuswb   xmm1, xmm1           // 8 bits, 2 pixels.
+    psrlw      xmm1, 7  // 8.7 fixed point to low 8 bits.
+    packuswb   xmm1, xmm1  // 8 bits, 2 pixels.
     movd       ebx, xmm1
     mov        [edi], bx
     lea        edi, [edi + 2]
-    sub        ecx, 2               // 2 pixels
+    sub        ecx, 2  // 2 pixels
     jge        xloop2
 
  xloop29:
     add        ecx, 2 - 1
     jl         xloop99
 
-    // 1 pixel remainder
+            // 1 pixel remainder
     movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
     movd       xmm0, ebx
-    psrlw      xmm2, 9              // 7 bit fractions.
-    pshufb     xmm2, xmm5           // 0011
+    psrlw      xmm2, 9  // 7 bit fractions.
+    pshufb     xmm2, xmm5  // 0011
     psubb      xmm0, xmmword ptr kFsub80  // make pixels signed.
-    pxor       xmm2, xmm6           // 0..7f and 7f..0
-    paddusb    xmm2, xmm7           // +1 so 0..7f and 80..1
-    pmaddubsw  xmm2, xmm0           // 16 bit
+    pxor       xmm2, xmm6  // 0..7f and 7f..0
+    paddusb    xmm2, xmm7  // +1 so 0..7f and 80..1
+    pmaddubsw  xmm2, xmm0  // 16 bit
     paddw      xmm2, xmmword ptr kFadd40  // make pixels unsigned and round.
-    psrlw      xmm2, 7              // 8.7 fixed point to low 8 bits.
-    packuswb   xmm2, xmm2           // 8 bits
+    psrlw      xmm2, 7  // 8.7 fixed point to low 8 bits.
+    packuswb   xmm2, xmm2  // 8 bits
     movd       ebx, xmm2
     mov        [edi], bl
 
@@ -955,13 +966,15 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
 }
 
 // Reads 16 pixels, duplicates them and writes 32 pixels.
-__declspec(naked)
-void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                       int dst_width, int x, int dx) {
+__declspec(naked) void ScaleColsUp2_SSE2(uint8_t* dst_ptr,
+                                         const uint8_t* src_ptr,
+                                         int dst_width,
+                                         int x,
+                                         int dx) {
   __asm {
-    mov        edx, [esp + 4]    // dst_ptr
-    mov        eax, [esp + 8]    // src_ptr
-    mov        ecx, [esp + 12]   // dst_width
+    mov        edx, [esp + 4]  // dst_ptr
+    mov        eax, [esp + 8]  // src_ptr
+    mov        ecx, [esp + 12]  // dst_width
 
   wloop:
     movdqu     xmm0, [eax]
@@ -980,15 +993,15 @@ void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
 }
 
 // Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
-__declspec(naked)
-void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
-                            ptrdiff_t src_stride,
-                            uint8* dst_argb, int dst_width) {
+__declspec(naked) void ScaleARGBRowDown2_SSE2(const uint8_t* src_argb,
+                                              ptrdiff_t src_stride,
+                                              uint8_t* dst_argb,
+                                              int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_argb
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_argb
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]   // src_argb
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_argb
+    mov        ecx, [esp + 16]  // dst_width
 
   wloop:
     movdqu     xmm0, [eax]
@@ -1005,23 +1018,23 @@ void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
 }
 
 // Blends 8x1 rectangle to 4x1.
-__declspec(naked)
-void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
-                                  ptrdiff_t src_stride,
-                                  uint8* dst_argb, int dst_width) {
+__declspec(naked) void ScaleARGBRowDown2Linear_SSE2(const uint8_t* src_argb,
+                                                    ptrdiff_t src_stride,
+                                                    uint8_t* dst_argb,
+                                                    int dst_width) {
   __asm {
-    mov        eax, [esp + 4]        // src_argb
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_argb
-    mov        ecx, [esp + 16]       // dst_width
+    mov        eax, [esp + 4]  // src_argb
+    // src_stride ignored
+    mov        edx, [esp + 12]  // dst_argb
+    mov        ecx, [esp + 16]  // dst_width
 
   wloop:
     movdqu     xmm0, [eax]
     movdqu     xmm1, [eax + 16]
     lea        eax,  [eax + 32]
     movdqa     xmm2, xmm0
-    shufps     xmm0, xmm1, 0x88      // even pixels
-    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    shufps     xmm0, xmm1, 0x88  // even pixels
+    shufps     xmm2, xmm1, 0xdd       // odd pixels
     pavgb      xmm0, xmm2
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
@@ -1033,16 +1046,16 @@ void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
 }
 
 // Blends 8x2 rectangle to 4x1.
-__declspec(naked)
-void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
-                               ptrdiff_t src_stride,
-                               uint8* dst_argb, int dst_width) {
+__declspec(naked) void ScaleARGBRowDown2Box_SSE2(const uint8_t* src_argb,
+                                                 ptrdiff_t src_stride,
+                                                 uint8_t* dst_argb,
+                                                 int dst_width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]    // src_argb
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_argb
-    mov        ecx, [esp + 4 + 16]   // dst_width
+    mov        eax, [esp + 4 + 4]  // src_argb
+    mov        esi, [esp + 4 + 8]  // src_stride
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // dst_width
 
   wloop:
     movdqu     xmm0, [eax]
@@ -1050,11 +1063,11 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
     movdqu     xmm2, [eax + esi]
     movdqu     xmm3, [eax + esi + 16]
     lea        eax,  [eax + 32]
-    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm0, xmm2  // average rows
     pavgb      xmm1, xmm3
-    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
-    shufps     xmm0, xmm1, 0x88      // even pixels
-    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    movdqa     xmm2, xmm0  // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88  // even pixels
+    shufps     xmm2, xmm1, 0xdd  // odd pixels
     pavgb      xmm0, xmm2
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
@@ -1067,18 +1080,19 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
 }
 
 // Reads 4 pixels at a time.
-__declspec(naked)
-void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
-                               int src_stepx,
-                               uint8* dst_argb, int dst_width) {
+__declspec(naked) void ScaleARGBRowDownEven_SSE2(const uint8_t* src_argb,
+                                                 ptrdiff_t src_stride,
+                                                 int src_stepx,
+                                                 uint8_t* dst_argb,
+                                                 int dst_width) {
   __asm {
     push       ebx
     push       edi
-    mov        eax, [esp + 8 + 4]    // src_argb
-                                     // src_stride ignored
-    mov        ebx, [esp + 8 + 12]   // src_stepx
-    mov        edx, [esp + 8 + 16]   // dst_argb
-    mov        ecx, [esp + 8 + 20]   // dst_width
+    mov        eax, [esp + 8 + 4]   // src_argb
+    // src_stride ignored
+    mov        ebx, [esp + 8 + 12]  // src_stepx
+    mov        edx, [esp + 8 + 16]  // dst_argb
+    mov        ecx, [esp + 8 + 20]  // dst_width
     lea        ebx, [ebx * 4]
     lea        edi, [ebx + ebx * 2]
 
@@ -1103,21 +1117,21 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
 }
 
 // Blends four 2x2 to 4x1.
-__declspec(naked)
-void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
-                                  ptrdiff_t src_stride,
-                                  int src_stepx,
-                                  uint8* dst_argb, int dst_width) {
+__declspec(naked) void ScaleARGBRowDownEvenBox_SSE2(const uint8_t* src_argb,
+                                                    ptrdiff_t src_stride,
+                                                    int src_stepx,
+                                                    uint8_t* dst_argb,
+                                                    int dst_width) {
   __asm {
     push       ebx
     push       esi
     push       edi
-    mov        eax, [esp + 12 + 4]    // src_argb
-    mov        esi, [esp + 12 + 8]    // src_stride
-    mov        ebx, [esp + 12 + 12]   // src_stepx
-    mov        edx, [esp + 12 + 16]   // dst_argb
-    mov        ecx, [esp + 12 + 20]   // dst_width
-    lea        esi, [eax + esi]       // row1 pointer
+    mov        eax, [esp + 12 + 4]  // src_argb
+    mov        esi, [esp + 12 + 8]  // src_stride
+    mov        ebx, [esp + 12 + 12]  // src_stepx
+    mov        edx, [esp + 12 + 16]  // dst_argb
+    mov        ecx, [esp + 12 + 20]  // dst_width
+    lea        esi, [eax + esi]  // row1 pointer
     lea        ebx, [ebx * 4]
     lea        edi, [ebx + ebx * 2]
 
@@ -1132,11 +1146,11 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
     movq       xmm3, qword ptr [esi + ebx * 2]
     movhps     xmm3, qword ptr [esi + edi]
     lea        esi,  [esi + ebx * 4]
-    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm0, xmm2  // average rows
     pavgb      xmm1, xmm3
-    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
-    shufps     xmm0, xmm1, 0x88      // even pixels
-    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    movdqa     xmm2, xmm0  // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88  // even pixels
+    shufps     xmm2, xmm1, 0xdd  // odd pixels
     pavgb      xmm0, xmm2
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
@@ -1151,64 +1165,66 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
 }
 
 // Column scaling unfiltered. SSE2 version.
-__declspec(naked)
-void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
-                        int dst_width, int x, int dx) {
+__declspec(naked) void ScaleARGBCols_SSE2(uint8_t* dst_argb,
+                                          const uint8_t* src_argb,
+                                          int dst_width,
+                                          int x,
+                                          int dx) {
   __asm {
     push       edi
     push       esi
-    mov        edi, [esp + 8 + 4]    // dst_argb
-    mov        esi, [esp + 8 + 8]    // src_argb
-    mov        ecx, [esp + 8 + 12]   // dst_width
+    mov        edi, [esp + 8 + 4]  // dst_argb
+    mov        esi, [esp + 8 + 8]  // src_argb
+    mov        ecx, [esp + 8 + 12]  // dst_width
     movd       xmm2, [esp + 8 + 16]  // x
     movd       xmm3, [esp + 8 + 20]  // dx
 
-    pshufd     xmm2, xmm2, 0         // x0 x0 x0 x0
-    pshufd     xmm0, xmm3, 0x11      // dx  0 dx  0
+    pshufd     xmm2, xmm2, 0  // x0 x0 x0 x0
+    pshufd     xmm0, xmm3, 0x11  // dx  0 dx  0
     paddd      xmm2, xmm0
-    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 2
-    pshufd     xmm0, xmm3, 0x05      // dx * 2, dx * 2, 0, 0
-    paddd      xmm2, xmm0            // x3 x2 x1 x0
-    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 4
-    pshufd     xmm3, xmm3, 0         // dx * 4, dx * 4, dx * 4, dx * 4
+    paddd      xmm3, xmm3  // 0, 0, 0,  dx * 2
+    pshufd     xmm0, xmm3, 0x05  // dx * 2, dx * 2, 0, 0
+    paddd      xmm2, xmm0  // x3 x2 x1 x0
+    paddd      xmm3, xmm3  // 0, 0, 0,  dx * 4
+    pshufd     xmm3, xmm3, 0  // dx * 4, dx * 4, dx * 4, dx * 4
 
-    pextrw     eax, xmm2, 1          // get x0 integer.
-    pextrw     edx, xmm2, 3          // get x1 integer.
+    pextrw     eax, xmm2, 1  // get x0 integer.
+    pextrw     edx, xmm2, 3  // get x1 integer.
 
     cmp        ecx, 0
     jle        xloop99
     sub        ecx, 4
     jl         xloop49
 
-    // 4 Pixel loop.
+        // 4 Pixel loop.
  xloop4:
     movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
     movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
-    pextrw     eax, xmm2, 5           // get x2 integer.
-    pextrw     edx, xmm2, 7           // get x3 integer.
-    paddd      xmm2, xmm3             // x += dx
-    punpckldq  xmm0, xmm1             // x0 x1
+    pextrw     eax, xmm2, 5  // get x2 integer.
+    pextrw     edx, xmm2, 7  // get x3 integer.
+    paddd      xmm2, xmm3  // x += dx
+    punpckldq  xmm0, xmm1  // x0 x1
 
     movd       xmm1, [esi + eax * 4]  // 1 source x2 pixels
     movd       xmm4, [esi + edx * 4]  // 1 source x3 pixels
-    pextrw     eax, xmm2, 1           // get x0 integer. next iteration.
-    pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
-    punpckldq  xmm1, xmm4             // x2 x3
-    punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
+    pextrw     eax, xmm2, 1  // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3  // get x1 integer. next iteration.
+    punpckldq  xmm1, xmm4  // x2 x3
+    punpcklqdq xmm0, xmm1  // x0 x1 x2 x3
     movdqu     [edi], xmm0
     lea        edi, [edi + 16]
-    sub        ecx, 4                 // 4 pixels
+    sub        ecx, 4  // 4 pixels
     jge        xloop4
 
  xloop49:
     test       ecx, 2
     je         xloop29
 
-    // 2 Pixels.
+        // 2 Pixels.
     movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
     movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
-    pextrw     eax, xmm2, 5           // get x2 integer.
-    punpckldq  xmm0, xmm1             // x0 x1
+    pextrw     eax, xmm2, 5  // get x2 integer.
+    punpckldq  xmm0, xmm1  // x0 x1
 
     movq       qword ptr [edi], xmm0
     lea        edi, [edi + 8]
@@ -1217,7 +1233,7 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
     test       ecx, 1
     je         xloop99
 
-    // 1 Pixels.
+        // 1 Pixels.
     movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
     movd       dword ptr [edi], xmm0
  xloop99:
@@ -1232,60 +1248,62 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
 // TODO(fbarchard): Port to Neon
 
 // Shuffle table for arranging 2 pixels into pairs for pmaddubsw
-static uvec8 kShuffleColARGB = {
-  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
-  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+static const uvec8 kShuffleColARGB = {
+    0u, 4u,  1u, 5u,  2u,  6u,  3u,  7u,  // bbggrraa 1st pixel
+    8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
 };
 
 // Shuffle table for duplicating 2 fractions into 8 bytes each
-static uvec8 kShuffleFractions = {
-  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+static const uvec8 kShuffleFractions = {
+    0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
 };
 
-__declspec(naked)
-void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
-                               int dst_width, int x, int dx) {
+__declspec(naked) void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
+                                                 const uint8_t* src_argb,
+                                                 int dst_width,
+                                                 int x,
+                                                 int dx) {
   __asm {
     push       esi
     push       edi
-    mov        edi, [esp + 8 + 4]    // dst_argb
-    mov        esi, [esp + 8 + 8]    // src_argb
-    mov        ecx, [esp + 8 + 12]   // dst_width
+    mov        edi, [esp + 8 + 4]  // dst_argb
+    mov        esi, [esp + 8 + 8]  // src_argb
+    mov        ecx, [esp + 8 + 12]  // dst_width
     movd       xmm2, [esp + 8 + 16]  // x
     movd       xmm3, [esp + 8 + 20]  // dx
     movdqa     xmm4, xmmword ptr kShuffleColARGB
     movdqa     xmm5, xmmword ptr kShuffleFractions
-    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    pcmpeqb    xmm6, xmm6  // generate 0x007f for inverting fraction.
     psrlw      xmm6, 9
-    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    pextrw     eax, xmm2, 1  // get x0 integer. preroll
     sub        ecx, 2
     jl         xloop29
 
-    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    movdqa     xmm0, xmm2  // x1 = x0 + dx
     paddd      xmm0, xmm3
-    punpckldq  xmm2, xmm0           // x0 x1
-    punpckldq  xmm3, xmm3           // dx dx
-    paddd      xmm3, xmm3           // dx * 2, dx * 2
-    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+    punpckldq  xmm2, xmm0  // x0 x1
+    punpckldq  xmm3, xmm3  // dx dx
+    paddd      xmm3, xmm3  // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3  // get x1 integer. preroll
 
     // 2 Pixel loop.
   xloop2:
-    movdqa     xmm1, xmm2           // x0, x1 fractions.
-    paddd      xmm2, xmm3           // x += dx
+    movdqa     xmm1, xmm2  // x0, x1 fractions.
+    paddd      xmm2, xmm3  // x += dx
     movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
-    psrlw      xmm1, 9              // 7 bit fractions.
+    psrlw      xmm1, 9  // 7 bit fractions.
     movhps     xmm0, qword ptr [esi + edx * 4]  // 2 source x1 pixels
-    pshufb     xmm1, xmm5           // 0000000011111111
-    pshufb     xmm0, xmm4           // arrange pixels into pairs
-    pxor       xmm1, xmm6           // 0..7f and 7f..0
-    pmaddubsw  xmm0, xmm1           // argb_argb 16 bit, 2 pixels.
-    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
-    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
-    psrlw      xmm0, 7              // argb 8.7 fixed point to low 8 bits.
-    packuswb   xmm0, xmm0           // argb_argb 8 bits, 2 pixels.
+    pshufb     xmm1, xmm5  // 0000000011111111
+    pshufb     xmm0, xmm4  // arrange pixels into pairs
+    pxor       xmm1, xmm6  // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1  // argb_argb 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1  // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3  // get x1 integer. next iteration.
+    psrlw      xmm0, 7  // argb 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0  // argb_argb 8 bits, 2 pixels.
     movq       qword ptr [edi], xmm0
     lea        edi, [edi + 8]
-    sub        ecx, 2               // 2 pixels
+    sub        ecx, 2  // 2 pixels
     jge        xloop2
 
  xloop29:
@@ -1293,15 +1311,15 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
     add        ecx, 2 - 1
     jl         xloop99
 
-    // 1 pixel remainder
-    psrlw      xmm2, 9              // 7 bit fractions.
+            // 1 pixel remainder
+    psrlw      xmm2, 9  // 7 bit fractions.
     movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
-    pshufb     xmm2, xmm5           // 00000000
-    pshufb     xmm0, xmm4           // arrange pixels into pairs
-    pxor       xmm2, xmm6           // 0..7f and 7f..0
-    pmaddubsw  xmm0, xmm2           // argb 16 bit, 1 pixel.
+    pshufb     xmm2, xmm5  // 00000000
+    pshufb     xmm0, xmm4  // arrange pixels into pairs
+    pxor       xmm2, xmm6  // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2  // argb 16 bit, 1 pixel.
     psrlw      xmm0, 7
-    packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
+    packuswb   xmm0, xmm0  // argb 8 bits, 1 pixel.
     movd       [edi], xmm0
 
  xloop99:
@@ -1313,13 +1331,15 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
 }
 
 // Reads 4 pixels, duplicates them and writes 8 pixels.
-__declspec(naked)
-void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
-                           int dst_width, int x, int dx) {
+__declspec(naked) void ScaleARGBColsUp2_SSE2(uint8_t* dst_argb,
+                                             const uint8_t* src_argb,
+                                             int dst_width,
+                                             int x,
+                                             int dx) {
   __asm {
-    mov        edx, [esp + 4]    // dst_argb
-    mov        eax, [esp + 8]    // src_argb
-    mov        ecx, [esp + 12]   // dst_width
+    mov        edx, [esp + 4]  // dst_argb
+    mov        eax, [esp + 8]  // src_argb
+    mov        ecx, [esp + 12]  // dst_width
 
   wloop:
     movdqu     xmm0, [eax]
@@ -1338,12 +1358,11 @@ void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
 }
 
 // Divide num by div and return as 16.16 fixed point result.
-__declspec(naked)
-int FixedDiv_X86(int num, int div) {
+__declspec(naked) int FixedDiv_X86(int num, int div) {
   __asm {
-    mov        eax, [esp + 4]    // num
-    cdq                          // extend num to 64 bits
-    shld       edx, eax, 16      // 32.16
+    mov        eax, [esp + 4]  // num
+    cdq  // extend num to 64 bits
+    shld       edx, eax, 16  // 32.16
     shl        eax, 16
     idiv       dword ptr [esp + 8]
     ret
@@ -1351,13 +1370,12 @@ int FixedDiv_X86(int num, int div) {
 }
 
 // Divide num by div and return as 16.16 fixed point result.
-__declspec(naked)
-int FixedDiv1_X86(int num, int div) {
+__declspec(naked) int FixedDiv1_X86(int num, int div) {
   __asm {
-    mov        eax, [esp + 4]    // num
-    mov        ecx, [esp + 8]    // denom
-    cdq                          // extend num to 64 bits
-    shld       edx, eax, 16      // 32.16
+    mov        eax, [esp + 4]  // num
+    mov        ecx, [esp + 8]  // denom
+    cdq  // extend num to 64 bits
+    shld       edx, eax, 16  // 32.16
     shl        eax, 16
     sub        eax, 0x00010001
     sbb        edx, 0
diff --git a/media/libyuv/source/test.sh b/media/libyuv/source/test.sh
new file mode 100755
index 0000000000..7f12c3c156
--- /dev/null
+++ b/media/libyuv/source/test.sh
@@ -0,0 +1,35 @@
+#!/bin/bash
+set -x
+
+function runbenchmark1 {
+  perf record /google/src/cloud/fbarchard/clean/google3/blaze-bin/third_party/libyuv/libyuv_test --gunit_filter=*$1 --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=1000 --libyuv_flags=-1 --libyuv_cpu_info=-1
+  perf report | grep AVX
+}
+
+runbenchmark1 ABGRToI420
+runbenchmark1 Android420ToI420
+runbenchmark1 ARGBToI420
+runbenchmark1 Convert16To8Plane
+runbenchmark1 ConvertToARGB
+runbenchmark1 ConvertToI420
+runbenchmark1 CopyPlane
+runbenchmark1 H010ToAB30
+runbenchmark1 H010ToAR30
+runbenchmark1 HalfFloatPlane
+runbenchmark1 I010ToAB30
+runbenchmark1 I010ToAR30
+runbenchmark1 I420Copy
+runbenchmark1 I420Psnr
+runbenchmark1 I420Scale
+runbenchmark1 I420Ssim
+runbenchmark1 I420ToARGB
+runbenchmark1 I420ToNV12
+runbenchmark1 I420ToUYVY
+runbenchmark1 I422ToI420
+runbenchmark1 InitCpuFlags
+runbenchmark1 J420ToARGB
+runbenchmark1 NV12ToARGB
+runbenchmark1 NV12ToI420
+runbenchmark1 NV12ToI420Rotate
+runbenchmark1 SetCpuFlags
+runbenchmark1 YUY2ToI420
diff --git a/media/libyuv/source/video_common.cc b/media/libyuv/source/video_common.cc
index 00fb71e18b..92384c050c 100644
--- a/media/libyuv/source/video_common.cc
+++ b/media/libyuv/source/video_common.cc
@@ -8,7 +8,6 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-
 #include "libyuv/video_common.h"
 
 #ifdef __cplusplus
@@ -16,40 +15,39 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
-
 struct FourCCAliasEntry {
-  uint32 alias;
-  uint32 canonical;
+  uint32_t alias;
+  uint32_t canonical;
 };
 
-static const struct FourCCAliasEntry kFourCCAliases[] = {
-  {FOURCC_IYUV, FOURCC_I420},
-  {FOURCC_YU12, FOURCC_I420},
-  {FOURCC_YU16, FOURCC_I422},
-  {FOURCC_YU24, FOURCC_I444},
-  {FOURCC_YUYV, FOURCC_YUY2},
-  {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
-  {FOURCC_HDYC, FOURCC_UYVY},
-  {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
-  {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
-  {FOURCC_DMB1, FOURCC_MJPG},
-  {FOURCC_BA81, FOURCC_BGGR},  // deprecated.
-  {FOURCC_RGB3, FOURCC_RAW },
-  {FOURCC_BGR3, FOURCC_24BG},
-  {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
-  {FOURCC_CM24, FOURCC_RAW },  // kCMPixelFormat_24RGB
-  {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
-  {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
-  {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
+#define NUM_ALIASES 18
+static const struct FourCCAliasEntry kFourCCAliases[NUM_ALIASES] = {
+    {FOURCC_IYUV, FOURCC_I420},
+    {FOURCC_YU12, FOURCC_I420},
+    {FOURCC_YU16, FOURCC_I422},
+    {FOURCC_YU24, FOURCC_I444},
+    {FOURCC_YUYV, FOURCC_YUY2},
+    {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
+    {FOURCC_HDYC, FOURCC_UYVY},
+    {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
+    {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
+    {FOURCC_DMB1, FOURCC_MJPG},
+    {FOURCC_BA81, FOURCC_BGGR},  // deprecated.
+    {FOURCC_RGB3, FOURCC_RAW},
+    {FOURCC_BGR3, FOURCC_24BG},
+    {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
+    {FOURCC_CM24, FOURCC_RAW},   // kCMPixelFormat_24RGB
+    {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
+    {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
+    {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
 };
 // TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB.
 //  {FOURCC_BGRA, FOURCC_ARGB},  // kCMPixelFormat_32BGRA
 
 LIBYUV_API
-uint32 CanonicalFourCC(uint32 fourcc) {
+uint32_t CanonicalFourCC(uint32_t fourcc) {
   int i;
-  for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) {
+  for (i = 0; i < NUM_ALIASES; ++i) {
     if (kFourCCAliases[i].alias == fourcc) {
       return kFourCCAliases[i].canonical;
     }
@@ -62,4 +60,3 @@ uint32 CanonicalFourCC(uint32 fourcc) {
 }  // extern "C"
 }  // namespace libyuv
 #endif
-
diff --git a/media/libyuv/sync_chromium.py b/media/libyuv/sync_chromium.py
deleted file mode 100644
index 4e51b6bd6b..0000000000
--- a/media/libyuv/sync_chromium.py
+++ /dev/null
@@ -1,154 +0,0 @@
-#!/usr/bin/env python
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Script to download a Chromium checkout into the workspace.
-
-The script downloads a full Chromium Git clone and its DEPS.
-
-The following environment variable can be used to alter the behavior:
-* CHROMIUM_NO_HISTORY - If set to 1, a Git checkout with no history will be
-  downloaded. This is consumes less bandwidth and disk space but is known to be
-  slower in general if you have a high-speed connection.
-
-After a successful sync has completed, a .last_sync_chromium file is written to
-the chromium directory. While it exists, no more gclient sync operations will be
-performed until the --target-revision changes or the SCRIPT_VERSION constant is
-incremented. The file can be removed manually to force a new sync.
-"""
-
-import argparse
-import os
-import subprocess
-import sys
-
-# Bump this whenever the algorithm changes and you need bots/devs to re-sync,
-# ignoring the .last_sync_chromium file
-SCRIPT_VERSION = 4
-
-ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
-CHROMIUM_NO_HISTORY = 'CHROMIUM_NO_HISTORY'
-
-def _parse_gclient_dict():
-  gclient_dict = {}
-  try:
-    main_gclient = os.path.join(os.path.dirname(ROOT_DIR), '.gclient')
-    with open(main_gclient, 'rb') as deps_content:
-      exec(deps_content, gclient_dict)
-  except Exception as e:
-    print >> sys.stderr, 'error while parsing .gclient:', e
-  return gclient_dict
-
-
-def get_cache_dir():
-  return _parse_gclient_dict().get('cache_dir')
-
-
-def get_target_os_list():
-  return ','.join(_parse_gclient_dict().get('target_os', []))
-
-
-def main():
-  CR_DIR = os.path.join(ROOT_DIR, 'chromium')
-
-  p = argparse.ArgumentParser()
-  p.add_argument('--target-revision', required=True,
-                 help='The target chromium git revision [REQUIRED]')
-  p.add_argument('--chromium-dir', default=CR_DIR,
-                 help=('The path to the chromium directory to sync '
-                       '(default: %(default)r)'))
-  opts = p.parse_args()
-  opts.chromium_dir = os.path.abspath(opts.chromium_dir)
-
-  target_os_list = get_target_os_list()
-
-  # Do a quick check to see if we were successful last time to make runhooks
-  # sooper fast.
-  flag_file = os.path.join(opts.chromium_dir, '.last_sync_chromium')
-  flag_file_content = '\n'.join([
-    str(SCRIPT_VERSION),
-    opts.target_revision,
-    repr(target_os_list),
-  ])
-  if (os.path.exists(os.path.join(opts.chromium_dir, 'src')) and
-      os.path.exists(flag_file)):
-    with open(flag_file, 'r') as f:
-      if f.read() == flag_file_content:
-        print 'Chromium already up to date: ', opts.target_revision
-        return 0
-    os.unlink(flag_file)
-
-  env = os.environ.copy()
-
-  # Avoid downloading NaCl toolchain as part of the Chromium hooks.
-  env['GYP_CHROMIUM_NO_ACTION'] = '1'
-  gclient_cmd = 'gclient.bat' if sys.platform.startswith('win') else 'gclient'
-  args = [
-      gclient_cmd, 'sync', '--force', '--revision', 'src@'+opts.target_revision
-  ]
-
-  if os.environ.get('CHROME_HEADLESS') == '1':
-    # Running on a buildbot.
-    args.append('-vvv')
-
-    if sys.platform.startswith('win'):
-      cache_path = os.path.join(os.path.splitdrive(ROOT_DIR)[0] + os.path.sep,
-                                'b', 'git-cache')
-    else:
-      cache_path = '/b/git-cache'
-  else:
-    # Support developers setting the cache_dir in .gclient.
-    cache_path = get_cache_dir()
-
-  # Allow for users with poor internet connections to download a Git clone
-  # without history (saves several gigs but is generally slower and doesn't work
-  # with the Git cache).
-  if os.environ.get(CHROMIUM_NO_HISTORY) == '1':
-    if cache_path:
-      print >> sys.stderr, (
-          'You cannot use "no-history" mode for syncing Chrome (i.e. set the '
-          '%s environment variable to 1) when you have cache_dir configured in '
-          'your .gclient.' % CHROMIUM_NO_HISTORY)
-      return 1
-    args.append('--no-history')
-    gclient_entries_file = os.path.join(opts.chromium_dir, '.gclient_entries')
-  else:
-    # Write a temporary .gclient file that has the cache_dir variable added.
-    gclientfile = os.path.join(opts.chromium_dir, '.gclient')
-    with open(gclientfile, 'rb') as spec:
-      spec = spec.read().splitlines()
-      spec[-1] = 'cache_dir = %r' % (cache_path,)
-    with open(gclientfile + '.tmp', 'wb') as f:
-      f.write('\n'.join(spec))
-
-    args += [
-      '--gclientfile', '.gclient.tmp',
-      '--delete_unversioned_trees', '--reset', '--upstream'
-    ]
-    gclient_entries_file = os.path.join(opts.chromium_dir,
-                                        '.gclient.tmp_entries')
-
-  # To avoid gclient sync problems when DEPS entries have been removed we must
-  # wipe the gclient's entries file that contains cached URLs for all DEPS.
-  if os.path.exists(gclient_entries_file):
-    os.unlink(gclient_entries_file)
-
-  if target_os_list:
-    args += ['--deps=' + target_os_list]
-
-  print 'Running "%s" in %s' % (' '.join(args), opts.chromium_dir)
-  ret = subprocess.call(args, cwd=opts.chromium_dir, env=env)
-  if ret == 0:
-    with open(flag_file, 'wb') as f:
-      f.write(flag_file_content)
-
-  return ret
-
-
-if __name__ == '__main__':
-  sys.exit(main())
diff --git a/media/libyuv/third_party/gflags/BUILD.gn b/media/libyuv/third_party/gflags/BUILD.gn
deleted file mode 100644
index 69a07232cc..0000000000
--- a/media/libyuv/third_party/gflags/BUILD.gn
+++ /dev/null
@@ -1,76 +0,0 @@
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# This is a copy of WebRTC's BUILD.gn.
-
-if (is_win) {
-  gflags_gen_arch_root = "gen/win"
-} else {
-  gflags_gen_arch_root = "gen/posix"
-}
-
-config("gflags_config") {
-  include_dirs = [
-    "$gflags_gen_arch_root/include",  # For configured files.
-    "src/src",  # For everything else.
-  ]
-
-  defines = [
-    # These macros exist so flags and symbols are properly exported when
-    # building DLLs. Since we don't build DLLs, we need to disable them.
-    "GFLAGS_DLL_DECL=",
-    "GFLAGS_DLL_DECLARE_FLAG=",
-    "GFLAGS_DLL_DEFINE_FLAG=",
-  ]
-
-  # GN orders flags on a target before flags from configs. The default config
-  # adds -Wall, and this flag have to be after -Wall -- so they need to
-  # come from a config and can't be on the target directly.
-  if (is_clang) {
-    cflags = [ "-Wno-unused-local-typedef" ]
-  }
-}
-
-source_set("gflags") {
-  cflags = []
-  sources = [
-    "src/src/gflags.cc",
-    "src/src/gflags_completions.cc",
-    "src/src/gflags_reporting.cc",
-  ]
-  if (is_win) {
-    sources += [ "src/src/windows_port.cc" ]
-
-    cflags += [
-      "/wd4005",  # WIN32_LEAN_AND_MEAN.
-      "/wd4267",  # Conversion from size_t to "type".
-    ]
-  }
-
-  include_dirs = [
-    "$gflags_gen_arch_root/include/gflags",  # For configured files.
-    "$gflags_gen_arch_root/include/private",  # For config.h
-  ]
-
-  public_configs = [ ":gflags_config" ]
-
-  configs -= [ "//build/config/compiler:chromium_code" ]
-  configs += [ "//build/config/compiler:no_chromium_code" ]
-
-  if (is_win) {
-    configs -= [ "//build/config/win:unicode" ]
-  }
-
-  if (is_clang) {
-    # TODO(andrew): Look into fixing this warning upstream:
-    # http://code.google.com/p/webrtc/issues/detail?id=760
-    configs -= [ "//build/config/clang:extra_warnings" ]
-    cflags += [ "-Wno-microsoft-include" ]
-  }
-}
diff --git a/media/libyuv/third_party/gflags/LICENSE b/media/libyuv/third_party/gflags/LICENSE
deleted file mode 100644
index d15b0c2413..0000000000
--- a/media/libyuv/third_party/gflags/LICENSE
+++ /dev/null
@@ -1,28 +0,0 @@
-Copyright (c) 2006, Google Inc.
-All rights reserved.
-
-Redistribution and use in source and binary forms, with or without
-modification, are permitted provided that the following conditions are
-met:
-
-    * Redistributions of source code must retain the above copyright
-notice, this list of conditions and the following disclaimer.
-    * Redistributions in binary form must reproduce the above
-copyright notice, this list of conditions and the following disclaimer
-in the documentation and/or other materials provided with the
-distribution.
-    * Neither the name of Google Inc. nor the names of its
-contributors may be used to endorse or promote products derived from
-this software without specific prior written permission.
-
-THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/media/libyuv/third_party/gflags/README.libyuv b/media/libyuv/third_party/gflags/README.libyuv
deleted file mode 100644
index 5b3bc2db71..0000000000
--- a/media/libyuv/third_party/gflags/README.libyuv
+++ /dev/null
@@ -1,28 +0,0 @@
-URL: https://github.com/gflags/gflags
-Version: 2.1.2
-License: New BSD
-License File: LICENSE
-
-Description:
-The gflags package contains a library that implements commandline
-flags processing. As such it's a replacement for getopt(). It has
-increased flexibility, including built-in support for C++ types like
-string, and the ability to define flags in the source file in which
-they're used.
-
-Local Modifications: None
-
-
-How to update platform configuration files:
-The gen/ directory contains pre-generated configuration header files.
-Historically, all operating systems and architectures have generated
-similar configurations except for Windows. This is why there's only
-posix and win directories below gen/.
-When rolling gflags to a newer version, it's a good idea to check if
-new configuration files needs to be generated as well.
-Do this by running ./configure in the newly checked out version of
-gflags. Then diff the generated files with the ones below gen/.
-If you notice a diff, update the files with the updated ones.
-If you suspect platform dependend changes other than Windows, you'll
-have to checkout gflags on the other platforms as well and run
-./configure there too.
diff --git a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags.h b/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags.h
deleted file mode 100644
index 0db38f5c86..0000000000
--- a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags.h
+++ /dev/null
@@ -1,573 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// or defines a command line flag or wants to parse command line flags
-// or print a program usage message (which will include information about
-// flags).  Executive summary, in the form of an example foo.cc file:
-//
-//    #include "foo.h"         // foo.h has a line "DECLARE_int32(start);"
-//    #include "validators.h"  // hypothetical file defining ValidateIsFile()
-//
-//    DEFINE_int32(end, 1000, "The last record to read");
-//
-//    DEFINE_string(filename, "my_file.txt", "The file to read");
-//    // Crash if the specified file does not exist.
-//    static bool dummy = RegisterFlagValidator(&FLAGS_filename,
-//                                              &ValidateIsFile);
-//
-//    DECLARE_bool(verbose); // some other file has a DEFINE_bool(verbose, ...)
-//
-//    void MyFunc() {
-//      if (FLAGS_verbose) printf("Records %d-%d\n", FLAGS_start, FLAGS_end);
-//    }
-//
-//    Then, at the command-line:
-//       ./foo --noverbose --start=5 --end=100
-//
-// For more details, see
-//    doc/gflags.html
-//
-// --- A note about thread-safety:
-//
-// We describe many functions in this routine as being thread-hostile,
-// thread-compatible, or thread-safe.  Here are the meanings we use:
-//
-// thread-safe: it is safe for multiple threads to call this routine
-//   (or, when referring to a class, methods of this class)
-//   concurrently.
-// thread-hostile: it is not safe for multiple threads to call this
-//   routine (or methods of this class) concurrently.  In gflags,
-//   most thread-hostile routines are intended to be called early in,
-//   or even before, main() -- that is, before threads are spawned.
-// thread-compatible: it is safe for multiple threads to read from
-//   this variable (when applied to variables), or to call const
-//   methods of this class (when applied to classes), as long as no
-//   other thread is writing to the variable or calling non-const
-//   methods of this class.
-
-#ifndef GFLAGS_GFLAGS_H_
-#define GFLAGS_GFLAGS_H_
-
-#include <string>
-#include <vector>
-
-#include "gflags_declare.h" // IWYU pragma: export
-
-
-// We always want to export variables defined in user code
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DEFINE_FLAG __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DEFINE_FLAG
-#  endif
-#endif
-
-
-namespace GFLAGS_NAMESPACE {
-
-
-// --------------------------------------------------------------------
-// To actually define a flag in a file, use DEFINE_bool,
-// DEFINE_string, etc. at the bottom of this file.  You may also find
-// it useful to register a validator with the flag.  This ensures that
-// when the flag is parsed from the commandline, or is later set via
-// SetCommandLineOption, we call the validation function. It is _not_
-// called when you assign the value to the flag directly using the = operator.
-//
-// The validation function should return true if the flag value is valid, and
-// false otherwise. If the function returns false for the new setting of the
-// flag, the flag will retain its current value. If it returns false for the
-// default value, ParseCommandLineFlags() will die.
-//
-// This function is safe to call at global construct time (as in the
-// example below).
-//
-// Example use:
-//    static bool ValidatePort(const char* flagname, int32 value) {
-//       if (value > 0 && value < 32768)   // value is ok
-//         return true;
-//       printf("Invalid value for --%s: %d\n", flagname, (int)value);
-//       return false;
-//    }
-//    DEFINE_int32(port, 0, "What port to listen on");
-//    static bool dummy = RegisterFlagValidator(&FLAGS_port, &ValidatePort);
-
-// Returns true if successfully registered, false if not (because the
-// first argument doesn't point to a command-line flag, or because a
-// validator is already registered for this flag).
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const bool*        flag, bool (*validate_fn)(const char*, bool));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int32*       flag, bool (*validate_fn)(const char*, int32));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int64*       flag, bool (*validate_fn)(const char*, int64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint64*      flag, bool (*validate_fn)(const char*, uint64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const double*      flag, bool (*validate_fn)(const char*, double));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const std::string* flag, bool (*validate_fn)(const char*, const std::string&));
-
-// Convenience macro for the registration of a flag validator
-#define DEFINE_validator(name, validator) \
-    static const bool name##_validator_registered = \
-            GFLAGS_NAMESPACE::RegisterFlagValidator(&FLAGS_##name, validator)
-
-
-// --------------------------------------------------------------------
-// These methods are the best way to get access to info about the
-// list of commandline flags.  Note that these routines are pretty slow.
-//   GetAllFlags: mostly-complete info about the list, sorted by file.
-//   ShowUsageWithFlags: pretty-prints the list to stdout (what --help does)
-//   ShowUsageWithFlagsRestrict: limit to filenames with restrict as a substr
-//
-// In addition to accessing flags, you can also access argv[0] (the program
-// name) and argv (the entire commandline), which we sock away a copy of.
-// These variables are static, so you should only set them once.
-//
-// No need to export this data only structure from DLL, avoiding VS warning 4251.
-struct CommandLineFlagInfo {
-  std::string name;            // the name of the flag
-  std::string type;            // the type of the flag: int32, etc
-  std::string description;     // the "help text" associated with the flag
-  std::string current_value;   // the current value, as a string
-  std::string default_value;   // the default value, as a string
-  std::string filename;        // 'cleaned' version of filename holding the flag
-  bool has_validator_fn;       // true if RegisterFlagValidator called on this flag
-  bool is_default;             // true if the flag has the default value and
-                               // has not been set explicitly from the cmdline
-                               // or via SetCommandLineOption
-  const void* flag_ptr;        // pointer to the flag's current value (i.e. FLAGS_foo)
-};
-
-// Using this inside of a validator is a recipe for a deadlock.
-// TODO(user) Fix locking when validators are running, to make it safe to
-// call validators during ParseAllFlags.
-// Also make sure then to uncomment the corresponding unit test in
-// gflags_unittest.sh
-extern GFLAGS_DLL_DECL void GetAllFlags(std::vector<CommandLineFlagInfo>* OUTPUT);
-// These two are actually defined in gflags_reporting.cc.
-extern GFLAGS_DLL_DECL void ShowUsageWithFlags(const char *argv0);  // what --help does
-extern GFLAGS_DLL_DECL void ShowUsageWithFlagsRestrict(const char *argv0, const char *restrict);
-
-// Create a descriptive string for a flag.
-// Goes to some trouble to make pretty line breaks.
-extern GFLAGS_DLL_DECL std::string DescribeOneFlag(const CommandLineFlagInfo& flag);
-
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetArgv(int argc, const char** argv);
-
-// The following functions are thread-safe as long as SetArgv() is
-// only called before any threads start.
-extern GFLAGS_DLL_DECL const std::vector<std::string>& GetArgvs();
-extern GFLAGS_DLL_DECL const char* GetArgv();                      // all of argv as a string
-extern GFLAGS_DLL_DECL const char* GetArgv0();                     // only argv0
-extern GFLAGS_DLL_DECL uint32 GetArgvSum();                        // simple checksum of argv
-extern GFLAGS_DLL_DECL const char* ProgramInvocationName();        // argv0, or "UNKNOWN" if not set
-extern GFLAGS_DLL_DECL const char* ProgramInvocationShortName();   // basename(argv0)
-
-// ProgramUsage() is thread-safe as long as SetUsageMessage() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* ProgramUsage();                 // string set by SetUsageMessage()
-
-// VersionString() is thread-safe as long as SetVersionString() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* VersionString();                // string set by SetVersionString()
-
-
-
-// --------------------------------------------------------------------
-// Normally you access commandline flags by just saying "if (FLAGS_foo)"
-// or whatever, and set them by calling "FLAGS_foo = bar" (or, more
-// commonly, via the DEFINE_foo macro).  But if you need a bit more
-// control, we have programmatic ways to get/set the flags as well.
-// These programmatic ways to access flags are thread-safe, but direct
-// access is only thread-compatible.
-
-// Return true iff the flagname was found.
-// OUTPUT is set to the flag's value, or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineOption(const char* name, std::string* OUTPUT);
-
-// Return true iff the flagname was found. OUTPUT is set to the flag's
-// CommandLineFlagInfo or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineFlagInfo(const char* name, CommandLineFlagInfo* OUTPUT);
-
-// Return the CommandLineFlagInfo of the flagname.  exit() if name not found.
-// Example usage, to check if a flag's value is currently the default value:
-//   if (GetCommandLineFlagInfoOrDie("foo").is_default) ...
-extern GFLAGS_DLL_DECL CommandLineFlagInfo GetCommandLineFlagInfoOrDie(const char* name);
-
-enum GFLAGS_DLL_DECL FlagSettingMode {
-  // update the flag's value (can call this multiple times).
-  SET_FLAGS_VALUE,
-  // update the flag's value, but *only if* it has not yet been updated
-  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
-  SET_FLAG_IF_DEFAULT,
-  // set the flag's default value to this.  If the flag has not yet updated
-  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
-  // change the flag's current value to the new default value as well.
-  SET_FLAGS_DEFAULT
-};
-
-// Set a particular flag ("command line option").  Returns a string
-// describing the new value that the option has been set to.  The
-// return value API is not well-specified, so basically just depend on
-// it to be empty if the setting failed for some reason -- the name is
-// not a valid flag name, or the value is not a valid value -- and
-// non-empty else.
-
-// SetCommandLineOption uses set_mode == SET_FLAGS_VALUE (the common case)
-extern GFLAGS_DLL_DECL std::string SetCommandLineOption        (const char* name, const char* value);
-extern GFLAGS_DLL_DECL std::string SetCommandLineOptionWithMode(const char* name, const char* value, FlagSettingMode set_mode);
-
-
-// --------------------------------------------------------------------
-// Saves the states (value, default value, whether the user has set
-// the flag, registered validators, etc) of all flags, and restores
-// them when the FlagSaver is destroyed.  This is very useful in
-// tests, say, when you want to let your tests change the flags, but
-// make sure that they get reverted to the original states when your
-// test is complete.
-//
-// Example usage:
-//   void TestFoo() {
-//     FlagSaver s1;
-//     FLAG_foo = false;
-//     FLAG_bar = "some value";
-//
-//     // test happens here.  You can return at any time
-//     // without worrying about restoring the FLAG values.
-//   }
-//
-// Note: This class is marked with GFLAGS_ATTRIBUTE_UNUSED because all
-// the work is done in the constructor and destructor, so in the standard
-// usage example above, the compiler would complain that it's an
-// unused variable.
-//
-// This class is thread-safe.  However, its destructor writes to
-// exactly the set of flags that have changed value during its
-// lifetime, so concurrent _direct_ access to those flags
-// (i.e. FLAGS_foo instead of {Get,Set}CommandLineOption()) is unsafe.
-
-class GFLAGS_DLL_DECL FlagSaver {
- public:
-  FlagSaver();
-  ~FlagSaver();
-
- private:
-  class FlagSaverImpl* impl_;   // we use pimpl here to keep API steady
-
-  FlagSaver(const FlagSaver&);  // no copying!
-  void operator=(const FlagSaver&);
-}__attribute((unused));
-
-// --------------------------------------------------------------------
-// Some deprecated or hopefully-soon-to-be-deprecated functions.
-
-// This is often used for logging.  TODO(csilvers): figure out a better way
-extern GFLAGS_DLL_DECL std::string CommandlineFlagsIntoString();
-// Usually where this is used, a FlagSaver should be used instead.
-extern GFLAGS_DLL_DECL
-bool ReadFlagsFromString(const std::string& flagfilecontents,
-                         const char* prog_name,
-                         bool errors_are_fatal);  // uses SET_FLAGS_VALUE
-
-// These let you manually implement --flagfile functionality.
-// DEPRECATED.
-extern GFLAGS_DLL_DECL bool AppendFlagsIntoFile(const std::string& filename, const char* prog_name);
-extern GFLAGS_DLL_DECL bool ReadFromFlagsFile(const std::string& filename, const char* prog_name, bool errors_are_fatal);   // uses SET_FLAGS_VALUE
-
-
-// --------------------------------------------------------------------
-// Useful routines for initializing flags from the environment.
-// In each case, if 'varname' does not exist in the environment
-// return defval.  If 'varname' does exist but is not valid
-// (e.g., not a number for an int32 flag), abort with an error.
-// Otherwise, return the value.  NOTE: for booleans, for true use
-// 't' or 'T' or 'true' or '1', for false 'f' or 'F' or 'false' or '0'.
-
-extern GFLAGS_DLL_DECL bool BoolFromEnv(const char *varname, bool defval);
-extern GFLAGS_DLL_DECL int32 Int32FromEnv(const char *varname, int32 defval);
-extern GFLAGS_DLL_DECL int64 Int64FromEnv(const char *varname, int64 defval);
-extern GFLAGS_DLL_DECL uint64 Uint64FromEnv(const char *varname, uint64 defval);
-extern GFLAGS_DLL_DECL double DoubleFromEnv(const char *varname, double defval);
-extern GFLAGS_DLL_DECL const char *StringFromEnv(const char *varname, const char *defval);
-
-
-// --------------------------------------------------------------------
-// The next two functions parse gflags from main():
-
-// Set the "usage" message for this program.  For example:
-//   string usage("This program does nothing.  Sample usage:\n");
-//   usage += argv[0] + " <uselessarg1> <uselessarg2>";
-//   SetUsageMessage(usage);
-// Do not include commandline flags in the usage: we do that for you!
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetUsageMessage(const std::string& usage);
-
-// Sets the version string, which is emitted with --version.
-// For instance: SetVersionString("1.3");
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetVersionString(const std::string& version);
-
-
-// Looks for flags in argv and parses them.  Rearranges argv to put
-// flags first, or removes them entirely if remove_flags is true.
-// If a flag is defined more than once in the command line or flag
-// file, the last definition is used.  Returns the index (into argv)
-// of the first non-flag argument.
-// See top-of-file for more details on this function.
-#ifndef SWIG   // In swig, use ParseCommandLineFlagsScript() instead.
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineFlags(int *argc, char*** argv, bool remove_flags);
-#endif
-
-
-// Calls to ParseCommandLineNonHelpFlags and then to
-// HandleCommandLineHelpFlags can be used instead of a call to
-// ParseCommandLineFlags during initialization, in order to allow for
-// changing default values for some FLAGS (via
-// e.g. SetCommandLineOptionWithMode calls) between the time of
-// command line parsing and the time of dumping help information for
-// the flags as a result of command line parsing.  If a flag is
-// defined more than once in the command line or flag file, the last
-// definition is used.  Returns the index (into argv) of the first
-// non-flag argument.  (If remove_flags is true, will always return 1.)
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineNonHelpFlags(int *argc, char*** argv, bool remove_flags);
-
-// This is actually defined in gflags_reporting.cc.
-// This function is misnamed (it also handles --version, etc.), but
-// it's too late to change that now. :-(
-extern GFLAGS_DLL_DECL void HandleCommandLineHelpFlags();   // in gflags_reporting.cc
-
-// Allow command line reparsing.  Disables the error normally
-// generated when an unknown flag is found, since it may be found in a
-// later parse.  Thread-hostile; meant to be called before any threads
-// are spawned.
-extern GFLAGS_DLL_DECL void AllowCommandLineReparsing();
-
-// Reparse the flags that have not yet been recognized.  Only flags
-// registered since the last parse will be recognized.  Any flag value
-// must be provided as part of the argument using "=", not as a
-// separate command line argument that follows the flag argument.
-// Intended for handling flags from dynamically loaded libraries,
-// since their flags are not registered until they are loaded.
-extern GFLAGS_DLL_DECL void ReparseCommandLineNonHelpFlags();
-
-// Clean up memory allocated by flags.  This is only needed to reduce
-// the quantity of "potentially leaked" reports emitted by memory
-// debugging tools such as valgrind.  It is not required for normal
-// operation, or for the google perftools heap-checker.  It must only
-// be called when the process is about to exit, and all threads that
-// might access flags are quiescent.  Referencing flags after this is
-// called will have unexpected consequences.  This is not safe to run
-// when multiple threads might be running: the function is
-// thread-hostile.
-extern GFLAGS_DLL_DECL void ShutDownCommandLineFlags();
-
-
-// --------------------------------------------------------------------
-// Now come the command line flag declaration/definition macros that
-// will actually be used.  They're kind of hairy.  A major reason
-// for this is initialization: we want people to be able to access
-// variables in global constructors and have that not crash, even if
-// their global constructor runs before the global constructor here.
-// (Obviously, we can't guarantee the flags will have the correct
-// default value in that case, but at least accessing them is safe.)
-// The only way to do that is have flags point to a static buffer.
-// So we make one, using a union to ensure proper alignment, and
-// then use placement-new to actually set up the flag with the
-// correct default value.  In the same vein, we have to worry about
-// flag access in global destructors, so FlagRegisterer has to be
-// careful never to destroy the flag-values it constructs.
-//
-// Note that when we define a flag variable FLAGS_<name>, we also
-// preemptively define a junk variable, FLAGS_no<name>.  This is to
-// cause a link-time error if someone tries to define 2 flags with
-// names like "logging" and "nologging".  We do this because a bool
-// flag FLAG can be set from the command line to true with a "-FLAG"
-// argument, and to false with a "-noFLAG" argument, and so this can
-// potentially avert confusion.
-//
-// We also put flags into their own namespace.  It is purposefully
-// named in an opaque way that people should have trouble typing
-// directly.  The idea is that DEFINE puts the flag in the weird
-// namespace, and DECLARE imports the flag from there into the current
-// namespace.  The net result is to force people to use DECLARE to get
-// access to a flag, rather than saying "extern GFLAGS_DLL_DECL bool FLAGS_whatever;"
-// or some such instead.  We want this so we can put extra
-// functionality (like sanity-checking) in DECLARE if we want, and
-// make sure it is picked up everywhere.
-//
-// We also put the type of the variable in the namespace, so that
-// people can't DECLARE_int32 something that they DEFINE_bool'd
-// elsewhere.
-
-class GFLAGS_DLL_DECL FlagRegisterer {
- public:
-  FlagRegisterer(const char* name, const char* type,
-                 const char* help, const char* filename,
-                 void* current_storage, void* defvalue_storage);
-};
-
-// If your application #defines STRIP_FLAG_HELP to a non-zero value
-// before #including this file, we remove the help message from the
-// binary file. This can reduce the size of the resulting binary
-// somewhat, and may also be useful for security reasons.
-
-extern GFLAGS_DLL_DECL const char kStrippedFlagHelp[];
-
-
-} // namespace GFLAGS_NAMESPACE
-
-
-#ifndef SWIG  // In swig, ignore the main flag declarations
-
-#if defined(STRIP_FLAG_HELP) && STRIP_FLAG_HELP > 0
-// Need this construct to avoid the 'defined but not used' warning.
-#define MAYBE_STRIPPED_HELP(txt) \
-   (false ? (txt) : GFLAGS_NAMESPACE::kStrippedFlagHelp)
-#else
-#define MAYBE_STRIPPED_HELP(txt) txt
-#endif
-
-// Each command-line flag has two variables associated with it: one
-// with the current value, and one with the default value.  However,
-// we have a third variable, which is where value is assigned; it's a
-// constant.  This guarantees that FLAG_##value is initialized at
-// static initialization time (e.g. before program-start) rather than
-// than global construction time (which is after program-start but
-// before main), at least when 'value' is a compile-time constant.  We
-// use a small trick for the "default value" variable, and call it
-// FLAGS_no<name>.  This serves the second purpose of assuring a
-// compile error if someone tries to define a flag named no<name>
-// which is illegal (--foo and --nofoo both affect the "foo" flag).
-#define DEFINE_VARIABLE(type, shorttype, name, value, help)             \
-  namespace fL##shorttype {                                             \
-    static const type FLAGS_nono##name = value;                         \
-    /* We always want to export defined variables, dll or no */         \
-    GFLAGS_DLL_DEFINE_FLAG type FLAGS_##name = FLAGS_nono##name;        \
-    type FLAGS_no##name = FLAGS_nono##name;                             \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                   \
-      #name, #type, MAYBE_STRIPPED_HELP(help), __FILE__,                \
-      &FLAGS_##name, &FLAGS_no##name);                                  \
-  }                                                                     \
-  using fL##shorttype::FLAGS_##name
-
-// For DEFINE_bool, we want to do the extra check that the passed-in
-// value is actually a bool, and not a string or something that can be
-// coerced to a bool.  These declarations (no definition needed!) will
-// help us do that, and never evaluate From, which is important.
-// We'll use 'sizeof(IsBool(val))' to distinguish. This code requires
-// that the compiler have different sizes for bool & double. Since
-// this is not guaranteed by the standard, we check it with a
-// COMPILE_ASSERT.
-namespace fLB {
-struct CompileAssert {};
-typedef CompileAssert expected_sizeof_double_neq_sizeof_bool[
-                      (sizeof(double) != sizeof(bool)) ? 1 : -1];
-template<typename From> double GFLAGS_DLL_DECL IsBoolFlag(const From& from);
-GFLAGS_DLL_DECL bool IsBoolFlag(bool from);
-}  // namespace fLB
-
-// Here are the actual DEFINE_*-macros. The respective DECLARE_*-macros
-// are in a separate include, gflags_declare.h, for reducing
-// the physical transitive size for DECLARE use.
-#define DEFINE_bool(name, val, txt)                                     \
-  namespace fLB {                                                       \
-    typedef ::fLB::CompileAssert FLAG_##name##_value_is_not_a_bool[     \
-            (sizeof(::fLB::IsBoolFlag(val)) != sizeof(double))? 1: -1]; \
-  }                                                                     \
-  DEFINE_VARIABLE(bool, B, name, val, txt)
-
-#define DEFINE_int32(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int32, I, \
-                   name, val, txt)
-
-#define DEFINE_int64(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int64, I64, \
-                   name, val, txt)
-
-#define DEFINE_uint64(name,val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint64, U64, \
-                   name, val, txt)
-
-#define DEFINE_double(name, val, txt) \
-   DEFINE_VARIABLE(double, D, name, val, txt)
-
-// Strings are trickier, because they're not a POD, so we can't
-// construct them at static-initialization time (instead they get
-// constructed at global-constructor time, which is much later).  To
-// try to avoid crashes in that case, we use a char buffer to store
-// the string, which we can static-initialize, and then placement-new
-// into it later.  It's not perfect, but the best we can do.
-
-namespace fLS {
-
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const char *value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const clstring &value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           int value);
-}  // namespace fLS
-
-// We need to define a var named FLAGS_no##name so people don't define
-// --string and --nostring.  And we need a temporary place to put val
-// so we don't have to evaluate it twice.  Two great needs that go
-// great together!
-// The weird 'using' + 'extern' inside the fLS namespace is to work around
-// an unknown compiler bug/issue with the gcc 4.2.1 on SUSE 10.  See
-//    http://code.google.com/p/google-gflags/issues/detail?id=20
-#define DEFINE_string(name, val, txt)                                       \
-  namespace fLS {                                                           \
-    using ::fLS::clstring;                                                  \
-    static union { void* align; char s[sizeof(clstring)]; } s_##name[2];    \
-    clstring* const FLAGS_no##name = ::fLS::                                \
-                                   dont_pass0toDEFINE_string(s_##name[0].s, \
-                                                             val);          \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                       \
-        #name, "string", MAYBE_STRIPPED_HELP(txt), __FILE__,                \
-        s_##name[0].s, new (s_##name[1].s) clstring(*FLAGS_no##name));      \
-    extern GFLAGS_DLL_DEFINE_FLAG clstring& FLAGS_##name;                   \
-    using fLS::FLAGS_##name;                                                \
-    clstring& FLAGS_##name = *FLAGS_no##name;                               \
-  }                                                                         \
-  using fLS::FLAGS_##name
-
-#endif  // SWIG
-
-
-// Import gflags library symbols into alternative/deprecated namespace(s)
-#include "gflags_gflags.h"
-
-
-#endif  // GFLAGS_GFLAGS_H_
diff --git a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_completions.h b/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_completions.h
deleted file mode 100644
index f951c1e02d..0000000000
--- a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_completions.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright (c) 2008, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// ---
-
-//
-// Implement helpful bash-style command line flag completions
-//
-// ** Functional API:
-// HandleCommandLineCompletions() should be called early during
-// program startup, but after command line flag code has been
-// initialized, such as the beginning of HandleCommandLineHelpFlags().
-// It checks the value of the flag --tab_completion_word.  If this
-// flag is empty, nothing happens here.  If it contains a string,
-// however, then HandleCommandLineCompletions() will hijack the
-// process, attempting to identify the intention behind this
-// completion.  Regardless of the outcome of this deduction, the
-// process will be terminated, similar to --helpshort flag
-// handling.
-//
-// ** Overview of Bash completions:
-// Bash can be told to programatically determine completions for the
-// current 'cursor word'.  It does this by (in this case) invoking a
-// command with some additional arguments identifying the command
-// being executed, the word being completed, and the previous word
-// (if any).  Bash then expects a sequence of output lines to be
-// printed to stdout.  If these lines all contain a common prefix
-// longer than the cursor word, bash will replace the cursor word
-// with that common prefix, and display nothing.  If there isn't such
-// a common prefix, bash will display the lines in pages using 'more'.
-//
-// ** Strategy taken for command line completions:
-// If we can deduce either the exact flag intended, or a common flag
-// prefix, we'll output exactly that.  Otherwise, if information
-// must be displayed to the user, we'll take the opportunity to add
-// some helpful information beyond just the flag name (specifically,
-// we'll include the default flag value and as much of the flag's
-// description as can fit on a single terminal line width, as specified
-// by the flag --tab_completion_columns).  Furthermore, we'll try to
-// make bash order the output such that the most useful or relevent
-// flags are the most likely to be shown at the top.
-//
-// ** Additional features:
-// To assist in finding that one really useful flag, substring matching
-// was implemented.  Before pressing a <TAB> to get completion for the
-// current word, you can append one or more '?' to the flag to do
-// substring matching.  Here's the semantics:
-//   --foo<TAB>     Show me all flags with names prefixed by 'foo'
-//   --foo?<TAB>    Show me all flags with 'foo' somewhere in the name
-//   --foo??<TAB>   Same as prior case, but also search in module
-//                  definition path for 'foo'
-//   --foo???<TAB>  Same as prior case, but also search in flag
-//                  descriptions for 'foo'
-// Finally, we'll trim the output to a relatively small number of
-// flags to keep bash quiet about the verbosity of output.  If one
-// really wanted to see all possible matches, appending a '+' to the
-// search word will force the exhaustive list of matches to be printed.
-//
-// ** How to have bash accept completions from a binary:
-// Bash requires that it be informed about each command that programmatic
-// completion should be enabled for.  Example addition to a .bashrc
-// file would be (your path to gflags_completions.sh file may differ):
-
-/*
-$ complete -o bashdefault -o default -o nospace -C                            \
- '/home/build/eng/bash/bash_completions.sh --tab_completion_columns $COLUMNS' \
-  time  env  binary_name  another_binary  [...]
-*/
-
-// This would allow the following to work:
-//   $ /path/to/binary_name --vmodule<TAB>
-// Or:
-//   $ ./bin/path/another_binary --gfs_u<TAB>
-// (etc)
-//
-// Sadly, it appears that bash gives no easy way to force this behavior for
-// all commands.  That's where the "time" in the above example comes in.
-// If you haven't specifically added a command to the list of completion
-// supported commands, you can still get completions by prefixing the
-// entire command with "env".
-//   $ env /some/brand/new/binary --vmod<TAB>
-// Assuming that "binary" is a newly compiled binary, this should still
-// produce the expected completion output.
-
-
-#ifndef GFLAGS_COMPLETIONS_H_
-#define GFLAGS_COMPLETIONS_H_
-
-namespace google {
-
-extern void HandleCommandLineCompletions(void);
-
-}
-
-#endif  // GFLAGS_COMPLETIONS_H_
diff --git a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_declare.h b/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_declare.h
deleted file mode 100644
index 935a20e775..0000000000
--- a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_declare.h
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright (c) 1999, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-//
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// command line flag.
-
-#ifndef GFLAGS_DECLARE_H_
-#define GFLAGS_DECLARE_H_
-
-
-// ---------------------------------------------------------------------------
-// Namespace of gflags library symbols.
-#define GFLAGS_NAMESPACE google
-
-// ---------------------------------------------------------------------------
-// Windows DLL import/export.
-
-// We always want to import the symbols of the gflags library
-#ifndef GFLAGS_DLL_DECL
-#  if 0 && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-
-// We always want to import variables declared in user code
-#ifndef GFLAGS_DLL_DECLARE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DECLARE_FLAG __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECLARE_FLAG
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Flag types
-#include <string>
-#if 1
-#  include <stdint.h>                   // the normal place uint32_t is defined
-#elif 1
-#  include <sys/types.h>                // the normal place u_int32_t is defined
-#elif 1
-#  include <inttypes.h>                 // a third place for uint32_t or u_int32_t
-#endif
-
-namespace GFLAGS_NAMESPACE {
-
-#if 1 // C99
-typedef int32_t          int32;
-typedef uint32_t         uint32;
-typedef int64_t          int64;
-typedef uint64_t         uint64;
-#elif 0 // BSD
-typedef int32_t          int32;
-typedef u_int32_t        uint32;
-typedef int64_t          int64;
-typedef u_int64_t        uint64;
-#elif 0 // Windows
-typedef __int32          int32;
-typedef unsigned __int32 uint32;
-typedef __int64          int64;
-typedef unsigned __int64 uint64;
-#else
-#  error Do not know how to define a 32-bit integer quantity on your system
-#endif
-
-} // namespace GFLAGS_NAMESPACE
-
-
-namespace fLS {
-
-// The meaning of "string" might be different between now and when the
-// macros below get invoked (e.g., if someone is experimenting with
-// other string implementations that get defined after this file is
-// included).  Save the current meaning now and use it in the macros.
-typedef std::string clstring;
-
-} // namespace fLS
-
-
-#define DECLARE_VARIABLE(type, shorttype, name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fL##shorttype { extern GFLAGS_DLL_DECLARE_FLAG type FLAGS_##name; } \
-  using fL##shorttype::FLAGS_##name
-
-#define DECLARE_bool(name) \
-  DECLARE_VARIABLE(bool, B, name)
-
-#define DECLARE_int32(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int32, I, name)
-
-#define DECLARE_int64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int64, I64, name)
-
-#define DECLARE_uint64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint64, U64, name)
-
-#define DECLARE_double(name) \
-  DECLARE_VARIABLE(double, D, name)
-
-#define DECLARE_string(name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fLS { \
-  using ::fLS::clstring; \
-  extern GFLAGS_DLL_DECLARE_FLAG ::fLS::clstring& FLAGS_##name; \
-  } \
-  using fLS::FLAGS_##name
-
-
-#endif  // GFLAGS_DECLARE_H_
diff --git a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_gflags.h b/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_gflags.h
deleted file mode 100644
index 0c17825dd6..0000000000
--- a/media/libyuv/third_party/gflags/gen/posix/include/gflags/gflags_gflags.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Copyright (c) 2014, Andreas Schuh
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// -----------------------------------------------------------------------------
-// Imports the gflags library symbols into an alternative/deprecated namespace.
-
-#ifndef GFLAGS_GFLAGS_H_
-#  error The internal header gflags_gflags.h may only be included by gflags.h
-#endif
-
-#ifndef GFLAGS_NS_GFLAGS_H_
-#define GFLAGS_NS_GFLAGS_H_
-
-
-namespace gflags {
-
-
-using GFLAGS_NAMESPACE::int32;
-using GFLAGS_NAMESPACE::uint32;
-using GFLAGS_NAMESPACE::int64;
-using GFLAGS_NAMESPACE::uint64;
-
-using GFLAGS_NAMESPACE::RegisterFlagValidator;
-using GFLAGS_NAMESPACE::CommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetAllFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlagsRestrict;
-using GFLAGS_NAMESPACE::DescribeOneFlag;
-using GFLAGS_NAMESPACE::SetArgv;
-using GFLAGS_NAMESPACE::GetArgvs;
-using GFLAGS_NAMESPACE::GetArgv;
-using GFLAGS_NAMESPACE::GetArgv0;
-using GFLAGS_NAMESPACE::GetArgvSum;
-using GFLAGS_NAMESPACE::ProgramInvocationName;
-using GFLAGS_NAMESPACE::ProgramInvocationShortName;
-using GFLAGS_NAMESPACE::ProgramUsage;
-using GFLAGS_NAMESPACE::VersionString;
-using GFLAGS_NAMESPACE::GetCommandLineOption;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfoOrDie;
-using GFLAGS_NAMESPACE::FlagSettingMode;
-using GFLAGS_NAMESPACE::SET_FLAGS_VALUE;
-using GFLAGS_NAMESPACE::SET_FLAG_IF_DEFAULT;
-using GFLAGS_NAMESPACE::SET_FLAGS_DEFAULT;
-using GFLAGS_NAMESPACE::SetCommandLineOption;
-using GFLAGS_NAMESPACE::SetCommandLineOptionWithMode;
-using GFLAGS_NAMESPACE::FlagSaver;
-using GFLAGS_NAMESPACE::CommandlineFlagsIntoString;
-using GFLAGS_NAMESPACE::ReadFlagsFromString;
-using GFLAGS_NAMESPACE::AppendFlagsIntoFile;
-using GFLAGS_NAMESPACE::ReadFromFlagsFile;
-using GFLAGS_NAMESPACE::BoolFromEnv;
-using GFLAGS_NAMESPACE::Int32FromEnv;
-using GFLAGS_NAMESPACE::Int64FromEnv;
-using GFLAGS_NAMESPACE::Uint64FromEnv;
-using GFLAGS_NAMESPACE::DoubleFromEnv;
-using GFLAGS_NAMESPACE::StringFromEnv;
-using GFLAGS_NAMESPACE::SetUsageMessage;
-using GFLAGS_NAMESPACE::SetVersionString;
-using GFLAGS_NAMESPACE::ParseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::HandleCommandLineHelpFlags;
-using GFLAGS_NAMESPACE::AllowCommandLineReparsing;
-using GFLAGS_NAMESPACE::ReparseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::ShutDownCommandLineFlags;
-using GFLAGS_NAMESPACE::FlagRegisterer;
-
-#ifndef SWIG
-using GFLAGS_NAMESPACE::ParseCommandLineFlags;
-#endif
-
-
-} // namespace gflags
-
-
-#endif  // GFLAGS_NS_GFLAGS_H_
diff --git a/media/libyuv/third_party/gflags/gen/posix/include/private/config.h b/media/libyuv/third_party/gflags/gen/posix/include/private/config.h
deleted file mode 100644
index 592d61c4c0..0000000000
--- a/media/libyuv/third_party/gflags/gen/posix/include/private/config.h
+++ /dev/null
@@ -1,112 +0,0 @@
-/* Generated from config.h.in during build configuration using CMake. */
-
-// Note: This header file is only used internally. It is not part of public interface!
-
-// ---------------------------------------------------------------------------
-// System checks
-
-// Define if you build this library for a MS Windows OS.
-/* #undef OS_WINDOWS */
-
-// Define if you have the <stdint.h> header file.
-#define HAVE_STDINT_H
-
-// Define if you have the <sys/types.h> header file.
-#define HAVE_SYS_TYPES_H
-
-// Define if you have the <inttypes.h> header file.
-#define HAVE_INTTYPES_H
-
-// Define if you have the <sys/stat.h> header file.
-#define HAVE_SYS_STAT_H
-
-// Define if you have the <unistd.h> header file.
-#define HAVE_UNISTD_H
-
-// Define if you have the <fnmatch.h> header file.
-#define HAVE_FNMATCH_H
-
-// Define if you have the <shlwapi.h> header file (Windows 2000/XP).
-/* #undef HAVE_SHLWAPI_H */
-
-// Define if you have the strtoll function.
-#define HAVE_STRTOLL
-
-// Define if you have the strtoq function.
-/* #undef HAVE_STRTOQ */
-
-// Define if you have the <pthread.h> header file.
-#define HAVE_PTHREAD
-
-// Define if your pthread library defines the type pthread_rwlock_t
-#define HAVE_RWLOCK
-
-// gcc requires this to get PRId64, etc.
-#if defined(HAVE_INTTYPES_H) && !defined(__STDC_FORMAT_MACROS)
-#  define __STDC_FORMAT_MACROS 1
-#endif
-
-// ---------------------------------------------------------------------------
-// Package information
-
-// Name of package.
-#define PACKAGE gflags
-
-// Define to the full name of this package.
-#define PACKAGE_NAME gflags
-
-// Define to the full name and version of this package.
-#define PACKAGE_STRING gflags 2.2.0
-
-// Define to the one symbol short name of this package.
-#define PACKAGE_TARNAME gflags-2.2.0
-
-// Define to the version of this package.
-#define PACKAGE_VERSION 2.2.0
-
-// Version number of package.
-#define VERSION PACKAGE_VERSION
-
-// Define to the address where bug reports for this package should be sent.
-#define PACKAGE_BUGREPORT https://github.com/schuhschuh/gflags/issues
-
-// ---------------------------------------------------------------------------
-// Path separator
-#ifndef PATH_SEPARATOR
-#  ifdef OS_WINDOWS
-#    define PATH_SEPARATOR  '\\'
-#  else
-#    define PATH_SEPARATOR  '/'
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Windows
-
-// Whether gflags library is a DLL.
-#ifndef GFLAGS_IS_A_DLL
-#  define GFLAGS_IS_A_DLL 0
-#endif
-
-// Always export symbols when compiling a shared library as this file is only
-// included by internal modules when building the gflags library itself.
-// The gflags_declare.h header file will set it to import these symbols otherwise.
-#ifndef GFLAGS_DLL_DECL
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-// Flags defined by the gflags library itself must be exported
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  define GFLAGS_DLL_DEFINE_FLAG GFLAGS_DLL_DECL
-#endif
-
-#ifdef OS_WINDOWS
-// The unittests import the symbols of the shared gflags library
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL_FOR_UNITTESTS __declspec(dllimport)
-#  endif
-#  include "windows_port.h"
-#endif
diff --git a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags.h b/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags.h
deleted file mode 100644
index 357eec6be7..0000000000
--- a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags.h
+++ /dev/null
@@ -1,573 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// or defines a command line flag or wants to parse command line flags
-// or print a program usage message (which will include information about
-// flags).  Executive summary, in the form of an example foo.cc file:
-//
-//    #include "foo.h"         // foo.h has a line "DECLARE_int32(start);"
-//    #include "validators.h"  // hypothetical file defining ValidateIsFile()
-//
-//    DEFINE_int32(end, 1000, "The last record to read");
-//
-//    DEFINE_string(filename, "my_file.txt", "The file to read");
-//    // Crash if the specified file does not exist.
-//    static bool dummy = RegisterFlagValidator(&FLAGS_filename,
-//                                              &ValidateIsFile);
-//
-//    DECLARE_bool(verbose); // some other file has a DEFINE_bool(verbose, ...)
-//
-//    void MyFunc() {
-//      if (FLAGS_verbose) printf("Records %d-%d\n", FLAGS_start, FLAGS_end);
-//    }
-//
-//    Then, at the command-line:
-//       ./foo --noverbose --start=5 --end=100
-//
-// For more details, see
-//    doc/gflags.html
-//
-// --- A note about thread-safety:
-//
-// We describe many functions in this routine as being thread-hostile,
-// thread-compatible, or thread-safe.  Here are the meanings we use:
-//
-// thread-safe: it is safe for multiple threads to call this routine
-//   (or, when referring to a class, methods of this class)
-//   concurrently.
-// thread-hostile: it is not safe for multiple threads to call this
-//   routine (or methods of this class) concurrently.  In gflags,
-//   most thread-hostile routines are intended to be called early in,
-//   or even before, main() -- that is, before threads are spawned.
-// thread-compatible: it is safe for multiple threads to read from
-//   this variable (when applied to variables), or to call const
-//   methods of this class (when applied to classes), as long as no
-//   other thread is writing to the variable or calling non-const
-//   methods of this class.
-
-#ifndef GFLAGS_GFLAGS_H_
-#define GFLAGS_GFLAGS_H_
-
-#include <string>
-#include <vector>
-
-#include "gflags_declare.h" // IWYU pragma: export
-
-
-// We always want to export variables defined in user code
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DEFINE_FLAG __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DEFINE_FLAG
-#  endif
-#endif
-
-
-namespace GFLAGS_NAMESPACE {
-
-
-// --------------------------------------------------------------------
-// To actually define a flag in a file, use DEFINE_bool,
-// DEFINE_string, etc. at the bottom of this file.  You may also find
-// it useful to register a validator with the flag.  This ensures that
-// when the flag is parsed from the commandline, or is later set via
-// SetCommandLineOption, we call the validation function. It is _not_
-// called when you assign the value to the flag directly using the = operator.
-//
-// The validation function should return true if the flag value is valid, and
-// false otherwise. If the function returns false for the new setting of the
-// flag, the flag will retain its current value. If it returns false for the
-// default value, ParseCommandLineFlags() will die.
-//
-// This function is safe to call at global construct time (as in the
-// example below).
-//
-// Example use:
-//    static bool ValidatePort(const char* flagname, int32 value) {
-//       if (value > 0 && value < 32768)   // value is ok
-//         return true;
-//       printf("Invalid value for --%s: %d\n", flagname, (int)value);
-//       return false;
-//    }
-//    DEFINE_int32(port, 0, "What port to listen on");
-//    static bool dummy = RegisterFlagValidator(&FLAGS_port, &ValidatePort);
-
-// Returns true if successfully registered, false if not (because the
-// first argument doesn't point to a command-line flag, or because a
-// validator is already registered for this flag).
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const bool*        flag, bool (*validate_fn)(const char*, bool));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int32*       flag, bool (*validate_fn)(const char*, int32));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int64*       flag, bool (*validate_fn)(const char*, int64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint64*      flag, bool (*validate_fn)(const char*, uint64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const double*      flag, bool (*validate_fn)(const char*, double));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const std::string* flag, bool (*validate_fn)(const char*, const std::string&));
-
-// Convenience macro for the registration of a flag validator
-#define DEFINE_validator(name, validator) \
-    static const bool name##_validator_registered = \
-            GFLAGS_NAMESPACE::RegisterFlagValidator(&FLAGS_##name, validator)
-
-
-// --------------------------------------------------------------------
-// These methods are the best way to get access to info about the
-// list of commandline flags.  Note that these routines are pretty slow.
-//   GetAllFlags: mostly-complete info about the list, sorted by file.
-//   ShowUsageWithFlags: pretty-prints the list to stdout (what --help does)
-//   ShowUsageWithFlagsRestrict: limit to filenames with restrict as a substr
-//
-// In addition to accessing flags, you can also access argv[0] (the program
-// name) and argv (the entire commandline), which we sock away a copy of.
-// These variables are static, so you should only set them once.
-//
-// No need to export this data only structure from DLL, avoiding VS warning 4251.
-struct CommandLineFlagInfo {
-  std::string name;            // the name of the flag
-  std::string type;            // the type of the flag: int32, etc
-  std::string description;     // the "help text" associated with the flag
-  std::string current_value;   // the current value, as a string
-  std::string default_value;   // the default value, as a string
-  std::string filename;        // 'cleaned' version of filename holding the flag
-  bool has_validator_fn;       // true if RegisterFlagValidator called on this flag
-  bool is_default;             // true if the flag has the default value and
-                               // has not been set explicitly from the cmdline
-                               // or via SetCommandLineOption
-  const void* flag_ptr;        // pointer to the flag's current value (i.e. FLAGS_foo)
-};
-
-// Using this inside of a validator is a recipe for a deadlock.
-// TODO(user) Fix locking when validators are running, to make it safe to
-// call validators during ParseAllFlags.
-// Also make sure then to uncomment the corresponding unit test in
-// gflags_unittest.sh
-extern GFLAGS_DLL_DECL void GetAllFlags(std::vector<CommandLineFlagInfo>* OUTPUT);
-// These two are actually defined in gflags_reporting.cc.
-extern GFLAGS_DLL_DECL void ShowUsageWithFlags(const char *argv0);  // what --help does
-extern GFLAGS_DLL_DECL void ShowUsageWithFlagsRestrict(const char *argv0, const char *restrict);
-
-// Create a descriptive string for a flag.
-// Goes to some trouble to make pretty line breaks.
-extern GFLAGS_DLL_DECL std::string DescribeOneFlag(const CommandLineFlagInfo& flag);
-
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetArgv(int argc, const char** argv);
-
-// The following functions are thread-safe as long as SetArgv() is
-// only called before any threads start.
-extern GFLAGS_DLL_DECL const std::vector<std::string>& GetArgvs();
-extern GFLAGS_DLL_DECL const char* GetArgv();                      // all of argv as a string
-extern GFLAGS_DLL_DECL const char* GetArgv0();                     // only argv0
-extern GFLAGS_DLL_DECL uint32 GetArgvSum();                        // simple checksum of argv
-extern GFLAGS_DLL_DECL const char* ProgramInvocationName();        // argv0, or "UNKNOWN" if not set
-extern GFLAGS_DLL_DECL const char* ProgramInvocationShortName();   // basename(argv0)
-
-// ProgramUsage() is thread-safe as long as SetUsageMessage() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* ProgramUsage();                 // string set by SetUsageMessage()
-
-// VersionString() is thread-safe as long as SetVersionString() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* VersionString();                // string set by SetVersionString()
-
-
-
-// --------------------------------------------------------------------
-// Normally you access commandline flags by just saying "if (FLAGS_foo)"
-// or whatever, and set them by calling "FLAGS_foo = bar" (or, more
-// commonly, via the DEFINE_foo macro).  But if you need a bit more
-// control, we have programmatic ways to get/set the flags as well.
-// These programmatic ways to access flags are thread-safe, but direct
-// access is only thread-compatible.
-
-// Return true iff the flagname was found.
-// OUTPUT is set to the flag's value, or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineOption(const char* name, std::string* OUTPUT);
-
-// Return true iff the flagname was found. OUTPUT is set to the flag's
-// CommandLineFlagInfo or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineFlagInfo(const char* name, CommandLineFlagInfo* OUTPUT);
-
-// Return the CommandLineFlagInfo of the flagname.  exit() if name not found.
-// Example usage, to check if a flag's value is currently the default value:
-//   if (GetCommandLineFlagInfoOrDie("foo").is_default) ...
-extern GFLAGS_DLL_DECL CommandLineFlagInfo GetCommandLineFlagInfoOrDie(const char* name);
-
-enum GFLAGS_DLL_DECL FlagSettingMode {
-  // update the flag's value (can call this multiple times).
-  SET_FLAGS_VALUE,
-  // update the flag's value, but *only if* it has not yet been updated
-  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
-  SET_FLAG_IF_DEFAULT,
-  // set the flag's default value to this.  If the flag has not yet updated
-  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
-  // change the flag's current value to the new default value as well.
-  SET_FLAGS_DEFAULT
-};
-
-// Set a particular flag ("command line option").  Returns a string
-// describing the new value that the option has been set to.  The
-// return value API is not well-specified, so basically just depend on
-// it to be empty if the setting failed for some reason -- the name is
-// not a valid flag name, or the value is not a valid value -- and
-// non-empty else.
-
-// SetCommandLineOption uses set_mode == SET_FLAGS_VALUE (the common case)
-extern GFLAGS_DLL_DECL std::string SetCommandLineOption        (const char* name, const char* value);
-extern GFLAGS_DLL_DECL std::string SetCommandLineOptionWithMode(const char* name, const char* value, FlagSettingMode set_mode);
-
-
-// --------------------------------------------------------------------
-// Saves the states (value, default value, whether the user has set
-// the flag, registered validators, etc) of all flags, and restores
-// them when the FlagSaver is destroyed.  This is very useful in
-// tests, say, when you want to let your tests change the flags, but
-// make sure that they get reverted to the original states when your
-// test is complete.
-//
-// Example usage:
-//   void TestFoo() {
-//     FlagSaver s1;
-//     FLAG_foo = false;
-//     FLAG_bar = "some value";
-//
-//     // test happens here.  You can return at any time
-//     // without worrying about restoring the FLAG values.
-//   }
-//
-// Note: This class is marked with GFLAGS_ATTRIBUTE_UNUSED because all
-// the work is done in the constructor and destructor, so in the standard
-// usage example above, the compiler would complain that it's an
-// unused variable.
-//
-// This class is thread-safe.  However, its destructor writes to
-// exactly the set of flags that have changed value during its
-// lifetime, so concurrent _direct_ access to those flags
-// (i.e. FLAGS_foo instead of {Get,Set}CommandLineOption()) is unsafe.
-
-class GFLAGS_DLL_DECL FlagSaver {
- public:
-  FlagSaver();
-  ~FlagSaver();
-
- private:
-  class FlagSaverImpl* impl_;   // we use pimpl here to keep API steady
-
-  FlagSaver(const FlagSaver&);  // no copying!
-  void operator=(const FlagSaver&);
-};
-
-// --------------------------------------------------------------------
-// Some deprecated or hopefully-soon-to-be-deprecated functions.
-
-// This is often used for logging.  TODO(csilvers): figure out a better way
-extern GFLAGS_DLL_DECL std::string CommandlineFlagsIntoString();
-// Usually where this is used, a FlagSaver should be used instead.
-extern GFLAGS_DLL_DECL
-bool ReadFlagsFromString(const std::string& flagfilecontents,
-                         const char* prog_name,
-                         bool errors_are_fatal);  // uses SET_FLAGS_VALUE
-
-// These let you manually implement --flagfile functionality.
-// DEPRECATED.
-extern GFLAGS_DLL_DECL bool AppendFlagsIntoFile(const std::string& filename, const char* prog_name);
-extern GFLAGS_DLL_DECL bool ReadFromFlagsFile(const std::string& filename, const char* prog_name, bool errors_are_fatal);   // uses SET_FLAGS_VALUE
-
-
-// --------------------------------------------------------------------
-// Useful routines for initializing flags from the environment.
-// In each case, if 'varname' does not exist in the environment
-// return defval.  If 'varname' does exist but is not valid
-// (e.g., not a number for an int32 flag), abort with an error.
-// Otherwise, return the value.  NOTE: for booleans, for true use
-// 't' or 'T' or 'true' or '1', for false 'f' or 'F' or 'false' or '0'.
-
-extern GFLAGS_DLL_DECL bool BoolFromEnv(const char *varname, bool defval);
-extern GFLAGS_DLL_DECL int32 Int32FromEnv(const char *varname, int32 defval);
-extern GFLAGS_DLL_DECL int64 Int64FromEnv(const char *varname, int64 defval);
-extern GFLAGS_DLL_DECL uint64 Uint64FromEnv(const char *varname, uint64 defval);
-extern GFLAGS_DLL_DECL double DoubleFromEnv(const char *varname, double defval);
-extern GFLAGS_DLL_DECL const char *StringFromEnv(const char *varname, const char *defval);
-
-
-// --------------------------------------------------------------------
-// The next two functions parse gflags from main():
-
-// Set the "usage" message for this program.  For example:
-//   string usage("This program does nothing.  Sample usage:\n");
-//   usage += argv[0] + " <uselessarg1> <uselessarg2>";
-//   SetUsageMessage(usage);
-// Do not include commandline flags in the usage: we do that for you!
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetUsageMessage(const std::string& usage);
-
-// Sets the version string, which is emitted with --version.
-// For instance: SetVersionString("1.3");
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetVersionString(const std::string& version);
-
-
-// Looks for flags in argv and parses them.  Rearranges argv to put
-// flags first, or removes them entirely if remove_flags is true.
-// If a flag is defined more than once in the command line or flag
-// file, the last definition is used.  Returns the index (into argv)
-// of the first non-flag argument.
-// See top-of-file for more details on this function.
-#ifndef SWIG   // In swig, use ParseCommandLineFlagsScript() instead.
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineFlags(int *argc, char*** argv, bool remove_flags);
-#endif
-
-
-// Calls to ParseCommandLineNonHelpFlags and then to
-// HandleCommandLineHelpFlags can be used instead of a call to
-// ParseCommandLineFlags during initialization, in order to allow for
-// changing default values for some FLAGS (via
-// e.g. SetCommandLineOptionWithMode calls) between the time of
-// command line parsing and the time of dumping help information for
-// the flags as a result of command line parsing.  If a flag is
-// defined more than once in the command line or flag file, the last
-// definition is used.  Returns the index (into argv) of the first
-// non-flag argument.  (If remove_flags is true, will always return 1.)
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineNonHelpFlags(int *argc, char*** argv, bool remove_flags);
-
-// This is actually defined in gflags_reporting.cc.
-// This function is misnamed (it also handles --version, etc.), but
-// it's too late to change that now. :-(
-extern GFLAGS_DLL_DECL void HandleCommandLineHelpFlags();   // in gflags_reporting.cc
-
-// Allow command line reparsing.  Disables the error normally
-// generated when an unknown flag is found, since it may be found in a
-// later parse.  Thread-hostile; meant to be called before any threads
-// are spawned.
-extern GFLAGS_DLL_DECL void AllowCommandLineReparsing();
-
-// Reparse the flags that have not yet been recognized.  Only flags
-// registered since the last parse will be recognized.  Any flag value
-// must be provided as part of the argument using "=", not as a
-// separate command line argument that follows the flag argument.
-// Intended for handling flags from dynamically loaded libraries,
-// since their flags are not registered until they are loaded.
-extern GFLAGS_DLL_DECL void ReparseCommandLineNonHelpFlags();
-
-// Clean up memory allocated by flags.  This is only needed to reduce
-// the quantity of "potentially leaked" reports emitted by memory
-// debugging tools such as valgrind.  It is not required for normal
-// operation, or for the google perftools heap-checker.  It must only
-// be called when the process is about to exit, and all threads that
-// might access flags are quiescent.  Referencing flags after this is
-// called will have unexpected consequences.  This is not safe to run
-// when multiple threads might be running: the function is
-// thread-hostile.
-extern GFLAGS_DLL_DECL void ShutDownCommandLineFlags();
-
-
-// --------------------------------------------------------------------
-// Now come the command line flag declaration/definition macros that
-// will actually be used.  They're kind of hairy.  A major reason
-// for this is initialization: we want people to be able to access
-// variables in global constructors and have that not crash, even if
-// their global constructor runs before the global constructor here.
-// (Obviously, we can't guarantee the flags will have the correct
-// default value in that case, but at least accessing them is safe.)
-// The only way to do that is have flags point to a static buffer.
-// So we make one, using a union to ensure proper alignment, and
-// then use placement-new to actually set up the flag with the
-// correct default value.  In the same vein, we have to worry about
-// flag access in global destructors, so FlagRegisterer has to be
-// careful never to destroy the flag-values it constructs.
-//
-// Note that when we define a flag variable FLAGS_<name>, we also
-// preemptively define a junk variable, FLAGS_no<name>.  This is to
-// cause a link-time error if someone tries to define 2 flags with
-// names like "logging" and "nologging".  We do this because a bool
-// flag FLAG can be set from the command line to true with a "-FLAG"
-// argument, and to false with a "-noFLAG" argument, and so this can
-// potentially avert confusion.
-//
-// We also put flags into their own namespace.  It is purposefully
-// named in an opaque way that people should have trouble typing
-// directly.  The idea is that DEFINE puts the flag in the weird
-// namespace, and DECLARE imports the flag from there into the current
-// namespace.  The net result is to force people to use DECLARE to get
-// access to a flag, rather than saying "extern GFLAGS_DLL_DECL bool FLAGS_whatever;"
-// or some such instead.  We want this so we can put extra
-// functionality (like sanity-checking) in DECLARE if we want, and
-// make sure it is picked up everywhere.
-//
-// We also put the type of the variable in the namespace, so that
-// people can't DECLARE_int32 something that they DEFINE_bool'd
-// elsewhere.
-
-class GFLAGS_DLL_DECL FlagRegisterer {
- public:
-  FlagRegisterer(const char* name, const char* type,
-                 const char* help, const char* filename,
-                 void* current_storage, void* defvalue_storage);
-};
-
-// If your application #defines STRIP_FLAG_HELP to a non-zero value
-// before #including this file, we remove the help message from the
-// binary file. This can reduce the size of the resulting binary
-// somewhat, and may also be useful for security reasons.
-
-extern GFLAGS_DLL_DECL const char kStrippedFlagHelp[];
-
-
-} // namespace GFLAGS_NAMESPACE
-
-
-#ifndef SWIG  // In swig, ignore the main flag declarations
-
-#if defined(STRIP_FLAG_HELP) && STRIP_FLAG_HELP > 0
-// Need this construct to avoid the 'defined but not used' warning.
-#define MAYBE_STRIPPED_HELP(txt) \
-   (false ? (txt) : GFLAGS_NAMESPACE::kStrippedFlagHelp)
-#else
-#define MAYBE_STRIPPED_HELP(txt) txt
-#endif
-
-// Each command-line flag has two variables associated with it: one
-// with the current value, and one with the default value.  However,
-// we have a third variable, which is where value is assigned; it's a
-// constant.  This guarantees that FLAG_##value is initialized at
-// static initialization time (e.g. before program-start) rather than
-// than global construction time (which is after program-start but
-// before main), at least when 'value' is a compile-time constant.  We
-// use a small trick for the "default value" variable, and call it
-// FLAGS_no<name>.  This serves the second purpose of assuring a
-// compile error if someone tries to define a flag named no<name>
-// which is illegal (--foo and --nofoo both affect the "foo" flag).
-#define DEFINE_VARIABLE(type, shorttype, name, value, help)             \
-  namespace fL##shorttype {                                             \
-    static const type FLAGS_nono##name = value;                         \
-    /* We always want to export defined variables, dll or no */         \
-    GFLAGS_DLL_DEFINE_FLAG type FLAGS_##name = FLAGS_nono##name;        \
-    type FLAGS_no##name = FLAGS_nono##name;                             \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                   \
-      #name, #type, MAYBE_STRIPPED_HELP(help), __FILE__,                \
-      &FLAGS_##name, &FLAGS_no##name);                                  \
-  }                                                                     \
-  using fL##shorttype::FLAGS_##name
-
-// For DEFINE_bool, we want to do the extra check that the passed-in
-// value is actually a bool, and not a string or something that can be
-// coerced to a bool.  These declarations (no definition needed!) will
-// help us do that, and never evaluate From, which is important.
-// We'll use 'sizeof(IsBool(val))' to distinguish. This code requires
-// that the compiler have different sizes for bool & double. Since
-// this is not guaranteed by the standard, we check it with a
-// COMPILE_ASSERT.
-namespace fLB {
-struct CompileAssert {};
-typedef CompileAssert expected_sizeof_double_neq_sizeof_bool[
-                      (sizeof(double) != sizeof(bool)) ? 1 : -1];
-template<typename From> double GFLAGS_DLL_DECL IsBoolFlag(const From& from);
-GFLAGS_DLL_DECL bool IsBoolFlag(bool from);
-}  // namespace fLB
-
-// Here are the actual DEFINE_*-macros. The respective DECLARE_*-macros
-// are in a separate include, gflags_declare.h, for reducing
-// the physical transitive size for DECLARE use.
-#define DEFINE_bool(name, val, txt)                                     \
-  namespace fLB {                                                       \
-    typedef ::fLB::CompileAssert FLAG_##name##_value_is_not_a_bool[     \
-            (sizeof(::fLB::IsBoolFlag(val)) != sizeof(double))? 1: -1]; \
-  }                                                                     \
-  DEFINE_VARIABLE(bool, B, name, val, txt)
-
-#define DEFINE_int32(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int32, I, \
-                   name, val, txt)
-
-#define DEFINE_int64(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int64, I64, \
-                   name, val, txt)
-
-#define DEFINE_uint64(name,val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint64, U64, \
-                   name, val, txt)
-
-#define DEFINE_double(name, val, txt) \
-   DEFINE_VARIABLE(double, D, name, val, txt)
-
-// Strings are trickier, because they're not a POD, so we can't
-// construct them at static-initialization time (instead they get
-// constructed at global-constructor time, which is much later).  To
-// try to avoid crashes in that case, we use a char buffer to store
-// the string, which we can static-initialize, and then placement-new
-// into it later.  It's not perfect, but the best we can do.
-
-namespace fLS {
-
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const char *value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const clstring &value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           int value);
-}  // namespace fLS
-
-// We need to define a var named FLAGS_no##name so people don't define
-// --string and --nostring.  And we need a temporary place to put val
-// so we don't have to evaluate it twice.  Two great needs that go
-// great together!
-// The weird 'using' + 'extern' inside the fLS namespace is to work around
-// an unknown compiler bug/issue with the gcc 4.2.1 on SUSE 10.  See
-//    http://code.google.com/p/google-gflags/issues/detail?id=20
-#define DEFINE_string(name, val, txt)                                       \
-  namespace fLS {                                                           \
-    using ::fLS::clstring;                                                  \
-    static union { void* align; char s[sizeof(clstring)]; } s_##name[2];    \
-    clstring* const FLAGS_no##name = ::fLS::                                \
-                                   dont_pass0toDEFINE_string(s_##name[0].s, \
-                                                             val);          \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                       \
-        #name, "string", MAYBE_STRIPPED_HELP(txt), __FILE__,                \
-        s_##name[0].s, new (s_##name[1].s) clstring(*FLAGS_no##name));      \
-    extern GFLAGS_DLL_DEFINE_FLAG clstring& FLAGS_##name;                   \
-    using fLS::FLAGS_##name;                                                \
-    clstring& FLAGS_##name = *FLAGS_no##name;                               \
-  }                                                                         \
-  using fLS::FLAGS_##name
-
-#endif  // SWIG
-
-
-// Import gflags library symbols into alternative/deprecated namespace(s)
-#include "gflags_gflags.h"
-
-
-#endif  // GFLAGS_GFLAGS_H_
diff --git a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_completions.h b/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_completions.h
deleted file mode 100644
index f951c1e02d..0000000000
--- a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_completions.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright (c) 2008, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// ---
-
-//
-// Implement helpful bash-style command line flag completions
-//
-// ** Functional API:
-// HandleCommandLineCompletions() should be called early during
-// program startup, but after command line flag code has been
-// initialized, such as the beginning of HandleCommandLineHelpFlags().
-// It checks the value of the flag --tab_completion_word.  If this
-// flag is empty, nothing happens here.  If it contains a string,
-// however, then HandleCommandLineCompletions() will hijack the
-// process, attempting to identify the intention behind this
-// completion.  Regardless of the outcome of this deduction, the
-// process will be terminated, similar to --helpshort flag
-// handling.
-//
-// ** Overview of Bash completions:
-// Bash can be told to programatically determine completions for the
-// current 'cursor word'.  It does this by (in this case) invoking a
-// command with some additional arguments identifying the command
-// being executed, the word being completed, and the previous word
-// (if any).  Bash then expects a sequence of output lines to be
-// printed to stdout.  If these lines all contain a common prefix
-// longer than the cursor word, bash will replace the cursor word
-// with that common prefix, and display nothing.  If there isn't such
-// a common prefix, bash will display the lines in pages using 'more'.
-//
-// ** Strategy taken for command line completions:
-// If we can deduce either the exact flag intended, or a common flag
-// prefix, we'll output exactly that.  Otherwise, if information
-// must be displayed to the user, we'll take the opportunity to add
-// some helpful information beyond just the flag name (specifically,
-// we'll include the default flag value and as much of the flag's
-// description as can fit on a single terminal line width, as specified
-// by the flag --tab_completion_columns).  Furthermore, we'll try to
-// make bash order the output such that the most useful or relevent
-// flags are the most likely to be shown at the top.
-//
-// ** Additional features:
-// To assist in finding that one really useful flag, substring matching
-// was implemented.  Before pressing a <TAB> to get completion for the
-// current word, you can append one or more '?' to the flag to do
-// substring matching.  Here's the semantics:
-//   --foo<TAB>     Show me all flags with names prefixed by 'foo'
-//   --foo?<TAB>    Show me all flags with 'foo' somewhere in the name
-//   --foo??<TAB>   Same as prior case, but also search in module
-//                  definition path for 'foo'
-//   --foo???<TAB>  Same as prior case, but also search in flag
-//                  descriptions for 'foo'
-// Finally, we'll trim the output to a relatively small number of
-// flags to keep bash quiet about the verbosity of output.  If one
-// really wanted to see all possible matches, appending a '+' to the
-// search word will force the exhaustive list of matches to be printed.
-//
-// ** How to have bash accept completions from a binary:
-// Bash requires that it be informed about each command that programmatic
-// completion should be enabled for.  Example addition to a .bashrc
-// file would be (your path to gflags_completions.sh file may differ):
-
-/*
-$ complete -o bashdefault -o default -o nospace -C                            \
- '/home/build/eng/bash/bash_completions.sh --tab_completion_columns $COLUMNS' \
-  time  env  binary_name  another_binary  [...]
-*/
-
-// This would allow the following to work:
-//   $ /path/to/binary_name --vmodule<TAB>
-// Or:
-//   $ ./bin/path/another_binary --gfs_u<TAB>
-// (etc)
-//
-// Sadly, it appears that bash gives no easy way to force this behavior for
-// all commands.  That's where the "time" in the above example comes in.
-// If you haven't specifically added a command to the list of completion
-// supported commands, you can still get completions by prefixing the
-// entire command with "env".
-//   $ env /some/brand/new/binary --vmod<TAB>
-// Assuming that "binary" is a newly compiled binary, this should still
-// produce the expected completion output.
-
-
-#ifndef GFLAGS_COMPLETIONS_H_
-#define GFLAGS_COMPLETIONS_H_
-
-namespace google {
-
-extern void HandleCommandLineCompletions(void);
-
-}
-
-#endif  // GFLAGS_COMPLETIONS_H_
diff --git a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_declare.h b/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_declare.h
deleted file mode 100644
index fbc8466fd4..0000000000
--- a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_declare.h
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright (c) 1999, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-//
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// command line flag.
-
-#ifndef GFLAGS_DECLARE_H_
-#define GFLAGS_DECLARE_H_
-
-
-// ---------------------------------------------------------------------------
-// Namespace of gflags library symbols.
-#define GFLAGS_NAMESPACE google
-
-// ---------------------------------------------------------------------------
-// Windows DLL import/export.
-
-// We always want to import the symbols of the gflags library
-#ifndef GFLAGS_DLL_DECL
-#  if 0 && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-
-// We always want to import variables declared in user code
-#ifndef GFLAGS_DLL_DECLARE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DECLARE_FLAG __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECLARE_FLAG
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Flag types
-#include <string>
-#if 1
-#  include <stdint.h>                   // the normal place uint32_t is defined
-#elif 1
-#  include <sys/types.h>                // the normal place u_int32_t is defined
-#elif 0
-#  include <inttypes.h>                 // a third place for uint32_t or u_int32_t
-#endif
-
-namespace GFLAGS_NAMESPACE {
-
-#if 0 // C99
-typedef int32_t          int32;
-typedef uint32_t         uint32;
-typedef int64_t          int64;
-typedef uint64_t         uint64;
-#elif 0 // BSD
-typedef int32_t          int32;
-typedef u_int32_t        uint32;
-typedef int64_t          int64;
-typedef u_int64_t        uint64;
-#elif 1 // Windows
-typedef __int32          int32;
-typedef unsigned __int32 uint32;
-typedef __int64          int64;
-typedef unsigned __int64 uint64;
-#else
-#  error Do not know how to define a 32-bit integer quantity on your system
-#endif
-
-} // namespace GFLAGS_NAMESPACE
-
-
-namespace fLS {
-
-// The meaning of "string" might be different between now and when the
-// macros below get invoked (e.g., if someone is experimenting with
-// other string implementations that get defined after this file is
-// included).  Save the current meaning now and use it in the macros.
-typedef std::string clstring;
-
-} // namespace fLS
-
-
-#define DECLARE_VARIABLE(type, shorttype, name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fL##shorttype { extern GFLAGS_DLL_DECLARE_FLAG type FLAGS_##name; } \
-  using fL##shorttype::FLAGS_##name
-
-#define DECLARE_bool(name) \
-  DECLARE_VARIABLE(bool, B, name)
-
-#define DECLARE_int32(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int32, I, name)
-
-#define DECLARE_int64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int64, I64, name)
-
-#define DECLARE_uint64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint64, U64, name)
-
-#define DECLARE_double(name) \
-  DECLARE_VARIABLE(double, D, name)
-
-#define DECLARE_string(name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fLS { \
-  using ::fLS::clstring; \
-  extern GFLAGS_DLL_DECLARE_FLAG ::fLS::clstring& FLAGS_##name; \
-  } \
-  using fLS::FLAGS_##name
-
-
-#endif  // GFLAGS_DECLARE_H_
diff --git a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_gflags.h b/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_gflags.h
deleted file mode 100644
index 0c17825dd6..0000000000
--- a/media/libyuv/third_party/gflags/gen/win/include/gflags/gflags_gflags.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Copyright (c) 2014, Andreas Schuh
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// -----------------------------------------------------------------------------
-// Imports the gflags library symbols into an alternative/deprecated namespace.
-
-#ifndef GFLAGS_GFLAGS_H_
-#  error The internal header gflags_gflags.h may only be included by gflags.h
-#endif
-
-#ifndef GFLAGS_NS_GFLAGS_H_
-#define GFLAGS_NS_GFLAGS_H_
-
-
-namespace gflags {
-
-
-using GFLAGS_NAMESPACE::int32;
-using GFLAGS_NAMESPACE::uint32;
-using GFLAGS_NAMESPACE::int64;
-using GFLAGS_NAMESPACE::uint64;
-
-using GFLAGS_NAMESPACE::RegisterFlagValidator;
-using GFLAGS_NAMESPACE::CommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetAllFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlagsRestrict;
-using GFLAGS_NAMESPACE::DescribeOneFlag;
-using GFLAGS_NAMESPACE::SetArgv;
-using GFLAGS_NAMESPACE::GetArgvs;
-using GFLAGS_NAMESPACE::GetArgv;
-using GFLAGS_NAMESPACE::GetArgv0;
-using GFLAGS_NAMESPACE::GetArgvSum;
-using GFLAGS_NAMESPACE::ProgramInvocationName;
-using GFLAGS_NAMESPACE::ProgramInvocationShortName;
-using GFLAGS_NAMESPACE::ProgramUsage;
-using GFLAGS_NAMESPACE::VersionString;
-using GFLAGS_NAMESPACE::GetCommandLineOption;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfoOrDie;
-using GFLAGS_NAMESPACE::FlagSettingMode;
-using GFLAGS_NAMESPACE::SET_FLAGS_VALUE;
-using GFLAGS_NAMESPACE::SET_FLAG_IF_DEFAULT;
-using GFLAGS_NAMESPACE::SET_FLAGS_DEFAULT;
-using GFLAGS_NAMESPACE::SetCommandLineOption;
-using GFLAGS_NAMESPACE::SetCommandLineOptionWithMode;
-using GFLAGS_NAMESPACE::FlagSaver;
-using GFLAGS_NAMESPACE::CommandlineFlagsIntoString;
-using GFLAGS_NAMESPACE::ReadFlagsFromString;
-using GFLAGS_NAMESPACE::AppendFlagsIntoFile;
-using GFLAGS_NAMESPACE::ReadFromFlagsFile;
-using GFLAGS_NAMESPACE::BoolFromEnv;
-using GFLAGS_NAMESPACE::Int32FromEnv;
-using GFLAGS_NAMESPACE::Int64FromEnv;
-using GFLAGS_NAMESPACE::Uint64FromEnv;
-using GFLAGS_NAMESPACE::DoubleFromEnv;
-using GFLAGS_NAMESPACE::StringFromEnv;
-using GFLAGS_NAMESPACE::SetUsageMessage;
-using GFLAGS_NAMESPACE::SetVersionString;
-using GFLAGS_NAMESPACE::ParseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::HandleCommandLineHelpFlags;
-using GFLAGS_NAMESPACE::AllowCommandLineReparsing;
-using GFLAGS_NAMESPACE::ReparseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::ShutDownCommandLineFlags;
-using GFLAGS_NAMESPACE::FlagRegisterer;
-
-#ifndef SWIG
-using GFLAGS_NAMESPACE::ParseCommandLineFlags;
-#endif
-
-
-} // namespace gflags
-
-
-#endif  // GFLAGS_NS_GFLAGS_H_
diff --git a/media/libyuv/third_party/gflags/gen/win/include/private/config.h b/media/libyuv/third_party/gflags/gen/win/include/private/config.h
deleted file mode 100644
index d541580eab..0000000000
--- a/media/libyuv/third_party/gflags/gen/win/include/private/config.h
+++ /dev/null
@@ -1,112 +0,0 @@
-/* Generated from config.h.in during build configuration using CMake. */
-
-// Note: This header file is only used internally. It is not part of public interface!
-
-// ---------------------------------------------------------------------------
-// System checks
-
-// Define if you build this library for a MS Windows OS.
-#define OS_WINDOWS
-
-// Define if you have the <stdint.h> header file.
-#define HAVE_STDINT_H
-
-// Define if you have the <sys/types.h> header file.
-#define HAVE_SYS_TYPES_H
-
-// Define if you have the <inttypes.h> header file.
-/* #undef HAVE_INTTYPES_H */
-
-// Define if you have the <sys/stat.h> header file.
-#define HAVE_SYS_STAT_H
-
-// Define if you have the <unistd.h> header file.
-/* #undef HAVE_UNISTD_H */
-
-// Define if you have the <fnmatch.h> header file.
-/* #undef HAVE_FNMATCH_H */
-
-// Define if you have the <shlwapi.h> header file (Windows 2000/XP).
-#define HAVE_SHLWAPI_H
-
-// Define if you have the strtoll function.
-/* #undef HAVE_STRTOLL */
-
-// Define if you have the strtoq function.
-/* #undef HAVE_STRTOQ */
-
-// Define if you have the <pthread.h> header file.
-/* #undef HAVE_PTHREAD */
-
-// Define if your pthread library defines the type pthread_rwlock_t
-/* #undef HAVE_RWLOCK */
-
-// gcc requires this to get PRId64, etc.
-#if defined(HAVE_INTTYPES_H) && !defined(__STDC_FORMAT_MACROS)
-#  define __STDC_FORMAT_MACROS 1
-#endif
-
-// ---------------------------------------------------------------------------
-// Package information
-
-// Name of package.
-#define PACKAGE gflags
-
-// Define to the full name of this package.
-#define PACKAGE_NAME gflags
-
-// Define to the full name and version of this package.
-#define PACKAGE_STRING gflags 2.2.0
-
-// Define to the one symbol short name of this package.
-#define PACKAGE_TARNAME gflags-2.2.0
-
-// Define to the version of this package.
-#define PACKAGE_VERSION 2.2.0
-
-// Version number of package.
-#define VERSION PACKAGE_VERSION
-
-// Define to the address where bug reports for this package should be sent.
-#define PACKAGE_BUGREPORT https://github.com/schuhschuh/gflags/issues
-
-// ---------------------------------------------------------------------------
-// Path separator
-#ifndef PATH_SEPARATOR
-#  ifdef OS_WINDOWS
-#    define PATH_SEPARATOR  '\\'
-#  else
-#    define PATH_SEPARATOR  '/'
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Windows
-
-// Whether gflags library is a DLL.
-#ifndef GFLAGS_IS_A_DLL
-#  define GFLAGS_IS_A_DLL 0
-#endif
-
-// Always export symbols when compiling a shared library as this file is only
-// included by internal modules when building the gflags library itself.
-// The gflags_declare.h header file will set it to import these symbols otherwise.
-#ifndef GFLAGS_DLL_DECL
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-// Flags defined by the gflags library itself must be exported
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  define GFLAGS_DLL_DEFINE_FLAG GFLAGS_DLL_DECL
-#endif
-
-#ifdef OS_WINDOWS
-// The unittests import the symbols of the shared gflags library
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL_FOR_UNITTESTS __declspec(dllimport)
-#  endif
-#  include "windows_port.h"
-#endif
diff --git a/media/libyuv/third_party/gflags/gflags.gyp b/media/libyuv/third_party/gflags/gflags.gyp
deleted file mode 100644
index 37f2815aa3..0000000000
--- a/media/libyuv/third_party/gflags/gflags.gyp
+++ /dev/null
@@ -1,92 +0,0 @@
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# This is a copy of WebRTC's gflags.gyp.
-
-{
-  'variables': {
-    'gflags_root': '<(DEPTH)/third_party/gflags',
-    'conditions': [
-      ['OS=="win"', {
-        'gflags_gen_arch_root': '<(gflags_root)/gen/win',
-      }, {
-        'gflags_gen_arch_root': '<(gflags_root)/gen/posix',
-      }],
-    ],
-  },
-  'targets': [
-    {
-      'target_name': 'gflags',
-      'type': 'static_library',
-      'include_dirs': [
-        '<(gflags_gen_arch_root)/include/gflags',  # For configured files.
-        '<(gflags_gen_arch_root)/include/private',  # For config.h
-        '<(gflags_root)/src/src',  # For everything else.
-      ],
-      'defines': [
-        # These macros exist so flags and symbols are properly
-        # exported when building DLLs. Since we don't build DLLs, we
-        # need to disable them.
-        'GFLAGS_DLL_DECL=',
-        'GFLAGS_DLL_DECLARE_FLAG=',
-        'GFLAGS_DLL_DEFINE_FLAG=',
-      ],
-      'direct_dependent_settings': {
-        'include_dirs': [
-          '<(gflags_gen_arch_root)/include',  # For configured files.
-          '<(gflags_root)/src/src',  # For everything else.
-        ],
-        'defines': [
-          'GFLAGS_DLL_DECL=',
-          'GFLAGS_DLL_DECLARE_FLAG=',
-          'GFLAGS_DLL_DEFINE_FLAG=',
-        ],
-      },
-      'sources': [
-        'src/src/gflags.cc',
-        'src/src/gflags_completions.cc',
-        'src/src/gflags_reporting.cc',
-      ],
-      'conditions': [
-        ['OS=="win"', {
-          'sources': [
-            'src/src/windows_port.cc',
-          ],
-          'msvs_disabled_warnings': [
-            4005,  # WIN32_LEAN_AND_MEAN redefinition.
-            4267,  # Conversion from size_t to "type".
-          ],
-          'configurations': {
-            'Common_Base': {
-              'msvs_configuration_attributes': {
-                'CharacterSet': '2',  # Use Multi-byte Character Set.
-              },
-            },
-          },
-        }],
-        # TODO(andrew): Look into fixing this warning upstream:
-        # http://code.google.com/p/webrtc/issues/detail?id=760
-        ['OS=="win" and clang==1', {
-          'msvs_settings': {
-            'VCCLCompilerTool': {
-              'AdditionalOptions': [
-                '-Wno-microsoft-include',
-              ],
-            },
-          },
-        }],
-        ['clang==1', {
-          'cflags': [
-            '-Wno-microsoft-include',
-          ],
-        }],
-      ],
-    },
-  ],
-}
diff --git a/media/libyuv/tools/OWNERS b/media/libyuv/tools/OWNERS
deleted file mode 100644
index aca046d45e..0000000000
--- a/media/libyuv/tools/OWNERS
+++ /dev/null
@@ -1 +0,0 @@
-kjellander@chromium.org
diff --git a/media/libyuv/tools/gritsettings/README b/media/libyuv/tools/gritsettings/README
deleted file mode 100644
index ca9cb31f74..0000000000
--- a/media/libyuv/tools/gritsettings/README
+++ /dev/null
@@ -1,7 +0,0 @@
-This is a dummy configuration file to workaround the assumption of the grit
-resource file being located here in the Chromium build toolchain.
-
-This is only needed for our Android native tests to be able to include the
-build/apk_test.gypi GYP file and depend on the
-<(DEPTH)/testing/android/native_test.gyp:native_test_native_code target in
-Chromium.
diff --git a/media/libyuv/tools/gritsettings/resource_ids b/media/libyuv/tools/gritsettings/resource_ids
deleted file mode 100644
index bfdfbbe90b..0000000000
--- a/media/libyuv/tools/gritsettings/resource_ids
+++ /dev/null
@@ -1,15 +0,0 @@
-# Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-{
-  "SRCDIR": "../..",
-
-  "chromium/src/net/base/net_resources.grd": {
-    "includes": [4000],
-  },
-}
diff --git a/media/libyuv/tools/msan/OWNERS b/media/libyuv/tools/msan/OWNERS
deleted file mode 100644
index 60351e7ea2..0000000000
--- a/media/libyuv/tools/msan/OWNERS
+++ /dev/null
@@ -1,3 +0,0 @@
-pbos@chromium.org
-kjellander@chromium.org
-
diff --git a/media/libyuv/tools/ubsan/OWNERS b/media/libyuv/tools/ubsan/OWNERS
deleted file mode 100644
index b608519abf..0000000000
--- a/media/libyuv/tools/ubsan/OWNERS
+++ /dev/null
@@ -1,4 +0,0 @@
-pbos@webrtc.org
-kjellander@webrtc.org
-fbarchard@chromium.org
-
diff --git a/media/libyuv/tools/valgrind-libyuv/libyuv_tests.bat b/media/libyuv/tools/valgrind-libyuv/libyuv_tests.bat
deleted file mode 100644
index e37f09eb25..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/libyuv_tests.bat
+++ /dev/null
@@ -1,79 +0,0 @@
-@echo off

-:: Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.

-::

-:: Use of this source code is governed by a BSD-style license

-:: that can be found in the LICENSE file in the root of the source

-:: tree. An additional intellectual property rights grant can be found

-:: in the file PATENTS.  All contributing project authors may

-:: be found in the AUTHORS file in the root of the source tree.

-

-:: This script is a copy of chrome_tests.bat with the following changes:

-:: - Invokes libyuv_tests.py instead of chrome_tests.py

-:: - Chromium's Valgrind scripts directory is added to the PYTHONPATH to make

-::   it possible to execute the Python scripts properly.

-

-:: TODO(timurrrr): batch files 'export' all the variables to the parent shell

-set THISDIR=%~dp0

-set TOOL_NAME="unknown"

-

-:: Get the tool name and put it into TOOL_NAME {{{1

-:: NB: SHIFT command doesn't modify %*

-:PARSE_ARGS_LOOP

-  if %1 == () GOTO:TOOLNAME_NOT_FOUND

-  if %1 == --tool GOTO:TOOLNAME_FOUND

-  SHIFT

-  goto :PARSE_ARGS_LOOP

-

-:TOOLNAME_NOT_FOUND

-echo "Please specify a tool (tsan or drmemory) by using --tool flag"

-exit /B 1

-

-:TOOLNAME_FOUND

-SHIFT

-set TOOL_NAME=%1

-:: }}}

-if "%TOOL_NAME%" == "drmemory"          GOTO :SETUP_DRMEMORY

-if "%TOOL_NAME%" == "drmemory_light"    GOTO :SETUP_DRMEMORY

-if "%TOOL_NAME%" == "drmemory_full"     GOTO :SETUP_DRMEMORY

-if "%TOOL_NAME%" == "drmemory_pattern"  GOTO :SETUP_DRMEMORY

-if "%TOOL_NAME%" == "tsan"     GOTO :SETUP_TSAN

-echo "Unknown tool: `%TOOL_NAME%`! Only tsan and drmemory are supported."

-exit /B 1

-

-:SETUP_DRMEMORY

-if NOT "%DRMEMORY_COMMAND%"=="" GOTO :RUN_TESTS

-:: Set up DRMEMORY_COMMAND to invoke Dr. Memory {{{1

-set DRMEMORY_PATH=%THISDIR%..\..\third_party\drmemory

-set DRMEMORY_SFX=%DRMEMORY_PATH%\drmemory-windows-sfx.exe

-if EXIST %DRMEMORY_SFX% GOTO DRMEMORY_BINARY_OK

-echo "Can't find Dr. Memory executables."

-echo "See http://www.chromium.org/developers/how-tos/using-valgrind/dr-memory"

-echo "for the instructions on how to get them."

-exit /B 1

-

-:DRMEMORY_BINARY_OK

-%DRMEMORY_SFX% -o%DRMEMORY_PATH%\unpacked -y

-set DRMEMORY_COMMAND=%DRMEMORY_PATH%\unpacked\bin\drmemory.exe

-:: }}}

-goto :RUN_TESTS

-

-:SETUP_TSAN

-:: Set up PIN_COMMAND to invoke TSan {{{1

-set TSAN_PATH=%THISDIR%..\..\third_party\tsan

-set TSAN_SFX=%TSAN_PATH%\tsan-x86-windows-sfx.exe

-if EXIST %TSAN_SFX% GOTO TSAN_BINARY_OK

-echo "Can't find ThreadSanitizer executables."

-echo "See http://www.chromium.org/developers/how-tos/using-valgrind/threadsanitizer/threadsanitizer-on-windows"

-echo "for the instructions on how to get them."

-exit /B 1

-

-:TSAN_BINARY_OK

-%TSAN_SFX% -o%TSAN_PATH%\unpacked -y

-set PIN_COMMAND=%TSAN_PATH%\unpacked\tsan-x86-windows\tsan.bat

-:: }}}

-goto :RUN_TESTS

-

-:RUN_TESTS

-set PYTHONPATH=%THISDIR%..\python\google;%THISDIR%..\valgrind

-set RUNNING_ON_VALGRIND=yes

-python %THISDIR%libyuv_tests.py %*

diff --git a/media/libyuv/tools/valgrind-libyuv/libyuv_tests.py b/media/libyuv/tools/valgrind-libyuv/libyuv_tests.py
deleted file mode 100755
index bd27cd5ca6..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/libyuv_tests.py
+++ /dev/null
@@ -1,138 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Runs various libyuv tests through valgrind_test.py.
-
-This script inherits the chrome_tests.py in Chrome, but allows running any test
-instead of only the hard-coded ones. It uses the -t cmdline flag to do this, and
-only supports specifying a single test for each run.
-
-Suppression files:
-The Chrome valgrind directory we use as a DEPS dependency contains the following
-suppression files:
-  valgrind/memcheck/suppressions.txt
-  valgrind/memcheck/suppressions_mac.txt
-  valgrind/tsan/suppressions.txt
-  valgrind/tsan/suppressions_mac.txt
-  valgrind/tsan/suppressions_win32.txt
-Since they're referenced from the chrome_tests.py script, we have similar files
-below the directory of this script. When executing, this script will setup both
-Chrome's suppression files and our own, so we can easily maintain libyuv
-specific suppressions in our own files.
-"""
-
-import logging
-import optparse
-import os
-import sys
-
-import logging_utils
-import path_utils
-
-import chrome_tests
-
-
-class LibyuvTest(chrome_tests.ChromeTests):
-  """Class that handles setup of suppressions for libyuv.
-
-  Everything else is inherited from chrome_tests.ChromeTests.
-  """
-
-  def _DefaultCommand(self, tool, exe=None, valgrind_test_args=None):
-    """Override command-building method so we can add more suppressions."""
-    cmd = chrome_tests.ChromeTests._DefaultCommand(self, tool, exe,
-                                                   valgrind_test_args)
-    # When ChromeTests._DefaultCommand has executed, it has setup suppression
-    # files based on what's found in the memcheck/ or tsan/ subdirectories of
-    # this script's location. If Mac or Windows is executing, additional
-    # platform specific files have also been added.
-    # Since only the ones located below this directory is added, we must also
-    # add the ones maintained by Chrome, located in ../valgrind.
-
-    # The idea is to look for --suppression arguments in the cmd list and add a
-    # modified copy of each suppression file, for the corresponding file in
-    # ../valgrind. If we would simply replace 'valgrind-libyuv' with 'valgrind'
-    # we may produce invalid paths if other parts of the path contain that
-    # string. That's why the code below only replaces the end of the path.
-    script_dir = path_utils.ScriptDir()
-    old_base, _ = os.path.split(script_dir)
-    new_dir = os.path.join(old_base, 'valgrind')
-    add_suppressions = []
-    for token in cmd:
-      if '--suppressions' in token:
-        add_suppressions.append(token.replace(script_dir, new_dir))
-    return add_suppressions + cmd
-
-
-def main(_):
-  parser = optparse.OptionParser('usage: %prog -b <dir> -t <test> <test args>')
-  parser.disable_interspersed_args()
-  parser.add_option('-b', '--build-dir',
-                    help=('Location of the compiler output. Can only be used '
-                          'when the test argument does not contain this path.'))
-  parser.add_option("--target", help="Debug or Release")
-  parser.add_option('-t', '--test', help='Test to run.')
-  parser.add_option('', '--baseline', action='store_true', default=False,
-                    help='Generate baseline data instead of validating')
-  parser.add_option('', '--gtest_filter',
-                    help='Additional arguments to --gtest_filter')
-  parser.add_option('', '--gtest_repeat',
-                    help='Argument for --gtest_repeat')
-  parser.add_option("--gtest_shuffle", action="store_true", default=False,
-                    help="Randomize tests' orders on every iteration.")
-  parser.add_option("--gtest_break_on_failure", action="store_true",
-                    default=False,
-                    help="Drop in to debugger on assertion failure. Also "
-                         "useful for forcing tests to exit with a stack dump "
-                         "on the first assertion failure when running with "
-                         "--gtest_repeat=-1")
-  parser.add_option('-v', '--verbose', action='store_true', default=False,
-                    help='Verbose output - enable debug log messages')
-  parser.add_option('', '--tool', dest='valgrind_tool', default='memcheck',
-                    help='Specify a valgrind tool to run the tests under')
-  parser.add_option('', '--tool_flags', dest='valgrind_tool_flags', default='',
-                    help='Specify custom flags for the selected valgrind tool')
-  parser.add_option('', '--keep_logs', action='store_true', default=False,
-                    help=('Store memory tool logs in the <tool>.logs directory '
-                          'instead of /tmp.\nThis can be useful for tool '
-                          'developers/maintainers.\nPlease note that the <tool>'
-                          '.logs directory will be clobbered on tool startup.'))
-  parser.add_option("--test-launcher-bot-mode", action="store_true",
-                    help="run the tests with --test-launcher-bot-mode")
-  parser.add_option("--test-launcher-total-shards", type=int,
-                    help="run the tests with --test-launcher-total-shards")
-  parser.add_option("--test-launcher-shard-index", type=int,
-                    help="run the tests with --test-launcher-shard-index")
-  options, args = parser.parse_args()
-
-  if options.verbose:
-    logging_utils.config_root(logging.DEBUG)
-  else:
-    logging_utils.config_root()
-
-  if not options.test:
-    parser.error('--test not specified')
-
-  # Support build dir both with and without the target.
-  if (options.target and options.build_dir and
-      not options.build_dir.endswith(options.target)):
-    options.build_dir = os.path.join(options.build_dir, options.target)
-
-  # If --build_dir is provided, prepend it to the test executable if needed.
-  test_executable = options.test
-  if options.build_dir and not test_executable.startswith(options.build_dir):
-    test_executable = os.path.join(options.build_dir, test_executable)
-  args = [test_executable] + args
-
-  test = LibyuvTest(options, args, 'cmdline')
-  return test.Run()
-
-if __name__ == '__main__':
-  return_code = main(sys.argv)
-  sys.exit(return_code)
diff --git a/media/libyuv/tools/valgrind-libyuv/libyuv_tests.sh b/media/libyuv/tools/valgrind-libyuv/libyuv_tests.sh
deleted file mode 100755
index 4fee7daed6..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/libyuv_tests.sh
+++ /dev/null
@@ -1,138 +0,0 @@
-#!/bin/bash
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# Set up some paths and re-direct the arguments to libyuv_tests.py
-
-# This script is a copy of the chrome_tests.sh wrapper script with the following
-# changes:
-# - The locate_valgrind.sh of Chromium's Valgrind scripts dir is used to locate
-#   the Valgrind framework install.
-# - libyuv_tests.py is invoked instead of chrome_tests.py.
-# - Chromium's Valgrind scripts directory is added to the PYTHONPATH to make it
-#   possible to execute the Python scripts properly.
-
-export THISDIR=`dirname $0`
-ARGV_COPY="$@"
-
-# We need to set CHROME_VALGRIND iff using Memcheck or TSan-Valgrind:
-#   tools/valgrind-libyuv/libyuv_tests.sh --tool memcheck
-# or
-#   tools/valgrind-libyuv/libyuv_tests.sh --tool=memcheck
-# (same for "--tool=tsan")
-tool="memcheck"  # Default to memcheck.
-while (( "$#" ))
-do
-  if [[ "$1" == "--tool" ]]
-  then
-    tool="$2"
-    shift
-  elif [[ "$1" =~ --tool=(.*) ]]
-  then
-    tool="${BASH_REMATCH[1]}"
-  fi
-  shift
-done
-
-NEEDS_VALGRIND=0
-NEEDS_DRMEMORY=0
-
-case "$tool" in
-  "memcheck")
-    NEEDS_VALGRIND=1
-    ;;
-  "tsan" | "tsan_rv")
-    if [ "`uname -s`" == CYGWIN* ]
-    then
-      NEEDS_PIN=1
-    else
-      NEEDS_VALGRIND=1
-    fi
-    ;;
-  "drmemory" | "drmemory_light" | "drmemory_full" | "drmemory_pattern")
-    NEEDS_DRMEMORY=1
-    ;;
-esac
-
-# For Libyuv, we'll use the locate_valgrind.sh script in Chromium's Valgrind
-# scripts dir to locate the Valgrind framework install
-CHROME_VALGRIND_SCRIPTS=$THISDIR/../valgrind
-
-if [ "$NEEDS_VALGRIND" == "1" ]
-then
-  CHROME_VALGRIND=`sh $CHROME_VALGRIND_SCRIPTS/locate_valgrind.sh`
-  if [ "$CHROME_VALGRIND" = "" ]
-  then
-    # locate_valgrind.sh failed
-    exit 1
-  fi
-  echo "Using valgrind binaries from ${CHROME_VALGRIND}"
-
-  PATH="${CHROME_VALGRIND}/bin:$PATH"
-  # We need to set these variables to override default lib paths hard-coded into
-  # Valgrind binary.
-  export VALGRIND_LIB="$CHROME_VALGRIND/lib/valgrind"
-  export VALGRIND_LIB_INNER="$CHROME_VALGRIND/lib/valgrind"
-
-  # Clean up some /tmp directories that might be stale due to interrupted
-  # chrome_tests.py execution.
-  # FYI:
-  #   -mtime +1  <- only print files modified more than 24h ago,
-  #   -print0/-0 are needed to handle possible newlines in the filenames.
-  echo "Cleanup /tmp from Valgrind stuff"
-  find /tmp -maxdepth 1 \(\
-        -name "vgdb-pipe-*" -or -name "vg_logs_*" -or -name "valgrind.*" \
-      \) -mtime +1 -print0 | xargs -0 rm -rf
-fi
-
-if [ "$NEEDS_DRMEMORY" == "1" ]
-then
-  if [ -z "$DRMEMORY_COMMAND" ]
-  then
-    DRMEMORY_PATH="$THISDIR/../../third_party/drmemory"
-    DRMEMORY_SFX="$DRMEMORY_PATH/drmemory-windows-sfx.exe"
-    if [ ! -f "$DRMEMORY_SFX" ]
-    then
-      echo "Can't find Dr. Memory executables."
-      echo "See http://www.chromium.org/developers/how-tos/using-valgrind/dr-memory"
-      echo "for the instructions on how to get them."
-      exit 1
-    fi
-
-    chmod +x "$DRMEMORY_SFX"  # Cygwin won't run it without +x.
-    "$DRMEMORY_SFX" -o"$DRMEMORY_PATH/unpacked" -y
-    export DRMEMORY_COMMAND="$DRMEMORY_PATH/unpacked/bin/drmemory.exe"
-  fi
-fi
-
-if [ "$NEEDS_PIN" == "1" ]
-then
-  if [ -z "$PIN_COMMAND" ]
-  then
-    # Set up PIN_COMMAND to invoke TSan.
-    TSAN_PATH="$THISDIR/../../third_party/tsan"
-    TSAN_SFX="$TSAN_PATH/tsan-x86-windows-sfx.exe"
-    echo "$TSAN_SFX"
-    if [ ! -f $TSAN_SFX ]
-    then
-      echo "Can't find ThreadSanitizer executables."
-      echo "See http://www.chromium.org/developers/how-tos/using-valgrind/threadsanitizer/threadsanitizer-on-windows"
-      echo "for the instructions on how to get them."
-      exit 1
-    fi
-
-    chmod +x "$TSAN_SFX"  # Cygwin won't run it without +x.
-    "$TSAN_SFX" -o"$TSAN_PATH"/unpacked -y
-    export PIN_COMMAND="$TSAN_PATH/unpacked/tsan-x86-windows/tsan.bat"
-  fi
-fi
-
-# Add Chrome's Valgrind scripts dir to the PYTHON_PATH since it contains
-# the scripts that are needed for this script to run
-PYTHONPATH=$THISDIR/../python/google:$CHROME_VALGRIND_SCRIPTS python \
-           "$THISDIR/libyuv_tests.py" $ARGV_COPY
diff --git a/media/libyuv/tools/valgrind-libyuv/memcheck/OWNERS b/media/libyuv/tools/valgrind-libyuv/memcheck/OWNERS
deleted file mode 100644
index 72e8ffc0db..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/memcheck/OWNERS
+++ /dev/null
@@ -1 +0,0 @@
-*
diff --git a/media/libyuv/tools/valgrind-libyuv/memcheck/PRESUBMIT.py b/media/libyuv/tools/valgrind-libyuv/memcheck/PRESUBMIT.py
deleted file mode 100644
index 46ff4cfcf1..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/memcheck/PRESUBMIT.py
+++ /dev/null
@@ -1,99 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""
-Copied from Chrome's src/tools/valgrind/memcheck/PRESUBMIT.py
-
-See http://dev.chromium.org/developers/how-tos/depottools/presubmit-scripts
-for more details on the presubmit API built into gcl.
-"""
-
-import os
-import re
-import sys
-
-def CheckChange(input_api, output_api):
-  """Checks the memcheck suppressions files for bad data."""
-
-  # Add the path to the Chrome valgrind dir to the import path:
-  tools_vg_path = os.path.join(input_api.PresubmitLocalPath(), '..', '..',
-                               'valgrind')
-  sys.path.append(tools_vg_path)
-  import suppressions
-
-  sup_regex = re.compile('suppressions.*\.txt$')
-  suppressions = {}
-  errors = []
-  check_for_memcheck = False
-  # skip_next_line has 3 possible values:
-  # - False: don't skip the next line.
-  # - 'skip_suppression_name': the next line is a suppression name, skip.
-  # - 'skip_param': the next line is a system call parameter error, skip.
-  skip_next_line = False
-  for f in filter(lambda x: sup_regex.search(x.LocalPath()),
-                  input_api.AffectedFiles()):
-    for line, line_num in zip(f.NewContents(),
-                              xrange(1, len(f.NewContents()) + 1)):
-      line = line.lstrip()
-      if line.startswith('#') or not line:
-        continue
-
-      if skip_next_line:
-        if skip_next_line == 'skip_suppression_name':
-          if 'insert_a_suppression_name_here' in line:
-            errors.append('"insert_a_suppression_name_here" is not a valid '
-                          'suppression name')
-          if suppressions.has_key(line):
-            if f.LocalPath() == suppressions[line][1]:
-              errors.append('suppression with name "%s" at %s line %s '
-                            'has already been defined at line %s' %
-                            (line, f.LocalPath(), line_num,
-                             suppressions[line][1]))
-            else:
-              errors.append('suppression with name "%s" at %s line %s '
-                            'has already been defined at %s line %s' %
-                            (line, f.LocalPath(), line_num,
-                             suppressions[line][0], suppressions[line][1]))
-          else:
-            suppressions[line] = (f, line_num)
-            check_for_memcheck = True;
-        skip_next_line = False
-        continue
-      if check_for_memcheck:
-        if not line.startswith('Memcheck:'):
-          errors.append('"%s" should be "Memcheck:..." in %s line %s' %
-                        (line, f.LocalPath(), line_num))
-        check_for_memcheck = False;
-      if line == '{':
-        skip_next_line = 'skip_suppression_name'
-        continue
-      if line == "Memcheck:Param":
-        skip_next_line = 'skip_param'
-        continue
-
-      if (line.startswith('fun:') or line.startswith('obj:') or
-          line.startswith('Memcheck:') or line == '}' or
-          line == '...'):
-        continue
-      errors.append('"%s" is probably wrong: %s line %s' % (line, f.LocalPath(),
-                                                            line_num))
-  if errors:
-    return [output_api.PresubmitError('\n'.join(errors))]
-  return []
-
-def CheckChangeOnUpload(input_api, output_api):
-  return CheckChange(input_api, output_api)
-
-def CheckChangeOnCommit(input_api, output_api):
-  return CheckChange(input_api, output_api)
-
-def GetPreferredTrySlaves():
-  # We don't have any memcheck slaves yet, so there's no use for this method.
-  # When we have, the slave name(s) should be put into this list.
-  return []
diff --git a/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions.txt b/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions.txt
deleted file mode 100644
index 3ad0c8ccc5..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_mac.txt b/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_mac.txt
deleted file mode 100644
index 3ad0c8ccc5..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_mac.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_win32.txt b/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_win32.txt
deleted file mode 100644
index 3ad0c8ccc5..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_win32.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/media/libyuv/tools/valgrind-libyuv/tsan/OWNERS b/media/libyuv/tools/valgrind-libyuv/tsan/OWNERS
deleted file mode 100644
index 72e8ffc0db..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/tsan/OWNERS
+++ /dev/null
@@ -1 +0,0 @@
-*
diff --git a/media/libyuv/tools/valgrind-libyuv/tsan/PRESUBMIT.py b/media/libyuv/tools/valgrind-libyuv/tsan/PRESUBMIT.py
deleted file mode 100644
index d25b6ebcee..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/tsan/PRESUBMIT.py
+++ /dev/null
@@ -1,41 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-import os
-import re
-import sys
-
-"""
-Copied from Chrome's src/tools/valgrind/tsan/PRESUBMIT.py
-
-See http://dev.chromium.org/developers/how-tos/depottools/presubmit-scripts
-for more details on the presubmit API built into gcl.
-"""
-
-def CheckChange(input_api, output_api):
-  """Checks the TSan suppressions files for bad suppressions."""
-
-  # Add the path to the Chrome valgrind dir to the import path:
-  tools_vg_path = os.path.join(input_api.PresubmitLocalPath(), '..', '..',
-                               'valgrind')
-  sys.path.append(tools_vg_path)
-  import suppressions
-
-  return suppressions.PresubmitCheck(input_api, output_api)
-
-def CheckChangeOnUpload(input_api, output_api):
-  return CheckChange(input_api, output_api)
-
-def CheckChangeOnCommit(input_api, output_api):
-  return CheckChange(input_api, output_api)
-
-def GetPreferredTrySlaves():
-  # We don't have any tsan slaves yet, so there's no use for this method.
-  # When we have, the slave name(s) should be put into this list.
-  return []
diff --git a/media/libyuv/tools/valgrind-libyuv/tsan/suppressions.txt b/media/libyuv/tools/valgrind-libyuv/tsan/suppressions.txt
deleted file mode 100644
index 3ad0c8ccc5..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/tsan/suppressions.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/media/libyuv/tools/valgrind-libyuv/tsan/suppressions_mac.txt b/media/libyuv/tools/valgrind-libyuv/tsan/suppressions_mac.txt
deleted file mode 100644
index 3ad0c8ccc5..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/tsan/suppressions_mac.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/media/libyuv/tools/valgrind-libyuv/tsan/suppressions_win32.txt b/media/libyuv/tools/valgrind-libyuv/tsan/suppressions_win32.txt
deleted file mode 100644
index 3ad0c8ccc5..0000000000
--- a/media/libyuv/tools/valgrind-libyuv/tsan/suppressions_win32.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/media/libyuv/tools_libyuv/OWNERS b/media/libyuv/tools_libyuv/OWNERS
new file mode 100644
index 0000000000..aae4fb6e02
--- /dev/null
+++ b/media/libyuv/tools_libyuv/OWNERS
@@ -0,0 +1,4 @@
+mbonadei@chromium.org
+fbarchard@chromium.org
+pbos@chromium.org
+
diff --git a/media/libyuv/tools_libyuv/autoroller/roll_deps.py b/media/libyuv/tools_libyuv/autoroller/roll_deps.py
new file mode 100755
index 0000000000..5496e42403
--- /dev/null
+++ b/media/libyuv/tools_libyuv/autoroller/roll_deps.py
@@ -0,0 +1,509 @@
+#!/usr/bin/env vpython3
+
+# Copyright 2017 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+# This is a modified copy of the script in
+# https://webrtc.googlesource.com/src/+/master/tools_webrtc/autoroller/roll_deps.py
+# customized for libyuv.
+
+"""Script to automatically roll dependencies in the libyuv DEPS file."""
+
+import argparse
+import base64
+import collections
+import logging
+import os
+import re
+import subprocess
+import sys
+import urllib.request
+
+
+# Skip these dependencies (list without solution name prefix).
+DONT_AUTOROLL_THESE = [
+  'src/third_party/gflags/src',
+]
+
+LIBYUV_URL = 'https://chromium.googlesource.com/libyuv/libyuv'
+CHROMIUM_SRC_URL = 'https://chromium.googlesource.com/chromium/src'
+CHROMIUM_COMMIT_TEMPLATE = CHROMIUM_SRC_URL + '/+/%s'
+CHROMIUM_LOG_TEMPLATE = CHROMIUM_SRC_URL + '/+log/%s'
+CHROMIUM_FILE_TEMPLATE = CHROMIUM_SRC_URL + '/+/%s/%s'
+
+COMMIT_POSITION_RE = re.compile('^Cr-Commit-Position: .*#([0-9]+).*$')
+CLANG_REVISION_RE = re.compile(r'^CLANG_REVISION = \'([0-9a-z-]+)\'$')
+ROLL_BRANCH_NAME = 'roll_chromium_revision'
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+CHECKOUT_SRC_DIR = os.path.realpath(os.path.join(SCRIPT_DIR, os.pardir,
+                                                 os.pardir))
+CHECKOUT_ROOT_DIR = os.path.realpath(os.path.join(CHECKOUT_SRC_DIR, os.pardir))
+
+sys.path.append(os.path.join(CHECKOUT_SRC_DIR, 'build'))
+import find_depot_tools  # pylint: disable=wrong-import-position
+find_depot_tools.add_depot_tools_to_path()
+
+CLANG_UPDATE_SCRIPT_URL_PATH = 'tools/clang/scripts/update.py'
+CLANG_UPDATE_SCRIPT_LOCAL_PATH = os.path.join(CHECKOUT_SRC_DIR, 'tools',
+                                              'clang', 'scripts', 'update.py')
+
+DepsEntry = collections.namedtuple('DepsEntry', 'path url revision')
+ChangedDep = collections.namedtuple('ChangedDep',
+                                    'path url current_rev new_rev')
+
+class RollError(Exception):
+  pass
+
+
+def VarLookup(local_scope):
+  return lambda var_name: local_scope['vars'][var_name]
+
+
+def ParseDepsDict(deps_content):
+  local_scope = {}
+  global_scope = {
+    'Var': VarLookup(local_scope),
+    'Str': lambda s: s,
+    'deps_os': {},
+  }
+  exec(deps_content, global_scope, local_scope)
+  return local_scope
+
+
+def ParseLocalDepsFile(filename):
+  with open(filename, 'rb') as f:
+    deps_content = f.read().decode('utf-8')
+  return ParseDepsDict(deps_content)
+
+
+def ParseRemoteCrDepsFile(revision):
+  deps_content = ReadRemoteCrFile('DEPS', revision)
+  return ParseDepsDict(deps_content)
+
+
+def ParseCommitPosition(commit_message):
+  for line in reversed(commit_message.splitlines()):
+    m = COMMIT_POSITION_RE.match(line.strip())
+    if m:
+      return int(m.group(1))
+  logging.error('Failed to parse commit position id from:\n%s\n',
+                commit_message)
+  sys.exit(-1)
+
+
+def _RunCommand(command, working_dir=None, ignore_exit_code=False,
+                extra_env=None, input_data=None):
+  """Runs a command and returns the output from that command.
+
+  If the command fails (exit code != 0), the function will exit the process.
+
+  Returns:
+    A tuple containing the stdout and stderr outputs as strings.
+  """
+  working_dir = working_dir or CHECKOUT_SRC_DIR
+  logging.debug('CMD: %s CWD: %s', ' '.join(command), working_dir)
+  env = os.environ.copy()
+  if extra_env:
+    assert all(isinstance(value, str) for value in extra_env.values())
+    logging.debug('extra env: %s', extra_env)
+    env.update(extra_env)
+  p = subprocess.Popen(command,
+                       stdin=subprocess.PIPE,
+                       stdout=subprocess.PIPE,
+                       stderr=subprocess.PIPE,
+                       env=env,
+                       cwd=working_dir,
+                       universal_newlines=True)
+  std_output, err_output = p.communicate(input_data)
+  p.stdout.close()
+  p.stderr.close()
+  if not ignore_exit_code and p.returncode != 0:
+    logging.error('Command failed: %s\n'
+                  'stdout:\n%s\n'
+                  'stderr:\n%s\n', ' '.join(command), std_output, err_output)
+    sys.exit(p.returncode)
+  return std_output, err_output
+
+
+def _GetBranches():
+  """Returns a tuple of active,branches.
+
+  The 'active' is the name of the currently active branch and 'branches' is a
+  list of all branches.
+  """
+  lines = _RunCommand(['git', 'branch'])[0].split('\n')
+  branches = []
+  active = ''
+  for line in lines:
+    if '*' in line:
+      # The assumption is that the first char will always be the '*'.
+      active = line[1:].strip()
+      branches.append(active)
+    else:
+      branch = line.strip()
+      if branch:
+        branches.append(branch)
+  return active, branches
+
+
+def _ReadGitilesContent(url):
+  # Download and decode BASE64 content until
+  # https://code.google.com/p/gitiles/issues/detail?id=7 is fixed.
+  base64_content = ReadUrlContent(url + '?format=TEXT')
+  return base64.b64decode(base64_content[0]).decode('utf-8')
+
+
+def ReadRemoteCrFile(path_below_src, revision):
+  """Reads a remote Chromium file of a specific revision. Returns a string."""
+  return _ReadGitilesContent(CHROMIUM_FILE_TEMPLATE % (revision,
+                                                       path_below_src))
+
+
+def ReadRemoteCrCommit(revision):
+  """Reads a remote Chromium commit message. Returns a string."""
+  return _ReadGitilesContent(CHROMIUM_COMMIT_TEMPLATE % revision)
+
+
+def ReadUrlContent(url):
+  """Connect to a remote host and read the contents. Returns a list of lines."""
+  conn = urllib.request.urlopen(url)
+  try:
+    return conn.readlines()
+  except IOError as e:
+    logging.exception('Error connecting to %s. Error: %s', url, e)
+    raise
+  finally:
+    conn.close()
+
+
+def GetMatchingDepsEntries(depsentry_dict, dir_path):
+  """Gets all deps entries matching the provided path.
+
+  This list may contain more than one DepsEntry object.
+  Example: dir_path='src/testing' would give results containing both
+  'src/testing/gtest' and 'src/testing/gmock' deps entries for Chromium's DEPS.
+  Example 2: dir_path='src/build' should return 'src/build' but not
+  'src/buildtools'.
+
+  Returns:
+    A list of DepsEntry objects.
+  """
+  result = []
+  for path, depsentry in depsentry_dict.items():
+    if path == dir_path:
+      result.append(depsentry)
+    else:
+      parts = path.split('/')
+      if all(part == parts[i]
+             for i, part in enumerate(dir_path.split('/'))):
+        result.append(depsentry)
+  return result
+
+def BuildDepsentryDict(deps_dict):
+  """Builds a dict of paths to DepsEntry objects from a raw deps dict."""
+  result = {}
+
+  def AddDepsEntries(deps_subdict):
+    for path, deps_url_spec in deps_subdict.items():
+      if isinstance(deps_url_spec, dict):
+        if deps_url_spec.get('dep_type') == 'cipd':
+          continue
+        deps_url = deps_url_spec['url']
+      else:
+        deps_url = deps_url_spec
+      if not path in result:
+        url, revision = deps_url.split('@') if deps_url else (None, None)
+        result[path] = DepsEntry(path, url, revision)
+
+  AddDepsEntries(deps_dict['deps'])
+  for deps_os in ['win', 'mac', 'linux', 'android', 'ios', 'unix']:
+    AddDepsEntries(deps_dict.get('deps_os', {}).get(deps_os, {}))
+  return result
+
+
+def CalculateChangedDeps(libyuv_deps, new_cr_deps):
+  """
+  Calculate changed deps entries based on entries defined in the libyuv DEPS
+  file:
+     - If a shared dependency with the Chromium DEPS file: roll it to the same
+       revision as Chromium (i.e. entry in the new_cr_deps dict)
+     - If it's a Chromium sub-directory, roll it to the HEAD revision (notice
+       this means it may be ahead of the chromium_revision, but generally these
+       should be close).
+     - If it's another DEPS entry (not shared with Chromium), roll it to HEAD
+       unless it's configured to be skipped.
+
+  Returns:
+    A list of ChangedDep objects representing the changed deps.
+  """
+  result = []
+  libyuv_entries = BuildDepsentryDict(libyuv_deps)
+  new_cr_entries = BuildDepsentryDict(new_cr_deps)
+  for path, libyuv_deps_entry in libyuv_entries.items():
+    if path in DONT_AUTOROLL_THESE:
+      continue
+    cr_deps_entry = new_cr_entries.get(path)
+    if cr_deps_entry:
+      # Use the revision from Chromium's DEPS file.
+      new_rev = cr_deps_entry.revision
+      assert libyuv_deps_entry.url == cr_deps_entry.url, (
+        'Libyuv DEPS entry %s has a different URL (%s) than Chromium (%s).' %
+        (path, libyuv_deps_entry.url, cr_deps_entry.url))
+    else:
+      # Use the HEAD of the deps repo.
+      stdout, _ = _RunCommand(['git', 'ls-remote', libyuv_deps_entry.url,
+                               'HEAD'])
+      new_rev = stdout.strip().split('\t')[0]
+
+    # Check if an update is necessary.
+    if libyuv_deps_entry.revision != new_rev:
+      logging.debug('Roll dependency %s to %s', path, new_rev)
+      result.append(ChangedDep(path, libyuv_deps_entry.url,
+                               libyuv_deps_entry.revision, new_rev))
+  return sorted(result)
+
+
+def CalculateChangedClang(new_cr_rev):
+  def GetClangRev(lines):
+    for line in lines:
+      match = CLANG_REVISION_RE.match(line)
+      if match:
+        return match.group(1)
+    raise RollError('Could not parse Clang revision from:\n' + '\n'.join('  ' + l for l in lines))
+
+  with open(CLANG_UPDATE_SCRIPT_LOCAL_PATH, 'r') as f:
+    current_lines = f.readlines()
+  current_rev = GetClangRev(current_lines)
+
+  new_clang_update_py = ReadRemoteCrFile(CLANG_UPDATE_SCRIPT_URL_PATH,
+                                             new_cr_rev).splitlines()
+  new_rev = GetClangRev(new_clang_update_py)
+  return ChangedDep(CLANG_UPDATE_SCRIPT_LOCAL_PATH, None, current_rev, new_rev)
+
+
+def GenerateCommitMessage(current_cr_rev, new_cr_rev, current_commit_pos,
+                          new_commit_pos, changed_deps_list, clang_change):
+  current_cr_rev = current_cr_rev[0:10]
+  new_cr_rev = new_cr_rev[0:10]
+  rev_interval = '%s..%s' % (current_cr_rev, new_cr_rev)
+  git_number_interval = '%s:%s' % (current_commit_pos, new_commit_pos)
+
+  commit_msg = ['Roll chromium_revision %s (%s)\n' % (rev_interval,
+                                                    git_number_interval)]
+  commit_msg.append('Change log: %s' % (CHROMIUM_LOG_TEMPLATE % rev_interval))
+  commit_msg.append('Full diff: %s\n' % (CHROMIUM_COMMIT_TEMPLATE %
+                                         rev_interval))
+  if changed_deps_list:
+    commit_msg.append('Changed dependencies:')
+
+    for c in changed_deps_list:
+      commit_msg.append('* %s: %s/+log/%s..%s' % (c.path, c.url,
+                                                  c.current_rev[0:10],
+                                                  c.new_rev[0:10]))
+    change_url = CHROMIUM_FILE_TEMPLATE % (rev_interval, 'DEPS')
+    commit_msg.append('DEPS diff: %s\n' % change_url)
+  else:
+    commit_msg.append('No dependencies changed.')
+
+  if clang_change.current_rev != clang_change.new_rev:
+    commit_msg.append('Clang version changed %s:%s' %
+                      (clang_change.current_rev, clang_change.new_rev))
+    change_url = CHROMIUM_FILE_TEMPLATE % (rev_interval,
+                                           CLANG_UPDATE_SCRIPT_URL_PATH)
+    commit_msg.append('Details: %s\n' % change_url)
+  else:
+    commit_msg.append('No update to Clang.\n')
+
+  # TBR needs to be non-empty for Gerrit to process it.
+  git_author = _RunCommand(['git', 'config', 'user.email'],
+                           working_dir=CHECKOUT_SRC_DIR)[0].strip()
+  commit_msg.append('TBR=%s' % git_author)
+
+  commit_msg.append('BUG=None')
+  return '\n'.join(commit_msg)
+
+
+def UpdateDepsFile(deps_filename, old_cr_revision, new_cr_revision,
+                   changed_deps):
+  """Update the DEPS file with the new revision."""
+
+  # Update the chromium_revision variable.
+  with open(deps_filename, 'rb') as deps_file:
+    deps_content = deps_file.read().decode('utf-8')
+  deps_content = deps_content.replace(old_cr_revision, new_cr_revision)
+  with open(deps_filename, 'wb') as deps_file:
+    deps_file.write(deps_content.encode('utf-8'))
+
+  # Update each individual DEPS entry.
+  for dep in changed_deps:
+    local_dep_dir = os.path.join(CHECKOUT_ROOT_DIR, dep.path)
+    if not os.path.isdir(local_dep_dir):
+      raise RollError(
+          'Cannot find local directory %s. Make sure the .gclient file\n'
+          'contains all platforms in the target_os list, i.e.\n'
+          'target_os = ["android", "unix", "mac", "ios", "win"];\n'
+          'Then run "gclient sync" again.' % local_dep_dir)
+    _RunCommand(
+      ['gclient', 'setdep', '--revision', '%s@%s' % (dep.path, dep.new_rev)],
+      working_dir=CHECKOUT_SRC_DIR)
+
+
+def _IsTreeClean():
+  stdout, _ = _RunCommand(['git', 'status', '--porcelain'])
+  if len(stdout) == 0:
+    return True
+
+  logging.error('Dirty/unversioned files:\n%s', stdout)
+  return False
+
+
+def _EnsureUpdatedMasterBranch(dry_run):
+  current_branch = _RunCommand(
+      ['git', 'rev-parse', '--abbrev-ref', 'HEAD'])[0].splitlines()[0]
+  if current_branch != 'main':
+    logging.error('Please checkout the main branch and re-run this script.')
+    if not dry_run:
+      sys.exit(-1)
+
+  logging.info('Updating main branch...')
+  _RunCommand(['git', 'pull'])
+
+
+def _CreateRollBranch(dry_run):
+  logging.info('Creating roll branch: %s', ROLL_BRANCH_NAME)
+  if not dry_run:
+    _RunCommand(['git', 'checkout', '-b', ROLL_BRANCH_NAME])
+
+
+def _RemovePreviousRollBranch(dry_run):
+  active_branch, branches = _GetBranches()
+  if active_branch == ROLL_BRANCH_NAME:
+    active_branch = 'main'
+  if ROLL_BRANCH_NAME in branches:
+    logging.info('Removing previous roll branch (%s)', ROLL_BRANCH_NAME)
+    if not dry_run:
+      _RunCommand(['git', 'checkout', active_branch])
+      _RunCommand(['git', 'branch', '-D', ROLL_BRANCH_NAME])
+
+
+def _LocalCommit(commit_msg, dry_run):
+  logging.info('Committing changes locally.')
+  if not dry_run:
+    _RunCommand(['git', 'add', '--update', '.'])
+    _RunCommand(['git', 'commit', '-m', commit_msg])
+
+
+def ChooseCQMode(skip_cq, cq_over, current_commit_pos, new_commit_pos):
+  if skip_cq:
+    return 0
+  if (new_commit_pos - current_commit_pos) < cq_over:
+    return 1
+  return 2
+
+
+def _UploadCL(commit_queue_mode):
+  """Upload the committed changes as a changelist to Gerrit.
+
+  commit_queue_mode:
+    - 2: Submit to commit queue.
+    - 1: Run trybots but do not submit to CQ.
+    - 0: Skip CQ, upload only.
+  """
+  cmd = ['git', 'cl', 'upload', '--force', '--bypass-hooks']
+  if commit_queue_mode >= 2:
+    logging.info('Sending the CL to the CQ...')
+    cmd.extend(['-o', 'label=Bot-Commit+1'])
+    cmd.extend(['-o', 'label=Commit-Queue+2'])
+  elif commit_queue_mode >= 1:
+    logging.info('Starting CQ dry run...')
+    cmd.extend(['-o', 'label=Commit-Queue+1'])
+  extra_env = {
+      'EDITOR': 'true',
+      'SKIP_GCE_AUTH_FOR_GIT': '1',
+  }
+  stdout, stderr = _RunCommand(cmd, extra_env=extra_env)
+  logging.debug('Output from "git cl upload":\nstdout:\n%s\n\nstderr:\n%s',
+      stdout, stderr)
+
+
+def main():
+  p = argparse.ArgumentParser()
+  p.add_argument('--clean', action='store_true', default=False,
+                 help='Removes any previous local roll branch.')
+  p.add_argument('-r', '--revision',
+                 help=('Chromium Git revision to roll to. Defaults to the '
+                       'Chromium HEAD revision if omitted.'))
+  p.add_argument('--dry-run', action='store_true', default=False,
+                 help=('Calculate changes and modify DEPS, but don\'t create '
+                       'any local branch, commit, upload CL or send any '
+                       'tryjobs.'))
+  p.add_argument('-i', '--ignore-unclean-workdir', action='store_true',
+                 default=False,
+                 help=('Ignore if the current branch is not main or if there '
+                       'are uncommitted changes (default: %(default)s).'))
+  grp = p.add_mutually_exclusive_group()
+  grp.add_argument('--skip-cq', action='store_true', default=False,
+                   help='Skip sending the CL to the CQ (default: %(default)s)')
+  grp.add_argument('--cq-over', type=int, default=1,
+                   help=('Commit queue dry run if the revision difference '
+                         'is below this number (default: %(default)s)'))
+  p.add_argument('-v', '--verbose', action='store_true', default=False,
+                 help='Be extra verbose in printing of log messages.')
+  opts = p.parse_args()
+
+  if opts.verbose:
+    logging.basicConfig(level=logging.DEBUG)
+  else:
+    logging.basicConfig(level=logging.INFO)
+
+  if not opts.ignore_unclean_workdir and not _IsTreeClean():
+    logging.error('Please clean your local checkout first.')
+    return 1
+
+  if opts.clean:
+    _RemovePreviousRollBranch(opts.dry_run)
+
+  if not opts.ignore_unclean_workdir:
+    _EnsureUpdatedMasterBranch(opts.dry_run)
+
+  new_cr_rev = opts.revision
+  if not new_cr_rev:
+    stdout, _ = _RunCommand(['git', 'ls-remote', CHROMIUM_SRC_URL, 'HEAD'])
+    head_rev = stdout.strip().split('\t')[0]
+    logging.info('No revision specified. Using HEAD: %s', head_rev)
+    new_cr_rev = head_rev
+
+  deps_filename = os.path.join(CHECKOUT_SRC_DIR, 'DEPS')
+  libyuv_deps = ParseLocalDepsFile(deps_filename)
+  current_cr_rev = libyuv_deps['vars']['chromium_revision']
+
+  current_commit_pos = ParseCommitPosition(ReadRemoteCrCommit(current_cr_rev))
+  new_commit_pos = ParseCommitPosition(ReadRemoteCrCommit(new_cr_rev))
+
+  new_cr_deps = ParseRemoteCrDepsFile(new_cr_rev)
+  changed_deps = CalculateChangedDeps(libyuv_deps, new_cr_deps)
+  clang_change = CalculateChangedClang(new_cr_rev)
+  commit_msg = GenerateCommitMessage(current_cr_rev, new_cr_rev,
+                                     current_commit_pos, new_commit_pos,
+                                     changed_deps, clang_change)
+  logging.debug('Commit message:\n%s', commit_msg)
+
+  _CreateRollBranch(opts.dry_run)
+  UpdateDepsFile(deps_filename, current_cr_rev, new_cr_rev, changed_deps)
+  _LocalCommit(commit_msg, opts.dry_run)
+  commit_queue_mode = ChooseCQMode(opts.skip_cq, opts.cq_over,
+                                   current_commit_pos, new_commit_pos)
+  logging.info('Uploading CL...')
+  if not opts.dry_run:
+    _UploadCL(commit_queue_mode)
+  return 0
+
+
+if __name__ == '__main__':
+  sys.exit(main())
diff --git a/media/libyuv/tools_libyuv/autoroller/unittests/roll_deps_test.py b/media/libyuv/tools_libyuv/autoroller/unittests/roll_deps_test.py
new file mode 100755
index 0000000000..af86bdd586
--- /dev/null
+++ b/media/libyuv/tools_libyuv/autoroller/unittests/roll_deps_test.py
@@ -0,0 +1,148 @@
+#!/usr/bin/env vpython3
+
+# Copyright 2017 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+import glob
+import os
+import shutil
+import sys
+import tempfile
+import unittest
+
+import roll_deps
+from roll_deps import CalculateChangedDeps, GetMatchingDepsEntries, \
+  ParseDepsDict, ParseLocalDepsFile, UpdateDepsFile
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+PARENT_DIR = os.path.join(SCRIPT_DIR, os.pardir)
+sys.path.append(PARENT_DIR)
+
+TEST_DATA_VARS = {
+  'chromium_git': 'https://chromium.googlesource.com',
+  'chromium_revision': '1b9c098a08e40114e44b6c1ec33ddf95c40b901d',
+}
+
+DEPS_ENTRIES = {
+  'src/build': 'https://build.com',
+  'src/buildtools': 'https://buildtools.com',
+  'src/testing/gtest': 'https://gtest.com',
+  'src/testing/gmock': 'https://gmock.com',
+}
+
+BUILD_OLD_REV = '52f7afeca991d96d68cf0507e20dbdd5b845691f'
+BUILD_NEW_REV = 'HEAD'
+BUILDTOOLS_OLD_REV = '64e38f0cebdde27aa0cfb405f330063582f9ac76'
+BUILDTOOLS_NEW_REV = '55ad626b08ef971fd82a62b7abb325359542952b'
+
+
+class TestError(Exception):
+  pass
+
+
+class FakeCmd():
+  def __init__(self):
+    self.expectations = []
+
+  def add_expectation(self, *args, **kwargs):
+    returns = kwargs.pop('_returns', None)
+    self.expectations.append((args, kwargs, returns))
+
+  def __call__(self, *args, **kwargs):
+    if not self.expectations:
+      raise TestError('Got unexpected\n%s\n%s' % (args, kwargs))
+    exp_args, exp_kwargs, exp_returns = self.expectations.pop(0)
+    if args != exp_args or kwargs != exp_kwargs:
+      message = 'Expected:\n  args: %s\n  kwargs: %s\n' % (exp_args, exp_kwargs)
+      message += 'Got:\n  args: %s\n  kwargs: %s\n' % (args, kwargs)
+      raise TestError(message)
+    return exp_returns
+
+
+class TestRollChromiumRevision(unittest.TestCase):
+  def setUp(self):
+    self._output_dir = tempfile.mkdtemp()
+    for test_file in glob.glob(os.path.join(SCRIPT_DIR, 'testdata', '*')):
+      shutil.copy(test_file, self._output_dir)
+    self._libyuv_depsfile = os.path.join(self._output_dir, 'DEPS')
+    self._old_cr_depsfile = os.path.join(self._output_dir, 'DEPS.chromium.old')
+    self._new_cr_depsfile = os.path.join(self._output_dir, 'DEPS.chromium.new')
+
+    self.fake = FakeCmd()
+    self.old_RunCommand = getattr(roll_deps, '_RunCommand')
+    setattr(roll_deps, '_RunCommand', self.fake)
+
+  def tearDown(self):
+    shutil.rmtree(self._output_dir, ignore_errors=True)
+    self.assertEqual(self.fake.expectations, [])
+    setattr(roll_deps, '_RunCommand', self.old_RunCommand)
+
+  def testVarLookup(self):
+    local_scope = {'foo': 'wrong', 'vars': {'foo': 'bar'}}
+    lookup = roll_deps.VarLookup(local_scope)
+    self.assertEqual(lookup('foo'), 'bar')
+
+  def testUpdateDepsFile(self):
+    new_rev = 'aaaaabbbbbcccccdddddeeeeefffff0000011111'
+
+    current_rev = TEST_DATA_VARS['chromium_revision']
+    UpdateDepsFile(self._libyuv_depsfile, current_rev, new_rev, [])
+    with open(self._libyuv_depsfile, 'r') as deps_file:
+      deps_contents = deps_file.read()
+      self.assertTrue(new_rev in deps_contents,
+                      'Failed to find %s in\n%s' % (new_rev, deps_contents))
+
+  def testParseDepsDict(self):
+    with open(self._libyuv_depsfile, 'r') as deps_file:
+      deps_contents = deps_file.read()
+    local_scope = ParseDepsDict(deps_contents)
+    vars_dict = local_scope['vars']
+
+    def assertVar(variable_name):
+      self.assertEqual(vars_dict[variable_name], TEST_DATA_VARS[variable_name])
+    assertVar('chromium_git')
+    assertVar('chromium_revision')
+    self.assertEqual(len(local_scope['deps']), 3)
+
+  def testGetMatchingDepsEntriesReturnsPathInSimpleCase(self):
+    entries = GetMatchingDepsEntries(DEPS_ENTRIES, 'src/testing/gtest')
+    self.assertEqual(len(entries), 1)
+    self.assertEqual(entries[0], DEPS_ENTRIES['src/testing/gtest'])
+
+  def testGetMatchingDepsEntriesHandlesSimilarStartingPaths(self):
+    entries = GetMatchingDepsEntries(DEPS_ENTRIES, 'src/testing')
+    self.assertEqual(len(entries), 2)
+
+  def testGetMatchingDepsEntriesHandlesTwoPathsWithIdenticalFirstParts(self):
+    entries = GetMatchingDepsEntries(DEPS_ENTRIES, 'src/build')
+    self.assertEqual(len(entries), 1)
+    self.assertEqual(entries[0], DEPS_ENTRIES['src/build'])
+
+  def testCalculateChangedDeps(self):
+    _SetupGitLsRemoteCall(self.fake,
+        'https://chromium.googlesource.com/chromium/src/build', BUILD_NEW_REV)
+    libyuv_deps = ParseLocalDepsFile(self._libyuv_depsfile)
+    new_cr_deps = ParseLocalDepsFile(self._new_cr_depsfile)
+    changed_deps = CalculateChangedDeps(libyuv_deps, new_cr_deps)
+    self.assertEqual(len(changed_deps), 2)
+    self.assertEqual(changed_deps[0].path, 'src/build')
+    self.assertEqual(changed_deps[0].current_rev, BUILD_OLD_REV)
+    self.assertEqual(changed_deps[0].new_rev, BUILD_NEW_REV)
+
+    self.assertEqual(changed_deps[1].path, 'src/buildtools')
+    self.assertEqual(changed_deps[1].current_rev, BUILDTOOLS_OLD_REV)
+    self.assertEqual(changed_deps[1].new_rev, BUILDTOOLS_NEW_REV)
+
+
+def _SetupGitLsRemoteCall(cmd_fake, url, revision):
+  cmd = ['git', 'ls-remote', url, revision]
+  cmd_fake.add_expectation(cmd, _returns=(revision, None))
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS b/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS
new file mode 100644
index 0000000000..4f45860cda
--- /dev/null
+++ b/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS
@@ -0,0 +1,21 @@
+# DEPS file for unit tests.
+
+vars = {
+  'chromium_git': 'https://chromium.googlesource.com',
+  'chromium_revision': '1b9c098a08e40114e44b6c1ec33ddf95c40b901d',
+  'ignored_str': Str(''),
+}
+
+deps = {
+  # Entry that is a directory in Chromium, so we're using a Git subtree mirror for it.
+  'src/build':
+    Var('chromium_git') + '/chromium/src/build' + '@' + '52f7afeca991d96d68cf0507e20dbdd5b845691f',
+
+  # Entry that's also a DEPS entry in the Chromium DEPS file.
+  'src/buildtools':
+    Var('chromium_git') + '/chromium/buildtools.git' + '@' + '64e38f0cebdde27aa0cfb405f330063582f9ac76',
+
+  # Entry only present in libyuv, not Chromium.
+  'src/third_party/gflags/src':
+    Var('chromium_git') + '/external/github.com/gflags/gflags@03bebcb065c83beff83d50ae025a55a4bf94dfca',
+}
diff --git a/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.new b/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.new
new file mode 100644
index 0000000000..d53083ceec
--- /dev/null
+++ b/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.new
@@ -0,0 +1,13 @@
+# DEPS file for unit tests.
+
+vars = {
+  'chromium_git': 'https://chromium.googlesource.com',
+
+  # This is updated compared to the DEPS.chromium.old file.
+  'buildtools_revision': '55ad626b08ef971fd82a62b7abb325359542952b',
+}
+
+deps = {
+  'src/buildtools':
+    Var('chromium_git') + '/chromium/buildtools.git' + '@' +  Var('buildtools_revision'),
+}
diff --git a/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.old b/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.old
new file mode 100644
index 0000000000..dd6ddaec81
--- /dev/null
+++ b/media/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.old
@@ -0,0 +1,13 @@
+# DEPS file for unit tests.
+
+vars = {
+  'chromium_git': 'https://chromium.googlesource.com',
+
+  # This is and older revision than DEPS.chromium.new file.
+  'buildtools_revision': '64e38f0cebdde27aa0cfb405f330063582f9ac76',
+}
+
+deps = {
+  'src/buildtools':
+    Var('chromium_git') + '/chromium/buildtools.git' + '@' +  Var('buildtools_revision'),
+}
diff --git a/media/libyuv/tools_libyuv/get_landmines.py b/media/libyuv/tools_libyuv/get_landmines.py
new file mode 100755
index 0000000000..8b33483e64
--- /dev/null
+++ b/media/libyuv/tools_libyuv/get_landmines.py
@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+
+# Copyright 2016 The LibYuv Project Authors. All rights reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS. All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+"""
+This file emits the list of reasons why a particular build needs to be clobbered
+(or a list of 'landmines').
+"""
+
+import sys
+
+
+def print_landmines():
+  """
+  ALL LANDMINES ARE EMITTED FROM HERE.
+  """
+  # DO NOT add landmines as part of a regular CL. Landmines are a last-effort
+  # bandaid fix if a CL that got landed has a build dependency bug and all bots
+  # need to be cleaned up. If you're writing a new CL that causes build
+  # dependency problems, fix the dependency problems instead of adding a
+  # landmine.
+  # See the Chromium version in src/build/get_landmines.py for usage examples.
+  print('Clobber to remove GYP artifacts after switching bots to GN.')
+  print('Another try to remove GYP artifacts after switching bots to GN.')
+
+
+def main():
+  print_landmines()
+  return 0
+
+
+if __name__ == '__main__':
+  sys.exit(main())
diff --git a/media/libyuv/tools_libyuv/msan/OWNERS b/media/libyuv/tools_libyuv/msan/OWNERS
new file mode 100644
index 0000000000..9b67a8f678
--- /dev/null
+++ b/media/libyuv/tools_libyuv/msan/OWNERS
@@ -0,0 +1,3 @@
+mbonadei@chromium.org
+fbarchard@chromium.org
+pbos@chromium.org
diff --git a/media/libyuv/tools/msan/blacklist.txt b/media/libyuv/tools_libyuv/msan/blacklist.txt
index 8b5e42a7b3..8b5e42a7b3 100644
--- a/media/libyuv/tools/msan/blacklist.txt
+++ b/media/libyuv/tools_libyuv/msan/blacklist.txt
diff --git a/media/libyuv/tools_libyuv/ubsan/OWNERS b/media/libyuv/tools_libyuv/ubsan/OWNERS
new file mode 100644
index 0000000000..9b67a8f678
--- /dev/null
+++ b/media/libyuv/tools_libyuv/ubsan/OWNERS
@@ -0,0 +1,3 @@
+mbonadei@chromium.org
+fbarchard@chromium.org
+pbos@chromium.org
diff --git a/media/libyuv/tools/ubsan/blacklist.txt b/media/libyuv/tools_libyuv/ubsan/blacklist.txt
index 8bcb29073b..8bcb29073b 100644
--- a/media/libyuv/tools/ubsan/blacklist.txt
+++ b/media/libyuv/tools_libyuv/ubsan/blacklist.txt
diff --git a/media/libyuv/tools/ubsan/vptr_blacklist.txt b/media/libyuv/tools_libyuv/ubsan/vptr_blacklist.txt
index 8ed070c05d..23cfca5373 100644
--- a/media/libyuv/tools/ubsan/vptr_blacklist.txt
+++ b/media/libyuv/tools_libyuv/ubsan/vptr_blacklist.txt
@@ -19,3 +19,7 @@
 # Example:
 # src:*/third_party/protobuf/src/google/protobuf/compiler/plugin.pb.cc
 
+#############################################################################
+# UBsan goes into an infinite recursion when __dynamic_cast instrumented with
+# "vptr". See crbug.com/609786.
+src:*/third_party/libc\+\+abi/trunk/src/private_typeinfo.cpp
diff --git a/media/libyuv/unit_test/basictypes_test.cc b/media/libyuv/unit_test/basictypes_test.cc
index 89f7644d58..9aaa2dcd98 100644
--- a/media/libyuv/unit_test/basictypes_test.cc
+++ b/media/libyuv/unit_test/basictypes_test.cc
@@ -13,25 +13,15 @@
 
 namespace libyuv {
 
-TEST_F(LibYUVBaseTest, Endian) {
-  uint16 v16 = 0x1234u;
-  uint8 first_byte = *reinterpret_cast<uint8*>(&v16);
-#if defined(LIBYUV_LITTLE_ENDIAN)
-  EXPECT_EQ(0x34u, first_byte);
-#else
-  EXPECT_EQ(0x12u, first_byte);
-#endif
-}
-
 TEST_F(LibYUVBaseTest, SizeOfTypes) {
-  int8 i8 = -1;
-  uint8 u8 = 1u;
-  int16 i16 = -1;
-  uint16 u16 = 1u;
-  int32 i32 = -1;
-  uint32 u32 = 1u;
-  int64 i64 = -1;
-  uint64 u64 = 1u;
+  int8_t i8 = -1;
+  uint8_t u8 = 1u;
+  int16_t i16 = -1;
+  uint16_t u16 = 1u;
+  int32_t i32 = -1;
+  uint32_t u32 = 1u;
+  int64_t i64 = -1;
+  uint64_t u64 = 1u;
   EXPECT_EQ(1u, sizeof(i8));
   EXPECT_EQ(1u, sizeof(u8));
   EXPECT_EQ(2u, sizeof(i16));
@@ -50,11 +40,4 @@ TEST_F(LibYUVBaseTest, SizeOfTypes) {
   EXPECT_LT(0u, u64);
 }
 
-TEST_F(LibYUVBaseTest, SizeOfConstants) {
-  EXPECT_EQ(8u, sizeof(INT64_C(0)));
-  EXPECT_EQ(8u, sizeof(UINT64_C(0)));
-  EXPECT_EQ(8u, sizeof(INT64_C(0x1234567887654321)));
-  EXPECT_EQ(8u, sizeof(UINT64_C(0x8765432112345678)));
-}
-
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/color_test.cc b/media/libyuv/unit_test/color_test.cc
index 36041d9982..01267ff1ed 100644
--- a/media/libyuv/unit_test/color_test.cc
+++ b/media/libyuv/unit_test/color_test.cc
@@ -10,138 +10,133 @@
 
 #include <stdlib.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/basic_types.h"
 #include "libyuv/convert.h"
 #include "libyuv/convert_argb.h"
 #include "libyuv/convert_from.h"
 #include "libyuv/convert_from_argb.h"
 #include "libyuv/cpu_id.h"
-#include "../unit_test/unit_test.h"
 
 namespace libyuv {
 
-// TODO(fbarchard): Port high accuracy YUV to RGB to Neon.
-#if !defined(LIBYUV_DISABLE_NEON) && \
-    (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
-#define ERROR_R 1
-#define ERROR_G 1
-#define ERROR_B 3
-#define ERROR_FULL 6
-#define ERROR_J420 5
+// TODO(fbarchard): clang x86 has a higher accuracy YUV to RGB.
+// Port to Visual C and other CPUs
+#if !defined(LIBYUV_BIT_EXACT) && !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || defined(__i386__))
+#define ERROR_FULL 5
+#define ERROR_J420 4
 #else
+#define ERROR_FULL 6
+#define ERROR_J420 6
+#endif
 #define ERROR_R 1
 #define ERROR_G 1
-#define ERROR_B 3
-#define ERROR_FULL 5
-#define ERROR_J420 3
+#ifdef LIBYUV_UNLIMITED_DATA
+#define ERROR_B 1
+#else
+#define ERROR_B 18
 #endif
 
 #define TESTCS(TESTNAME, YUVTOARGB, ARGBTOYUV, HS1, HS, HN, DIFF)              \
   TEST_F(LibYUVColorTest, TESTNAME) {                                          \
-  const int kPixels = benchmark_width_ * benchmark_height_;                    \
-  const int kHalfPixels = ((benchmark_width_ + 1) / 2) *                       \
-      ((benchmark_height_ + HS1) / HS);                                        \
-  align_buffer_page_end(orig_y, kPixels);                                      \
-  align_buffer_page_end(orig_u, kHalfPixels);                                  \
-  align_buffer_page_end(orig_v, kHalfPixels);                                  \
-  align_buffer_page_end(orig_pixels, kPixels * 4);                             \
-  align_buffer_page_end(temp_y, kPixels);                                      \
-  align_buffer_page_end(temp_u, kHalfPixels);                                  \
-  align_buffer_page_end(temp_v, kHalfPixels);                                  \
-  align_buffer_page_end(dst_pixels_opt, kPixels * 4);                          \
-  align_buffer_page_end(dst_pixels_c, kPixels * 4);                            \
+    const int kPixels = benchmark_width_ * benchmark_height_;                  \
+    const int kHalfPixels =                                                    \
+        ((benchmark_width_ + 1) / 2) * ((benchmark_height_ + HS1) / HS);       \
+    align_buffer_page_end(orig_y, kPixels);                                    \
+    align_buffer_page_end(orig_u, kHalfPixels);                                \
+    align_buffer_page_end(orig_v, kHalfPixels);                                \
+    align_buffer_page_end(orig_pixels, kPixels * 4);                           \
+    align_buffer_page_end(temp_y, kPixels);                                    \
+    align_buffer_page_end(temp_u, kHalfPixels);                                \
+    align_buffer_page_end(temp_v, kHalfPixels);                                \
+    align_buffer_page_end(dst_pixels_opt, kPixels * 4);                        \
+    align_buffer_page_end(dst_pixels_c, kPixels * 4);                          \
                                                                                \
-  MemRandomize(orig_pixels, kPixels * 4);                                      \
-  MemRandomize(orig_y, kPixels);                                               \
-  MemRandomize(orig_u, kHalfPixels);                                           \
-  MemRandomize(orig_v, kHalfPixels);                                           \
-  MemRandomize(temp_y, kPixels);                                               \
-  MemRandomize(temp_u, kHalfPixels);                                           \
-  MemRandomize(temp_v, kHalfPixels);                                           \
-  MemRandomize(dst_pixels_opt, kPixels * 4);                                   \
-  MemRandomize(dst_pixels_c, kPixels * 4);                                     \
+    MemRandomize(orig_pixels, kPixels * 4);                                    \
+    MemRandomize(orig_y, kPixels);                                             \
+    MemRandomize(orig_u, kHalfPixels);                                         \
+    MemRandomize(orig_v, kHalfPixels);                                         \
+    MemRandomize(temp_y, kPixels);                                             \
+    MemRandomize(temp_u, kHalfPixels);                                         \
+    MemRandomize(temp_v, kHalfPixels);                                         \
+    MemRandomize(dst_pixels_opt, kPixels * 4);                                 \
+    MemRandomize(dst_pixels_c, kPixels * 4);                                   \
                                                                                \
-  /* The test is overall for color conversion matrix being reversible, so */   \
-  /* this initializes the pixel with 2x2 blocks to eliminate subsampling. */   \
-  uint8* p = orig_y;                                                           \
-  for (int y = 0; y < benchmark_height_ - HS1; y += HS) {                      \
-    for (int x = 0; x < benchmark_width_ - 1; x += 2) {                        \
-      uint8 r = static_cast<uint8>(fastrand());                                \
-      p[0] = r;                                                                \
-      p[1] = r;                                                                \
-      p[HN] = r;                                                               \
-      p[HN + 1] = r;                                                           \
-      p += 2;                                                                  \
-    }                                                                          \
-    if (benchmark_width_ & 1) {                                                \
-      uint8 r = static_cast<uint8>(fastrand());                                \
-      p[0] = r;                                                                \
-      p[HN] = r;                                                               \
-      p += 1;                                                                  \
+    /* The test is overall for color conversion matrix being reversible, so */ \
+    /* this initializes the pixel with 2x2 blocks to eliminate subsampling. */ \
+    uint8_t* p = orig_y;                                                       \
+    for (int y = 0; y < benchmark_height_ - HS1; y += HS) {                    \
+      for (int x = 0; x < benchmark_width_ - 1; x += 2) {                      \
+        uint8_t r = static_cast<uint8_t>(fastrand());                          \
+        p[0] = r;                                                              \
+        p[1] = r;                                                              \
+        p[HN] = r;                                                             \
+        p[HN + 1] = r;                                                         \
+        p += 2;                                                                \
+      }                                                                        \
+      if (benchmark_width_ & 1) {                                              \
+        uint8_t r = static_cast<uint8_t>(fastrand());                          \
+        p[0] = r;                                                              \
+        p[HN] = r;                                                             \
+        p += 1;                                                                \
+      }                                                                        \
+      p += HN;                                                                 \
     }                                                                          \
-    p += HN;                                                                   \
-  }                                                                            \
-  if ((benchmark_height_ & 1) && HS == 2) {                                    \
-    for (int x = 0; x < benchmark_width_ - 1; x += 2) {                        \
-      uint8 r = static_cast<uint8>(fastrand());                                \
-      p[0] = r;                                                                \
-      p[1] = r;                                                                \
-      p += 2;                                                                  \
+    if ((benchmark_height_ & 1) && HS == 2) {                                  \
+      for (int x = 0; x < benchmark_width_ - 1; x += 2) {                      \
+        uint8_t r = static_cast<uint8_t>(fastrand());                          \
+        p[0] = r;                                                              \
+        p[1] = r;                                                              \
+        p += 2;                                                                \
+      }                                                                        \
+      if (benchmark_width_ & 1) {                                              \
+        uint8_t r = static_cast<uint8_t>(fastrand());                          \
+        p[0] = r;                                                              \
+        p += 1;                                                                \
+      }                                                                        \
     }                                                                          \
-    if (benchmark_width_ & 1) {                                                \
-      uint8 r = static_cast<uint8>(fastrand());                                \
-      p[0] = r;                                                                \
-      p += 1;                                                                  \
-    }                                                                          \
-  }                                                                            \
-  /* Start with YUV converted to ARGB. */                                      \
-  YUVTOARGB(orig_y, benchmark_width_,                                          \
-            orig_u, (benchmark_width_ + 1) / 2,                                \
-            orig_v, (benchmark_width_ + 1) / 2,                                \
-            orig_pixels, benchmark_width_ * 4,                                 \
-            benchmark_width_, benchmark_height_);                              \
+    /* Start with YUV converted to ARGB. */                                    \
+    YUVTOARGB(orig_y, benchmark_width_, orig_u, (benchmark_width_ + 1) / 2,    \
+              orig_v, (benchmark_width_ + 1) / 2, orig_pixels,                 \
+              benchmark_width_ * 4, benchmark_width_, benchmark_height_);      \
                                                                                \
-  ARGBTOYUV(orig_pixels, benchmark_width_ * 4,                                 \
-            temp_y, benchmark_width_,                                          \
-            temp_u, (benchmark_width_ + 1) / 2,                                \
-            temp_v, (benchmark_width_ + 1) / 2,                                \
-            benchmark_width_, benchmark_height_);                              \
+    ARGBTOYUV(orig_pixels, benchmark_width_ * 4, temp_y, benchmark_width_,     \
+              temp_u, (benchmark_width_ + 1) / 2, temp_v,                      \
+              (benchmark_width_ + 1) / 2, benchmark_width_,                    \
+              benchmark_height_);                                              \
                                                                                \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  YUVTOARGB(temp_y, benchmark_width_,                                          \
-            temp_u, (benchmark_width_ + 1) / 2,                                \
-            temp_v, (benchmark_width_ + 1) / 2,                                \
-            dst_pixels_c, benchmark_width_ * 4,                                \
-            benchmark_width_, benchmark_height_);                              \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    YUVTOARGB(temp_y, benchmark_width_, temp_u, (benchmark_width_ + 1) / 2,    \
+              temp_v, (benchmark_width_ + 1) / 2, dst_pixels_c,                \
+              benchmark_width_ * 4, benchmark_width_, benchmark_height_);      \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
                                                                                \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    YUVTOARGB(temp_y, benchmark_width_,                                        \
-              temp_u, (benchmark_width_ + 1) / 2,                              \
-              temp_v, (benchmark_width_ + 1) / 2,                              \
-              dst_pixels_opt, benchmark_width_ * 4,                            \
-              benchmark_width_, benchmark_height_);                            \
-  }                                                                            \
-  /* Test C and SIMD match. */                                                 \
-  for (int i = 0; i < kPixels * 4; ++i) {                                      \
-    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                             \
-  }                                                                            \
-  /* Test SIMD is close to original. */                                        \
-  for (int i = 0; i < kPixels * 4; ++i) {                                      \
-    EXPECT_NEAR(static_cast<int>(orig_pixels[i]),                              \
-                static_cast<int>(dst_pixels_opt[i]), DIFF);                    \
-  }                                                                            \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      YUVTOARGB(temp_y, benchmark_width_, temp_u, (benchmark_width_ + 1) / 2,  \
+                temp_v, (benchmark_width_ + 1) / 2, dst_pixels_opt,            \
+                benchmark_width_ * 4, benchmark_width_, benchmark_height_);    \
+    }                                                                          \
+    /* Test C and SIMD match. */                                               \
+    for (int i = 0; i < kPixels * 4; ++i) {                                    \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                           \
+    }                                                                          \
+    /* Test SIMD is close to original. */                                      \
+    for (int i = 0; i < kPixels * 4; ++i) {                                    \
+      EXPECT_NEAR(static_cast<int>(orig_pixels[i]),                            \
+                  static_cast<int>(dst_pixels_opt[i]), DIFF);                  \
+    }                                                                          \
                                                                                \
-  free_aligned_buffer_page_end(orig_pixels);                                   \
-  free_aligned_buffer_page_end(orig_y);                                        \
-  free_aligned_buffer_page_end(orig_u);                                        \
-  free_aligned_buffer_page_end(orig_v);                                        \
-  free_aligned_buffer_page_end(temp_y);                                        \
-  free_aligned_buffer_page_end(temp_u);                                        \
-  free_aligned_buffer_page_end(temp_v);                                        \
-  free_aligned_buffer_page_end(dst_pixels_opt);                                \
-  free_aligned_buffer_page_end(dst_pixels_c);                                  \
-}                                                                              \
+    free_aligned_buffer_page_end(orig_pixels);                                 \
+    free_aligned_buffer_page_end(orig_y);                                      \
+    free_aligned_buffer_page_end(orig_u);                                      \
+    free_aligned_buffer_page_end(orig_v);                                      \
+    free_aligned_buffer_page_end(temp_y);                                      \
+    free_aligned_buffer_page_end(temp_u);                                      \
+    free_aligned_buffer_page_end(temp_v);                                      \
+    free_aligned_buffer_page_end(dst_pixels_opt);                              \
+    free_aligned_buffer_page_end(dst_pixels_c);                                \
+  }
 
 TESTCS(TestI420, I420ToARGB, ARGBToI420, 1, 2, benchmark_width_, ERROR_FULL)
 TESTCS(TestI422, I422ToARGB, ARGBToI422, 0, 1, 0, ERROR_FULL)
@@ -154,20 +149,17 @@ static void YUVToRGB(int y, int u, int v, int* r, int* g, int* b) {
   const int kPixels = kWidth * kHeight;
   const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
 
-  SIMD_ALIGNED(uint8 orig_y[16]);
-  SIMD_ALIGNED(uint8 orig_u[8]);
-  SIMD_ALIGNED(uint8 orig_v[8]);
-  SIMD_ALIGNED(uint8 orig_pixels[16 * 4]);
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
   memset(orig_y, y, kPixels);
   memset(orig_u, u, kHalfPixels);
   memset(orig_v, v, kHalfPixels);
 
   /* YUV converted to ARGB. */
-  I422ToARGB(orig_y, kWidth,
-             orig_u, (kWidth + 1) / 2,
-             orig_v, (kWidth + 1) / 2,
-             orig_pixels, kWidth * 4,
-             kWidth, kHeight);
+  I422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
 
   *b = orig_pixels[0];
   *g = orig_pixels[1];
@@ -180,20 +172,115 @@ static void YUVJToRGB(int y, int u, int v, int* r, int* g, int* b) {
   const int kPixels = kWidth * kHeight;
   const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
 
-  SIMD_ALIGNED(uint8 orig_y[16]);
-  SIMD_ALIGNED(uint8 orig_u[8]);
-  SIMD_ALIGNED(uint8 orig_v[8]);
-  SIMD_ALIGNED(uint8 orig_pixels[16 * 4]);
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  J422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+static void YUVHToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  H422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+#define F422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+
+static void YUVFToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  F422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+static void YUVUToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  U422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+#define V422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+
+static void YUVVToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
   memset(orig_y, y, kPixels);
   memset(orig_u, u, kHalfPixels);
   memset(orig_v, v, kHalfPixels);
 
   /* YUV converted to ARGB. */
-  J422ToARGB(orig_y, kWidth,
-             orig_u, (kWidth + 1) / 2,
-             orig_v, (kWidth + 1) / 2,
-             orig_pixels, kWidth * 4,
-             kWidth, kHeight);
+  V422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
 
   *b = orig_pixels[0];
   *g = orig_pixels[1];
@@ -205,8 +292,8 @@ static void YToRGB(int y, int* r, int* g, int* b) {
   const int kHeight = 1;
   const int kPixels = kWidth * kHeight;
 
-  SIMD_ALIGNED(uint8 orig_y[16]);
-  SIMD_ALIGNED(uint8 orig_pixels[16 * 4]);
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
   memset(orig_y, y, kPixels);
 
   /* YUV converted to ARGB. */
@@ -222,8 +309,8 @@ static void YJToRGB(int y, int* r, int* g, int* b) {
   const int kHeight = 1;
   const int kPixels = kWidth * kHeight;
 
-  SIMD_ALIGNED(uint8 orig_y[16]);
-  SIMD_ALIGNED(uint8 orig_pixels[16 * 4]);
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
   memset(orig_y, y, kPixels);
 
   /* YUV converted to ARGB. */
@@ -248,7 +335,7 @@ static void YJToRGB(int y, int* r, int* g, int* b) {
 
 #if defined(CLAMPMETHOD_IF)
 static int RoundToByte(float f) {
-  int i =  ROUND(f);
+  int i = ROUND(f);
   if (i < 0) {
     i = 0;
   }
@@ -259,52 +346,61 @@ static int RoundToByte(float f) {
 }
 #elif defined(CLAMPMETHOD_TABLE)
 static const unsigned char clamptable[811] = {
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
-  10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28,
-  29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
-  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66,
-  67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85,
-  86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103,
-  104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118,
-  119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133,
-  134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148,
-  149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163,
-  164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178,
-  179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193,
-  194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208,
-  209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223,
-  224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238,
-  239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253,
-  254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255
-};
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
+    0,   0,   0,   0,   0,   0,   0,   1,   2,   3,   4,   5,   6,   7,   8,
+    9,   10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  20,  21,  22,  23,
+    24,  25,  26,  27,  28,  29,  30,  31,  32,  33,  34,  35,  36,  37,  38,
+    39,  40,  41,  42,  43,  44,  45,  46,  47,  48,  49,  50,  51,  52,  53,
+    54,  55,  56,  57,  58,  59,  60,  61,  62,  63,  64,  65,  66,  67,  68,
+    69,  70,  71,  72,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,
+    84,  85,  86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  98,
+    99,  100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113,
+    114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128,
+    129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143,
+    144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158,
+    159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173,
+    174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188,
+    189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203,
+    204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218,
+    219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233,
+    234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248,
+    249, 250, 251, 252, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255};
 
 static int RoundToByte(float f) {
   return clamptable[ROUND(f) + 276];
@@ -317,7 +413,7 @@ static int RoundToByte(float f) {
 #elif defined(CLAMPMETHOD_MASK)
 static int RoundToByte(float f) {
   int i = ROUND(f);
-  i =  ((-(i) >> 31) & (i));  // clamp to 0.
+  i = ((-(i) >> 31) & (i));                  // clamp to 0.
   return (((255 - (i)) >> 31) | (i)) & 255;  // clamp to 255.
 }
 #endif
@@ -339,18 +435,50 @@ TEST_F(LibYUVColorTest, TestRoundToByte) {
   EXPECT_LE(allb, 255);
 }
 
+// BT.601 limited range YUV to RGB reference
 static void YUVToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
   *r = RoundToByte((y - 16) * 1.164 - (v - 128) * -1.596);
   *g = RoundToByte((y - 16) * 1.164 - (u - 128) * 0.391 - (v - 128) * 0.813);
   *b = RoundToByte((y - 16) * 1.164 - (u - 128) * -2.018);
 }
 
+// BT.601 full range YUV to RGB reference (aka JPEG)
 static void YUVJToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
   *r = RoundToByte(y - (v - 128) * -1.40200);
   *g = RoundToByte(y - (u - 128) * 0.34414 - (v - 128) * 0.71414);
   *b = RoundToByte(y - (u - 128) * -1.77200);
 }
 
+// BT.709 limited range YUV to RGB reference
+// See also http://www.equasys.de/colorconversion.html
+static void YUVHToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte((y - 16) * 1.164 - (v - 128) * -1.793);
+  *g = RoundToByte((y - 16) * 1.164 - (u - 128) * 0.213 - (v - 128) * 0.533);
+  *b = RoundToByte((y - 16) * 1.164 - (u - 128) * -2.112);
+}
+
+// BT.709 full range YUV to RGB reference
+static void YUVFToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte(y - (v - 128) * -1.5748);
+  *g = RoundToByte(y - (u - 128) * 0.18732 - (v - 128) * 0.46812);
+  *b = RoundToByte(y - (u - 128) * -1.8556);
+}
+
+// BT.2020 limited range YUV to RGB reference
+static void YUVUToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte((y - 16) * 1.164384 - (v - 128) * -1.67867);
+  *g = RoundToByte((y - 16) * 1.164384 - (u - 128) * 0.187326 -
+                   (v - 128) * 0.65042);
+  *b = RoundToByte((y - 16) * 1.164384 - (u - 128) * -2.14177);
+}
+
+// BT.2020 full range YUV to RGB reference
+static void YUVVToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte(y + (v - 128) * 1.474600);
+  *g = RoundToByte(y - (u - 128) * 0.164553 - (v - 128) * 0.571353);
+  *b = RoundToByte(y + (u - 128) * 1.881400);
+}
+
 TEST_F(LibYUVColorTest, TestYUV) {
   int r0, g0, b0, r1, g1, b1;
 
@@ -374,7 +502,11 @@ TEST_F(LibYUVColorTest, TestYUV) {
   YUVToRGB(240, 0, 0, &r1, &g1, &b1);
   EXPECT_EQ(57, r1);
   EXPECT_EQ(255, g1);
+#ifdef LIBYUV_UNLIMITED_DATA
+  EXPECT_EQ(3, b1);
+#else
   EXPECT_EQ(5, b1);
+#endif
 
   for (int i = 0; i < 256; ++i) {
     YUVToRGBReference(i, 128, 128, &r0, &g0, &b0);
@@ -433,7 +565,6 @@ TEST_F(LibYUVColorTest, TestGreyYUV) {
   EXPECT_EQ(130, g1);
   EXPECT_EQ(130, b1);
 
-
   for (int y = 0; y < 256; ++y) {
     YUVToRGBReference(y, 128, 128, &r0, &g0, &b0);
     YUVToRGB(y, 128, 128, &r1, &g1, &b1);
@@ -449,38 +580,55 @@ TEST_F(LibYUVColorTest, TestGreyYUV) {
 
 static void PrintHistogram(int rh[256], int gh[256], int bh[256]) {
   int i;
-  printf("hist");
+  printf("hist ");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", i - 128);
+      printf(" %8d", i - 128);
     }
   }
-  printf("\nred");
+  printf("\nred  ");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", rh[i]);
+      printf(" %8d", rh[i]);
     }
   }
   printf("\ngreen");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", gh[i]);
+      printf(" %8d", gh[i]);
     }
   }
-  printf("\nblue");
+  printf("\nblue ");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", bh[i]);
+      printf(" %8d", bh[i]);
     }
   }
   printf("\n");
 }
 
+// Step by 5 on inner loop goes from 0 to 255 inclusive.
+// Set to 1 for better converage.  3, 5 or 17 for faster testing.
+#ifdef DISABLE_SLOW_TESTS
+#define FASTSTEP 5
+#else
+#define FASTSTEP 1
+#endif
+
+// BT.601 limited range.
 TEST_F(LibYUVColorTest, TestFullYUV) {
-  int rh[256] = { 0, }, gh[256] = { 0, }, bh[256] = { 0, };
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
   for (int u = 0; u < 256; ++u) {
     for (int v = 0; v < 256; ++v) {
-      for (int y2 = 0; y2 < 256; ++y2) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
         int r0, g0, b0, r1, g1, b1;
         int y = RANDOM256(y2);
         YUVToRGBReference(y, u, v, &r0, &g0, &b0);
@@ -497,18 +645,147 @@ TEST_F(LibYUVColorTest, TestFullYUV) {
   PrintHistogram(rh, gh, bh);
 }
 
+// BT.601 full range.
 TEST_F(LibYUVColorTest, TestFullYUVJ) {
-  int rh[256] = { 0, }, gh[256] = { 0, }, bh[256] = { 0, };
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
   for (int u = 0; u < 256; ++u) {
     for (int v = 0; v < 256; ++v) {
-      for (int y2 = 0; y2 < 256; ++y2) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
         int r0, g0, b0, r1, g1, b1;
         int y = RANDOM256(y2);
         YUVJToRGBReference(y, u, v, &r0, &g0, &b0);
         YUVJToRGB(y, u, v, &r1, &g1, &b1);
-        EXPECT_NEAR(r0, r1, 1);
-        EXPECT_NEAR(g0, g1, 1);
-        EXPECT_NEAR(b0, b1, 1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.709 limited range.
+TEST_F(LibYUVColorTest, TestFullYUVH) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVHToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVHToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.709 full range.
+TEST_F(LibYUVColorTest, TestFullYUVF) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVFToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVFToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.2020 limited range.
+TEST_F(LibYUVColorTest, TestFullYUVU) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVUToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVUToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.2020 full range.
+TEST_F(LibYUVColorTest, TestFullYUVV) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVVToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVVToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, 2);
+        EXPECT_NEAR(b0, b1, ERROR_B);
         ++rh[r1 - r0 + 128];
         ++gh[g1 - g0 + 128];
         ++bh[b1 - b0 + 128];
@@ -517,6 +794,7 @@ TEST_F(LibYUVColorTest, TestFullYUVJ) {
   }
   PrintHistogram(rh, gh, bh);
 }
+#undef FASTSTEP
 
 TEST_F(LibYUVColorTest, TestGreyYUVJ) {
   int r0, g0, b0, r1, g1, b1, r2, g2, b2;
diff --git a/media/libyuv/unit_test/compare_test.cc b/media/libyuv/unit_test/compare_test.cc
index a8ce671d69..c29562cb86 100644
--- a/media/libyuv/unit_test/compare_test.cc
+++ b/media/libyuv/unit_test/compare_test.cc
@@ -18,11 +18,17 @@
 #include "libyuv/cpu_id.h"
 #include "libyuv/video_common.h"
 
+#ifdef ENABLE_ROW_TESTS
+#include "libyuv/compare_row.h" /* For HammingDistance_C */
+#endif
+
 namespace libyuv {
 
 // hash seed of 5381 recommended.
-static uint32 ReferenceHashDjb2(const uint8* src, uint64 count, uint32 seed) {
-  uint32 hash = seed;
+static uint32_t ReferenceHashDjb2(const uint8_t* src,
+                                  uint64_t count,
+                                  uint32_t seed) {
+  uint32_t hash = seed;
   if (count > 0) {
     do {
       hash = hash * 33 + *src++;
@@ -31,16 +37,17 @@ static uint32 ReferenceHashDjb2(const uint8* src, uint64 count, uint32 seed) {
   return hash;
 }
 
-TEST_F(LibYUVBaseTest, Djb2_Test) {
+TEST_F(LibYUVCompareTest, Djb2_Test) {
   const int kMaxTest = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_a, kMaxTest);
   align_buffer_page_end(src_b, kMaxTest);
 
-  const char* fox = "The quick brown fox jumps over the lazy dog"
+  const char* fox =
+      "The quick brown fox jumps over the lazy dog"
       " and feels as if he were in the seventh heaven of typography"
       " together with Hermann Zapf";
-  uint32 foxhash = HashDjb2(reinterpret_cast<const uint8*>(fox), 131, 5381);
-  const uint32 kExpectedFoxHash = 2611006483u;
+  uint32_t foxhash = HashDjb2(reinterpret_cast<const uint8_t*>(fox), 131, 5381);
+  const uint32_t kExpectedFoxHash = 2611006483u;
   EXPECT_EQ(kExpectedFoxHash, foxhash);
 
   for (int i = 0; i < kMaxTest; ++i) {
@@ -48,8 +55,8 @@ TEST_F(LibYUVBaseTest, Djb2_Test) {
     src_b[i] = (fastrand() & 0xff);
   }
   // Compare different buffers. Expect hash is different.
-  uint32 h1 = HashDjb2(src_a, kMaxTest, 5381);
-  uint32 h2 = HashDjb2(src_b, kMaxTest, 5381);
+  uint32_t h1 = HashDjb2(src_a, kMaxTest, 5381);
+  uint32_t h2 = HashDjb2(src_b, kMaxTest, 5381);
   EXPECT_NE(h1, h2);
 
   // Make last half same. Expect hash is different.
@@ -115,15 +122,15 @@ TEST_F(LibYUVBaseTest, Djb2_Test) {
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, BenchmarkDjb2_Opt) {
+TEST_F(LibYUVCompareTest, BenchmarkDjb2_Opt) {
   const int kMaxTest = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_a, kMaxTest);
 
   for (int i = 0; i < kMaxTest; ++i) {
     src_a[i] = i;
   }
-  uint32 h2 = ReferenceHashDjb2(src_a, kMaxTest, 5381);
-  uint32 h1;
+  uint32_t h2 = ReferenceHashDjb2(src_a, kMaxTest, 5381);
+  uint32_t h1;
   for (int i = 0; i < benchmark_iterations_; ++i) {
     h1 = HashDjb2(src_a, kMaxTest, 5381);
   }
@@ -131,14 +138,14 @@ TEST_F(LibYUVBaseTest, BenchmarkDjb2_Opt) {
   free_aligned_buffer_page_end(src_a);
 }
 
-TEST_F(LibYUVBaseTest, BenchmarkDjb2_Unaligned) {
+TEST_F(LibYUVCompareTest, BenchmarkDjb2_Unaligned) {
   const int kMaxTest = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_a, kMaxTest + 1);
   for (int i = 0; i < kMaxTest; ++i) {
     src_a[i + 1] = i;
   }
-  uint32 h2 = ReferenceHashDjb2(src_a + 1, kMaxTest, 5381);
-  uint32 h1;
+  uint32_t h2 = ReferenceHashDjb2(src_a + 1, kMaxTest, 5381);
+  uint32_t h1;
   for (int i = 0; i < benchmark_iterations_; ++i) {
     h1 = HashDjb2(src_a + 1, kMaxTest, 5381);
   }
@@ -146,8 +153,8 @@ TEST_F(LibYUVBaseTest, BenchmarkDjb2_Unaligned) {
   free_aligned_buffer_page_end(src_a);
 }
 
-TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Opt) {
-  uint32 fourcc;
+TEST_F(LibYUVCompareTest, BenchmarkARGBDetect_Opt) {
+  uint32_t fourcc;
   const int kMaxTest = benchmark_width_ * benchmark_height_ * 4;
   align_buffer_page_end(src_a, kMaxTest);
   for (int i = 0; i < kMaxTest; ++i) {
@@ -155,27 +162,27 @@ TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Opt) {
   }
 
   src_a[0] = 0;
-  fourcc = ARGBDetect(src_a, benchmark_width_ * 4,
-                      benchmark_width_, benchmark_height_);
-  EXPECT_EQ(libyuv::FOURCC_BGRA, fourcc);
+  fourcc = ARGBDetect(src_a, benchmark_width_ * 4, benchmark_width_,
+                      benchmark_height_);
+  EXPECT_EQ(static_cast<uint32_t>(libyuv::FOURCC_BGRA), fourcc);
   src_a[0] = 255;
   src_a[3] = 0;
-  fourcc = ARGBDetect(src_a, benchmark_width_ * 4,
-                      benchmark_width_, benchmark_height_);
-  EXPECT_EQ(libyuv::FOURCC_ARGB, fourcc);
+  fourcc = ARGBDetect(src_a, benchmark_width_ * 4, benchmark_width_,
+                      benchmark_height_);
+  EXPECT_EQ(static_cast<uint32_t>(libyuv::FOURCC_ARGB), fourcc);
   src_a[3] = 255;
 
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    fourcc = ARGBDetect(src_a, benchmark_width_ * 4,
-                        benchmark_width_, benchmark_height_);
+    fourcc = ARGBDetect(src_a, benchmark_width_ * 4, benchmark_width_,
+                        benchmark_height_);
   }
-  EXPECT_EQ(0, fourcc);
+  EXPECT_EQ(0u, fourcc);
 
   free_aligned_buffer_page_end(src_a);
 }
 
-TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Unaligned) {
-  uint32 fourcc;
+TEST_F(LibYUVCompareTest, BenchmarkARGBDetect_Unaligned) {
+  uint32_t fourcc;
   const int kMaxTest = benchmark_width_ * benchmark_height_ * 4 + 1;
   align_buffer_page_end(src_a, kMaxTest);
   for (int i = 1; i < kMaxTest; ++i) {
@@ -183,25 +190,243 @@ TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Unaligned) {
   }
 
   src_a[0 + 1] = 0;
-  fourcc = ARGBDetect(src_a + 1, benchmark_width_ * 4,
-                      benchmark_width_, benchmark_height_);
-  EXPECT_EQ(libyuv::FOURCC_BGRA, fourcc);
+  fourcc = ARGBDetect(src_a + 1, benchmark_width_ * 4, benchmark_width_,
+                      benchmark_height_);
+  EXPECT_EQ(static_cast<uint32_t>(libyuv::FOURCC_BGRA), fourcc);
   src_a[0 + 1] = 255;
   src_a[3 + 1] = 0;
-  fourcc = ARGBDetect(src_a + 1, benchmark_width_ * 4,
-                      benchmark_width_, benchmark_height_);
-  EXPECT_EQ(libyuv::FOURCC_ARGB, fourcc);
+  fourcc = ARGBDetect(src_a + 1, benchmark_width_ * 4, benchmark_width_,
+                      benchmark_height_);
+  EXPECT_EQ(static_cast<uint32_t>(libyuv::FOURCC_ARGB), fourcc);
   src_a[3 + 1] = 255;
 
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    fourcc = ARGBDetect(src_a + 1, benchmark_width_ * 4,
-                        benchmark_width_, benchmark_height_);
+    fourcc = ARGBDetect(src_a + 1, benchmark_width_ * 4, benchmark_width_,
+                        benchmark_height_);
+  }
+  EXPECT_EQ(0u, fourcc);
+
+  free_aligned_buffer_page_end(src_a);
+}
+
+#ifdef ENABLE_ROW_TESTS
+TEST_F(LibYUVCompareTest, BenchmarkHammingDistance_Opt) {
+  const int kMaxWidth = 4096 * 3;
+  align_buffer_page_end(src_a, kMaxWidth);
+  align_buffer_page_end(src_b, kMaxWidth);
+  memset(src_a, 0, kMaxWidth);
+  memset(src_b, 0, kMaxWidth);
+
+  // Test known value
+  memcpy(src_a, "test0123test4567", 16);
+  memcpy(src_b, "tick0123tock4567", 16);
+  uint32_t h1 = HammingDistance_C(src_a, src_b, 16);
+  EXPECT_EQ(16u, h1);
+
+  // Test C vs OPT on random buffer
+  MemRandomize(src_a, kMaxWidth);
+  MemRandomize(src_b, kMaxWidth);
+
+  uint32_t h0 = HammingDistance_C(src_a, src_b, kMaxWidth);
+
+  int count =
+      benchmark_iterations_ *
+      ((benchmark_width_ * benchmark_height_ + kMaxWidth - 1) / kMaxWidth);
+  for (int i = 0; i < count; ++i) {
+#if defined(HAS_HAMMINGDISTANCE_NEON)
+    h1 = HammingDistance_NEON(src_a, src_b, kMaxWidth);
+#elif defined(HAS_HAMMINGDISTANCE_AVX2)
+    int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+    if (has_avx2) {
+      h1 = HammingDistance_AVX2(src_a, src_b, kMaxWidth);
+    } else {
+      int has_sse42 = TestCpuFlag(kCpuHasSSE42);
+      if (has_sse42) {
+        h1 = HammingDistance_SSE42(src_a, src_b, kMaxWidth);
+      } else {
+        int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+        if (has_ssse3) {
+          h1 = HammingDistance_SSSE3(src_a, src_b, kMaxWidth);
+        } else {
+          h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+        }
+      }
+    }
+#elif defined(HAS_HAMMINGDISTANCE_SSE42)
+    int has_sse42 = TestCpuFlag(kCpuHasSSE42);
+    if (has_sse42) {
+      h1 = HammingDistance_SSE42(src_a, src_b, kMaxWidth);
+    } else {
+      h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+    }
+#else
+    h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+#endif
+  }
+  EXPECT_EQ(h0, h1);
+
+  free_aligned_buffer_page_end(src_a);
+  free_aligned_buffer_page_end(src_b);
+}
+
+TEST_F(LibYUVCompareTest, BenchmarkHammingDistance_C) {
+  const int kMaxWidth = 4096 * 3;
+  align_buffer_page_end(src_a, kMaxWidth);
+  align_buffer_page_end(src_b, kMaxWidth);
+  memset(src_a, 0, kMaxWidth);
+  memset(src_b, 0, kMaxWidth);
+
+  // Test known value
+  memcpy(src_a, "test0123test4567", 16);
+  memcpy(src_b, "tick0123tock4567", 16);
+  uint32_t h1 = HammingDistance_C(src_a, src_b, 16);
+  EXPECT_EQ(16u, h1);
+
+  // Test C vs OPT on random buffer
+  MemRandomize(src_a, kMaxWidth);
+  MemRandomize(src_b, kMaxWidth);
+
+  uint32_t h0 = HammingDistance_C(src_a, src_b, kMaxWidth);
+
+  int count =
+      benchmark_iterations_ *
+      ((benchmark_width_ * benchmark_height_ + kMaxWidth - 1) / kMaxWidth);
+  for (int i = 0; i < count; ++i) {
+    h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+  }
+
+  EXPECT_EQ(h0, h1);
+
+  free_aligned_buffer_page_end(src_a);
+  free_aligned_buffer_page_end(src_b);
+}
+
+TEST_F(LibYUVCompareTest, BenchmarkHammingDistance) {
+  const int kMaxWidth = 4096 * 3;
+  align_buffer_page_end(src_a, kMaxWidth);
+  align_buffer_page_end(src_b, kMaxWidth);
+  memset(src_a, 0, kMaxWidth);
+  memset(src_b, 0, kMaxWidth);
+
+  memcpy(src_a, "test0123test4567", 16);
+  memcpy(src_b, "tick0123tock4567", 16);
+  uint64_t h1 = ComputeHammingDistance(src_a, src_b, 16);
+  EXPECT_EQ(16u, h1);
+
+  // Test C vs OPT on random buffer
+  MemRandomize(src_a, kMaxWidth);
+  MemRandomize(src_b, kMaxWidth);
+
+  uint32_t h0 = HammingDistance_C(src_a, src_b, kMaxWidth);
+
+  int count =
+      benchmark_iterations_ *
+      ((benchmark_width_ * benchmark_height_ + kMaxWidth - 1) / kMaxWidth);
+  for (int i = 0; i < count; ++i) {
+    h1 = ComputeHammingDistance(src_a, src_b, kMaxWidth);
+  }
+
+  EXPECT_EQ(h0, h1);
+
+  free_aligned_buffer_page_end(src_a);
+  free_aligned_buffer_page_end(src_b);
+}
+
+// Tests low levels match reference C for specified size.
+// The opt implementations have size limitations
+// For NEON the counters are 16 bit so the shorts overflow after 65536 bytes.
+// So doing one less iteration of the loop is the maximum.
+#if defined(HAS_HAMMINGDISTANCE_NEON)
+static const int kMaxOptCount = 65536 - 32;  // 65504
+#else
+static const int kMaxOptCount = (1 << (32 - 3)) - 64;  // 536870848
+#endif
+
+TEST_F(LibYUVCompareTest, TestHammingDistance_Opt) {
+  uint32_t h1 = 0;
+  const int kMaxWidth = (benchmark_width_ * benchmark_height_ + 63) & ~63;
+  align_buffer_page_end(src_a, kMaxWidth);
+  align_buffer_page_end(src_b, kMaxWidth);
+  memset(src_a, 255u, kMaxWidth);
+  memset(src_b, 0u, kMaxWidth);
+
+  uint64_t h0 = ComputeHammingDistance(src_a, src_b, kMaxWidth);
+  EXPECT_EQ(kMaxWidth * 8ULL, h0);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+#if defined(HAS_HAMMINGDISTANCE_NEON)
+    h1 = HammingDistance_NEON(src_a, src_b, kMaxWidth);
+#elif defined(HAS_HAMMINGDISTANCE_AVX2)
+    int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+    if (has_avx2) {
+      h1 = HammingDistance_AVX2(src_a, src_b, kMaxWidth);
+    } else {
+      int has_sse42 = TestCpuFlag(kCpuHasSSE42);
+      if (has_sse42) {
+        h1 = HammingDistance_SSE42(src_a, src_b, kMaxWidth);
+      } else {
+        int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+        if (has_ssse3) {
+          h1 = HammingDistance_SSSE3(src_a, src_b, kMaxWidth);
+        } else {
+          h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+        }
+      }
+    }
+#elif defined(HAS_HAMMINGDISTANCE_SSE42)
+    int has_sse42 = TestCpuFlag(kCpuHasSSE42);
+    if (has_sse42) {
+      h1 = HammingDistance_SSE42(src_a, src_b, kMaxWidth);
+    } else {
+      h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+    }
+#else
+    h1 = HammingDistance_C(src_a, src_b, kMaxWidth);
+#endif
+  }
+
+  // A large count will cause the low level to potentially overflow so the
+  // result can not be expected to be correct.
+  // TODO(fbarchard): Consider expecting the low 16 bits to match.
+  if (kMaxWidth <= kMaxOptCount) {
+    EXPECT_EQ(kMaxWidth * 8U, h1);
+  } else {
+    if (kMaxWidth * 8ULL != static_cast<uint64_t>(h1)) {
+      printf(
+          "warning - HammingDistance_Opt %u does not match %llu "
+          "but length of %u is longer than guaranteed.\n",
+          h1, kMaxWidth * 8ULL, kMaxWidth);
+    } else {
+      printf(
+          "warning - HammingDistance_Opt %u matches but length of %u "
+          "is longer than guaranteed.\n",
+          h1, kMaxWidth);
+    }
   }
-  EXPECT_EQ(0, fourcc);
 
   free_aligned_buffer_page_end(src_a);
+  free_aligned_buffer_page_end(src_b);
 }
-TEST_F(LibYUVBaseTest, BenchmarkSumSquareError_Opt) {
+#endif  // ENABLE_ROW_TESTS
+
+TEST_F(LibYUVCompareTest, TestHammingDistance) {
+  align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_);
+  memset(src_a, 255u, benchmark_width_ * benchmark_height_);
+  memset(src_b, 0, benchmark_width_ * benchmark_height_);
+
+  uint64_t h1 = 0;
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    h1 = ComputeHammingDistance(src_a, src_b,
+                                benchmark_width_ * benchmark_height_);
+  }
+  EXPECT_EQ(benchmark_width_ * benchmark_height_ * 8ULL, h1);
+
+  free_aligned_buffer_page_end(src_a);
+  free_aligned_buffer_page_end(src_b);
+}
+
+TEST_F(LibYUVCompareTest, BenchmarkSumSquareError_Opt) {
   const int kMaxWidth = 4096 * 3;
   align_buffer_page_end(src_a, kMaxWidth);
   align_buffer_page_end(src_b, kMaxWidth);
@@ -210,7 +435,7 @@ TEST_F(LibYUVBaseTest, BenchmarkSumSquareError_Opt) {
 
   memcpy(src_a, "test0123test4567", 16);
   memcpy(src_b, "tick0123tock4567", 16);
-  uint64 h1 = ComputeSumSquareError(src_a, src_b, 16);
+  uint64_t h1 = ComputeSumSquareError(src_a, src_b, 16);
   EXPECT_EQ(790u, h1);
 
   for (int i = 0; i < kMaxWidth; ++i) {
@@ -220,40 +445,41 @@ TEST_F(LibYUVBaseTest, BenchmarkSumSquareError_Opt) {
   memset(src_a, 0, kMaxWidth);
   memset(src_b, 0, kMaxWidth);
 
-  int count = benchmark_iterations_ *
-    ((benchmark_width_ * benchmark_height_ + kMaxWidth - 1) / kMaxWidth);
+  int count =
+      benchmark_iterations_ *
+      ((benchmark_width_ * benchmark_height_ + kMaxWidth - 1) / kMaxWidth);
   for (int i = 0; i < count; ++i) {
     h1 = ComputeSumSquareError(src_a, src_b, kMaxWidth);
   }
 
-  EXPECT_EQ(0, h1);
+  EXPECT_EQ(0u, h1);
 
   free_aligned_buffer_page_end(src_a);
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, SumSquareError) {
+TEST_F(LibYUVCompareTest, SumSquareError) {
   const int kMaxWidth = 4096 * 3;
   align_buffer_page_end(src_a, kMaxWidth);
   align_buffer_page_end(src_b, kMaxWidth);
   memset(src_a, 0, kMaxWidth);
   memset(src_b, 0, kMaxWidth);
 
-  uint64 err;
+  uint64_t err;
   err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
 
-  EXPECT_EQ(0, err);
+  EXPECT_EQ(0u, err);
 
   memset(src_a, 1, kMaxWidth);
   err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
 
-  EXPECT_EQ(err, kMaxWidth);
+  EXPECT_EQ(static_cast<int>(err), kMaxWidth);
 
   memset(src_a, 190, kMaxWidth);
   memset(src_b, 193, kMaxWidth);
   err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
 
-  EXPECT_EQ(kMaxWidth * 3 * 3, err);
+  EXPECT_EQ(static_cast<int>(err), kMaxWidth * 3 * 3);
 
   for (int i = 0; i < kMaxWidth; ++i) {
     src_a[i] = (fastrand() & 0xff);
@@ -261,10 +487,10 @@ TEST_F(LibYUVBaseTest, SumSquareError) {
   }
 
   MaskCpuFlags(disable_cpu_flags_);
-  uint64 c_err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
+  uint64_t c_err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
 
   MaskCpuFlags(benchmark_cpu_info_);
-  uint64 opt_err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
+  uint64_t opt_err = ComputeSumSquareError(src_a, src_b, kMaxWidth);
 
   EXPECT_EQ(c_err, opt_err);
 
@@ -272,7 +498,7 @@ TEST_F(LibYUVBaseTest, SumSquareError) {
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, BenchmarkPsnr_Opt) {
+TEST_F(LibYUVCompareTest, BenchmarkPsnr_Opt) {
   align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_);
   align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_);
   for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
@@ -283,10 +509,10 @@ TEST_F(LibYUVBaseTest, BenchmarkPsnr_Opt) {
   MaskCpuFlags(benchmark_cpu_info_);
 
   double opt_time = get_time();
-  for (int i = 0; i < benchmark_iterations_; ++i)
-    CalcFramePsnr(src_a, benchmark_width_,
-                  src_b, benchmark_width_,
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    CalcFramePsnr(src_a, benchmark_width_, src_b, benchmark_width_,
                   benchmark_width_, benchmark_height_);
+  }
 
   opt_time = (get_time() - opt_time) / benchmark_iterations_;
   printf("BenchmarkPsnr_Opt - %8.2f us opt\n", opt_time * 1e6);
@@ -297,7 +523,7 @@ TEST_F(LibYUVBaseTest, BenchmarkPsnr_Opt) {
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, BenchmarkPsnr_Unaligned) {
+TEST_F(LibYUVCompareTest, BenchmarkPsnr_Unaligned) {
   align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_ + 1);
   align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_);
   for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
@@ -308,10 +534,10 @@ TEST_F(LibYUVBaseTest, BenchmarkPsnr_Unaligned) {
   MaskCpuFlags(benchmark_cpu_info_);
 
   double opt_time = get_time();
-  for (int i = 0; i < benchmark_iterations_; ++i)
-    CalcFramePsnr(src_a + 1, benchmark_width_,
-                  src_b, benchmark_width_,
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    CalcFramePsnr(src_a + 1, benchmark_width_, src_b, benchmark_width_,
                   benchmark_width_, benchmark_height_);
+  }
 
   opt_time = (get_time() - opt_time) / benchmark_iterations_;
   printf("BenchmarkPsnr_Opt - %8.2f us opt\n", opt_time * 1e6);
@@ -322,7 +548,7 @@ TEST_F(LibYUVBaseTest, BenchmarkPsnr_Unaligned) {
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, Psnr) {
+TEST_F(LibYUVCompareTest, Psnr) {
   const int kSrcWidth = benchmark_width_;
   const int kSrcHeight = benchmark_height_;
   const int b = 128;
@@ -335,24 +561,24 @@ TEST_F(LibYUVBaseTest, Psnr) {
 
   double err;
   err = CalcFramePsnr(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   EXPECT_EQ(err, kMaxPsnr);
 
   memset(src_a, 255, kSrcPlaneSize);
 
   err = CalcFramePsnr(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   EXPECT_EQ(err, 0.0);
 
   memset(src_a, 1, kSrcPlaneSize);
 
   err = CalcFramePsnr(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   EXPECT_GT(err, 48.0);
   EXPECT_LT(err, 49.0);
@@ -362,8 +588,8 @@ TEST_F(LibYUVBaseTest, Psnr) {
   }
 
   err = CalcFramePsnr(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   EXPECT_GT(err, 2.0);
   if (kSrcWidth * kSrcHeight >= 256) {
@@ -384,14 +610,14 @@ TEST_F(LibYUVBaseTest, Psnr) {
   double c_err, opt_err;
 
   c_err = CalcFramePsnr(src_a + kSrcStride * b + b, kSrcStride,
-                        src_b + kSrcStride * b + b, kSrcStride,
-                        kSrcWidth, kSrcHeight);
+                        src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                        kSrcHeight);
 
   MaskCpuFlags(benchmark_cpu_info_);
 
   opt_err = CalcFramePsnr(src_a + kSrcStride * b + b, kSrcStride,
-                          src_b + kSrcStride * b + b, kSrcStride,
-                          kSrcWidth, kSrcHeight);
+                          src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                          kSrcHeight);
 
   EXPECT_EQ(opt_err, c_err);
 
@@ -399,7 +625,7 @@ TEST_F(LibYUVBaseTest, Psnr) {
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, DISABLED_BenchmarkSsim_Opt) {
+TEST_F(LibYUVCompareTest, DISABLED_BenchmarkSsim_Opt) {
   align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_);
   align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_);
   for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
@@ -410,10 +636,10 @@ TEST_F(LibYUVBaseTest, DISABLED_BenchmarkSsim_Opt) {
   MaskCpuFlags(benchmark_cpu_info_);
 
   double opt_time = get_time();
-  for (int i = 0; i < benchmark_iterations_; ++i)
-    CalcFrameSsim(src_a, benchmark_width_,
-                  src_b, benchmark_width_,
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    CalcFrameSsim(src_a, benchmark_width_, src_b, benchmark_width_,
                   benchmark_width_, benchmark_height_);
+  }
 
   opt_time = (get_time() - opt_time) / benchmark_iterations_;
   printf("BenchmarkSsim_Opt - %8.2f us opt\n", opt_time * 1e6);
@@ -424,7 +650,7 @@ TEST_F(LibYUVBaseTest, DISABLED_BenchmarkSsim_Opt) {
   free_aligned_buffer_page_end(src_b);
 }
 
-TEST_F(LibYUVBaseTest, Ssim) {
+TEST_F(LibYUVCompareTest, Ssim) {
   const int kSrcWidth = benchmark_width_;
   const int kSrcHeight = benchmark_height_;
   const int b = 128;
@@ -435,14 +661,14 @@ TEST_F(LibYUVBaseTest, Ssim) {
   memset(src_a, 0, kSrcPlaneSize);
   memset(src_b, 0, kSrcPlaneSize);
 
-  if (kSrcWidth <=8 || kSrcHeight <= 8) {
+  if (kSrcWidth <= 8 || kSrcHeight <= 8) {
     printf("warning - Ssim size too small.  Testing function executes.\n");
   }
 
   double err;
   err = CalcFrameSsim(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   if (kSrcWidth > 8 && kSrcHeight > 8) {
     EXPECT_EQ(err, 1.0);
@@ -451,8 +677,8 @@ TEST_F(LibYUVBaseTest, Ssim) {
   memset(src_a, 255, kSrcPlaneSize);
 
   err = CalcFrameSsim(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   if (kSrcWidth > 8 && kSrcHeight > 8) {
     EXPECT_LT(err, 0.0001);
@@ -461,8 +687,8 @@ TEST_F(LibYUVBaseTest, Ssim) {
   memset(src_a, 1, kSrcPlaneSize);
 
   err = CalcFrameSsim(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   if (kSrcWidth > 8 && kSrcHeight > 8) {
     EXPECT_GT(err, 0.0001);
@@ -474,8 +700,8 @@ TEST_F(LibYUVBaseTest, Ssim) {
   }
 
   err = CalcFrameSsim(src_a + kSrcStride * b + b, kSrcStride,
-                      src_b + kSrcStride * b + b, kSrcStride,
-                      kSrcWidth, kSrcHeight);
+                      src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                      kSrcHeight);
 
   if (kSrcWidth > 8 && kSrcHeight > 8) {
     EXPECT_GT(err, 0.0);
@@ -493,14 +719,14 @@ TEST_F(LibYUVBaseTest, Ssim) {
   double c_err, opt_err;
 
   c_err = CalcFrameSsim(src_a + kSrcStride * b + b, kSrcStride,
-                        src_b + kSrcStride * b + b, kSrcStride,
-                        kSrcWidth, kSrcHeight);
+                        src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                        kSrcHeight);
 
   MaskCpuFlags(benchmark_cpu_info_);
 
   opt_err = CalcFrameSsim(src_a + kSrcStride * b + b, kSrcStride,
-                          src_b + kSrcStride * b + b, kSrcStride,
-                          kSrcWidth, kSrcHeight);
+                          src_b + kSrcStride * b + b, kSrcStride, kSrcWidth,
+                          kSrcHeight);
 
   if (kSrcWidth > 8 && kSrcHeight > 8) {
     EXPECT_EQ(opt_err, c_err);
diff --git a/media/libyuv/unit_test/convert_test.cc b/media/libyuv/unit_test/convert_test.cc
index 56a2bfd82d..f94a7d31cc 100644
--- a/media/libyuv/unit_test/convert_test.cc
+++ b/media/libyuv/unit_test/convert_test.cc
@@ -8,6 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <assert.h>
 #include <stdlib.h>
 #include <time.h>
 
@@ -21,1161 +22,1769 @@
 #ifdef HAVE_JPEG
 #include "libyuv/mjpeg_decoder.h"
 #endif
+#include "../unit_test/unit_test.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
 #include "libyuv/video_common.h"
-#include "../unit_test/unit_test.h"
+
+#ifdef ENABLE_ROW_TESTS
+#include "libyuv/row.h" /* For ARGBToAR30Row_AVX2 */
+#endif
+
+// Some functions fail on big endian. Enable these tests on all cpus except
+// PowerPC, but they are not optimized so disabled by default.
+#if !defined(DISABLE_SLOW_TESTS) && !defined(__powerpc__)
+#define LITTLE_ENDIAN_ONLY_TEST 1
+#endif
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
 
 namespace libyuv {
 
-#define SUBSAMPLE(v, a) ((((v) + (a) - 1)) / (a))
-
-#define TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,           \
-                       FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF)   \
-TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {                 \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u,                                                 \
-                        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *                     \
-                        SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF);              \
-  align_buffer_page_end(src_v,                                                 \
-                        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *                     \
-                        SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF);              \
-  align_buffer_page_end(dst_y_c, kWidth * kHeight);                            \
-  align_buffer_page_end(dst_u_c,                                               \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_v_c,                                               \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_y_opt, kWidth * kHeight);                          \
-  align_buffer_page_end(dst_u_opt,                                             \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_v_opt,                                             \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  for (int i = 0; i < kHeight; ++i)                                            \
-    for (int j = 0; j < kWidth; ++j)                                           \
-      src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                       \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {                \
-    for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {               \
-      src_u[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =                \
-          (fastrand() & 0xff);                                                 \
-      src_v[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =                \
-          (fastrand() & 0xff);                                                 \
-    }                                                                          \
-  }                                                                            \
-  memset(dst_y_c, 1, kWidth * kHeight);                                        \
-  memset(dst_u_c, 2, SUBSAMPLE(kWidth, SUBSAMP_X) *                            \
-                     SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_v_c, 3, SUBSAMPLE(kWidth, SUBSAMP_X) *                            \
-                     SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_y_opt, 101, kWidth * kHeight);                                    \
-  memset(dst_u_opt, 102, SUBSAMPLE(kWidth, SUBSAMP_X) *                        \
-                         SUBSAMPLE(kHeight, SUBSAMP_Y));                       \
-  memset(dst_v_opt, 103, SUBSAMPLE(kWidth, SUBSAMP_X) *                        \
-                         SUBSAMPLE(kHeight, SUBSAMP_Y));                       \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  SRC_FMT_PLANAR##To##FMT_PLANAR(src_y + OFF, kWidth,                          \
-                                 src_u + OFF,                                  \
-                                 SUBSAMPLE(kWidth, SRC_SUBSAMP_X),             \
-                                 src_v + OFF,                                  \
-                                 SUBSAMPLE(kWidth, SRC_SUBSAMP_X),             \
-                                 dst_y_c, kWidth,                              \
-                                 dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X),        \
-                                 dst_v_c, SUBSAMPLE(kWidth, SUBSAMP_X),        \
-                                 kWidth, NEG kHeight);                         \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    SRC_FMT_PLANAR##To##FMT_PLANAR(src_y + OFF, kWidth,                        \
-                                   src_u + OFF,                                \
-                                       SUBSAMPLE(kWidth, SRC_SUBSAMP_X),       \
-                                   src_v + OFF,                                \
-                                       SUBSAMPLE(kWidth, SRC_SUBSAMP_X),       \
-                                   dst_y_opt, kWidth,                          \
-                                   dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X),    \
-                                   dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X),    \
-                                   kWidth, NEG kHeight);                       \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kHeight; ++i) {                                          \
-    for (int j = 0; j < kWidth; ++j) {                                         \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_y_c[i * kWidth + j]) -                      \
-              static_cast<int>(dst_y_opt[i * kWidth + j]));                    \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_EQ(0, max_diff);                                                      \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                   \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_u_c[i *                                     \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]) -            \
-              static_cast<int>(dst_u_opt[i *                                   \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]));            \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 3);                                                      \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                   \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_v_c[i *                                     \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]) -            \
-              static_cast<int>(dst_v_opt[i *                                   \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]));            \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 3);                                                      \
-  free_aligned_buffer_page_end(dst_y_c);                                       \
-  free_aligned_buffer_page_end(dst_u_c);                                       \
-  free_aligned_buffer_page_end(dst_v_c);                                       \
-  free_aligned_buffer_page_end(dst_y_opt);                                     \
-  free_aligned_buffer_page_end(dst_u_opt);                                     \
-  free_aligned_buffer_page_end(dst_v_opt);                                     \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
+// Alias to copy pixels as is
+#define AR30ToAR30 ARGBCopy
+#define ABGRToABGR ARGBCopy
+
+// subsample amount uses a divide.
+#define SUBSAMPLE(v, a) ((((v) + (a)-1)) / (a))
+
+// Planar test
+
+#define TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,         \
+                       SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,             \
+                       DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF,      \
+                       SRC_DEPTH)                                             \
+  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {              \
+    static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");       \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");       \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                   \
+                  "SRC_SUBSAMP_X unsupported");                               \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                   \
+                  "SRC_SUBSAMP_Y unsupported");                               \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                   \
+                  "DST_SUBSAMP_X unsupported");                               \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                   \
+                  "DST_SUBSAMP_Y unsupported");                               \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);               \
+    const int kSrcHalfHeight = SUBSAMPLE(kHeight, SRC_SUBSAMP_Y);             \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);               \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);             \
+    align_buffer_page_end(src_y, kWidth* kHeight* SRC_BPC + OFF);             \
+    align_buffer_page_end(src_u,                                              \
+                          kSrcHalfWidth* kSrcHalfHeight* SRC_BPC + OFF);      \
+    align_buffer_page_end(src_v,                                              \
+                          kSrcHalfWidth* kSrcHalfHeight* SRC_BPC + OFF);      \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                 \
+    align_buffer_page_end(dst_u_c, kDstHalfWidth* kDstHalfHeight* DST_BPC);   \
+    align_buffer_page_end(dst_v_c, kDstHalfWidth* kDstHalfHeight* DST_BPC);   \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);               \
+    align_buffer_page_end(dst_u_opt, kDstHalfWidth* kDstHalfHeight* DST_BPC); \
+    align_buffer_page_end(dst_v_opt, kDstHalfWidth* kDstHalfHeight* DST_BPC); \
+    MemRandomize(src_y + OFF, kWidth * kHeight * SRC_BPC);                    \
+    MemRandomize(src_u + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    MemRandomize(src_v + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                   \
+    SRC_T* src_u_p = reinterpret_cast<SRC_T*>(src_u + OFF);                   \
+    SRC_T* src_v_p = reinterpret_cast<SRC_T*>(src_v + OFF);                   \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      src_y_p[i] = src_y_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
+    for (int i = 0; i < kSrcHalfWidth * kSrcHalfHeight; ++i) {                \
+      src_u_p[i] = src_u_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+      src_v_p[i] = src_v_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                             \
+    memset(dst_u_c, 2, kDstHalfWidth* kDstHalfHeight* DST_BPC);               \
+    memset(dst_v_c, 3, kDstHalfWidth* kDstHalfHeight* DST_BPC);               \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                         \
+    memset(dst_u_opt, 102, kDstHalfWidth* kDstHalfHeight* DST_BPC);           \
+    memset(dst_v_opt, 103, kDstHalfWidth* kDstHalfHeight* DST_BPC);           \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    SRC_FMT_PLANAR##To##FMT_PLANAR(                                           \
+        src_y_p, kWidth, src_u_p, kSrcHalfWidth, src_v_p, kSrcHalfWidth,      \
+        reinterpret_cast<DST_T*>(dst_y_c), kWidth,                            \
+        reinterpret_cast<DST_T*>(dst_u_c), kDstHalfWidth,                     \
+        reinterpret_cast<DST_T*>(dst_v_c), kDstHalfWidth, kWidth,             \
+        NEG kHeight);                                                         \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      SRC_FMT_PLANAR##To##FMT_PLANAR(                                         \
+          src_y_p, kWidth, src_u_p, kSrcHalfWidth, src_v_p, kSrcHalfWidth,    \
+          reinterpret_cast<DST_T*>(dst_y_opt), kWidth,                        \
+          reinterpret_cast<DST_T*>(dst_u_opt), kDstHalfWidth,                 \
+          reinterpret_cast<DST_T*>(dst_v_opt), kDstHalfWidth, kWidth,         \
+          NEG kHeight);                                                       \
+    }                                                                         \
+    for (int i = 0; i < kHeight * kWidth * DST_BPC; ++i) {                    \
+      EXPECT_EQ(dst_y_c[i], dst_y_opt[i]);                                    \
+    }                                                                         \
+    for (int i = 0; i < kDstHalfWidth * kDstHalfHeight * DST_BPC; ++i) {      \
+      EXPECT_EQ(dst_u_c[i], dst_u_opt[i]);                                    \
+      EXPECT_EQ(dst_v_c[i], dst_v_opt[i]);                                    \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_u_c);                                    \
+    free_aligned_buffer_page_end(dst_v_c);                                    \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_u_opt);                                  \
+    free_aligned_buffer_page_end(dst_v_opt);                                  \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_u);                                      \
+    free_aligned_buffer_page_end(src_v);                                      \
+  }
+
+#define TESTPLANARTOP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,           \
+                      SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,               \
+                      DST_SUBSAMP_X, DST_SUBSAMP_Y, SRC_DEPTH)                 \
+  TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
+                 FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
+                 benchmark_width_ + 1, _Any, +, 0, SRC_DEPTH)                  \
+  TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
+                 FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
+                 benchmark_width_, _Unaligned, +, 2, SRC_DEPTH)                \
+  TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
+                 FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
+                 benchmark_width_, _Invert, -, 0, SRC_DEPTH)                   \
+  TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
+                 FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
+                 benchmark_width_, _Opt, +, 0, SRC_DEPTH)
+
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I420, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I444, uint8_t, 1, 1, 1, I420, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I422, uint8_t, 1, 2, 1, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I444, uint8_t, 1, 1, 1, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I420Mirror, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I422, uint8_t, 1, 2, 1, 8)
+TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I444, uint8_t, 1, 1, 1, 8)
+TESTPLANARTOP(I444, uint8_t, 1, 1, 1, I444, uint8_t, 1, 1, 1, 8)
+TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I010, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I012, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(H010, uint16_t, 2, 2, 2, H010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(H010, uint16_t, 2, 2, 2, H420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H010, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H012, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I410, uint16_t, 2, 1, 1, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I410, uint16_t, 2, 1, 1, 10)
+TESTPLANARTOP(I012, uint16_t, 2, 2, 2, I412, uint16_t, 2, 1, 1, 12)
+TESTPLANARTOP(I212, uint16_t, 2, 2, 1, I412, uint16_t, 2, 1, 1, 12)
+TESTPLANARTOP(I410, uint16_t, 2, 1, 1, I010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(I412, uint16_t, 2, 1, 1, I012, uint16_t, 2, 2, 2, 12)
+TESTPLANARTOP(I212, uint16_t, 2, 2, 1, I012, uint16_t, 2, 2, 2, 12)
+TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I422, uint8_t, 1, 2, 1, 10)
+TESTPLANARTOP(I410, uint16_t, 2, 1, 1, I420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(I410, uint16_t, 2, 1, 1, I444, uint8_t, 1, 1, 1, 10)
+TESTPLANARTOP(I012, uint16_t, 2, 2, 2, I420, uint8_t, 1, 2, 2, 12)
+TESTPLANARTOP(I212, uint16_t, 2, 2, 1, I422, uint8_t, 1, 2, 1, 12)
+TESTPLANARTOP(I412, uint16_t, 2, 1, 1, I444, uint8_t, 1, 1, 1, 12)
+
+// Test Android 420 to I420
+#define TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X,          \
+                        SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                        W1280, N, NEG, OFF, PN, OFF_U, OFF_V)                 \
+  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##To##PN##N) {      \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kSizeUV =                                                       \
+        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                      \
+    align_buffer_page_end(src_uv,                                             \
+                          kSizeUV*((PIXEL_STRIDE == 3) ? 3 : 2) + OFF);       \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight);                          \
+    align_buffer_page_end(dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X) *             \
+                                       SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    align_buffer_page_end(dst_v_c, SUBSAMPLE(kWidth, SUBSAMP_X) *             \
+                                       SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                        \
+    align_buffer_page_end(dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *           \
+                                         SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    align_buffer_page_end(dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *           \
+                                         SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    uint8_t* src_u = src_uv + OFF_U;                                          \
+    uint8_t* src_v = src_uv + (PIXEL_STRIDE == 1 ? kSizeUV : OFF_V);          \
+    int src_stride_uv = SUBSAMPLE(kWidth, SUBSAMP_X) * PIXEL_STRIDE;          \
+    for (int i = 0; i < kHeight; ++i)                                         \
+      for (int j = 0; j < kWidth; ++j)                                        \
+        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                    \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {             \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {            \
+        src_u[(i * src_stride_uv) + j * PIXEL_STRIDE + OFF] =                 \
+            (fastrand() & 0xff);                                              \
+        src_v[(i * src_stride_uv) + j * PIXEL_STRIDE + OFF] =                 \
+            (fastrand() & 0xff);                                              \
+      }                                                                       \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight);                                      \
+    memset(dst_u_c, 2,                                                        \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_v_c, 3,                                                        \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_y_opt, 101, kWidth* kHeight);                                  \
+    memset(dst_u_opt, 102,                                                    \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_v_opt, 103,                                                    \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    SRC_FMT_PLANAR##To##FMT_PLANAR(                                           \
+        src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X),   \
+        src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), PIXEL_STRIDE, dst_y_c, \
+        kWidth, dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X), dst_v_c,               \
+        SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight);                   \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      SRC_FMT_PLANAR##To##FMT_PLANAR(                                         \
+          src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), \
+          src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), PIXEL_STRIDE,        \
+          dst_y_opt, kWidth, dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X),         \
+          dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight);      \
+    }                                                                         \
+    for (int i = 0; i < kHeight; ++i) {                                       \
+      for (int j = 0; j < kWidth; ++j) {                                      \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);        \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
+        EXPECT_EQ(dst_u_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_u_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
+        EXPECT_EQ(dst_v_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_v_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
+      }                                                                       \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_u_c);                                    \
+    free_aligned_buffer_page_end(dst_v_c);                                    \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_u_opt);                                  \
+    free_aligned_buffer_page_end(dst_v_opt);                                  \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_uv);                                     \
+  }
+
+#define TESTAPLANARTOP(SRC_FMT_PLANAR, PN, PIXEL_STRIDE, OFF_U, OFF_V,         \
+                       SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X,    \
+                       SUBSAMP_Y)                                              \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_ + 1,      \
+                  _Any, +, 0, PN, OFF_U, OFF_V)                                \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_,          \
+                  _Unaligned, +, 2, PN, OFF_U, OFF_V)                          \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, \
+                  -, 0, PN, OFF_U, OFF_V)                                      \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, \
+                  0, PN, OFF_U, OFF_V)
+
+TESTAPLANARTOP(Android420, I420, 1, 0, 0, 2, 2, I420, 2, 2)
+TESTAPLANARTOP(Android420, NV12, 2, 0, 1, 2, 2, I420, 2, 2)
+TESTAPLANARTOP(Android420, NV21, 2, 1, 0, 2, 2, I420, 2, 2)
+#undef TESTAPLANARTOP
+#undef TESTAPLANARTOPI
+
+// wrapper to keep API the same
+int I400ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* /* src_u */,
+               int /* src_stride_u */,
+               const uint8_t* /* src_v */,
+               int /* src_stride_v */,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  return I400ToNV21(src_y, src_stride_y, dst_y, dst_stride_y, dst_vu,
+                    dst_stride_vu, width, height);
 }
 
-#define TESTPLANARTOP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,            \
-                      FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y)                        \
-    TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,               \
-                   FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                           \
-                   benchmark_width_ - 4, _Any, +, 0)                           \
-    TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,               \
-                   FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                           \
-                   benchmark_width_, _Unaligned, +, 1)                         \
-    TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,               \
-                   FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                           \
-                   benchmark_width_, _Invert, -, 0)                            \
-    TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,               \
-                   FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                           \
-                   benchmark_width_, _Opt, +, 0)
-
-TESTPLANARTOP(I420, 2, 2, I420, 2, 2)
-TESTPLANARTOP(I422, 2, 1, I420, 2, 2)
-TESTPLANARTOP(I444, 1, 1, I420, 2, 2)
-TESTPLANARTOP(I411, 4, 1, I420, 2, 2)
-TESTPLANARTOP(I420, 2, 2, I422, 2, 1)
-TESTPLANARTOP(I420, 2, 2, I444, 1, 1)
-TESTPLANARTOP(I420, 2, 2, I411, 4, 1)
-TESTPLANARTOP(I420, 2, 2, I420Mirror, 2, 2)
-TESTPLANARTOP(I422, 2, 1, I422, 2, 1)
-TESTPLANARTOP(I444, 1, 1, I444, 1, 1)
-
-#define TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,          \
-                       FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF)   \
-TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {                 \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u,                                                 \
-                        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *                     \
-                        SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF);              \
-  align_buffer_page_end(src_v,                                                 \
-                        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *                     \
-                        SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF);              \
-  align_buffer_page_end(dst_y_c, kWidth * kHeight);                            \
-  align_buffer_page_end(dst_uv_c, SUBSAMPLE(kWidth * 2, SUBSAMP_X) *           \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_y_opt, kWidth * kHeight);                          \
-  align_buffer_page_end(dst_uv_opt, SUBSAMPLE(kWidth * 2, SUBSAMP_X) *         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  for (int i = 0; i < kHeight; ++i)                                            \
-    for (int j = 0; j < kWidth; ++j)                                           \
-      src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                       \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {                \
-    for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {               \
-      src_u[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =                \
-          (fastrand() & 0xff);                                                 \
-      src_v[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =                \
-          (fastrand() & 0xff);                                                 \
+#define TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,        \
+                        SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,            \
+                        DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF,     \
+                        SRC_DEPTH)                                            \
+  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {              \
+    static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");       \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");       \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                   \
+                  "SRC_SUBSAMP_X unsupported");                               \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                   \
+                  "SRC_SUBSAMP_Y unsupported");                               \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                   \
+                  "DST_SUBSAMP_X unsupported");                               \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                   \
+                  "DST_SUBSAMP_Y unsupported");                               \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);               \
+    const int kSrcHalfHeight = SUBSAMPLE(kHeight, SRC_SUBSAMP_Y);             \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);               \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);             \
+    align_buffer_page_end(src_y, kWidth* kHeight* SRC_BPC + OFF);             \
+    align_buffer_page_end(src_u,                                              \
+                          kSrcHalfWidth* kSrcHalfHeight* SRC_BPC + OFF);      \
+    align_buffer_page_end(src_v,                                              \
+                          kSrcHalfWidth* kSrcHalfHeight* SRC_BPC + OFF);      \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                 \
+    align_buffer_page_end(dst_uv_c,                                           \
+                          kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);        \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);               \
+    align_buffer_page_end(dst_uv_opt,                                         \
+                          kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);        \
+    MemRandomize(src_y + OFF, kWidth * kHeight * SRC_BPC);                    \
+    MemRandomize(src_u + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    MemRandomize(src_v + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                   \
+    SRC_T* src_u_p = reinterpret_cast<SRC_T*>(src_u + OFF);                   \
+    SRC_T* src_v_p = reinterpret_cast<SRC_T*>(src_v + OFF);                   \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      src_y_p[i] = src_y_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
+    for (int i = 0; i < kSrcHalfWidth * kSrcHalfHeight; ++i) {                \
+      src_u_p[i] = src_u_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+      src_v_p[i] = src_v_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                             \
+    memset(dst_uv_c, 2, kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);          \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                         \
+    memset(dst_uv_opt, 102, kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);      \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    SRC_FMT_PLANAR##To##FMT_PLANAR(src_y_p, kWidth, src_u_p, kSrcHalfWidth,   \
+                                   src_v_p, kSrcHalfWidth,                    \
+                                   reinterpret_cast<DST_T*>(dst_y_c), kWidth, \
+                                   reinterpret_cast<DST_T*>(dst_uv_c),        \
+                                   kDstHalfWidth * 2, kWidth, NEG kHeight);   \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      SRC_FMT_PLANAR##To##FMT_PLANAR(                                         \
+          src_y_p, kWidth, src_u_p, kSrcHalfWidth, src_v_p, kSrcHalfWidth,    \
+          reinterpret_cast<DST_T*>(dst_y_opt), kWidth,                        \
+          reinterpret_cast<DST_T*>(dst_uv_opt), kDstHalfWidth * 2, kWidth,    \
+          NEG kHeight);                                                       \
+    }                                                                         \
+    for (int i = 0; i < kHeight * kWidth * DST_BPC; ++i) {                    \
+      EXPECT_EQ(dst_y_c[i], dst_y_opt[i]);                                    \
+    }                                                                         \
+    for (int i = 0; i < kDstHalfWidth * kDstHalfHeight * DST_BPC * 2; ++i) {  \
+      EXPECT_EQ(dst_uv_c[i], dst_uv_opt[i]);                                  \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_uv_c);                                   \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_uv_opt);                                 \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_u);                                      \
+    free_aligned_buffer_page_end(src_v);                                      \
+  }
+
+#define TESTPLANARTOBP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,         \
+                       SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,             \
+                       DST_SUBSAMP_X, DST_SUBSAMP_Y, SRC_DEPTH)               \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_ + 1, _Any, +, 0, SRC_DEPTH) \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_, _Unaligned, +, 2,          \
+                  SRC_DEPTH)                                                  \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_, _Invert, -, 0, SRC_DEPTH)  \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_, _Opt, +, 0, SRC_DEPTH)
+
+TESTPLANARTOBP(I420, uint8_t, 1, 2, 2, NV12, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I420, uint8_t, 1, 2, 2, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I422, uint8_t, 1, 2, 1, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I444, uint8_t, 1, 1, 1, NV12, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I444, uint8_t, 1, 1, 1, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I400, uint8_t, 1, 2, 2, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I010, uint16_t, 2, 2, 2, P010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOBP(I210, uint16_t, 2, 2, 1, P210, uint16_t, 2, 2, 1, 10)
+TESTPLANARTOBP(I012, uint16_t, 2, 2, 2, P012, uint16_t, 2, 2, 2, 12)
+TESTPLANARTOBP(I212, uint16_t, 2, 2, 1, P212, uint16_t, 2, 2, 1, 12)
+
+#define TESTBPTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,            \
+                    SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, \
+                    DST_SUBSAMP_Y, W1280, N, NEG, OFF, DOY, SRC_DEPTH,        \
+                    TILE_WIDTH, TILE_HEIGHT)                                  \
+  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {              \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");       \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                   \
+                  "SRC_SUBSAMP_X unsupported");                               \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                   \
+                  "SRC_SUBSAMP_Y unsupported");                               \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                   \
+                  "DST_SUBSAMP_X unsupported");                               \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                   \
+                  "DST_SUBSAMP_Y unsupported");                               \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);               \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);               \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);             \
+    const int kPaddedWidth = (kWidth + (TILE_WIDTH - 1)) & ~(TILE_WIDTH - 1); \
+    const int kPaddedHeight =                                                 \
+        (kHeight + (TILE_HEIGHT - 1)) & ~(TILE_HEIGHT - 1);                   \
+    const int kSrcHalfPaddedWidth = SUBSAMPLE(kPaddedWidth, SRC_SUBSAMP_X);   \
+    const int kSrcHalfPaddedHeight = SUBSAMPLE(kPaddedHeight, SRC_SUBSAMP_Y); \
+    align_buffer_page_end(src_y, kPaddedWidth* kPaddedHeight* SRC_BPC + OFF); \
+    align_buffer_page_end(                                                    \
+        src_uv,                                                               \
+        2 * kSrcHalfPaddedWidth * kSrcHalfPaddedHeight * SRC_BPC + OFF);      \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                 \
+    align_buffer_page_end(dst_uv_c,                                           \
+                          2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);      \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);               \
+    align_buffer_page_end(dst_uv_opt,                                         \
+                          2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);      \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                   \
+    SRC_T* src_uv_p = reinterpret_cast<SRC_T*>(src_uv + OFF);                 \
+    for (int i = 0;                                                           \
+         i < kPaddedWidth * kPaddedHeight * SRC_BPC / (int)sizeof(SRC_T);     \
+         ++i) {                                                               \
+      src_y_p[i] =                                                            \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));      \
+    }                                                                         \
+    for (int i = 0; i < kSrcHalfPaddedWidth * kSrcHalfPaddedHeight * 2 *      \
+                            SRC_BPC / (int)sizeof(SRC_T);                     \
+         ++i) {                                                               \
+      src_uv_p[i] =                                                           \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));      \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                             \
+    memset(dst_uv_c, 2, 2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);        \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                         \
+    memset(dst_uv_opt, 102, 2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);    \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    SRC_FMT_PLANAR##To##FMT_PLANAR(                                           \
+        src_y_p, kWidth* SRC_BPC / (int)sizeof(SRC_T), src_uv_p,              \
+        2 * kSrcHalfWidth * SRC_BPC / (int)sizeof(SRC_T),                     \
+        DOY ? reinterpret_cast<DST_T*>(dst_y_c) : NULL, kWidth,               \
+        reinterpret_cast<DST_T*>(dst_uv_c), 2 * kDstHalfWidth, kWidth,        \
+        NEG kHeight);                                                         \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      SRC_FMT_PLANAR##To##FMT_PLANAR(                                         \
+          src_y_p, kWidth* SRC_BPC / (int)sizeof(SRC_T), src_uv_p,            \
+          2 * kSrcHalfWidth * SRC_BPC / (int)sizeof(SRC_T),                   \
+          DOY ? reinterpret_cast<DST_T*>(dst_y_opt) : NULL, kWidth,           \
+          reinterpret_cast<DST_T*>(dst_uv_opt), 2 * kDstHalfWidth, kWidth,    \
+          NEG kHeight);                                                       \
+    }                                                                         \
+    if (DOY) {                                                                \
+      for (int i = 0; i < kHeight; ++i) {                                     \
+        for (int j = 0; j < kWidth; ++j) {                                    \
+          EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);      \
+        }                                                                     \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < kDstHalfHeight; ++i) {                                \
+      for (int j = 0; j < 2 * kDstHalfWidth; ++j) {                           \
+        EXPECT_EQ(dst_uv_c[i * 2 * kDstHalfWidth + j],                        \
+                  dst_uv_opt[i * 2 * kDstHalfWidth + j]);                     \
+      }                                                                       \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_uv_c);                                   \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_uv_opt);                                 \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_uv);                                     \
+  }
+
+#define TESTBPTOBP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,            \
+                   SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, \
+                   DST_SUBSAMP_Y, SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)        \
+  TESTBPTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+              FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+              benchmark_width_ + 1, _Any, +, 0, 1, SRC_DEPTH, TILE_WIDTH,    \
+              TILE_HEIGHT)                                                   \
+  TESTBPTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+              FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+              benchmark_width_, _Unaligned, +, 2, 1, SRC_DEPTH, TILE_WIDTH,  \
+              TILE_HEIGHT)                                                   \
+  TESTBPTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+              FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+              benchmark_width_, _Invert, -, 0, 1, SRC_DEPTH, TILE_WIDTH,     \
+              TILE_HEIGHT)                                                   \
+  TESTBPTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+              FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+              benchmark_width_, _Opt, +, 0, 1, SRC_DEPTH, TILE_WIDTH,        \
+              TILE_HEIGHT)                                                   \
+  TESTBPTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+              FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+              benchmark_width_, _NullY, +, 0, 0, SRC_DEPTH, TILE_WIDTH,      \
+              TILE_HEIGHT)
+
+TESTBPTOBP(NV21, uint8_t, 1, 2, 2, NV12, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBPTOBP(NV12, uint8_t, 1, 2, 2, NV12Mirror, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBPTOBP(NV12, uint8_t, 1, 2, 2, NV24, uint8_t, 1, 1, 1, 8, 1, 1)
+TESTBPTOBP(NV16, uint8_t, 1, 2, 1, NV24, uint8_t, 1, 1, 1, 8, 1, 1)
+TESTBPTOBP(P010, uint16_t, 2, 2, 2, P410, uint16_t, 2, 1, 1, 10, 1, 1)
+TESTBPTOBP(P210, uint16_t, 2, 2, 1, P410, uint16_t, 2, 1, 1, 10, 1, 1)
+TESTBPTOBP(P012, uint16_t, 2, 2, 2, P412, uint16_t, 2, 1, 1, 10, 1, 1)
+TESTBPTOBP(P212, uint16_t, 2, 2, 1, P412, uint16_t, 2, 1, 1, 12, 1, 1)
+TESTBPTOBP(P016, uint16_t, 2, 2, 2, P416, uint16_t, 2, 1, 1, 12, 1, 1)
+TESTBPTOBP(P216, uint16_t, 2, 2, 1, P416, uint16_t, 2, 1, 1, 12, 1, 1)
+TESTBPTOBP(MM21, uint8_t, 1, 2, 2, NV12, uint8_t, 1, 2, 2, 8, 16, 32)
+TESTBPTOBP(MT2T, uint8_t, 10 / 8, 2, 2, P010, uint16_t, 2, 2, 2, 10, 16, 32)
+
+#define TESTBPTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                   SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                   DST_SUBSAMP_Y, W1280, N, NEG, OFF, SRC_DEPTH, TILE_WIDTH,   \
+                   TILE_HEIGHT)                                                \
+  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {               \
+    static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");        \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");        \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                    \
+                  "SRC_SUBSAMP_X unsupported");                                \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                    \
+                  "SRC_SUBSAMP_Y unsupported");                                \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                    \
+                  "DST_SUBSAMP_X unsupported");                                \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                    \
+                  "DST_SUBSAMP_Y unsupported");                                \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);                \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);                \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);              \
+    const int kPaddedWidth = (kWidth + (TILE_WIDTH - 1)) & ~(TILE_WIDTH - 1);  \
+    const int kPaddedHeight =                                                  \
+        (kHeight + (TILE_HEIGHT - 1)) & ~(TILE_HEIGHT - 1);                    \
+    const int kSrcHalfPaddedWidth = SUBSAMPLE(kPaddedWidth, SRC_SUBSAMP_X);    \
+    const int kSrcHalfPaddedHeight = SUBSAMPLE(kPaddedHeight, SRC_SUBSAMP_Y);  \
+    align_buffer_page_end(src_y, kPaddedWidth* kPaddedHeight* SRC_BPC + OFF);  \
+    align_buffer_page_end(                                                     \
+        src_uv, kSrcHalfPaddedWidth* kSrcHalfPaddedHeight* SRC_BPC * 2 + OFF); \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                  \
+    align_buffer_page_end(dst_u_c, kDstHalfWidth* kDstHalfHeight* DST_BPC);    \
+    align_buffer_page_end(dst_v_c, kDstHalfWidth* kDstHalfHeight* DST_BPC);    \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);                \
+    align_buffer_page_end(dst_u_opt, kDstHalfWidth* kDstHalfHeight* DST_BPC);  \
+    align_buffer_page_end(dst_v_opt, kDstHalfWidth* kDstHalfHeight* DST_BPC);  \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                    \
+    SRC_T* src_uv_p = reinterpret_cast<SRC_T*>(src_uv + OFF);                  \
+    for (int i = 0; i < kPaddedWidth * kPaddedHeight; ++i) {                   \
+      src_y_p[i] =                                                             \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));       \
     }                                                                          \
-  }                                                                            \
-  memset(dst_y_c, 1, kWidth * kHeight);                                        \
-  memset(dst_uv_c, 2, SUBSAMPLE(kWidth * 2, SUBSAMP_X) *                       \
-                      SUBSAMPLE(kHeight, SUBSAMP_Y));                          \
-  memset(dst_y_opt, 101, kWidth * kHeight);                                    \
-  memset(dst_uv_opt, 102, SUBSAMPLE(kWidth * 2, SUBSAMP_X) *                   \
-                          SUBSAMPLE(kHeight, SUBSAMP_Y));                      \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  SRC_FMT_PLANAR##To##FMT_PLANAR(src_y + OFF, kWidth,                          \
-                                 src_u + OFF,                                  \
-                                 SUBSAMPLE(kWidth, SRC_SUBSAMP_X),             \
-                                 src_v + OFF,                                  \
-                                 SUBSAMPLE(kWidth, SRC_SUBSAMP_X),             \
-                                 dst_y_c, kWidth,                              \
-                                 dst_uv_c, SUBSAMPLE(kWidth * 2, SUBSAMP_X),   \
-                                 kWidth, NEG kHeight);                         \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    SRC_FMT_PLANAR##To##FMT_PLANAR(src_y + OFF, kWidth,                        \
-                                   src_u + OFF,                                \
-                                   SUBSAMPLE(kWidth, SRC_SUBSAMP_X),           \
-                                   src_v + OFF,                                \
-                                   SUBSAMPLE(kWidth, SRC_SUBSAMP_X),           \
-                                   dst_y_opt, kWidth,                          \
-                                   dst_uv_opt,                                 \
-                                   SUBSAMPLE(kWidth * 2, SUBSAMP_X),           \
-                                   kWidth, NEG kHeight);                       \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kHeight; ++i) {                                          \
-    for (int j = 0; j < kWidth; ++j) {                                         \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_y_c[i * kWidth + j]) -                      \
-              static_cast<int>(dst_y_opt[i * kWidth + j]));                    \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    for (int i = 0; i < kSrcHalfPaddedWidth * kSrcHalfPaddedHeight * 2; ++i) { \
+      src_uv_p[i] =                                                            \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));       \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 1);                                                      \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < SUBSAMPLE(kWidth * 2, SUBSAMP_X); ++j) {               \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_uv_c[i *                                    \
-                               SUBSAMPLE(kWidth * 2, SUBSAMP_X) + j]) -        \
-              static_cast<int>(dst_uv_opt[i *                                  \
-                               SUBSAMPLE(kWidth * 2, SUBSAMP_X) + j]));        \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                              \
+    memset(dst_u_c, 2, kDstHalfWidth* kDstHalfHeight* DST_BPC);                \
+    memset(dst_v_c, 3, kDstHalfWidth* kDstHalfHeight* DST_BPC);                \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                          \
+    memset(dst_u_opt, 102, kDstHalfWidth* kDstHalfHeight* DST_BPC);            \
+    memset(dst_v_opt, 103, kDstHalfWidth* kDstHalfHeight* DST_BPC);            \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    SRC_FMT_PLANAR##To##FMT_PLANAR(                                            \
+        src_y_p, kWidth, src_uv_p, kSrcHalfWidth * 2,                          \
+        reinterpret_cast<DST_T*>(dst_y_c), kWidth,                             \
+        reinterpret_cast<DST_T*>(dst_u_c), kDstHalfWidth,                      \
+        reinterpret_cast<DST_T*>(dst_v_c), kDstHalfWidth, kWidth,              \
+        NEG kHeight);                                                          \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      SRC_FMT_PLANAR##To##FMT_PLANAR(                                          \
+          src_y_p, kWidth, src_uv_p, kSrcHalfWidth * 2,                        \
+          reinterpret_cast<DST_T*>(dst_y_opt), kWidth,                         \
+          reinterpret_cast<DST_T*>(dst_u_opt), kDstHalfWidth,                  \
+          reinterpret_cast<DST_T*>(dst_v_opt), kDstHalfWidth, kWidth,          \
+          NEG kHeight);                                                        \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 1);                                                      \
-  free_aligned_buffer_page_end(dst_y_c);                                       \
-  free_aligned_buffer_page_end(dst_uv_c);                                      \
-  free_aligned_buffer_page_end(dst_y_opt);                                     \
-  free_aligned_buffer_page_end(dst_uv_opt);                                    \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
-}
+    for (int i = 0; i < kHeight * kWidth * DST_BPC; ++i) {                     \
+      EXPECT_EQ(dst_y_c[i], dst_y_opt[i]);                                     \
+    }                                                                          \
+    for (int i = 0; i < kDstHalfWidth * kDstHalfHeight * DST_BPC; ++i) {       \
+      EXPECT_EQ(dst_u_c[i], dst_u_opt[i]);                                     \
+      EXPECT_EQ(dst_v_c[i], dst_v_opt[i]);                                     \
+    }                                                                          \
+    free_aligned_buffer_page_end(dst_y_c);                                     \
+    free_aligned_buffer_page_end(dst_u_c);                                     \
+    free_aligned_buffer_page_end(dst_v_c);                                     \
+    free_aligned_buffer_page_end(dst_y_opt);                                   \
+    free_aligned_buffer_page_end(dst_u_opt);                                   \
+    free_aligned_buffer_page_end(dst_v_opt);                                   \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_uv);                                      \
+  }
 
-#define TESTPLANARTOBP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,           \
-                       FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y)                       \
-    TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,              \
-                    FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                          \
-                    benchmark_width_ - 4, _Any, +, 0)                          \
-    TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,              \
-                    FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                          \
-                    benchmark_width_, _Unaligned, +, 1)                        \
-    TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,              \
-                    FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                          \
-                    benchmark_width_, _Invert, -, 0)                           \
-    TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,              \
-                    FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                          \
-                    benchmark_width_, _Opt, +, 0)
-
-TESTPLANARTOBP(I420, 2, 2, NV12, 2, 2)
-TESTPLANARTOBP(I420, 2, 2, NV21, 2, 2)
-
-#define TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,         \
-                         FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF) \
-TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {                 \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_uv, 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *         \
-                        SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF);              \
-  align_buffer_page_end(dst_y_c, kWidth * kHeight);                            \
-  align_buffer_page_end(dst_u_c,                                               \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_v_c,                                               \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_y_opt, kWidth * kHeight);                          \
-  align_buffer_page_end(dst_u_opt,                                             \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  align_buffer_page_end(dst_v_opt,                                             \
-                        SUBSAMPLE(kWidth, SUBSAMP_X) *                         \
-                        SUBSAMPLE(kHeight, SUBSAMP_Y));                        \
-  for (int i = 0; i < kHeight; ++i)                                            \
-    for (int j = 0; j < kWidth; ++j)                                           \
-      src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                       \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {                \
-    for (int j = 0; j < 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {           \
-      src_uv[(i * 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =           \
-          (fastrand() & 0xff);                                                 \
+#define TESTBPTOP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,            \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, \
+                  DST_SUBSAMP_Y, SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)        \
+  TESTBPTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+             FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+             benchmark_width_ + 1, _Any, +, 0, SRC_DEPTH, TILE_WIDTH,       \
+             TILE_HEIGHT)                                                   \
+  TESTBPTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+             FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+             benchmark_width_, _Unaligned, +, 2, SRC_DEPTH, TILE_WIDTH,     \
+             TILE_HEIGHT)                                                   \
+  TESTBPTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+             FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+             benchmark_width_, _Invert, -, 0, SRC_DEPTH, TILE_WIDTH,        \
+             TILE_HEIGHT)                                                   \
+  TESTBPTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+             FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,      \
+             benchmark_width_, _Opt, +, 0, SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)
+
+TESTBPTOP(NV12, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBPTOP(NV21, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBPTOP(MM21, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8, 16, 32)
+TESTBPTOP(P010, uint16_t, 2, 2, 2, I010, uint16_t, 2, 2, 2, 10, 1, 1)
+TESTBPTOP(P012, uint16_t, 2, 2, 2, I012, uint16_t, 2, 2, 2, 12, 1, 1)
+
+// Provide matrix wrappers for full range bt.709
+#define F420ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuF709Constants, i, j)
+#define F420ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+#define F422ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuF709Constants, i, j)
+#define F422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+#define F444ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuF709Constants, i, j)
+#define F444ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+
+// Provide matrix wrappers for full range bt.2020
+#define V420ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuV2020Constants, i, j)
+#define V420ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+#define V422ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuV2020Constants, i, j)
+#define V422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+#define V444ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuV2020Constants, i, j)
+#define V444ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+
+#define I420ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I420ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I422ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I422ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I420ToRGB24Filter(a, b, c, d, e, f, g, h, i, j)                     \
+  I420ToRGB24MatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                          kFilterBilinear)
+#define I422ToRGB24Filter(a, b, c, d, e, f, g, h, i, j)                     \
+  I420ToRGB24MatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                          kFilterBilinear)
+
+#define ALIGNINT(V, ALIGN) (((V) + (ALIGN)-1) / (ALIGN) * (ALIGN))
+
+#define TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                       YALIGN, W1280, N, NEG, OFF)                            \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                       \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                  \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                     \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);            \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                      \
+    align_buffer_page_end(src_u, kSizeUV + OFF);                              \
+    align_buffer_page_end(src_v, kSizeUV + OFF);                              \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + OFF);               \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + OFF);             \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      src_y[i + OFF] = (fastrand() & 0xff);                                   \
+    }                                                                         \
+    for (int i = 0; i < kSizeUV; ++i) {                                       \
+      src_u[i + OFF] = (fastrand() & 0xff);                                   \
+      src_v[i + OFF] = (fastrand() & 0xff);                                   \
+    }                                                                         \
+    memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                          \
+    memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                      \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    double time0 = get_time();                                                \
+    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,        \
+                          src_v + OFF, kStrideUV, dst_argb_c + OFF, kStrideB, \
+                          kWidth, NEG kHeight);                               \
+    double time1 = get_time();                                                \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,      \
+                            src_v + OFF, kStrideUV, dst_argb_opt + OFF,       \
+                            kStrideB, kWidth, NEG kHeight);                   \
+    }                                                                         \
+    double time2 = get_time();                                                \
+    printf(" %8d us C - %8d us OPT\n",                                        \
+           static_cast<int>((time1 - time0) * 1e6),                           \
+           static_cast<int>((time2 - time1) * 1e6 / benchmark_iterations_));  \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                      \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_opt[i + OFF]);                  \
+    }                                                                         \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_u);                                      \
+    free_aligned_buffer_page_end(src_v);                                      \
+    free_aligned_buffer_page_end(dst_argb_c);                                 \
+    free_aligned_buffer_page_end(dst_argb_opt);                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+#define TESTPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                      YALIGN)                                                \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_ + 1, _Any, +, 0)                   \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_, _Unaligned, +, 4)                 \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_, _Invert, -, 0)                    \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_, _Opt, +, 0)
+#else
+#define TESTPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                      YALIGN)                                                \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_ + 1, _Any, +, 0)                   \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_, _Opt, +, 0)
+#endif
+
+#if defined(ENABLE_FULL_TESTS)
+TESTPLANARTOB(I420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(J420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(J420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(F420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(F420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(H420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(H420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(U420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(U420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(V420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(V420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, BGRA, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, RGBA, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, RAW, 3, 3, 1)
+TESTPLANARTOB(I420, 2, 2, RGB24, 3, 3, 1)
+TESTPLANARTOB(J420, 2, 2, RAW, 3, 3, 1)
+TESTPLANARTOB(J420, 2, 2, RGB24, 3, 3, 1)
+TESTPLANARTOB(H420, 2, 2, RAW, 3, 3, 1)
+TESTPLANARTOB(H420, 2, 2, RGB24, 3, 3, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANARTOB(I420, 2, 2, RGB565, 2, 2, 1)
+TESTPLANARTOB(J420, 2, 2, RGB565, 2, 2, 1)
+TESTPLANARTOB(H420, 2, 2, RGB565, 2, 2, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB1555, 2, 2, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB4444, 2, 2, 1)
+TESTPLANARTOB(I422, 2, 1, RGB565, 2, 2, 1)
+#endif
+TESTPLANARTOB(I422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(J422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(J422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(H422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(H422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(U422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(U422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(V422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(V422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, BGRA, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, RGBA, 4, 4, 1)
+TESTPLANARTOB(I422, 1, 1, RGB24, 3, 3, 1)
+TESTPLANARTOB(I422, 1, 1, RAW, 3, 3, 1)
+TESTPLANARTOB(I444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(I444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I444, 1, 1, RGB24, 3, 3, 1)
+TESTPLANARTOB(I444, 1, 1, RAW, 3, 3, 1)
+TESTPLANARTOB(J444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(J444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(H444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(H444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(U444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(U444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(V444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(V444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, YUY2, 2, 4, 1)
+TESTPLANARTOB(I420, 2, 2, UYVY, 2, 4, 1)
+TESTPLANARTOB(I422, 2, 1, YUY2, 2, 4, 1)
+TESTPLANARTOB(I422, 2, 1, UYVY, 2, 4, 1)
+TESTPLANARTOB(I420, 2, 2, I400, 1, 1, 1)
+TESTPLANARTOB(J420, 2, 2, J400, 1, 1, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANARTOB(I420, 2, 2, AR30, 4, 4, 1)
+TESTPLANARTOB(H420, 2, 2, AR30, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, AB30, 4, 4, 1)
+TESTPLANARTOB(H420, 2, 2, AB30, 4, 4, 1)
+#endif
+TESTPLANARTOB(I420, 2, 2, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, RGB24Filter, 3, 3, 1)
+TESTPLANARTOB(I422, 2, 2, RGB24Filter, 3, 3, 1)
+#else
+TESTPLANARTOB(I420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, BGRA, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, RAW, 3, 3, 1)
+TESTPLANARTOB(I420, 2, 2, RGB24, 3, 3, 1)
+TESTPLANARTOB(I420, 2, 2, RGBA, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANARTOB(I420, 2, 2, RGB565, 2, 2, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB1555, 2, 2, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB4444, 2, 2, 1)
+TESTPLANARTOB(I422, 2, 1, RGB565, 2, 2, 1)
+#endif
+TESTPLANARTOB(I420, 2, 2, I400, 1, 1, 1)
+TESTPLANARTOB(I420, 2, 2, UYVY, 2, 4, 1)
+TESTPLANARTOB(I420, 2, 2, YUY2, 2, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, BGRA, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, RGBA, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, UYVY, 2, 4, 1)
+TESTPLANARTOB(I422, 2, 1, YUY2, 2, 4, 1)
+TESTPLANARTOB(I420, 2, 2, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, RGB24Filter, 3, 3, 1)
+TESTPLANARTOB(I444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I444, 1, 1, ARGB, 4, 4, 1)
+#endif
+
+#define TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                        YALIGN, W1280, N, NEG, OFF, ATTEN)                     \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);             \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                       \
+    align_buffer_page_end(src_u, kSizeUV + OFF);                               \
+    align_buffer_page_end(src_v, kSizeUV + OFF);                               \
+    align_buffer_page_end(src_a, kWidth* kHeight + OFF);                       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + OFF);              \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      src_y[i + OFF] = (fastrand() & 0xff);                                    \
+      src_a[i + OFF] = (fastrand() & 0xff);                                    \
     }                                                                          \
-  }                                                                            \
-  memset(dst_y_c, 1, kWidth * kHeight);                                        \
-  memset(dst_u_c, 2, SUBSAMPLE(kWidth, SUBSAMP_X) *                            \
-                     SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_v_c, 3, SUBSAMPLE(kWidth, SUBSAMP_X) *                            \
-                     SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_y_opt, 101, kWidth * kHeight);                                    \
-  memset(dst_u_opt, 102, SUBSAMPLE(kWidth, SUBSAMP_X) *                        \
-                         SUBSAMPLE(kHeight, SUBSAMP_Y));                       \
-  memset(dst_v_opt, 103, SUBSAMPLE(kWidth, SUBSAMP_X) *                        \
-                         SUBSAMPLE(kHeight, SUBSAMP_Y));                       \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  SRC_FMT_PLANAR##To##FMT_PLANAR(src_y + OFF, kWidth,                          \
-                                 src_uv + OFF,                                 \
-                                 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X),         \
-                                 dst_y_c, kWidth,                              \
-                                 dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X),        \
-                                 dst_v_c, SUBSAMPLE(kWidth, SUBSAMP_X),        \
-                                 kWidth, NEG kHeight);                         \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    SRC_FMT_PLANAR##To##FMT_PLANAR(src_y + OFF, kWidth,                        \
-                                   src_uv + OFF,                               \
-                                   2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X),       \
-                                   dst_y_opt, kWidth,                          \
-                                   dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X),    \
-                                   dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X),    \
-                                   kWidth, NEG kHeight);                       \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kHeight; ++i) {                                          \
-    for (int j = 0; j < kWidth; ++j) {                                         \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_y_c[i * kWidth + j]) -                      \
-              static_cast<int>(dst_y_opt[i * kWidth + j]));                    \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      src_u[i + OFF] = (fastrand() & 0xff);                                    \
+      src_v[i + OFF] = (fastrand() & 0xff);                                    \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 1);                                                      \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                   \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_u_c[i *                                     \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]) -            \
-              static_cast<int>(dst_u_opt[i *                                   \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]));            \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                           \
+    memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                       \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,         \
+                          src_v + OFF, kStrideUV, src_a + OFF, kWidth,         \
+                          dst_argb_c + OFF, kStrideB, kWidth, NEG kHeight,     \
+                          ATTEN);                                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,       \
+                            src_v + OFF, kStrideUV, src_a + OFF, kWidth,       \
+                            dst_argb_opt + OFF, kStrideB, kWidth, NEG kHeight, \
+                            ATTEN);                                            \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 1);                                                      \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                   \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_v_c[i *                                     \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]) -            \
-              static_cast<int>(dst_v_opt[i *                                   \
-                               SUBSAMPLE(kWidth, SUBSAMP_X) + j]));            \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                       \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_opt[i + OFF]);                   \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 1);                                                      \
-  free_aligned_buffer_page_end(dst_y_c);                                       \
-  free_aligned_buffer_page_end(dst_u_c);                                       \
-  free_aligned_buffer_page_end(dst_v_c);                                       \
-  free_aligned_buffer_page_end(dst_y_opt);                                     \
-  free_aligned_buffer_page_end(dst_u_opt);                                     \
-  free_aligned_buffer_page_end(dst_v_opt);                                     \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_uv);                                        \
-}
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_u);                                       \
+    free_aligned_buffer_page_end(src_v);                                       \
+    free_aligned_buffer_page_end(src_a);                                       \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+  }
 
-#define TESTBIPLANARTOP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,          \
-                        FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y)                      \
-    TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,             \
-                     FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                         \
-                     benchmark_width_ - 4, _Any, +, 0)                         \
-    TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,             \
-                     FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                         \
-                     benchmark_width_, _Unaligned, +, 1)                       \
-    TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,             \
-                     FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                         \
-                     benchmark_width_, _Invert, -, 0)                          \
-    TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,             \
-                     FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,                         \
-                     benchmark_width_, _Opt, +, 0)
-
-TESTBIPLANARTOP(NV12, 2, 2, I420, 2, 2)
-TESTBIPLANARTOP(NV21, 2, 2, I420, 2, 2)
-
-#define ALIGNINT(V, ALIGN) (((V) + (ALIGN) - 1) / (ALIGN) * (ALIGN))
-
-#define TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
-                       YALIGN, W1280, DIFF, N, NEG, OFF, FMT_C, BPP_C)         \
-TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                          \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                     \
-  const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                        \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);               \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_v, kSizeUV + OFF);                                 \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeight + OFF);                 \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeight + OFF);               \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    src_y[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  for (int i = 0; i < kSizeUV; ++i) {                                          \
-    src_u[i + OFF] = (fastrand() & 0xff);                                      \
-    src_v[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                             \
-  memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                         \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                   \
-                        src_u + OFF, kStrideUV,                                \
-                        src_v + OFF, kStrideUV,                                \
-                        dst_argb_c + OFF, kStrideB,                            \
-                        kWidth, NEG kHeight);                                  \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                 \
-                          src_u + OFF, kStrideUV,                              \
-                          src_v + OFF, kStrideUV,                              \
-                          dst_argb_opt + OFF, kStrideB,                        \
-                          kWidth, NEG kHeight);                                \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  /* Convert to ARGB so 565 is expanded to bytes that can be compared. */      \
-  align_buffer_page_end(dst_argb32_c, kWidth * BPP_C  * kHeight);              \
-  align_buffer_page_end(dst_argb32_opt, kWidth * BPP_C  * kHeight);            \
-  memset(dst_argb32_c, 2, kWidth * BPP_C  * kHeight);                          \
-  memset(dst_argb32_opt, 102, kWidth * BPP_C  * kHeight);                      \
-  FMT_B##To##FMT_C(dst_argb_c + OFF, kStrideB,                                 \
-                   dst_argb32_c, kWidth * BPP_C ,                              \
-                   kWidth, kHeight);                                           \
-  FMT_B##To##FMT_C(dst_argb_opt + OFF, kStrideB,                               \
-                   dst_argb32_opt, kWidth * BPP_C ,                            \
-                   kWidth, kHeight);                                           \
-  for (int i = 0; i < kWidth * BPP_C * kHeight; ++i) {                         \
-    int abs_diff =                                                             \
-        abs(static_cast<int>(dst_argb32_c[i]) -                                \
-            static_cast<int>(dst_argb32_opt[i]));                              \
-    if (abs_diff > max_diff) {                                                 \
-      max_diff = abs_diff;                                                     \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, DIFF);                                                   \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-  free_aligned_buffer_page_end(dst_argb32_c);                                  \
-  free_aligned_buffer_page_end(dst_argb32_opt);                                \
-}
+#if defined(ENABLE_FULL_TESTS)
+#define TESTQPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                       YALIGN)                                                \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_ + 1, _Any, +, 0, 0)                \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Unaligned, +, 2, 0)              \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Invert, -, 0, 0)                 \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Opt, +, 0, 0)                    \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Premult, +, 0, 1)
+#else
+#define TESTQPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                       YALIGN)                                                \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Opt, +, 0, 0)
+#endif
 
-#define TESTPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,   \
-                      YALIGN, DIFF, FMT_C, BPP_C)                              \
-    TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-        YALIGN, benchmark_width_ - 4, DIFF, _Any, +, 0, FMT_C, BPP_C)          \
-    TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-        YALIGN, benchmark_width_, DIFF, _Unaligned, +, 1, FMT_C, BPP_C)        \
-    TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-        YALIGN, benchmark_width_, DIFF, _Invert, -, 0, FMT_C, BPP_C)           \
-    TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-        YALIGN, benchmark_width_, DIFF, _Opt, +, 0, FMT_C, BPP_C)
-
-TESTPLANARTOB(I420, 2, 2, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(J420, 2, 2, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(J420, 2, 2, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(H420, 2, 2, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(H420, 2, 2, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, BGRA, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, RGBA, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, RAW, 3, 3, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, RGB24, 3, 3, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, RGB565, 2, 2, 1, 9, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, ARGB1555, 2, 2, 1, 9, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, ARGB4444, 2, 2, 1, 17, ARGB, 4)
-TESTPLANARTOB(I422, 2, 1, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(J422, 2, 1, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(J422, 2, 1, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(H422, 2, 1, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(H422, 2, 1, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I422, 2, 1, BGRA, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I422, 2, 1, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I422, 2, 1, RGBA, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I411, 4, 1, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I444, 1, 1, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(J444, 1, 1, ARGB, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I444, 1, 1, ABGR, 4, 4, 1, 2, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, YUY2, 2, 4, 1, 1, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, UYVY, 2, 4, 1, 1, ARGB, 4)
-TESTPLANARTOB(I422, 2, 1, YUY2, 2, 4, 1, 0, ARGB, 4)
-TESTPLANARTOB(I422, 2, 1, UYVY, 2, 4, 1, 0, ARGB, 4)
-TESTPLANARTOB(I420, 2, 2, I400, 1, 1, 1, 0, ARGB, 4)
-TESTPLANARTOB(J420, 2, 2, J400, 1, 1, 1, 0, ARGB, 4)
+#define J420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define J422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define J444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
 
-#define TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
-                       YALIGN, W1280, DIFF, N, NEG, OFF, ATTEN)                \
-TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                          \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                     \
-  const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                        \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);               \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_v, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_a, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeight + OFF);                 \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeight + OFF);               \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    src_y[i + OFF] = (fastrand() & 0xff);                                      \
-    src_a[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  for (int i = 0; i < kSizeUV; ++i) {                                          \
-    src_u[i + OFF] = (fastrand() & 0xff);                                      \
-    src_v[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                             \
-  memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                         \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                   \
-                        src_u + OFF, kStrideUV,                                \
-                        src_v + OFF, kStrideUV,                                \
-                        src_a + OFF, kWidth,                                   \
-                        dst_argb_c + OFF, kStrideB,                            \
-                        kWidth, NEG kHeight, ATTEN);                           \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                 \
-                          src_u + OFF, kStrideUV,                              \
-                          src_v + OFF, kStrideUV,                              \
-                          src_a + OFF, kWidth,                                 \
-                          dst_argb_opt + OFF, kStrideB,                        \
-                          kWidth, NEG kHeight, ATTEN);                         \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                         \
-    int abs_diff =                                                             \
-        abs(static_cast<int>(dst_argb_c[i + OFF]) -                            \
-            static_cast<int>(dst_argb_opt[i + OFF]));                          \
-    if (abs_diff > max_diff) {                                                 \
-      max_diff = abs_diff;                                                     \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, DIFF);                                                   \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
-  free_aligned_buffer_page_end(src_a);                                         \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-}
+#define I420AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I420AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+#define I422AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I422AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
 
-#define TESTQPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
-                       YALIGN, DIFF)                                           \
-    TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_ - 4, DIFF, _Any, +, 0, 0)                     \
-    TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Unaligned, +, 1, 0)                   \
-    TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Invert, -, 0, 0)                      \
-    TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Opt, +, 0, 0)                         \
-    TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Premult, +, 0, 1)
-
-TESTQPLANARTOB(I420Alpha, 2, 2, ARGB, 4, 4, 1, 2)
-TESTQPLANARTOB(I420Alpha, 2, 2, ABGR, 4, 4, 1, 2)
-
-#define TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,       \
-                         W1280, DIFF, N, NEG, OFF)                             \
-TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                          \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kStrideB = kWidth * BPP_B;                                         \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_uv,                                                \
-                        kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2 + OFF);  \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeight);                       \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeight);                     \
-  for (int i = 0; i < kHeight; ++i)                                            \
-    for (int j = 0; j < kWidth; ++j)                                           \
-      src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                       \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < kStrideUV * 2; ++j) {                                  \
-      src_uv[i * kStrideUV * 2 + j + OFF] = (fastrand() & 0xff);               \
+#if defined(ENABLE_FULL_TESTS)
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(J420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(J420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(H420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(H420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(F420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(F420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(U420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(U420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(V420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(V420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(J422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(J422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(H422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(H422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(F422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(F422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(U422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(U422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(V422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(V422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(I444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(J444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(J444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(H444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(H444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(F444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(F444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(U444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(U444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(V444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(V444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGBFilter, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGBFilter, 4, 4, 1)
+#else
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGBFilter, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGBFilter, 4, 4, 1)
+#endif
+
+#define TESTBPTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
+                   W1280, N, NEG, OFF)                                         \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
+    const int kStrideB = kWidth * BPP_B;                                       \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                       \
+    align_buffer_page_end(src_uv,                                              \
+                          kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y) * 2 + OFF); \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight);                      \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight);                    \
+    for (int i = 0; i < kHeight; ++i)                                          \
+      for (int j = 0; j < kWidth; ++j)                                         \
+        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                     \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                  \
+      for (int j = 0; j < kStrideUV * 2; ++j) {                                \
+        src_uv[i * kStrideUV * 2 + j + OFF] = (fastrand() & 0xff);             \
+      }                                                                        \
     }                                                                          \
-  }                                                                            \
-  memset(dst_argb_c, 1, kStrideB * kHeight);                                   \
-  memset(dst_argb_opt, 101, kStrideB * kHeight);                               \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                   \
-                        src_uv + OFF, kStrideUV * 2,                           \
-                        dst_argb_c, kWidth * BPP_B,                            \
-                        kWidth, NEG kHeight);                                  \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                 \
-                          src_uv + OFF, kStrideUV * 2,                         \
-                          dst_argb_opt, kWidth * BPP_B,                        \
-                          kWidth, NEG kHeight);                                \
-  }                                                                            \
-  /* Convert to ARGB so 565 is expanded to bytes that can be compared. */      \
-  align_buffer_page_end(dst_argb32_c, kWidth * 4 * kHeight);                   \
-  align_buffer_page_end(dst_argb32_opt, kWidth * 4 * kHeight);                 \
-  memset(dst_argb32_c, 2, kWidth * 4 * kHeight);                               \
-  memset(dst_argb32_opt, 102, kWidth * 4 * kHeight);                           \
-  FMT_B##ToARGB(dst_argb_c, kStrideB,                                          \
-                dst_argb32_c, kWidth * 4,                                      \
-                kWidth, kHeight);                                              \
-  FMT_B##ToARGB(dst_argb_opt, kStrideB,                                        \
-                dst_argb32_opt, kWidth * 4,                                    \
-                kWidth, kHeight);                                              \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kHeight; ++i) {                                          \
-    for (int j = 0; j < kWidth * 4; ++j) {                                     \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_argb32_c[i * kWidth * 4 + j]) -             \
-              static_cast<int>(dst_argb32_opt[i * kWidth * 4 + j]));           \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
+    memset(dst_argb_c, 1, kStrideB* kHeight);                                  \
+    memset(dst_argb_opt, 101, kStrideB* kHeight);                              \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_uv + OFF, kStrideUV * 2,    \
+                          dst_argb_c, kWidth * BPP_B, kWidth, NEG kHeight);    \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_uv + OFF, kStrideUV * 2,  \
+                            dst_argb_opt, kWidth * BPP_B, kWidth,              \
+                            NEG kHeight);                                      \
+    }                                                                          \
+    /* Convert to ARGB so 565 is expanded to bytes that can be compared. */    \
+    align_buffer_page_end(dst_argb32_c, kWidth * 4 * kHeight);                 \
+    align_buffer_page_end(dst_argb32_opt, kWidth * 4 * kHeight);               \
+    memset(dst_argb32_c, 2, kWidth * 4 * kHeight);                             \
+    memset(dst_argb32_opt, 102, kWidth * 4 * kHeight);                         \
+    FMT_C##ToARGB(dst_argb_c, kStrideB, dst_argb32_c, kWidth * 4, kWidth,      \
+                  kHeight);                                                    \
+    FMT_C##ToARGB(dst_argb_opt, kStrideB, dst_argb32_opt, kWidth * 4, kWidth,  \
+                  kHeight);                                                    \
+    for (int i = 0; i < kHeight; ++i) {                                        \
+      for (int j = 0; j < kWidth * 4; ++j) {                                   \
+        EXPECT_EQ(dst_argb32_c[i * kWidth * 4 + j],                            \
+                  dst_argb32_opt[i * kWidth * 4 + j]);                         \
       }                                                                        \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, DIFF);                                                   \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_uv);                                        \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-  free_aligned_buffer_page_end(dst_argb32_c);                                  \
-  free_aligned_buffer_page_end(dst_argb32_opt);                                \
-}
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_uv);                                      \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+    free_aligned_buffer_page_end(dst_argb32_c);                                \
+    free_aligned_buffer_page_end(dst_argb32_opt);                              \
+  }
 
-#define TESTBIPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, DIFF)  \
-    TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,           \
-                     benchmark_width_ - 4, DIFF, _Any, +, 0)                   \
-    TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,           \
-                     benchmark_width_, DIFF, _Unaligned, +, 1)                 \
-    TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,           \
-                     benchmark_width_, DIFF, _Invert, -, 0)                    \
-    TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,           \
-                     benchmark_width_, DIFF, _Opt, +, 0)
+#define TESTBPTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B) \
+  TESTBPTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
+             benchmark_width_ + 1, _Any, +, 0)                           \
+  TESTBPTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
+             benchmark_width_, _Unaligned, +, 2)                         \
+  TESTBPTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
+             benchmark_width_, _Invert, -, 0)                            \
+  TESTBPTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
+             benchmark_width_, _Opt, +, 0)
 
-TESTBIPLANARTOB(NV12, 2, 2, ARGB, 4, 2)
-TESTBIPLANARTOB(NV21, 2, 2, ARGB, 4, 2)
-TESTBIPLANARTOB(NV12, 2, 2, RGB565, 2, 9)
+#define JNV12ToARGB(a, b, c, d, e, f, g, h) \
+  NV12ToARGBMatrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV21ToARGB(a, b, c, d, e, f, g, h) \
+  NV21ToARGBMatrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV12ToABGR(a, b, c, d, e, f, g, h) \
+  NV21ToARGBMatrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV21ToABGR(a, b, c, d, e, f, g, h) \
+  NV12ToARGBMatrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV12ToRGB24(a, b, c, d, e, f, g, h) \
+  NV12ToRGB24Matrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV21ToRGB24(a, b, c, d, e, f, g, h) \
+  NV21ToRGB24Matrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV12ToRAW(a, b, c, d, e, f, g, h) \
+  NV21ToRGB24Matrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV21ToRAW(a, b, c, d, e, f, g, h) \
+  NV12ToRGB24Matrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV12ToRGB565(a, b, c, d, e, f, g, h) \
+  NV12ToRGB565Matrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+
+TESTBPTOB(JNV12, 2, 2, ARGB, ARGB, 4)
+TESTBPTOB(JNV21, 2, 2, ARGB, ARGB, 4)
+TESTBPTOB(JNV12, 2, 2, ABGR, ABGR, 4)
+TESTBPTOB(JNV21, 2, 2, ABGR, ABGR, 4)
+TESTBPTOB(JNV12, 2, 2, RGB24, RGB24, 3)
+TESTBPTOB(JNV21, 2, 2, RGB24, RGB24, 3)
+TESTBPTOB(JNV12, 2, 2, RAW, RAW, 3)
+TESTBPTOB(JNV21, 2, 2, RAW, RAW, 3)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTBPTOB(JNV12, 2, 2, RGB565, RGB565, 2)
+#endif
+
+TESTBPTOB(NV12, 2, 2, ARGB, ARGB, 4)
+TESTBPTOB(NV21, 2, 2, ARGB, ARGB, 4)
+TESTBPTOB(NV12, 2, 2, ABGR, ABGR, 4)
+TESTBPTOB(NV21, 2, 2, ABGR, ABGR, 4)
+TESTBPTOB(NV12, 2, 2, RGB24, RGB24, 3)
+TESTBPTOB(NV21, 2, 2, RGB24, RGB24, 3)
+TESTBPTOB(NV12, 2, 2, RAW, RAW, 3)
+TESTBPTOB(NV21, 2, 2, RAW, RAW, 3)
+TESTBPTOB(NV21, 2, 2, YUV24, RAW, 3)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTBPTOB(NV12, 2, 2, RGB565, RGB565, 2)
+#endif
 
 #define TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, \
-                       W1280, DIFF, N, NEG, OFF)                               \
-TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                          \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                     \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  const int kStride =                                                          \
-      (kStrideUV * SUBSAMP_X * 8 * BPP_A + 7) / 8;                             \
-  align_buffer_page_end(src_argb, kStride * kHeight + OFF);                    \
-  align_buffer_page_end(dst_y_c, kWidth * kHeight);                            \
-  align_buffer_page_end(dst_u_c,                                               \
-                        kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));            \
-  align_buffer_page_end(dst_v_c,                                               \
-                        kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));            \
-  align_buffer_page_end(dst_y_opt, kWidth * kHeight);                          \
-  align_buffer_page_end(dst_u_opt,                                             \
-                        kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));            \
-  align_buffer_page_end(dst_v_opt,                                             \
-                  kStrideUV *                                                  \
-                  SUBSAMPLE(kHeight, SUBSAMP_Y));                              \
-  memset(dst_y_c, 1, kWidth * kHeight);                                        \
-  memset(dst_u_c, 2,                                                           \
-         kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_v_c, 3,                                                           \
-         kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_y_opt, 101, kWidth * kHeight);                                    \
-  memset(dst_u_opt, 102,                                                       \
-         kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  memset(dst_v_opt, 103,                                                       \
-         kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y));                           \
-  for (int i = 0; i < kHeight; ++i)                                            \
-    for (int j = 0; j < kStride; ++j)                                          \
-      src_argb[(i * kStride) + j + OFF] = (fastrand() & 0xff);                 \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride,                               \
-                        dst_y_c, kWidth,                                       \
-                        dst_u_c, kStrideUV,                                    \
-                        dst_v_c, kStrideUV,                                    \
-                        kWidth, NEG kHeight);                                  \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride,                             \
-                          dst_y_opt, kWidth,                                   \
-                          dst_u_opt, kStrideUV,                                \
-                          dst_v_opt, kStrideUV,                                \
+                       W1280, N, NEG, OFF)                                     \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
+    const int kStride = (kStrideUV * SUBSAMP_X * 8 * BPP_A + 7) / 8;           \
+    align_buffer_page_end(src_argb, kStride* kHeight + OFF);                   \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight);                           \
+    align_buffer_page_end(dst_uv_c,                                            \
+                          kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                         \
+    align_buffer_page_end(dst_uv_opt,                                          \
+                          kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    memset(dst_y_c, 1, kWidth* kHeight);                                       \
+    memset(dst_uv_c, 2, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    memset(dst_y_opt, 101, kWidth* kHeight);                                   \
+    memset(dst_uv_opt, 102, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));    \
+    for (int i = 0; i < kHeight; ++i)                                          \
+      for (int j = 0; j < kStride; ++j)                                        \
+        src_argb[(i * kStride) + j + OFF] = (fastrand() & 0xff);               \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_c, kWidth, dst_uv_c,  \
+                          kStrideUV * 2, dst_uv_c + kStrideUV, kStrideUV * 2,  \
                           kWidth, NEG kHeight);                                \
-  }                                                                            \
-  for (int i = 0; i < kHeight; ++i) {                                          \
-    for (int j = 0; j < kWidth; ++j) {                                         \
-      EXPECT_NEAR(static_cast<int>(dst_y_c[i * kWidth + j]),                   \
-                  static_cast<int>(dst_y_opt[i * kWidth + j]), DIFF);          \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_opt, kWidth,        \
+                            dst_uv_opt, kStrideUV * 2, dst_uv_opt + kStrideUV, \
+                            kStrideUV * 2, kWidth, NEG kHeight);               \
     }                                                                          \
-  }                                                                            \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < kStrideUV; ++j) {                                      \
-      EXPECT_NEAR(static_cast<int>(dst_u_c[i * kStrideUV + j]),                \
-                  static_cast<int>(dst_u_opt[i * kStrideUV + j]), DIFF);       \
+    for (int i = 0; i < kHeight; ++i) {                                        \
+      for (int j = 0; j < kWidth; ++j) {                                       \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);         \
+      }                                                                        \
     }                                                                          \
-  }                                                                            \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < kStrideUV; ++j) {                                      \
-      EXPECT_NEAR(static_cast<int>(dst_v_c[i *                                 \
-                                   kStrideUV + j]),                            \
-                  static_cast<int>(dst_v_opt[i *                               \
-                                   kStrideUV + j]), DIFF);                     \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y) * 2; ++i) {              \
+      for (int j = 0; j < kStrideUV; ++j) {                                    \
+        EXPECT_EQ(dst_uv_c[i * kStrideUV + j], dst_uv_opt[i * kStrideUV + j]); \
+      }                                                                        \
     }                                                                          \
-  }                                                                            \
-  free_aligned_buffer_page_end(dst_y_c);                                       \
-  free_aligned_buffer_page_end(dst_u_c);                                       \
-  free_aligned_buffer_page_end(dst_v_c);                                       \
-  free_aligned_buffer_page_end(dst_y_opt);                                     \
-  free_aligned_buffer_page_end(dst_u_opt);                                     \
-  free_aligned_buffer_page_end(dst_v_opt);                                     \
-  free_aligned_buffer_page_end(src_argb);                                      \
-}
+    free_aligned_buffer_page_end(dst_y_c);                                     \
+    free_aligned_buffer_page_end(dst_uv_c);                                    \
+    free_aligned_buffer_page_end(dst_y_opt);                                   \
+    free_aligned_buffer_page_end(dst_uv_opt);                                  \
+    free_aligned_buffer_page_end(src_argb);                                    \
+  }
 
-#define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,  \
-                      DIFF)                                                    \
-    TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,     \
-                   benchmark_width_ - 4, DIFF, _Any, +, 0)                     \
-    TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,     \
-                   benchmark_width_, DIFF, _Unaligned, +, 1)                   \
-    TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,     \
-                   benchmark_width_, DIFF, _Invert, -, 0)                      \
-    TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,     \
-                   benchmark_width_, DIFF, _Opt, +, 0)
-
-TESTATOPLANAR(ARGB, 4, 1, I420, 2, 2, 4)
-#if defined(__arm__) || defined (__aarch64__)
-// arm version subsamples by summing 4 pixels then multiplying by matrix with
-// 4x smaller coefficients which are rounded to nearest integer.
-TESTATOPLANAR(ARGB, 4, 1, J420, 2, 2, 4)
-TESTATOPLANAR(ARGB, 4, 1, J422, 2, 1, 4)
+#if defined(ENABLE_FULL_TESTS)
+#define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_ + 1, _Any, +, 0)                            \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_, _Unaligned, +, 2)                          \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_, _Invert, -, 0)                             \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_, _Opt, +, 0)
 #else
-TESTATOPLANAR(ARGB, 4, 1, J420, 2, 2, 0)
-TESTATOPLANAR(ARGB, 4, 1, J422, 2, 1, 0)
+#define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_ + 1, _Any, +, 0)                            \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_, _Opt, +, 0)
 #endif
-TESTATOPLANAR(BGRA, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(ABGR, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(RGBA, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(RAW, 3, 1, I420, 2, 2, 4)
-TESTATOPLANAR(RGB24, 3, 1, I420, 2, 2, 4)
-TESTATOPLANAR(RGB565, 2, 1, I420, 2, 2, 5)
-// TODO(fbarchard): Make 1555 neon work same as C code, reduce to diff 9.
-TESTATOPLANAR(ARGB1555, 2, 1, I420, 2, 2, 15)
-TESTATOPLANAR(ARGB4444, 2, 1, I420, 2, 2, 17)
-TESTATOPLANAR(ARGB, 4, 1, I411, 4, 1, 4)
-TESTATOPLANAR(ARGB, 4, 1, I422, 2, 1, 2)
-TESTATOPLANAR(ARGB, 4, 1, I444, 1, 1, 2)
-TESTATOPLANAR(YUY2, 2, 1, I420, 2, 2, 2)
-TESTATOPLANAR(UYVY, 2, 1, I420, 2, 2, 2)
-TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1, 2)
-TESTATOPLANAR(UYVY, 2, 1, I422, 2, 1, 2)
-TESTATOPLANAR(I400, 1, 1, I420, 2, 2, 2)
-TESTATOPLANAR(J400, 1, 1, J420, 2, 2, 2)
-
-#define TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR,                      \
-                         SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF)             \
-TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                          \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kStride = SUBSAMPLE(kWidth, SUB_A) * BPP_A;                        \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  align_buffer_page_end(src_argb, kStride * kHeight + OFF);                    \
-  align_buffer_page_end(dst_y_c, kWidth * kHeight);                            \
-  align_buffer_page_end(dst_uv_c,                                              \
-                        kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));        \
-  align_buffer_page_end(dst_y_opt, kWidth * kHeight);                          \
-  align_buffer_page_end(dst_uv_opt,                                            \
-                        kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));        \
-  for (int i = 0; i < kHeight; ++i)                                            \
-    for (int j = 0; j < kStride; ++j)                                          \
-      src_argb[(i * kStride) + j + OFF] = (fastrand() & 0xff);                 \
-  memset(dst_y_c, 1, kWidth * kHeight);                                        \
-  memset(dst_uv_c, 2, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));          \
-  memset(dst_y_opt, 101, kWidth * kHeight);                                    \
-  memset(dst_uv_opt, 102, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride,                               \
-                        dst_y_c, kWidth, dst_uv_c, kStrideUV * 2,              \
-                        kWidth, NEG kHeight);                                  \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride,                             \
-                          dst_y_opt, kWidth,                                   \
-                          dst_uv_opt, kStrideUV * 2, kWidth, NEG kHeight);     \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kHeight; ++i) {                                          \
-    for (int j = 0; j < kWidth; ++j) {                                         \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_y_c[i * kWidth + j]) -                      \
-              static_cast<int>(dst_y_opt[i * kWidth + j]));                    \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 4);                                                      \
-  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                    \
-    for (int j = 0; j < kStrideUV * 2; ++j) {                                  \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_uv_c[i * kStrideUV * 2 + j]) -              \
-              static_cast<int>(dst_uv_opt[i * kStrideUV * 2 + j]));            \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, 4);                                                      \
-  free_aligned_buffer_page_end(dst_y_c);                                       \
-  free_aligned_buffer_page_end(dst_uv_c);                                      \
-  free_aligned_buffer_page_end(dst_y_opt);                                     \
-  free_aligned_buffer_page_end(dst_uv_opt);                                    \
-  free_aligned_buffer_page_end(src_argb);                                      \
-}
 
-#define TESTATOBIPLANAR(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
-    TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,    \
-                     benchmark_width_ - 4, _Any, +, 0)                         \
-    TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,    \
-                     benchmark_width_, _Unaligned, +, 1)                       \
-    TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,    \
-                     benchmark_width_, _Invert, -, 0)                          \
-    TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,    \
-                     benchmark_width_, _Opt, +, 0)
-
-TESTATOBIPLANAR(ARGB, 1, 4, NV12, 2, 2)
-TESTATOBIPLANAR(ARGB, 1, 4, NV21, 2, 2)
-TESTATOBIPLANAR(YUY2, 2, 4, NV12, 2, 2)
-TESTATOBIPLANAR(UYVY, 2, 4, NV12, 2, 2)
-
-#define TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                            \
-                  FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                            \
-                  W1280, DIFF, N, NEG, OFF)                                    \
-TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) {                               \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;         \
-  const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;         \
-  const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;  \
-  const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;  \
-  align_buffer_page_end(src_argb, kStrideA * kHeightA + OFF);                  \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeightB);                      \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeightB);                    \
-  for (int i = 0; i < kStrideA * kHeightA; ++i) {                              \
-    src_argb[i + OFF] = (fastrand() & 0xff);                                   \
-  }                                                                            \
-  memset(dst_argb_c, 1, kStrideB * kHeightB);                                  \
-  memset(dst_argb_opt, 101, kStrideB * kHeightB);                              \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_A##To##FMT_B(src_argb + OFF, kStrideA,                                   \
-                   dst_argb_c, kStrideB,                                       \
-                   kWidth, NEG kHeight);                                       \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_A##To##FMT_B(src_argb + OFF, kStrideA,                                 \
-                     dst_argb_opt, kStrideB,                                   \
-                     kWidth, NEG kHeight);                                     \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kStrideB * kHeightB; ++i) {                              \
-    int abs_diff =                                                             \
-        abs(static_cast<int>(dst_argb_c[i]) -                                  \
-            static_cast<int>(dst_argb_opt[i]));                                \
-    if (abs_diff > max_diff) {                                                 \
-      max_diff = abs_diff;                                                     \
-    }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, DIFF);                                                   \
-  free_aligned_buffer_page_end(src_argb);                                      \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-}
+TESTATOPLANAR(ABGR, 4, 1, I420, 2, 2)
+TESTATOPLANAR(ARGB, 4, 1, I420, 2, 2)
+TESTATOPLANAR(ARGB, 4, 1, I422, 2, 1)
+TESTATOPLANAR(ARGB, 4, 1, I444, 1, 1)
+TESTATOPLANAR(ARGB, 4, 1, J420, 2, 2)
+TESTATOPLANAR(ARGB, 4, 1, J422, 2, 1)
+TESTATOPLANAR(ABGR, 4, 1, J420, 2, 2)
+TESTATOPLANAR(ABGR, 4, 1, J422, 2, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOPLANAR(ARGB4444, 2, 1, I420, 2, 2)
+TESTATOPLANAR(RGB565, 2, 1, I420, 2, 2)
+TESTATOPLANAR(ARGB1555, 2, 1, I420, 2, 2)
+#endif
+TESTATOPLANAR(BGRA, 4, 1, I420, 2, 2)
+TESTATOPLANAR(I400, 1, 1, I420, 2, 2)
+TESTATOPLANAR(J400, 1, 1, J420, 2, 2)
+TESTATOPLANAR(RAW, 3, 1, I420, 2, 2)
+TESTATOPLANAR(RAW, 3, 1, J420, 2, 2)
+TESTATOPLANAR(RGB24, 3, 1, I420, 2, 2)
+TESTATOPLANAR(RGB24, 3, 1, J420, 2, 2)
+TESTATOPLANAR(RGBA, 4, 1, I420, 2, 2)
+TESTATOPLANAR(UYVY, 2, 1, I420, 2, 2)
+TESTATOPLANAR(UYVY, 2, 1, I422, 2, 1)
+TESTATOPLANAR(YUY2, 2, 1, I420, 2, 2)
+TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1)
+
+#define TESTATOBPI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,     \
+                   W1280, N, NEG, OFF)                                        \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                       \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kStride = SUBSAMPLE(kWidth, SUB_A) * BPP_A;                     \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
+    align_buffer_page_end(src_argb, kStride* kHeight + OFF);                  \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight);                          \
+    align_buffer_page_end(dst_uv_c,                                           \
+                          kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                        \
+    align_buffer_page_end(dst_uv_opt,                                         \
+                          kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    for (int i = 0; i < kHeight; ++i)                                         \
+      for (int j = 0; j < kStride; ++j)                                       \
+        src_argb[(i * kStride) + j + OFF] = (fastrand() & 0xff);              \
+    memset(dst_y_c, 1, kWidth* kHeight);                                      \
+    memset(dst_uv_c, 2, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));       \
+    memset(dst_y_opt, 101, kWidth* kHeight);                                  \
+    memset(dst_uv_opt, 102, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));   \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_c, kWidth, dst_uv_c, \
+                          kStrideUV * 2, kWidth, NEG kHeight);                \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_opt, kWidth,       \
+                            dst_uv_opt, kStrideUV * 2, kWidth, NEG kHeight);  \
+    }                                                                         \
+    for (int i = 0; i < kHeight; ++i) {                                       \
+      for (int j = 0; j < kWidth; ++j) {                                      \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);        \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < kStrideUV * 2; ++j) {                               \
+        EXPECT_EQ(dst_uv_c[i * kStrideUV * 2 + j],                            \
+                  dst_uv_opt[i * kStrideUV * 2 + j]);                         \
+      }                                                                       \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_uv_c);                                   \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_uv_opt);                                 \
+    free_aligned_buffer_page_end(src_argb);                                   \
+  }
+
+#define TESTATOBP(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
+  TESTATOBPI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+             benchmark_width_ + 1, _Any, +, 0)                           \
+  TESTATOBPI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+             benchmark_width_, _Unaligned, +, 2)                         \
+  TESTATOBPI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+             benchmark_width_, _Invert, -, 0)                            \
+  TESTATOBPI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+             benchmark_width_, _Opt, +, 0)
+
+TESTATOBP(ARGB, 1, 4, NV12, 2, 2)
+TESTATOBP(ARGB, 1, 4, NV21, 2, 2)
+TESTATOBP(ABGR, 1, 4, NV12, 2, 2)
+TESTATOBP(ABGR, 1, 4, NV21, 2, 2)
+TESTATOBP(RAW, 1, 3, JNV21, 2, 2)
+TESTATOBP(YUY2, 2, 4, NV12, 2, 2)
+TESTATOBP(UYVY, 2, 4, NV12, 2, 2)
+TESTATOBP(AYUV, 1, 4, NV12, 2, 2)
+TESTATOBP(AYUV, 1, 4, NV21, 2, 2)
 
-#define TESTATOBRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                       \
-                       FMT_B, BPP_B, STRIDE_B, HEIGHT_B, DIFF)                 \
-TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##_Random) {                         \
-  for (int times = 0; times < benchmark_iterations_; ++times) {                \
-    const int kWidth = (fastrand() & 63) + 1;                                  \
-    const int kHeight = (fastrand() & 31) + 1;                                 \
+#define TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,     \
+                  EPP_B, STRIDE_B, HEIGHT_B, W1280, N, NEG, OFF)               \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) {                             \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
     const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;       \
     const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;       \
-    const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;\
-    const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;\
-    align_buffer_page_end(src_argb, kStrideA * kHeightA);                      \
-    align_buffer_page_end(dst_argb_c, kStrideB * kHeightB);                    \
-    align_buffer_page_end(dst_argb_opt, kStrideB * kHeightB);                  \
-    for (int i = 0; i < kStrideA * kHeightA; ++i) {                            \
-      src_argb[i] = (fastrand() & 0xff);                                       \
+    const int kStrideA =                                                       \
+        (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                 \
+    const int kStrideB =                                                       \
+        (kWidth * EPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;                 \
+    align_buffer_page_end(src_argb,                                            \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeightB*(int)sizeof(TYPE_B)); \
+    align_buffer_page_end(dst_argb_opt,                                        \
+                          kStrideB* kHeightB*(int)sizeof(TYPE_B));             \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {      \
+      src_argb[i + OFF] = (fastrand() & 0xff);                                 \
     }                                                                          \
-    memset(dst_argb_c, 123, kStrideB * kHeightB);                              \
-    memset(dst_argb_opt, 123, kStrideB * kHeightB);                            \
+    memset(dst_argb_c, 1, kStrideB* kHeightB);                                 \
+    memset(dst_argb_opt, 101, kStrideB* kHeightB);                             \
     MaskCpuFlags(disable_cpu_flags_);                                          \
-    FMT_A##To##FMT_B(src_argb, kStrideA,                                       \
-                     dst_argb_c, kStrideB,                                     \
-                     kWidth, kHeight);                                         \
+    FMT_A##To##FMT_B((TYPE_A*)(src_argb + OFF), kStrideA, (TYPE_B*)dst_argb_c, \
+                     kStrideB, kWidth, NEG kHeight);                           \
     MaskCpuFlags(benchmark_cpu_info_);                                         \
-    FMT_A##To##FMT_B(src_argb, kStrideA,                                       \
-                     dst_argb_opt, kStrideB,                                   \
-                     kWidth, kHeight);                                         \
-    int max_diff = 0;                                                          \
-    for (int i = 0; i < kStrideB * kHeightB; ++i) {                            \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_argb_c[i]) -                                \
-              static_cast<int>(dst_argb_opt[i]));                              \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_A##To##FMT_B((TYPE_A*)(src_argb + OFF), kStrideA,                    \
+                       (TYPE_B*)dst_argb_opt, kStrideB, kWidth, NEG kHeight);  \
+    }                                                                          \
+    for (int i = 0; i < kStrideB * kHeightB * (int)sizeof(TYPE_B); ++i) {      \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                               \
     }                                                                          \
-    EXPECT_LE(max_diff, DIFF);                                                 \
     free_aligned_buffer_page_end(src_argb);                                    \
     free_aligned_buffer_page_end(dst_argb_c);                                  \
     free_aligned_buffer_page_end(dst_argb_opt);                                \
-  }                                                                            \
-}
+  }
 
-#define TESTATOB(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                             \
-                 FMT_B, BPP_B, STRIDE_B, HEIGHT_B, DIFF)                       \
-    TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                                \
-              FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                                \
-              benchmark_width_ - 4, DIFF, _Any, +, 0)                          \
-    TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                                \
-              FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                                \
-              benchmark_width_, DIFF, _Unaligned, +, 1)                        \
-    TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                                \
-              FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                                \
-              benchmark_width_, DIFF, _Invert, -, 0)                           \
-    TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                                \
-              FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                                \
-              benchmark_width_, DIFF, _Opt, +, 0)                              \
-    TESTATOBRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                           \
-                   FMT_B, BPP_B, STRIDE_B, HEIGHT_B, DIFF)
-
-TESTATOB(ARGB, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, BGRA, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ABGR, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RGBA, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RAW, 3, 3, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RGB24, 3, 3, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RGB565, 2, 2, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGB1555, 2, 2, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGB4444, 2, 2, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, YUY2, 2, 4, 1, 4)
-TESTATOB(ARGB, 4, 4, 1, UYVY, 2, 4, 1, 4)
-TESTATOB(ARGB, 4, 4, 1, I400, 1, 1, 1, 2)
-TESTATOB(ARGB, 4, 4, 1, J400, 1, 1, 1, 2)
-TESTATOB(BGRA, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ABGR, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RGBA, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RAW, 3, 3, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RAW, 3, 3, 1, RGB24, 3, 3, 1, 0)
-TESTATOB(RGB24, 3, 3, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RGB565, 2, 2, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ARGB1555, 2, 2, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ARGB4444, 2, 2, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(YUY2, 2, 4, 1, ARGB, 4, 4, 1, 4)
-TESTATOB(UYVY, 2, 4, 1, ARGB, 4, 4, 1, 4)
-TESTATOB(I400, 1, 1, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(J400, 1, 1, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(I400, 1, 1, 1, I400, 1, 1, 1, 0)
-TESTATOB(J400, 1, 1, 1, J400, 1, 1, 1, 0)
-TESTATOB(I400, 1, 1, 1, I400Mirror, 1, 1, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGBMirror, 4, 4, 1, 0)
-
-#define TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                           \
-                   FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                           \
-                   W1280, DIFF, N, NEG, OFF)                                   \
-TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither##N) {                       \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;         \
-  const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;         \
-  const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;  \
-  const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;  \
-  align_buffer_page_end(src_argb, kStrideA * kHeightA + OFF);                  \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeightB);                      \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeightB);                    \
-  for (int i = 0; i < kStrideA * kHeightA; ++i) {                              \
-    src_argb[i + OFF] = (fastrand() & 0xff);                                   \
-  }                                                                            \
-  memset(dst_argb_c, 1, kStrideB * kHeightB);                                  \
-  memset(dst_argb_opt, 101, kStrideB * kHeightB);                              \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_A##To##FMT_B##Dither(src_argb + OFF, kStrideA,                           \
-                           dst_argb_c, kStrideB,                               \
-                           NULL, kWidth, NEG kHeight);                         \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_A##To##FMT_B##Dither(src_argb + OFF, kStrideA,                         \
-                             dst_argb_opt, kStrideB,                           \
-                             NULL, kWidth, NEG kHeight);                       \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  for (int i = 0; i < kStrideB * kHeightB; ++i) {                              \
-    int abs_diff =                                                             \
-        abs(static_cast<int>(dst_argb_c[i]) -                                  \
-            static_cast<int>(dst_argb_opt[i]));                                \
-    if (abs_diff > max_diff) {                                                 \
-      max_diff = abs_diff;                                                     \
+#define TESTATOBRANDOM(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B,        \
+                       TYPE_B, EPP_B, STRIDE_B, HEIGHT_B)                      \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##_Random) {                       \
+    for (int times = 0; times < benchmark_iterations_; ++times) {              \
+      const int kWidth = (fastrand() & 63) + 1;                                \
+      const int kHeight = (fastrand() & 31) + 1;                               \
+      const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;     \
+      const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;     \
+      const int kStrideA =                                                     \
+          (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;               \
+      const int kStrideB =                                                     \
+          (kWidth * EPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;               \
+      align_buffer_page_end(src_argb, kStrideA* kHeightA*(int)sizeof(TYPE_A)); \
+      align_buffer_page_end(dst_argb_c,                                        \
+                            kStrideB* kHeightB*(int)sizeof(TYPE_B));           \
+      align_buffer_page_end(dst_argb_opt,                                      \
+                            kStrideB* kHeightB*(int)sizeof(TYPE_B));           \
+      for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {    \
+        src_argb[i] = 0xfe;                                                    \
+      }                                                                        \
+      memset(dst_argb_c, 123, kStrideB* kHeightB);                             \
+      memset(dst_argb_opt, 123, kStrideB* kHeightB);                           \
+      MaskCpuFlags(disable_cpu_flags_);                                        \
+      FMT_A##To##FMT_B((TYPE_A*)src_argb, kStrideA, (TYPE_B*)dst_argb_c,       \
+                       kStrideB, kWidth, kHeight);                             \
+      MaskCpuFlags(benchmark_cpu_info_);                                       \
+      FMT_A##To##FMT_B((TYPE_A*)src_argb, kStrideA, (TYPE_B*)dst_argb_opt,     \
+                       kStrideB, kWidth, kHeight);                             \
+      for (int i = 0; i < kStrideB * kHeightB * (int)sizeof(TYPE_B); ++i) {    \
+        EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                             \
+      }                                                                        \
+      free_aligned_buffer_page_end(src_argb);                                  \
+      free_aligned_buffer_page_end(dst_argb_c);                                \
+      free_aligned_buffer_page_end(dst_argb_opt);                              \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, DIFF);                                                   \
-  free_aligned_buffer_page_end(src_argb);                                      \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-}
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+#define TESTATOB(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)                                 \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_ + 1, _Any, +, 0)           \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Unaligned, +, 4)         \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Invert, -, 0)            \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Opt, +, 0)               \
+  TESTATOBRANDOM(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)
+#else
+#define TESTATOB(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)                                 \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Opt, +, 0)
+#endif
+
+TESTATOB(AB30, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOB(AB30, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(ABGR, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(ABGR, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(AR30, uint8_t, 4, 4, 1, AB30, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(AR30, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(AR30, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+TESTATOB(AR30, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(ARGB, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGB1555, uint8_t, 2, 2, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGB4444, uint8_t, 2, 2, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGBMirror, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, BGRA, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, I400, uint8_t, 1, 1, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(RGBA, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RAW, uint8_t, 3, 3, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RGB24, uint8_t, 3, 3, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, RAW, uint8_t, 3, 3, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, RGB24, uint8_t, 3, 3, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RGB565, uint8_t, 2, 2, 1)
+#endif
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RGBA, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, UYVY, uint8_t, 2, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, YUY2, uint8_t, 2, 4, 1)
+TESTATOB(ARGB1555, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(ARGB4444, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(BGRA, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(I400, uint8_t, 1, 1, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(I400, uint8_t, 1, 1, 1, I400, uint8_t, 1, 1, 1)
+TESTATOB(I400, uint8_t, 1, 1, 1, I400Mirror, uint8_t, 1, 1, 1)
+TESTATOB(J400, uint8_t, 1, 1, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(J400, uint8_t, 1, 1, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, RGBA, uint8_t, 4, 4, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, RGB24, uint8_t, 3, 3, 1)
+TESTATOB(RGB24, uint8_t, 3, 3, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(RGB24, uint8_t, 3, 3, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(RGB24, uint8_t, 3, 3, 1, RGB24Mirror, uint8_t, 3, 3, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, J400, uint8_t, 1, 1, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(RGB565, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(RGBA, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(UYVY, uint8_t, 2, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(YUY2, uint8_t, 2, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(YUY2, uint8_t, 2, 4, 1, Y, uint8_t, 1, 1, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOB(AR64, uint16_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(AB64, uint16_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(AR64, uint16_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOB(AB64, uint16_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOB(AR64, uint16_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOB(AB64, uint16_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+
+// in place test
+#define TESTATOAI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,    \
+                  EPP_B, STRIDE_B, HEIGHT_B, W1280, N, NEG, OFF)              \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) {                            \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;      \
+    const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;      \
+    const int kStrideA =                                                      \
+        (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                \
+    const int kStrideB =                                                      \
+        (kWidth * EPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;                \
+    align_buffer_page_end(src_argb,                                           \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);      \
+    align_buffer_page_end(dst_argb_c,                                         \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);      \
+    align_buffer_page_end(dst_argb_opt,                                       \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);      \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {     \
+      src_argb[i + OFF] = (fastrand() & 0xff);                                \
+    }                                                                         \
+    memcpy(dst_argb_c + OFF, src_argb,                                        \
+           kStrideA * kHeightA * (int)sizeof(TYPE_A));                        \
+    memcpy(dst_argb_opt + OFF, src_argb,                                      \
+           kStrideA * kHeightA * (int)sizeof(TYPE_A));                        \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    FMT_A##To##FMT_B((TYPE_A*)(dst_argb_c /* src */ + OFF), kStrideA,         \
+                     (TYPE_B*)dst_argb_c, kStrideB, kWidth, NEG kHeight);     \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_A##To##FMT_B((TYPE_A*)(dst_argb_opt /* src */ + OFF), kStrideA,     \
+                       (TYPE_B*)dst_argb_opt, kStrideB, kWidth, NEG kHeight); \
+    }                                                                         \
+    memcpy(dst_argb_opt + OFF, src_argb,                                      \
+           kStrideA * kHeightA * (int)sizeof(TYPE_A));                        \
+    FMT_A##To##FMT_B((TYPE_A*)(dst_argb_opt /* src */ + OFF), kStrideA,       \
+                     (TYPE_B*)dst_argb_opt, kStrideB, kWidth, NEG kHeight);   \
+    for (int i = 0; i < kStrideB * kHeightB * (int)sizeof(TYPE_B); ++i) {     \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                              \
+    }                                                                         \
+    free_aligned_buffer_page_end(src_argb);                                   \
+    free_aligned_buffer_page_end(dst_argb_c);                                 \
+    free_aligned_buffer_page_end(dst_argb_opt);                               \
+  }
+
+#define TESTATOA(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)                                 \
+  TESTATOAI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Inplace, +, 0)
+
+TESTATOA(AB30, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOA(AB30, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOA(ABGR, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+#endif
+TESTATOA(ABGR, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOA(AR30, uint8_t, 4, 4, 1, AB30, uint8_t, 4, 4, 1)
+#endif
+TESTATOA(AR30, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOA(AR30, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+TESTATOA(AR30, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#endif
+TESTATOA(ARGB, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOA(ARGB, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+#endif
+TESTATOA(ARGB, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, ARGB1555, uint8_t, 2, 2, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, ARGB4444, uint8_t, 2, 2, 1)
+// TODO(fbarchard): Support in place for mirror.
+// TESTATOA(ARGB, uint8_t, 4, 4, 1, ARGBMirror, uint8_t, 4, 4, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, BGRA, uint8_t, 4, 4, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, I400, uint8_t, 1, 1, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOA(RGBA, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, RAW, uint8_t, 3, 3, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, RGB24, uint8_t, 3, 3, 1)
+TESTATOA(ABGR, uint8_t, 4, 4, 1, RAW, uint8_t, 3, 3, 1)
+TESTATOA(ABGR, uint8_t, 4, 4, 1, RGB24, uint8_t, 3, 3, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOA(ARGB, uint8_t, 4, 4, 1, RGB565, uint8_t, 2, 2, 1)
+#endif
+TESTATOA(ARGB, uint8_t, 4, 4, 1, RGBA, uint8_t, 4, 4, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, UYVY, uint8_t, 2, 4, 1)
+TESTATOA(ARGB, uint8_t, 4, 4, 1, YUY2, uint8_t, 2, 4, 1)
+// TODO(fbarchard): Support in place for conversions that increase bpp.
+// TESTATOA(ARGB1555, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+// TESTATOA(ARGB4444, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(BGRA, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+// TESTATOA(I400, uint8_t, 1, 1, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(I400, uint8_t, 1, 1, 1, I400, uint8_t, 1, 1, 1)
+// TESTATOA(I400, uint8_t, 1, 1, 1, I400Mirror, uint8_t, 1, 1, 1)
+// TESTATOA(J400, uint8_t, 1, 1, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(J400, uint8_t, 1, 1, 1, J400, uint8_t, 1, 1, 1)
+// TESTATOA(RAW, uint8_t, 3, 3, 1, ARGB, uint8_t, 4, 4, 1)
+// TESTATOA(RAW, uint8_t, 3, 3, 1, RGBA, uint8_t, 4, 4, 1)
+TESTATOA(RAW, uint8_t, 3, 3, 1, RGB24, uint8_t, 3, 3, 1)
+// TESTATOA(RGB24, uint8_t, 3, 3, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(RGB24, uint8_t, 3, 3, 1, J400, uint8_t, 1, 1, 1)
+// TESTATOA(RGB24, uint8_t, 3, 3, 1, RGB24Mirror, uint8_t, 3, 3, 1)
+TESTATOA(RAW, uint8_t, 3, 3, 1, J400, uint8_t, 1, 1, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+// TESTATOA(RGB565, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+#endif
+TESTATOA(RGBA, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+// TESTATOA(UYVY, uint8_t, 2, 4, 1, ARGB, uint8_t, 4, 4, 1)
+// TESTATOA(YUY2, uint8_t, 2, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(YUY2, uint8_t, 2, 4, 1, Y, uint8_t, 1, 1, 1)
+// TESTATOA(ARGB, uint8_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+// TESTATOA(ARGB, uint8_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+// TESTATOA(ABGR, uint8_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+// TESTATOA(ABGR, uint8_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOA(AR64, uint16_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(AB64, uint16_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOA(AR64, uint16_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOA(AB64, uint16_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOA(AR64, uint16_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOA(AB64, uint16_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+
+#define TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
+                   HEIGHT_B, W1280, N, NEG, OFF)                             \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither##N) {                   \
+    const int kWidth = W1280;                                                \
+    const int kHeight = benchmark_height_;                                   \
+    const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;     \
+    const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;     \
+    const int kStrideA =                                                     \
+        (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;               \
+    const int kStrideB =                                                     \
+        (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;               \
+    align_buffer_page_end(src_argb, kStrideA* kHeightA + OFF);               \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeightB);                   \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeightB);                 \
+    for (int i = 0; i < kStrideA * kHeightA; ++i) {                          \
+      src_argb[i + OFF] = (fastrand() & 0xff);                               \
+    }                                                                        \
+    memset(dst_argb_c, 1, kStrideB* kHeightB);                               \
+    memset(dst_argb_opt, 101, kStrideB* kHeightB);                           \
+    MaskCpuFlags(disable_cpu_flags_);                                        \
+    FMT_A##To##FMT_B##Dither(src_argb + OFF, kStrideA, dst_argb_c, kStrideB, \
+                             NULL, kWidth, NEG kHeight);                     \
+    MaskCpuFlags(benchmark_cpu_info_);                                       \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                        \
+      FMT_A##To##FMT_B##Dither(src_argb + OFF, kStrideA, dst_argb_opt,       \
+                               kStrideB, NULL, kWidth, NEG kHeight);         \
+    }                                                                        \
+    for (int i = 0; i < kStrideB * kHeightB; ++i) {                          \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                             \
+    }                                                                        \
+    free_aligned_buffer_page_end(src_argb);                                  \
+    free_aligned_buffer_page_end(dst_argb_c);                                \
+    free_aligned_buffer_page_end(dst_argb_opt);                              \
+  }
+
+#define TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B,        \
+                        STRIDE_B, HEIGHT_B)                                    \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither_Random) {                 \
+    for (int times = 0; times < benchmark_iterations_; ++times) {              \
+      const int kWidth = (fastrand() & 63) + 1;                                \
+      const int kHeight = (fastrand() & 31) + 1;                               \
+      const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;     \
+      const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;     \
+      const int kStrideA =                                                     \
+          (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;               \
+      const int kStrideB =                                                     \
+          (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;               \
+      align_buffer_page_end(src_argb, kStrideA* kHeightA);                     \
+      align_buffer_page_end(dst_argb_c, kStrideB* kHeightB);                   \
+      align_buffer_page_end(dst_argb_opt, kStrideB* kHeightB);                 \
+      for (int i = 0; i < kStrideA * kHeightA; ++i) {                          \
+        src_argb[i] = (fastrand() & 0xff);                                     \
+      }                                                                        \
+      memset(dst_argb_c, 123, kStrideB* kHeightB);                             \
+      memset(dst_argb_opt, 123, kStrideB* kHeightB);                           \
+      MaskCpuFlags(disable_cpu_flags_);                                        \
+      FMT_A##To##FMT_B##Dither(src_argb, kStrideA, dst_argb_c, kStrideB, NULL, \
+                               kWidth, kHeight);                               \
+      MaskCpuFlags(benchmark_cpu_info_);                                       \
+      FMT_A##To##FMT_B##Dither(src_argb, kStrideA, dst_argb_opt, kStrideB,     \
+                               NULL, kWidth, kHeight);                         \
+      for (int i = 0; i < kStrideB * kHeightB; ++i) {                          \
+        EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                             \
+      }                                                                        \
+      free_aligned_buffer_page_end(src_argb);                                  \
+      free_aligned_buffer_page_end(dst_argb_c);                                \
+      free_aligned_buffer_page_end(dst_argb_opt);                              \
+    }                                                                          \
+  }
+
+#define TESTATOBD(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
+                  HEIGHT_B)                                                 \
+  TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
+             HEIGHT_B, benchmark_width_ + 1, _Any, +, 0)                    \
+  TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
+             HEIGHT_B, benchmark_width_, _Unaligned, +, 2)                  \
+  TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
+             HEIGHT_B, benchmark_width_, _Invert, -, 0)                     \
+  TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
+             HEIGHT_B, benchmark_width_, _Opt, +, 0)                        \
+  TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
+                  HEIGHT_B)
 
-#define TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                      \
-                       FMT_B, BPP_B, STRIDE_B, HEIGHT_B, DIFF)                 \
-TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither_Random) {                   \
-  for (int times = 0; times < benchmark_iterations_; ++times) {                \
-    const int kWidth = (fastrand() & 63) + 1;                                  \
-    const int kHeight = (fastrand() & 31) + 1;                                 \
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOBD(ARGB, 4, 4, 1, RGB565, 2, 2, 1)
+#endif
+
+// These conversions called twice, produce the original result.
+// e.g. endian swap twice.
+#define TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, W1280, N, NEG,   \
+                 OFF)                                                          \
+  TEST_F(LibYUVConvertTest, FMT_ATOB##_Endswap##N) {                           \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
     const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;       \
-    const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;       \
-    const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;\
-    const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;\
-    align_buffer_page_end(src_argb, kStrideA * kHeightA);                      \
-    align_buffer_page_end(dst_argb_c, kStrideB * kHeightB);                    \
-    align_buffer_page_end(dst_argb_opt, kStrideB * kHeightB);                  \
-    for (int i = 0; i < kStrideA * kHeightA; ++i) {                            \
-      src_argb[i] = (fastrand() & 0xff);                                       \
+    const int kStrideA =                                                       \
+        (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                 \
+    align_buffer_page_end(src_argb,                                            \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);       \
+    align_buffer_page_end(dst_argb_c, kStrideA* kHeightA*(int)sizeof(TYPE_A)); \
+    align_buffer_page_end(dst_argb_opt,                                        \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A));             \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {      \
+      src_argb[i + OFF] = (fastrand() & 0xff);                                 \
     }                                                                          \
-    memset(dst_argb_c, 123, kStrideB * kHeightB);                              \
-    memset(dst_argb_opt, 123, kStrideB * kHeightB);                            \
+    memset(dst_argb_c, 1, kStrideA* kHeightA);                                 \
+    memset(dst_argb_opt, 101, kStrideA* kHeightA);                             \
     MaskCpuFlags(disable_cpu_flags_);                                          \
-    FMT_A##To##FMT_B##Dither(src_argb, kStrideA,                               \
-                             dst_argb_c, kStrideB,                             \
-                             NULL, kWidth, kHeight);                           \
+    FMT_ATOB((TYPE_A*)(src_argb + OFF), kStrideA, (TYPE_A*)dst_argb_c,         \
+             kStrideA, kWidth, NEG kHeight);                                   \
     MaskCpuFlags(benchmark_cpu_info_);                                         \
-    FMT_A##To##FMT_B##Dither(src_argb, kStrideA,                               \
-                             dst_argb_opt, kStrideB,                           \
-                             NULL, kWidth, kHeight);                           \
-    int max_diff = 0;                                                          \
-    for (int i = 0; i < kStrideB * kHeightB; ++i) {                            \
-      int abs_diff =                                                           \
-          abs(static_cast<int>(dst_argb_c[i]) -                                \
-              static_cast<int>(dst_argb_opt[i]));                              \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_ATOB((TYPE_A*)(src_argb + OFF), kStrideA, (TYPE_A*)dst_argb_opt,     \
+               kStrideA, kWidth, NEG kHeight);                                 \
+    }                                                                          \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_ATOB((TYPE_A*)dst_argb_c, kStrideA, (TYPE_A*)dst_argb_c, kStrideA,     \
+             kWidth, NEG kHeight);                                             \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    FMT_ATOB((TYPE_A*)dst_argb_opt, kStrideA, (TYPE_A*)dst_argb_opt, kStrideA, \
+             kWidth, NEG kHeight);                                             \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {      \
+      EXPECT_EQ(src_argb[i + OFF], dst_argb_opt[i]);                           \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                               \
     }                                                                          \
-    EXPECT_LE(max_diff, DIFF);                                                 \
     free_aligned_buffer_page_end(src_argb);                                    \
     free_aligned_buffer_page_end(dst_argb_c);                                  \
     free_aligned_buffer_page_end(dst_argb_opt);                                \
-  }                                                                            \
-}
-
-#define TESTATOBD(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                            \
-                  FMT_B, BPP_B, STRIDE_B, HEIGHT_B, DIFF)                      \
-    TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                               \
-               FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                               \
-               benchmark_width_ - 4, DIFF, _Any, +, 0)                         \
-    TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                               \
-               FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                               \
-               benchmark_width_, DIFF, _Unaligned, +, 1)                       \
-    TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                               \
-               FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                               \
-               benchmark_width_, DIFF, _Invert, -, 0)                          \
-    TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                               \
-               FMT_B, BPP_B, STRIDE_B, HEIGHT_B,                               \
-               benchmark_width_, DIFF, _Opt, +, 0)                             \
-    TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A,                          \
-                    FMT_B, BPP_B, STRIDE_B, HEIGHT_B, DIFF)
-
-TESTATOBD(ARGB, 4, 4, 1, RGB565, 2, 2, 1, 0)
-
-#define TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A,                          \
-                 W1280, N, NEG, OFF)                                           \
-TEST_F(LibYUVConvertTest, FMT_ATOB##_Symetric##N) {                            \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;         \
-  const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;  \
-  align_buffer_page_end(src_argb, kStrideA * kHeightA + OFF);                  \
-  align_buffer_page_end(dst_argb_c, kStrideA * kHeightA);                      \
-  align_buffer_page_end(dst_argb_opt, kStrideA * kHeightA);                    \
-  for (int i = 0; i < kStrideA * kHeightA; ++i) {                              \
-    src_argb[i + OFF] = (fastrand() & 0xff);                                   \
-  }                                                                            \
-  memset(dst_argb_c, 1, kStrideA * kHeightA);                                  \
-  memset(dst_argb_opt, 101, kStrideA * kHeightA);                              \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_ATOB(src_argb + OFF, kStrideA,                                           \
-           dst_argb_c, kStrideA,                                               \
-           kWidth, NEG kHeight);                                               \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_ATOB(src_argb + OFF, kStrideA,                                         \
-             dst_argb_opt, kStrideA,                                           \
-             kWidth, NEG kHeight);                                             \
-  }                                                                            \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_ATOB(dst_argb_c, kStrideA,                                               \
-           dst_argb_c, kStrideA,                                               \
-           kWidth, NEG kHeight);                                               \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  FMT_ATOB(dst_argb_opt, kStrideA,                                             \
-           dst_argb_opt, kStrideA,                                             \
-           kWidth, NEG kHeight);                                               \
-  for (int i = 0; i < kStrideA * kHeightA; ++i) {                              \
-    EXPECT_EQ(src_argb[i + OFF], dst_argb_opt[i]);                             \
-    EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                                 \
-  }                                                                            \
-  free_aligned_buffer_page_end(src_argb);                                      \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-}
-
-#define TESTSYM(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A)                           \
-    TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A,                              \
-             benchmark_width_ - 4, _Any, +, 0)                                 \
-    TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A,                              \
-             benchmark_width_, _Unaligned, +, 1)                               \
-    TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A,                              \
-             benchmark_width_, _Opt, +, 0)
+  }
 
-TESTSYM(ARGBToARGB, 4, 4, 1)
-TESTSYM(ARGBToBGRA, 4, 4, 1)
-TESTSYM(ARGBToABGR, 4, 4, 1)
-TESTSYM(BGRAToARGB, 4, 4, 1)
-TESTSYM(ABGRToARGB, 4, 4, 1)
+#if defined(ENABLE_FULL_TESTS)
+#define TESTEND(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A)                  \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_ + 1, \
+           _Any, +, 0)                                                        \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_,     \
+           _Unaligned, +, 2)                                                  \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_,     \
+           _Opt, +, 0)
+#else
+#define TESTEND(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A)              \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_, \
+           _Opt, +, 0)
+#endif
 
-TEST_F(LibYUVConvertTest, Test565) {
-  SIMD_ALIGNED(uint8 orig_pixels[256][4]);
-  SIMD_ALIGNED(uint8 pixels565[256][2]);
-
-  for (int i = 0; i < 256; ++i) {
-    for (int j = 0; j < 4; ++j) {
-      orig_pixels[i][j] = i;
-    }
-  }
-  ARGBToRGB565(&orig_pixels[0][0], 0, &pixels565[0][0], 0, 256, 1);
-  uint32 checksum = HashDjb2(&pixels565[0][0], sizeof(pixels565), 5381);
-  EXPECT_EQ(610919429u, checksum);
-}
+TESTEND(ARGBToBGRA, uint8_t, 4, 4, 1)
+TESTEND(ARGBToABGR, uint8_t, 4, 4, 1)
+TESTEND(BGRAToARGB, uint8_t, 4, 4, 1)
+TESTEND(ABGRToARGB, uint8_t, 4, 4, 1)
+TESTEND(AB64ToAR64, uint16_t, 4, 4, 1)
 
 #ifdef HAVE_JPEG
 TEST_F(LibYUVConvertTest, ValidateJpeg) {
   const int kOff = 10;
   const int kMinJpeg = 64;
-  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg ?
-    benchmark_width_ * benchmark_height_ : kMinJpeg;
+  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg
+                             ? benchmark_width_ * benchmark_height_
+                             : kMinJpeg;
   const int kSize = kImageSize + kOff;
   align_buffer_page_end(orig_pixels, kSize);
 
@@ -1190,6 +1799,7 @@ TEST_F(LibYUVConvertTest, ValidateJpeg) {
   // EOI, SOI. Expect pass.
   orig_pixels[0] = 0xff;
   orig_pixels[1] = 0xd8;  // SOI.
+  orig_pixels[2] = 0xff;
   orig_pixels[kSize - kOff + 0] = 0xff;
   orig_pixels[kSize - kOff + 1] = 0xd9;  // EOI.
   for (int times = 0; times < benchmark_iterations_; ++times) {
@@ -1201,8 +1811,9 @@ TEST_F(LibYUVConvertTest, ValidateJpeg) {
 TEST_F(LibYUVConvertTest, ValidateJpegLarge) {
   const int kOff = 10;
   const int kMinJpeg = 64;
-  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg ?
-    benchmark_width_ * benchmark_height_ : kMinJpeg;
+  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg
+                             ? benchmark_width_ * benchmark_height_
+                             : kMinJpeg;
   const int kSize = kImageSize + kOff;
   const int kMultiple = 10;
   const int kBufSize = kImageSize * kMultiple + kOff;
@@ -1215,6 +1826,7 @@ TEST_F(LibYUVConvertTest, ValidateJpegLarge) {
   // EOI, SOI. Expect pass.
   orig_pixels[0] = 0xff;
   orig_pixels[1] = 0xd8;  // SOI.
+  orig_pixels[2] = 0xff;
   orig_pixels[kSize - kOff + 0] = 0xff;
   orig_pixels[kSize - kOff + 1] = 0xd9;  // EOI.
   for (int times = 0; times < benchmark_iterations_; ++times) {
@@ -1226,8 +1838,9 @@ TEST_F(LibYUVConvertTest, ValidateJpegLarge) {
 TEST_F(LibYUVConvertTest, InvalidateJpeg) {
   const int kOff = 10;
   const int kMinJpeg = 64;
-  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg ?
-    benchmark_width_ * benchmark_height_ : kMinJpeg;
+  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg
+                             ? benchmark_width_ * benchmark_height_
+                             : kMinJpeg;
   const int kSize = kImageSize + kOff;
   align_buffer_page_end(orig_pixels, kSize);
 
@@ -1247,6 +1860,7 @@ TEST_F(LibYUVConvertTest, InvalidateJpeg) {
   // SOI but no EOI. Expect fail.
   orig_pixels[0] = 0xff;
   orig_pixels[1] = 0xd8;  // SOI.
+  orig_pixels[2] = 0xff;
   for (int times = 0; times < benchmark_iterations_; ++times) {
     EXPECT_FALSE(ValidateJpeg(orig_pixels, kSize));
   }
@@ -1264,87 +1878,828 @@ TEST_F(LibYUVConvertTest, InvalidateJpeg) {
 TEST_F(LibYUVConvertTest, FuzzJpeg) {
   // SOI but no EOI. Expect fail.
   for (int times = 0; times < benchmark_iterations_; ++times) {
-    const int kSize = fastrand() % 5000 + 2;
+    const int kSize = fastrand() % 5000 + 3;
     align_buffer_page_end(orig_pixels, kSize);
     MemRandomize(orig_pixels, kSize);
 
     // Add SOI so frame will be scanned.
     orig_pixels[0] = 0xff;
     orig_pixels[1] = 0xd8;  // SOI.
+    orig_pixels[2] = 0xff;
     orig_pixels[kSize - 1] = 0xff;
-    ValidateJpeg(orig_pixels, kSize);  // Failure normally expected.
+    ValidateJpeg(orig_pixels,
+                 kSize);  // Failure normally expected.
     free_aligned_buffer_page_end(orig_pixels);
   }
 }
 
-TEST_F(LibYUVConvertTest, MJPGToI420) {
-  const int kOff = 10;
-  const int kMinJpeg = 64;
-  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg ?
-    benchmark_width_ * benchmark_height_ : kMinJpeg;
-  const int kSize = kImageSize + kOff;
-  align_buffer_page_end(orig_pixels, kSize);
-  align_buffer_page_end(dst_y_opt, benchmark_width_ * benchmark_height_);
-  align_buffer_page_end(dst_u_opt,
-                        SUBSAMPLE(benchmark_width_, 2) *
-                        SUBSAMPLE(benchmark_height_, 2));
-  align_buffer_page_end(dst_v_opt,
-                        SUBSAMPLE(benchmark_width_, 2) *
-                        SUBSAMPLE(benchmark_height_, 2));
-
-  // EOI, SOI to make MJPG appear valid.
-  memset(orig_pixels, 0, kSize);
-  orig_pixels[0] = 0xff;
-  orig_pixels[1] = 0xd8;  // SOI.
-  orig_pixels[kSize - kOff + 0] = 0xff;
-  orig_pixels[kSize - kOff + 1] = 0xd9;  // EOI.
+// Test data created in GIMP.  In export jpeg, disable
+// thumbnails etc, choose a subsampling, and use low quality
+// (50) to keep size small. Generated with xxd -i test.jpg
+// test 0 is J400
+static const uint8_t kTest0Jpg[] = {
+    0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 0x4a, 0x46, 0x49, 0x46, 0x00, 0x01,
+    0x01, 0x01, 0x00, 0x48, 0x00, 0x48, 0x00, 0x00, 0xff, 0xdb, 0x00, 0x43,
+    0x00, 0x10, 0x0b, 0x0c, 0x0e, 0x0c, 0x0a, 0x10, 0x0e, 0x0d, 0x0e, 0x12,
+    0x11, 0x10, 0x13, 0x18, 0x28, 0x1a, 0x18, 0x16, 0x16, 0x18, 0x31, 0x23,
+    0x25, 0x1d, 0x28, 0x3a, 0x33, 0x3d, 0x3c, 0x39, 0x33, 0x38, 0x37, 0x40,
+    0x48, 0x5c, 0x4e, 0x40, 0x44, 0x57, 0x45, 0x37, 0x38, 0x50, 0x6d, 0x51,
+    0x57, 0x5f, 0x62, 0x67, 0x68, 0x67, 0x3e, 0x4d, 0x71, 0x79, 0x70, 0x64,
+    0x78, 0x5c, 0x65, 0x67, 0x63, 0xff, 0xc2, 0x00, 0x0b, 0x08, 0x00, 0x10,
+    0x00, 0x20, 0x01, 0x01, 0x11, 0x00, 0xff, 0xc4, 0x00, 0x17, 0x00, 0x01,
+    0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x03, 0x04, 0x01, 0x02, 0xff, 0xda, 0x00, 0x08, 0x01,
+    0x01, 0x00, 0x00, 0x00, 0x01, 0x43, 0x7e, 0xa7, 0x97, 0x57, 0xff, 0xc4,
+    0x00, 0x1b, 0x10, 0x00, 0x03, 0x00, 0x02, 0x03, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x11, 0x00, 0x03,
+    0x10, 0x12, 0x13, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x05,
+    0x02, 0x3b, 0xc0, 0x6f, 0x66, 0x76, 0x56, 0x23, 0x87, 0x99, 0x0d, 0x26,
+    0x62, 0xf6, 0xbf, 0xff, 0xc4, 0x00, 0x1e, 0x10, 0x00, 0x02, 0x01, 0x03,
+    0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x11, 0x21, 0x02, 0x12, 0x32, 0x10, 0x31, 0x71, 0x81, 0xa1, 0xff,
+    0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x06, 0x3f, 0x02, 0x4b, 0xb3, 0x28,
+    0x32, 0xd2, 0xed, 0xf9, 0x1d, 0x3e, 0x13, 0x51, 0x73, 0x83, 0xff, 0xc4,
+    0x00, 0x1c, 0x10, 0x01, 0x01, 0x01, 0x00, 0x02, 0x03, 0x01, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x00, 0x21, 0x51,
+    0x31, 0x61, 0x81, 0xf0, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01,
+    0x3f, 0x21, 0x65, 0x6e, 0x31, 0x86, 0x28, 0xf9, 0x30, 0xdc, 0x27, 0xdb,
+    0xa9, 0x01, 0xf3, 0xde, 0x02, 0xa0, 0xed, 0x1e, 0x34, 0x68, 0x23, 0xf9,
+    0xc6, 0x48, 0x5d, 0x7a, 0x35, 0x02, 0xf5, 0x6f, 0xff, 0xda, 0x00, 0x08,
+    0x01, 0x01, 0x00, 0x00, 0x00, 0x10, 0x35, 0xff, 0xc4, 0x00, 0x1f, 0x10,
+    0x01, 0x00, 0x02, 0x01, 0x04, 0x03, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x11, 0x31, 0x41, 0x61, 0x71, 0x91,
+    0x21, 0x81, 0xd1, 0xb1, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01,
+    0x3f, 0x10, 0x0b, 0x30, 0xe9, 0x58, 0xbe, 0x1a, 0xfd, 0x88, 0xab, 0x8b,
+    0x34, 0x74, 0x80, 0x4b, 0xb5, 0xd5, 0xab, 0xcd, 0x46, 0x96, 0x2e, 0xec,
+    0xbd, 0xaa, 0x78, 0x47, 0x5c, 0x47, 0xa7, 0x30, 0x49, 0xad, 0x88, 0x7c,
+    0x40, 0x74, 0x30, 0xff, 0x00, 0x23, 0x1d, 0x03, 0x0b, 0xb7, 0xd4, 0xff,
+    0xd9};
+static const size_t kTest0JpgLen = 421;
 
-  for (int times = 0; times < benchmark_iterations_; ++times) {
-    int ret = MJPGToI420(orig_pixels, kSize,
-                         dst_y_opt, benchmark_width_,
-                         dst_u_opt, SUBSAMPLE(benchmark_width_, 2),
-                         dst_v_opt, SUBSAMPLE(benchmark_width_, 2),
-                         benchmark_width_, benchmark_height_,
-                         benchmark_width_, benchmark_height_);
-    // Expect failure because image is not really valid.
-    EXPECT_EQ(1, ret);
-  }
-
-  free_aligned_buffer_page_end(dst_y_opt);
-  free_aligned_buffer_page_end(dst_u_opt);
-  free_aligned_buffer_page_end(dst_v_opt);
-  free_aligned_buffer_page_end(orig_pixels);
+// test 1 is J444
+static const uint8_t kTest1Jpg[] = {
+    0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 0x4a, 0x46, 0x49, 0x46, 0x00, 0x01,
+    0x01, 0x01, 0x00, 0x48, 0x00, 0x48, 0x00, 0x00, 0xff, 0xdb, 0x00, 0x43,
+    0x00, 0x10, 0x0b, 0x0c, 0x0e, 0x0c, 0x0a, 0x10, 0x0e, 0x0d, 0x0e, 0x12,
+    0x11, 0x10, 0x13, 0x18, 0x28, 0x1a, 0x18, 0x16, 0x16, 0x18, 0x31, 0x23,
+    0x25, 0x1d, 0x28, 0x3a, 0x33, 0x3d, 0x3c, 0x39, 0x33, 0x38, 0x37, 0x40,
+    0x48, 0x5c, 0x4e, 0x40, 0x44, 0x57, 0x45, 0x37, 0x38, 0x50, 0x6d, 0x51,
+    0x57, 0x5f, 0x62, 0x67, 0x68, 0x67, 0x3e, 0x4d, 0x71, 0x79, 0x70, 0x64,
+    0x78, 0x5c, 0x65, 0x67, 0x63, 0xff, 0xdb, 0x00, 0x43, 0x01, 0x11, 0x12,
+    0x12, 0x18, 0x15, 0x18, 0x2f, 0x1a, 0x1a, 0x2f, 0x63, 0x42, 0x38, 0x42,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0xff, 0xc2, 0x00, 0x11, 0x08, 0x00, 0x10, 0x00, 0x20, 0x03,
+    0x01, 0x11, 0x00, 0x02, 0x11, 0x01, 0x03, 0x11, 0x01, 0xff, 0xc4, 0x00,
+    0x17, 0x00, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x04, 0x01, 0x02, 0xff, 0xc4,
+    0x00, 0x16, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x01, 0x03, 0xff, 0xda,
+    0x00, 0x0c, 0x03, 0x01, 0x00, 0x02, 0x10, 0x03, 0x10, 0x00, 0x00, 0x01,
+    0x40, 0x8f, 0x26, 0xe8, 0xf4, 0xcc, 0xf9, 0x69, 0x2b, 0x1b, 0x2a, 0xcb,
+    0xff, 0xc4, 0x00, 0x1b, 0x10, 0x00, 0x03, 0x00, 0x02, 0x03, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x11,
+    0x00, 0x03, 0x10, 0x12, 0x13, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00,
+    0x01, 0x05, 0x02, 0x3b, 0x80, 0x6f, 0x56, 0x76, 0x56, 0x23, 0x87, 0x99,
+    0x0d, 0x26, 0x62, 0xf6, 0xbf, 0xff, 0xc4, 0x00, 0x19, 0x11, 0x01, 0x00,
+    0x03, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x01, 0x00, 0x10, 0x11, 0x02, 0x12, 0xff, 0xda, 0x00, 0x08,
+    0x01, 0x03, 0x01, 0x01, 0x3f, 0x01, 0xf1, 0x00, 0x27, 0x45, 0xbb, 0x31,
+    0xaf, 0xff, 0xc4, 0x00, 0x1a, 0x11, 0x00, 0x02, 0x03, 0x01, 0x01, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01,
+    0x02, 0x10, 0x11, 0x41, 0x12, 0xff, 0xda, 0x00, 0x08, 0x01, 0x02, 0x01,
+    0x01, 0x3f, 0x01, 0xf6, 0x4b, 0x5f, 0x48, 0xb3, 0x69, 0x63, 0x35, 0x72,
+    0xbf, 0xff, 0xc4, 0x00, 0x1e, 0x10, 0x00, 0x02, 0x01, 0x03, 0x05, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x11,
+    0x21, 0x02, 0x12, 0x32, 0x10, 0x31, 0x71, 0x81, 0xa1, 0xff, 0xda, 0x00,
+    0x08, 0x01, 0x01, 0x00, 0x06, 0x3f, 0x02, 0x4b, 0xb3, 0x28, 0x32, 0xd2,
+    0xed, 0xf9, 0x1d, 0x3e, 0x13, 0x51, 0x73, 0x83, 0xff, 0xc4, 0x00, 0x1c,
+    0x10, 0x01, 0x01, 0x01, 0x00, 0x02, 0x03, 0x01, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x00, 0x21, 0x51, 0x31, 0x61,
+    0x81, 0xf0, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x3f, 0x21,
+    0x75, 0x6e, 0x31, 0x94, 0x28, 0xf9, 0x30, 0xdc, 0x27, 0xdb, 0xa9, 0x01,
+    0xf3, 0xde, 0x02, 0xa0, 0xed, 0x1e, 0x34, 0x68, 0x23, 0xf9, 0xc6, 0x48,
+    0x5d, 0x7a, 0x35, 0x02, 0xf5, 0x6f, 0xff, 0xda, 0x00, 0x0c, 0x03, 0x01,
+    0x00, 0x02, 0x00, 0x03, 0x00, 0x00, 0x00, 0x10, 0x26, 0x61, 0xd4, 0xff,
+    0xc4, 0x00, 0x1a, 0x11, 0x00, 0x03, 0x01, 0x00, 0x03, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x21,
+    0x31, 0x41, 0x51, 0xff, 0xda, 0x00, 0x08, 0x01, 0x03, 0x01, 0x01, 0x3f,
+    0x10, 0x54, 0xa8, 0xbf, 0x50, 0x87, 0xb0, 0x9d, 0x8b, 0xc4, 0x6a, 0x26,
+    0x6b, 0x2a, 0x9c, 0x1f, 0xff, 0xc4, 0x00, 0x18, 0x11, 0x01, 0x01, 0x01,
+    0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x01, 0x00, 0x11, 0x21, 0x51, 0xff, 0xda, 0x00, 0x08, 0x01, 0x02,
+    0x01, 0x01, 0x3f, 0x10, 0x70, 0xe1, 0x3e, 0xd1, 0x8e, 0x0d, 0xe1, 0xb5,
+    0xd5, 0x91, 0x76, 0x43, 0x82, 0x45, 0x4c, 0x7b, 0x7f, 0xff, 0xc4, 0x00,
+    0x1f, 0x10, 0x01, 0x00, 0x02, 0x01, 0x04, 0x03, 0x01, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x11, 0x31, 0x41, 0x61,
+    0x71, 0x91, 0x21, 0x81, 0xd1, 0xb1, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01,
+    0x00, 0x01, 0x3f, 0x10, 0x1b, 0x30, 0xe9, 0x58, 0xbe, 0x1a, 0xfd, 0x8a,
+    0xeb, 0x8b, 0x34, 0x74, 0x80, 0x4b, 0xb5, 0xd5, 0xab, 0xcd, 0x46, 0x96,
+    0x2e, 0xec, 0xbd, 0xaa, 0x78, 0x47, 0x5c, 0x47, 0xa7, 0x30, 0x49, 0xad,
+    0x88, 0x7c, 0x40, 0x74, 0x30, 0xff, 0x00, 0x23, 0x1d, 0x03, 0x0b, 0xb7,
+    0xd4, 0xff, 0xd9};
+static const size_t kTest1JpgLen = 735;
+
+// test 2 is J420
+static const uint8_t kTest2Jpg[] = {
+    0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 0x4a, 0x46, 0x49, 0x46, 0x00, 0x01,
+    0x01, 0x01, 0x00, 0x48, 0x00, 0x48, 0x00, 0x00, 0xff, 0xdb, 0x00, 0x43,
+    0x00, 0x10, 0x0b, 0x0c, 0x0e, 0x0c, 0x0a, 0x10, 0x0e, 0x0d, 0x0e, 0x12,
+    0x11, 0x10, 0x13, 0x18, 0x28, 0x1a, 0x18, 0x16, 0x16, 0x18, 0x31, 0x23,
+    0x25, 0x1d, 0x28, 0x3a, 0x33, 0x3d, 0x3c, 0x39, 0x33, 0x38, 0x37, 0x40,
+    0x48, 0x5c, 0x4e, 0x40, 0x44, 0x57, 0x45, 0x37, 0x38, 0x50, 0x6d, 0x51,
+    0x57, 0x5f, 0x62, 0x67, 0x68, 0x67, 0x3e, 0x4d, 0x71, 0x79, 0x70, 0x64,
+    0x78, 0x5c, 0x65, 0x67, 0x63, 0xff, 0xdb, 0x00, 0x43, 0x01, 0x11, 0x12,
+    0x12, 0x18, 0x15, 0x18, 0x2f, 0x1a, 0x1a, 0x2f, 0x63, 0x42, 0x38, 0x42,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0xff, 0xc2, 0x00, 0x11, 0x08, 0x00, 0x10, 0x00, 0x20, 0x03,
+    0x01, 0x22, 0x00, 0x02, 0x11, 0x01, 0x03, 0x11, 0x01, 0xff, 0xc4, 0x00,
+    0x18, 0x00, 0x00, 0x02, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x05, 0x01, 0x02, 0x04, 0xff,
+    0xc4, 0x00, 0x16, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x01, 0x02, 0xff,
+    0xda, 0x00, 0x0c, 0x03, 0x01, 0x00, 0x02, 0x10, 0x03, 0x10, 0x00, 0x00,
+    0x01, 0x20, 0xe7, 0x28, 0xa3, 0x0b, 0x2e, 0x2d, 0xcf, 0xff, 0xc4, 0x00,
+    0x1b, 0x10, 0x00, 0x03, 0x00, 0x02, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x11, 0x00, 0x03, 0x10,
+    0x12, 0x13, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x05, 0x02,
+    0x3b, 0x80, 0x6f, 0x56, 0x76, 0x56, 0x23, 0x87, 0x99, 0x0d, 0x26, 0x62,
+    0xf6, 0xbf, 0xff, 0xc4, 0x00, 0x17, 0x11, 0x01, 0x00, 0x03, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x01, 0x11, 0x21, 0xff, 0xda, 0x00, 0x08, 0x01, 0x03, 0x01, 0x01, 0x3f,
+    0x01, 0xc8, 0x53, 0xff, 0xc4, 0x00, 0x16, 0x11, 0x01, 0x01, 0x01, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x11, 0x32, 0xff, 0xda, 0x00, 0x08, 0x01, 0x02, 0x01, 0x01, 0x3f,
+    0x01, 0xd2, 0xc7, 0xff, 0xc4, 0x00, 0x1e, 0x10, 0x00, 0x02, 0x01, 0x03,
+    0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x11, 0x21, 0x02, 0x12, 0x32, 0x10, 0x31, 0x71, 0x81, 0xa1, 0xff,
+    0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x06, 0x3f, 0x02, 0x4b, 0xb3, 0x28,
+    0x32, 0xd2, 0xed, 0xf9, 0x1d, 0x3e, 0x13, 0x51, 0x73, 0x83, 0xff, 0xc4,
+    0x00, 0x1c, 0x10, 0x01, 0x01, 0x01, 0x00, 0x02, 0x03, 0x01, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x00, 0x21, 0x51,
+    0x31, 0x61, 0x81, 0xf0, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01,
+    0x3f, 0x21, 0x75, 0x6e, 0x31, 0x94, 0x28, 0xf9, 0x30, 0xdc, 0x27, 0xdb,
+    0xa9, 0x01, 0xf3, 0xde, 0x02, 0xa0, 0xed, 0x1e, 0x34, 0x68, 0x23, 0xf9,
+    0xc6, 0x48, 0x5d, 0x7a, 0x35, 0x02, 0xf5, 0x6f, 0xff, 0xda, 0x00, 0x0c,
+    0x03, 0x01, 0x00, 0x02, 0x00, 0x03, 0x00, 0x00, 0x00, 0x10, 0x13, 0x5f,
+    0xff, 0xc4, 0x00, 0x17, 0x11, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x11,
+    0x21, 0xff, 0xda, 0x00, 0x08, 0x01, 0x03, 0x01, 0x01, 0x3f, 0x10, 0x0e,
+    0xa1, 0x3a, 0x76, 0xff, 0xc4, 0x00, 0x17, 0x11, 0x01, 0x01, 0x01, 0x01,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x01, 0x00, 0x21, 0x11, 0xff, 0xda, 0x00, 0x08, 0x01, 0x02, 0x01, 0x01,
+    0x3f, 0x10, 0x57, 0x0b, 0x08, 0x70, 0xdb, 0xff, 0xc4, 0x00, 0x1f, 0x10,
+    0x01, 0x00, 0x02, 0x01, 0x04, 0x03, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x11, 0x31, 0x41, 0x61, 0x71, 0x91,
+    0x21, 0x81, 0xd1, 0xb1, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01,
+    0x3f, 0x10, 0x1b, 0x30, 0xe9, 0x58, 0xbe, 0x1a, 0xfd, 0x8a, 0xeb, 0x8b,
+    0x34, 0x74, 0x80, 0x4b, 0xb5, 0xd5, 0xab, 0xcd, 0x46, 0x96, 0x2e, 0xec,
+    0xbd, 0xaa, 0x78, 0x47, 0x5c, 0x47, 0xa7, 0x30, 0x49, 0xad, 0x88, 0x7c,
+    0x40, 0x74, 0x30, 0xff, 0x00, 0x23, 0x1d, 0x03, 0x0b, 0xb7, 0xd4, 0xff,
+    0xd9};
+static const size_t kTest2JpgLen = 685;
+
+// test 3 is J422
+static const uint8_t kTest3Jpg[] = {
+    0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 0x4a, 0x46, 0x49, 0x46, 0x00, 0x01,
+    0x01, 0x01, 0x00, 0x48, 0x00, 0x48, 0x00, 0x00, 0xff, 0xdb, 0x00, 0x43,
+    0x00, 0x10, 0x0b, 0x0c, 0x0e, 0x0c, 0x0a, 0x10, 0x0e, 0x0d, 0x0e, 0x12,
+    0x11, 0x10, 0x13, 0x18, 0x28, 0x1a, 0x18, 0x16, 0x16, 0x18, 0x31, 0x23,
+    0x25, 0x1d, 0x28, 0x3a, 0x33, 0x3d, 0x3c, 0x39, 0x33, 0x38, 0x37, 0x40,
+    0x48, 0x5c, 0x4e, 0x40, 0x44, 0x57, 0x45, 0x37, 0x38, 0x50, 0x6d, 0x51,
+    0x57, 0x5f, 0x62, 0x67, 0x68, 0x67, 0x3e, 0x4d, 0x71, 0x79, 0x70, 0x64,
+    0x78, 0x5c, 0x65, 0x67, 0x63, 0xff, 0xdb, 0x00, 0x43, 0x01, 0x11, 0x12,
+    0x12, 0x18, 0x15, 0x18, 0x2f, 0x1a, 0x1a, 0x2f, 0x63, 0x42, 0x38, 0x42,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0xff, 0xc2, 0x00, 0x11, 0x08, 0x00, 0x10, 0x00, 0x20, 0x03,
+    0x01, 0x21, 0x00, 0x02, 0x11, 0x01, 0x03, 0x11, 0x01, 0xff, 0xc4, 0x00,
+    0x17, 0x00, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x04, 0x01, 0x02, 0xff, 0xc4,
+    0x00, 0x17, 0x01, 0x00, 0x03, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x03, 0x00, 0xff,
+    0xda, 0x00, 0x0c, 0x03, 0x01, 0x00, 0x02, 0x10, 0x03, 0x10, 0x00, 0x00,
+    0x01, 0x43, 0x8d, 0x1f, 0xa2, 0xb3, 0xca, 0x1b, 0x57, 0x0f, 0xff, 0xc4,
+    0x00, 0x1b, 0x10, 0x00, 0x03, 0x00, 0x02, 0x03, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x11, 0x00, 0x03,
+    0x10, 0x12, 0x13, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x05,
+    0x02, 0x3b, 0x80, 0x6f, 0x56, 0x76, 0x56, 0x23, 0x87, 0x99, 0x0d, 0x26,
+    0x62, 0xf6, 0xbf, 0xff, 0xc4, 0x00, 0x19, 0x11, 0x00, 0x02, 0x03, 0x01,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x01, 0x02, 0x10, 0x11, 0x21, 0xff, 0xda, 0x00, 0x08, 0x01, 0x03,
+    0x01, 0x01, 0x3f, 0x01, 0x51, 0xce, 0x8c, 0x75, 0xff, 0xc4, 0x00, 0x18,
+    0x11, 0x00, 0x03, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x61, 0x21, 0xff, 0xda,
+    0x00, 0x08, 0x01, 0x02, 0x01, 0x01, 0x3f, 0x01, 0xa6, 0xd9, 0x2f, 0x84,
+    0xe8, 0xf0, 0xff, 0xc4, 0x00, 0x1e, 0x10, 0x00, 0x02, 0x01, 0x03, 0x05,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x11, 0x21, 0x02, 0x12, 0x32, 0x10, 0x31, 0x71, 0x81, 0xa1, 0xff, 0xda,
+    0x00, 0x08, 0x01, 0x01, 0x00, 0x06, 0x3f, 0x02, 0x4b, 0xb3, 0x28, 0x32,
+    0xd2, 0xed, 0xf9, 0x1d, 0x3e, 0x13, 0x51, 0x73, 0x83, 0xff, 0xc4, 0x00,
+    0x1c, 0x10, 0x01, 0x01, 0x01, 0x00, 0x02, 0x03, 0x01, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x00, 0x21, 0x51, 0x31,
+    0x61, 0x81, 0xf0, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x3f,
+    0x21, 0x75, 0x6e, 0x31, 0x94, 0x28, 0xf9, 0x30, 0xdc, 0x27, 0xdb, 0xa9,
+    0x01, 0xf3, 0xde, 0x02, 0xa0, 0xed, 0x1e, 0x34, 0x68, 0x23, 0xf9, 0xc6,
+    0x48, 0x5d, 0x7a, 0x35, 0x02, 0xf5, 0x6f, 0xff, 0xda, 0x00, 0x0c, 0x03,
+    0x01, 0x00, 0x02, 0x00, 0x03, 0x00, 0x00, 0x00, 0x10, 0x2e, 0x45, 0xff,
+    0xc4, 0x00, 0x18, 0x11, 0x00, 0x03, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x21,
+    0x31, 0xff, 0xda, 0x00, 0x08, 0x01, 0x03, 0x01, 0x01, 0x3f, 0x10, 0x53,
+    0x50, 0xba, 0x54, 0xc1, 0x67, 0x4f, 0xff, 0xc4, 0x00, 0x18, 0x11, 0x00,
+    0x03, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x01, 0x11, 0x21, 0x00, 0x10, 0xff, 0xda, 0x00, 0x08,
+    0x01, 0x02, 0x01, 0x01, 0x3f, 0x10, 0x18, 0x81, 0x5c, 0x04, 0x1a, 0xca,
+    0x91, 0xbf, 0xff, 0xc4, 0x00, 0x1f, 0x10, 0x01, 0x00, 0x02, 0x01, 0x04,
+    0x03, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01,
+    0x00, 0x11, 0x31, 0x41, 0x61, 0x71, 0x91, 0x21, 0x81, 0xd1, 0xb1, 0xff,
+    0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x3f, 0x10, 0x1b, 0x30, 0xe9,
+    0x58, 0xbe, 0x1a, 0xfd, 0x8a, 0xeb, 0x8b, 0x34, 0x74, 0x80, 0x4b, 0xb5,
+    0xd5, 0xab, 0xcd, 0x46, 0x96, 0x2e, 0xec, 0xbd, 0xaa, 0x78, 0x47, 0x5c,
+    0x47, 0xa7, 0x30, 0x49, 0xad, 0x88, 0x7c, 0x40, 0x74, 0x30, 0xff, 0x00,
+    0x23, 0x1d, 0x03, 0x0b, 0xb7, 0xd4, 0xff, 0xd9};
+static const size_t kTest3JpgLen = 704;
+
+// test 4 is J422 vertical - not supported
+static const uint8_t kTest4Jpg[] = {
+    0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 0x4a, 0x46, 0x49, 0x46, 0x00, 0x01,
+    0x01, 0x01, 0x00, 0x48, 0x00, 0x48, 0x00, 0x00, 0xff, 0xdb, 0x00, 0x43,
+    0x00, 0x10, 0x0b, 0x0c, 0x0e, 0x0c, 0x0a, 0x10, 0x0e, 0x0d, 0x0e, 0x12,
+    0x11, 0x10, 0x13, 0x18, 0x28, 0x1a, 0x18, 0x16, 0x16, 0x18, 0x31, 0x23,
+    0x25, 0x1d, 0x28, 0x3a, 0x33, 0x3d, 0x3c, 0x39, 0x33, 0x38, 0x37, 0x40,
+    0x48, 0x5c, 0x4e, 0x40, 0x44, 0x57, 0x45, 0x37, 0x38, 0x50, 0x6d, 0x51,
+    0x57, 0x5f, 0x62, 0x67, 0x68, 0x67, 0x3e, 0x4d, 0x71, 0x79, 0x70, 0x64,
+    0x78, 0x5c, 0x65, 0x67, 0x63, 0xff, 0xdb, 0x00, 0x43, 0x01, 0x11, 0x12,
+    0x12, 0x18, 0x15, 0x18, 0x2f, 0x1a, 0x1a, 0x2f, 0x63, 0x42, 0x38, 0x42,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63, 0x63,
+    0x63, 0x63, 0xff, 0xc2, 0x00, 0x11, 0x08, 0x00, 0x10, 0x00, 0x20, 0x03,
+    0x01, 0x12, 0x00, 0x02, 0x11, 0x01, 0x03, 0x11, 0x01, 0xff, 0xc4, 0x00,
+    0x18, 0x00, 0x00, 0x02, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x05, 0x01, 0x02, 0x03, 0xff,
+    0xc4, 0x00, 0x16, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x03, 0xff,
+    0xda, 0x00, 0x0c, 0x03, 0x01, 0x00, 0x02, 0x10, 0x03, 0x10, 0x00, 0x00,
+    0x01, 0xd2, 0x98, 0xe9, 0x03, 0x0c, 0x00, 0x46, 0x21, 0xd9, 0xff, 0xc4,
+    0x00, 0x1b, 0x10, 0x00, 0x03, 0x00, 0x02, 0x03, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x11, 0x00, 0x03,
+    0x10, 0x12, 0x13, 0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x05,
+    0x02, 0x3b, 0x80, 0x6f, 0x56, 0x76, 0x56, 0x23, 0x87, 0x99, 0x0d, 0x26,
+    0x62, 0xf6, 0xbf, 0xff, 0xc4, 0x00, 0x17, 0x11, 0x01, 0x01, 0x01, 0x01,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x11, 0x01, 0x21, 0xff, 0xda, 0x00, 0x08, 0x01, 0x03, 0x01, 0x01,
+    0x3f, 0x01, 0x98, 0xb1, 0xbd, 0x47, 0xff, 0xc4, 0x00, 0x18, 0x11, 0x00,
+    0x03, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x01, 0x12, 0x11, 0x21, 0xff, 0xda, 0x00, 0x08,
+    0x01, 0x02, 0x01, 0x01, 0x3f, 0x01, 0xb6, 0x35, 0xa2, 0xe1, 0x47, 0xff,
+    0xc4, 0x00, 0x1e, 0x10, 0x00, 0x02, 0x01, 0x03, 0x05, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x11, 0x21, 0x02,
+    0x12, 0x32, 0x10, 0x31, 0x71, 0x81, 0xa1, 0xff, 0xda, 0x00, 0x08, 0x01,
+    0x01, 0x00, 0x06, 0x3f, 0x02, 0x4b, 0xb3, 0x28, 0x32, 0xd2, 0xed, 0xf9,
+    0x1d, 0x3e, 0x13, 0x51, 0x73, 0x83, 0xff, 0xc4, 0x00, 0x1c, 0x10, 0x01,
+    0x01, 0x01, 0x00, 0x02, 0x03, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x01, 0x11, 0x00, 0x21, 0x51, 0x31, 0x61, 0x81, 0xf0,
+    0xff, 0xda, 0x00, 0x08, 0x01, 0x01, 0x00, 0x01, 0x3f, 0x21, 0x75, 0x6e,
+    0x31, 0x94, 0x28, 0xf9, 0x30, 0xdc, 0x27, 0xdb, 0xa9, 0x01, 0xf3, 0xde,
+    0x02, 0xa0, 0xed, 0x1e, 0x34, 0x68, 0x23, 0xf9, 0xc6, 0x48, 0x5d, 0x7a,
+    0x35, 0x02, 0xf5, 0x6f, 0xff, 0xda, 0x00, 0x0c, 0x03, 0x01, 0x00, 0x02,
+    0x00, 0x03, 0x00, 0x00, 0x00, 0x10, 0x24, 0xaf, 0xff, 0xc4, 0x00, 0x19,
+    0x11, 0x00, 0x03, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x11, 0x51, 0x21, 0x31, 0xff,
+    0xda, 0x00, 0x08, 0x01, 0x03, 0x01, 0x01, 0x3f, 0x10, 0x59, 0x11, 0xca,
+    0x42, 0x60, 0x9f, 0x69, 0xff, 0xc4, 0x00, 0x19, 0x11, 0x00, 0x02, 0x03,
+    0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x01, 0x11, 0x21, 0x31, 0x61, 0xff, 0xda, 0x00, 0x08, 0x01,
+    0x02, 0x01, 0x01, 0x3f, 0x10, 0xb0, 0xd7, 0x27, 0x51, 0xb6, 0x41, 0xff,
+    0xc4, 0x00, 0x1f, 0x10, 0x01, 0x00, 0x02, 0x01, 0x04, 0x03, 0x01, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x11, 0x31,
+    0x41, 0x61, 0x71, 0x91, 0x21, 0x81, 0xd1, 0xb1, 0xff, 0xda, 0x00, 0x08,
+    0x01, 0x01, 0x00, 0x01, 0x3f, 0x10, 0x1b, 0x30, 0xe9, 0x58, 0xbe, 0x1a,
+    0xfd, 0x8a, 0xeb, 0x8b, 0x34, 0x74, 0x80, 0x4b, 0xb5, 0xd5, 0xab, 0xcd,
+    0x46, 0x96, 0x2e, 0xec, 0xbd, 0xaa, 0x78, 0x47, 0x5c, 0x47, 0xa7, 0x30,
+    0x49, 0xad, 0x88, 0x7c, 0x40, 0x74, 0x30, 0xff, 0x00, 0x23, 0x1d, 0x03,
+    0x0b, 0xb7, 0xd4, 0xff, 0xd9};
+static const size_t kTest4JpgLen = 701;
+
+TEST_F(LibYUVConvertTest, TestMJPGSize) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  printf("test jpeg size %d x %d\n", width, height);
 }
 
-TEST_F(LibYUVConvertTest, MJPGToARGB) {
-  const int kOff = 10;
-  const int kMinJpeg = 64;
-  const int kImageSize = benchmark_width_ * benchmark_height_ >= kMinJpeg ?
-    benchmark_width_ * benchmark_height_ : kMinJpeg;
-  const int kSize = kImageSize + kOff;
-  align_buffer_page_end(orig_pixels, kSize);
-  align_buffer_page_end(dst_argb_opt, benchmark_width_ * benchmark_height_ * 4);
+TEST_F(LibYUVConvertTest, TestMJPGToI420) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
 
-  // EOI, SOI to make MJPG appear valid.
-  memset(orig_pixels, 0, kSize);
-  orig_pixels[0] = 0xff;
-  orig_pixels[1] = 0xd8;  // SOI.
-  orig_pixels[kSize - kOff + 0] = 0xff;
-  orig_pixels[kSize - kOff + 1] = 0xd9;  // EOI.
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
 
-  for (int times = 0; times < benchmark_iterations_; ++times) {
-    int ret = MJPGToARGB(orig_pixels, kSize,
-                         dst_argb_opt, benchmark_width_ * 4,
-                         benchmark_width_, benchmark_height_,
-                         benchmark_width_, benchmark_height_);
-    // Expect failure because image is not really valid.
-    EXPECT_EQ(1, ret);
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_u, half_width * half_height);
+  align_buffer_page_end(dst_v, half_width * half_height);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToI420(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_u, half_width,
+                     dst_v, half_width, width, height, width, height);
   }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
 
-  free_aligned_buffer_page_end(dst_argb_opt);
-  free_aligned_buffer_page_end(orig_pixels);
+  // Test result matches known hash value.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  uint32_t dst_u_hash = HashDjb2(dst_u, half_width * half_height, 5381);
+  uint32_t dst_v_hash = HashDjb2(dst_v, half_width * half_height, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_u_hash, 2501859930u);
+  EXPECT_EQ(dst_v_hash, 2126459123u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_u);
+  free_aligned_buffer_page_end(dst_v);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToI420_NV21) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  // Convert to NV21
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV21(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_vu,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Convert to I420
+  align_buffer_page_end(dst2_y, width * height);
+  align_buffer_page_end(dst2_u, half_width * half_height);
+  align_buffer_page_end(dst2_v, half_width * half_height);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToI420(kTest2Jpg, kTest2JpgLen, dst2_y, width, dst2_u, half_width,
+                     dst2_v, half_width, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Convert I420 to NV21
+  align_buffer_page_end(dst3_y, width * height);
+  align_buffer_page_end(dst3_vu, half_width * half_height * 2);
+
+  I420ToNV21(dst2_y, width, dst2_u, half_width, dst2_v, half_width, dst3_y,
+             width, dst3_vu, half_width * 2, width, height);
+
+  for (int i = 0; i < width * height; ++i) {
+    EXPECT_EQ(dst_y[i], dst3_y[i]);
+  }
+  for (int i = 0; i < half_width * half_height * 2; ++i) {
+    EXPECT_EQ(dst_vu[i], dst3_vu[i]);
+    EXPECT_EQ(dst_vu[i], dst3_vu[i]);
+  }
+
+  free_aligned_buffer_page_end(dst3_y);
+  free_aligned_buffer_page_end(dst3_vu);
+
+  free_aligned_buffer_page_end(dst2_y);
+  free_aligned_buffer_page_end(dst2_u);
+  free_aligned_buffer_page_end(dst2_v);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToI420_NV12) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  // Convert to NV12
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Convert to I420
+  align_buffer_page_end(dst2_y, width * height);
+  align_buffer_page_end(dst2_u, half_width * half_height);
+  align_buffer_page_end(dst2_v, half_width * half_height);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToI420(kTest2Jpg, kTest2JpgLen, dst2_y, width, dst2_u, half_width,
+                     dst2_v, half_width, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Convert I420 to NV12
+  align_buffer_page_end(dst3_y, width * height);
+  align_buffer_page_end(dst3_uv, half_width * half_height * 2);
+
+  I420ToNV12(dst2_y, width, dst2_u, half_width, dst2_v, half_width, dst3_y,
+             width, dst3_uv, half_width * 2, width, height);
+
+  for (int i = 0; i < width * height; ++i) {
+    EXPECT_EQ(dst_y[i], dst3_y[i]);
+  }
+  for (int i = 0; i < half_width * half_height * 2; ++i) {
+    EXPECT_EQ(dst_uv[i], dst3_uv[i]);
+    EXPECT_EQ(dst_uv[i], dst3_uv[i]);
+  }
+
+  free_aligned_buffer_page_end(dst3_y);
+  free_aligned_buffer_page_end(dst3_uv);
+
+  free_aligned_buffer_page_end(dst2_y);
+  free_aligned_buffer_page_end(dst2_u);
+  free_aligned_buffer_page_end(dst2_v);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToNV21_420) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV21(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  uint32_t dst_uv_hash = HashDjb2(dst_uv, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_uv_hash, 1069662856u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToNV12_420) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_vu_hash, 1069662856u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
+// TODO(fbarchard): Improve test to compare against I422, not checksum
+TEST_F(LibYUVConvertTest, DISABLED_TestMJPGToNV21_422) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest3Jpg, kTest3JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV21(kTest3Jpg, kTest3JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  uint32_t dst_uv_hash = HashDjb2(dst_uv, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_uv_hash, 493520167u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVConvertTest, DISABLED_TestMJPGToNV12_422) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest3Jpg, kTest3JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest3Jpg, kTest3JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_vu_hash, 493520167u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToNV21_400) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest0Jpg, kTest0JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV21(kTest0Jpg, kTest0JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  uint32_t dst_uv_hash = HashDjb2(dst_uv, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 330644005u);
+  EXPECT_EQ(dst_uv_hash, 135214341u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToNV12_400) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest0Jpg, kTest0JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest0Jpg, kTest0JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 330644005u);
+  EXPECT_EQ(dst_vu_hash, 135214341u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToNV21_444) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest1Jpg, kTest1JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV21(kTest1Jpg, kTest1JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  uint32_t dst_uv_hash = HashDjb2(dst_uv, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_uv_hash, 506143297u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToNV12_444) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest1Jpg, kTest1JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest1Jpg, kTest1JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_vu_hash, 506143297u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
+TEST_F(LibYUVConvertTest, TestMJPGToARGB) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest3Jpg, kTest3JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_argb, width * height * 4);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToARGB(kTest3Jpg, kTest3JpgLen, dst_argb, width * 4, width,
+                     height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value.
+  uint32_t dst_argb_hash = HashDjb2(dst_argb, width * height, 5381);
+#ifdef LIBYUV_UNLIMITED_DATA
+  EXPECT_EQ(dst_argb_hash, 3900633302u);
+#else
+  EXPECT_EQ(dst_argb_hash, 2355976473u);
+#endif
+
+  free_aligned_buffer_page_end(dst_argb);
+}
+
+static int ShowJPegInfo(const uint8_t* sample, size_t sample_size) {
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+
+  int width = mjpeg_decoder.GetWidth();
+  int height = mjpeg_decoder.GetHeight();
+
+  // YUV420
+  if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+      mjpeg_decoder.GetNumComponents() == 3 &&
+      mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+      mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+      mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+      mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+      mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+      mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+    printf("JPeg is J420, %dx%d %d bytes\n", width, height,
+           static_cast<int>(sample_size));
+    // YUV422
+  } else if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+             mjpeg_decoder.GetNumComponents() == 3 &&
+             mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+             mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+             mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+    printf("JPeg is J422, %dx%d %d bytes\n", width, height,
+           static_cast<int>(sample_size));
+    // YUV444
+  } else if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+             mjpeg_decoder.GetNumComponents() == 3 &&
+             mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+             mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+             mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+    printf("JPeg is J444, %dx%d %d bytes\n", width, height,
+           static_cast<int>(sample_size));
+    // YUV400
+  } else if (mjpeg_decoder.GetColorSpace() ==
+                 MJpegDecoder::kColorSpaceGrayscale &&
+             mjpeg_decoder.GetNumComponents() == 1 &&
+             mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+             mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+    printf("JPeg is J400, %dx%d %d bytes\n", width, height,
+           static_cast<int>(sample_size));
+  } else {
+    // Unknown colorspace.
+    printf("JPeg is Unknown colorspace.\n");
+  }
+  mjpeg_decoder.UnloadFrame();
+  return ret;
 }
 
+TEST_F(LibYUVConvertTest, TestMJPGInfo) {
+  EXPECT_EQ(1, ShowJPegInfo(kTest0Jpg, kTest0JpgLen));
+  EXPECT_EQ(1, ShowJPegInfo(kTest1Jpg, kTest1JpgLen));
+  EXPECT_EQ(1, ShowJPegInfo(kTest2Jpg, kTest2JpgLen));
+  EXPECT_EQ(1, ShowJPegInfo(kTest3Jpg, kTest3JpgLen));
+  EXPECT_EQ(1, ShowJPegInfo(kTest4Jpg,
+                            kTest4JpgLen));  // Valid but unsupported.
+}
 #endif  // HAVE_JPEG
 
 TEST_F(LibYUVConvertTest, NV12Crop) {
@@ -1353,66 +2708,53 @@ TEST_F(LibYUVConvertTest, NV12Crop) {
   const int kWidth = benchmark_width_;
   const int kHeight = benchmark_height_;
   const int crop_y =
-    ((benchmark_height_ - (benchmark_height_ * 360 / 480)) / 2 + 1) & ~1;
+      ((benchmark_height_ - (benchmark_height_ * 360 / 480)) / 2 + 1) & ~1;
   const int kDestWidth = benchmark_width_;
   const int kDestHeight = benchmark_height_ - crop_y * 2;
   const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);
-  const int sample_size = kWidth * kHeight +
-    kStrideUV *
-    SUBSAMPLE(kHeight, SUBSAMP_Y) * 2;
+  const int sample_size =
+      kWidth * kHeight + kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2;
   align_buffer_page_end(src_y, sample_size);
-  uint8* src_uv = src_y + kWidth * kHeight;
+  uint8_t* src_uv = src_y + kWidth * kHeight;
 
   align_buffer_page_end(dst_y, kDestWidth * kDestHeight);
-  align_buffer_page_end(dst_u,
-                  SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                  SUBSAMPLE(kDestHeight, SUBSAMP_Y));
-  align_buffer_page_end(dst_v,
-                  SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                  SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  align_buffer_page_end(dst_u, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
+                                   SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  align_buffer_page_end(dst_v, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
+                                   SUBSAMPLE(kDestHeight, SUBSAMP_Y));
 
   align_buffer_page_end(dst_y_2, kDestWidth * kDestHeight);
-  align_buffer_page_end(dst_u_2,
-                  SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                  SUBSAMPLE(kDestHeight, SUBSAMP_Y));
-  align_buffer_page_end(dst_v_2,
-                  SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                  SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  align_buffer_page_end(dst_u_2, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
+                                     SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  align_buffer_page_end(dst_v_2, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
+                                     SUBSAMPLE(kDestHeight, SUBSAMP_Y));
 
   for (int i = 0; i < kHeight * kWidth; ++i) {
     src_y[i] = (fastrand() & 0xff);
   }
-  for (int i = 0; i < (SUBSAMPLE(kHeight, SUBSAMP_Y) *
-       kStrideUV) * 2; ++i) {
+  for (int i = 0; i < (SUBSAMPLE(kHeight, SUBSAMP_Y) * kStrideUV) * 2; ++i) {
     src_uv[i] = (fastrand() & 0xff);
   }
   memset(dst_y, 1, kDestWidth * kDestHeight);
-  memset(dst_u, 2, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                   SUBSAMPLE(kDestHeight, SUBSAMP_Y));
-  memset(dst_v, 3, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                   SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  memset(dst_u, 2,
+         SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  memset(dst_v, 3,
+         SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y));
   memset(dst_y_2, 1, kDestWidth * kDestHeight);
-  memset(dst_u_2, 2, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                     SUBSAMPLE(kDestHeight, SUBSAMP_Y));
-  memset(dst_v_2, 3, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
-                     SUBSAMPLE(kDestHeight, SUBSAMP_Y));
-
-  ConvertToI420(src_y, sample_size,
-                dst_y_2, kDestWidth,
-                dst_u_2, SUBSAMPLE(kDestWidth, SUBSAMP_X),
-                dst_v_2, SUBSAMPLE(kDestWidth, SUBSAMP_X),
-                0, crop_y,
-                kWidth, kHeight,
-                kDestWidth, kDestHeight,
-                libyuv::kRotate0, libyuv::FOURCC_NV12);
+  memset(dst_u_2, 2,
+         SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  memset(dst_v_2, 3,
+         SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+
+  ConvertToI420(src_y, sample_size, dst_y_2, kDestWidth, dst_u_2,
+                SUBSAMPLE(kDestWidth, SUBSAMP_X), dst_v_2,
+                SUBSAMPLE(kDestWidth, SUBSAMP_X), 0, crop_y, kWidth, kHeight,
+                kDestWidth, kDestHeight, libyuv::kRotate0, libyuv::FOURCC_NV12);
 
   NV12ToI420(src_y + crop_y * kWidth, kWidth,
-             src_uv + (crop_y / 2) * kStrideUV * 2,
-               kStrideUV * 2,
-             dst_y, kDestWidth,
-             dst_u, SUBSAMPLE(kDestWidth, SUBSAMP_X),
-             dst_v, SUBSAMPLE(kDestWidth, SUBSAMP_X),
-             kDestWidth, kDestHeight);
+             src_uv + (crop_y / 2) * kStrideUV * 2, kStrideUV * 2, dst_y,
+             kDestWidth, dst_u, SUBSAMPLE(kDestWidth, SUBSAMP_X), dst_v,
+             SUBSAMPLE(kDestWidth, SUBSAMP_X), kDestWidth, kDestHeight);
 
   for (int i = 0; i < kDestHeight; ++i) {
     for (int j = 0; j < kDestWidth; ++j) {
@@ -1440,33 +2782,99 @@ TEST_F(LibYUVConvertTest, NV12Crop) {
   free_aligned_buffer_page_end(src_y);
 }
 
+TEST_F(LibYUVConvertTest, I420CropOddY) {
+  const int SUBSAMP_X = 2;
+  const int SUBSAMP_Y = 2;
+  const int kWidth = benchmark_width_;
+  const int kHeight = benchmark_height_;
+  const int crop_y = benchmark_height_ > 1 ? 1 : 0;
+  const int kDestWidth = benchmark_width_;
+  const int kDestHeight = benchmark_height_ - crop_y * 2;
+  const int kStrideU = SUBSAMPLE(kWidth, SUBSAMP_X);
+  const int kStrideV = SUBSAMPLE(kWidth, SUBSAMP_X);
+  const int sample_size = kWidth * kHeight +
+                          kStrideU * SUBSAMPLE(kHeight, SUBSAMP_Y) +
+                          kStrideV * SUBSAMPLE(kHeight, SUBSAMP_Y);
+  align_buffer_page_end(src_y, sample_size);
+  uint8_t* src_u = src_y + kWidth * kHeight;
+  uint8_t* src_v = src_u + kStrideU * SUBSAMPLE(kHeight, SUBSAMP_Y);
+
+  align_buffer_page_end(dst_y, kDestWidth * kDestHeight);
+  align_buffer_page_end(dst_u, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
+                                   SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  align_buffer_page_end(dst_v, SUBSAMPLE(kDestWidth, SUBSAMP_X) *
+                                   SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+
+  for (int i = 0; i < kHeight * kWidth; ++i) {
+    src_y[i] = (fastrand() & 0xff);
+  }
+  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y) * kStrideU; ++i) {
+    src_u[i] = (fastrand() & 0xff);
+  }
+  for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y) * kStrideV; ++i) {
+    src_v[i] = (fastrand() & 0xff);
+  }
+  memset(dst_y, 1, kDestWidth * kDestHeight);
+  memset(dst_u, 2,
+         SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+  memset(dst_v, 3,
+         SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y));
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ConvertToI420(src_y, sample_size, dst_y, kDestWidth, dst_u,
+                  SUBSAMPLE(kDestWidth, SUBSAMP_X), dst_v,
+                  SUBSAMPLE(kDestWidth, SUBSAMP_X), 0, crop_y, kWidth, kHeight,
+                  kDestWidth, kDestHeight, libyuv::kRotate0,
+                  libyuv::FOURCC_I420);
+  }
+
+  for (int i = 0; i < kDestHeight; ++i) {
+    for (int j = 0; j < kDestWidth; ++j) {
+      EXPECT_EQ(src_y[crop_y * kWidth + i * kWidth + j],
+                dst_y[i * kDestWidth + j]);
+    }
+  }
+  for (int i = 0; i < SUBSAMPLE(kDestHeight, SUBSAMP_Y); ++i) {
+    for (int j = 0; j < SUBSAMPLE(kDestWidth, SUBSAMP_X); ++j) {
+      EXPECT_EQ(src_u[(crop_y / 2 + i) * kStrideU + j],
+                dst_u[i * SUBSAMPLE(kDestWidth, SUBSAMP_X) + j]);
+    }
+  }
+  for (int i = 0; i < SUBSAMPLE(kDestHeight, SUBSAMP_Y); ++i) {
+    for (int j = 0; j < SUBSAMPLE(kDestWidth, SUBSAMP_X); ++j) {
+      EXPECT_EQ(src_v[(crop_y / 2 + i) * kStrideV + j],
+                dst_v[i * SUBSAMPLE(kDestWidth, SUBSAMP_X) + j]);
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_u);
+  free_aligned_buffer_page_end(dst_v);
+  free_aligned_buffer_page_end(src_y);
+}
+
 TEST_F(LibYUVConvertTest, TestYToARGB) {
-  uint8 y[32];
-  uint8 expectedg[32];
+  uint8_t y[32];
+  uint8_t expectedg[32];
   for (int i = 0; i < 32; ++i) {
     y[i] = i * 5 + 17;
     expectedg[i] = static_cast<int>((y[i] - 16) * 1.164f + 0.5f);
   }
-  uint8 argb[32 * 4];
+  uint8_t argb[32 * 4];
   YToARGB(y, 0, argb, 0, 32, 1);
 
   for (int i = 0; i < 32; ++i) {
     printf("%2d %d: %d <-> %d,%d,%d,%d\n", i, y[i], expectedg[i],
-           argb[i * 4 + 0],
-           argb[i * 4 + 1],
-           argb[i * 4 + 2],
-           argb[i * 4 + 3]);
+           argb[i * 4 + 0], argb[i * 4 + 1], argb[i * 4 + 2], argb[i * 4 + 3]);
   }
   for (int i = 0; i < 32; ++i) {
     EXPECT_EQ(expectedg[i], argb[i * 4 + 0]);
   }
 }
 
-static const uint8 kNoDither4x4[16] = {
-  0, 0, 0, 0,
-  0, 0, 0, 0,
-  0, 0, 0, 0,
-  0, 0, 0, 0,
+static const uint8_t kNoDither4x4[16] = {
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
 };
 
 TEST_F(LibYUVConvertTest, TestNoDither) {
@@ -1477,12 +2885,11 @@ TEST_F(LibYUVConvertTest, TestNoDither) {
   MemRandomize(src_argb, benchmark_width_ * benchmark_height_ * 4);
   MemRandomize(dst_rgb565, benchmark_width_ * benchmark_height_ * 2);
   MemRandomize(dst_rgb565dither, benchmark_width_ * benchmark_height_ * 2);
-  ARGBToRGB565(src_argb, benchmark_width_ * 4,
-               dst_rgb565, benchmark_width_ * 2,
+  ARGBToRGB565(src_argb, benchmark_width_ * 4, dst_rgb565, benchmark_width_ * 2,
                benchmark_width_, benchmark_height_);
-  ARGBToRGB565Dither(src_argb, benchmark_width_ * 4,
-                     dst_rgb565dither, benchmark_width_ * 2,
-                     kNoDither4x4, benchmark_width_, benchmark_height_);
+  ARGBToRGB565Dither(src_argb, benchmark_width_ * 4, dst_rgb565dither,
+                     benchmark_width_ * 2, kNoDither4x4, benchmark_width_,
+                     benchmark_height_);
   for (int i = 0; i < benchmark_width_ * benchmark_height_ * 2; ++i) {
     EXPECT_EQ(dst_rgb565[i], dst_rgb565dither[i]);
   }
@@ -1493,11 +2900,8 @@ TEST_F(LibYUVConvertTest, TestNoDither) {
 }
 
 // Ordered 4x4 dither for 888 to 565.  Values from 0 to 7.
-static const uint8 kDither565_4x4[16] = {
-  0, 4, 1, 5,
-  6, 2, 7, 3,
-  1, 5, 0, 4,
-  7, 3, 6, 2,
+static const uint8_t kDither565_4x4[16] = {
+    0, 4, 1, 5, 6, 2, 7, 3, 1, 5, 0, 4, 7, 3, 6, 2,
 };
 
 TEST_F(LibYUVConvertTest, TestDither) {
@@ -1513,18 +2917,15 @@ TEST_F(LibYUVConvertTest, TestDither) {
   MemRandomize(dst_rgb565dither, benchmark_width_ * benchmark_height_ * 2);
   MemRandomize(dst_argb, benchmark_width_ * benchmark_height_ * 4);
   MemRandomize(dst_argbdither, benchmark_width_ * benchmark_height_ * 4);
-  ARGBToRGB565(src_argb, benchmark_width_ * 4,
-               dst_rgb565, benchmark_width_ * 2,
-               benchmark_width_, benchmark_height_);
-  ARGBToRGB565Dither(src_argb, benchmark_width_ * 4,
-                     dst_rgb565dither, benchmark_width_ * 2,
-                     kDither565_4x4, benchmark_width_, benchmark_height_);
-  RGB565ToARGB(dst_rgb565, benchmark_width_ * 2,
-               dst_argb, benchmark_width_ * 4,
+  ARGBToRGB565(src_argb, benchmark_width_ * 4, dst_rgb565, benchmark_width_ * 2,
                benchmark_width_, benchmark_height_);
-  RGB565ToARGB(dst_rgb565dither, benchmark_width_ * 2,
-               dst_argbdither, benchmark_width_ * 4,
+  ARGBToRGB565Dither(src_argb, benchmark_width_ * 4, dst_rgb565dither,
+                     benchmark_width_ * 2, kDither565_4x4, benchmark_width_,
+                     benchmark_height_);
+  RGB565ToARGB(dst_rgb565, benchmark_width_ * 2, dst_argb, benchmark_width_ * 4,
                benchmark_width_, benchmark_height_);
+  RGB565ToARGB(dst_rgb565dither, benchmark_width_ * 2, dst_argbdither,
+               benchmark_width_ * 4, benchmark_width_, benchmark_height_);
 
   for (int i = 0; i < benchmark_width_ * benchmark_height_ * 4; ++i) {
     EXPECT_NEAR(dst_argb[i], dst_argbdither[i], 9);
@@ -1537,325 +2938,1593 @@ TEST_F(LibYUVConvertTest, TestDither) {
 }
 
 #define TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
-                       YALIGN, W1280, DIFF, N, NEG, OFF, FMT_C, BPP_C)         \
-TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##Dither##N) {                  \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                     \
-  const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                        \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);               \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_v, kSizeUV + OFF);                                 \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeight + OFF);                 \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeight + OFF);               \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    src_y[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  for (int i = 0; i < kSizeUV; ++i) {                                          \
-    src_u[i + OFF] = (fastrand() & 0xff);                                      \
-    src_v[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                             \
-  memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                         \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  FMT_PLANAR##To##FMT_B##Dither(src_y + OFF, kWidth,                           \
-                        src_u + OFF, kStrideUV,                                \
-                        src_v + OFF, kStrideUV,                                \
-                        dst_argb_c + OFF, kStrideB,                            \
-                        NULL, kWidth, NEG kHeight);                            \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_PLANAR##To##FMT_B##Dither(src_y + OFF, kWidth,                         \
-                          src_u + OFF, kStrideUV,                              \
-                          src_v + OFF, kStrideUV,                              \
-                          dst_argb_opt + OFF, kStrideB,                        \
-                          NULL, kWidth, NEG kHeight);                          \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  /* Convert to ARGB so 565 is expanded to bytes that can be compared. */      \
-  align_buffer_page_end(dst_argb32_c, kWidth * BPP_C  * kHeight);              \
-  align_buffer_page_end(dst_argb32_opt, kWidth * BPP_C  * kHeight);            \
-  memset(dst_argb32_c, 2, kWidth * BPP_C  * kHeight);                          \
-  memset(dst_argb32_opt, 102, kWidth * BPP_C  * kHeight);                      \
-  FMT_B##To##FMT_C(dst_argb_c + OFF, kStrideB,                                 \
-                   dst_argb32_c, kWidth * BPP_C ,                              \
-                   kWidth, kHeight);                                           \
-  FMT_B##To##FMT_C(dst_argb_opt + OFF, kStrideB,                               \
-                   dst_argb32_opt, kWidth * BPP_C ,                            \
-                   kWidth, kHeight);                                           \
-  for (int i = 0; i < kWidth * BPP_C * kHeight; ++i) {                         \
-    int abs_diff =                                                             \
-        abs(static_cast<int>(dst_argb32_c[i]) -                                \
-            static_cast<int>(dst_argb32_opt[i]));                              \
-    if (abs_diff > max_diff) {                                                 \
-      max_diff = abs_diff;                                                     \
+                        YALIGN, W1280, N, NEG, OFF, FMT_C, BPP_C)              \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##Dither##N) {                \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);             \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                       \
+    align_buffer_page_end(src_u, kSizeUV + OFF);                               \
+    align_buffer_page_end(src_v, kSizeUV + OFF);                               \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + OFF);              \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      src_y[i + OFF] = (fastrand() & 0xff);                                    \
     }                                                                          \
-  }                                                                            \
-  EXPECT_LE(max_diff, DIFF);                                                   \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-  free_aligned_buffer_page_end(dst_argb32_c);                                  \
-  free_aligned_buffer_page_end(dst_argb32_opt);                                \
-}
-
-#define TESTPLANARTOBD(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
-                      YALIGN, DIFF, FMT_C, BPP_C)                              \
-    TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_ - 4, DIFF, _Any, +, 0, FMT_C, BPP_C)          \
-    TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Unaligned, +, 1, FMT_C, BPP_C)        \
-    TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Invert, -, 0, FMT_C, BPP_C)           \
-    TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,     \
-        YALIGN, benchmark_width_, DIFF, _Opt, +, 0, FMT_C, BPP_C)
-
-TESTPLANARTOBD(I420, 2, 2, RGB565, 2, 2, 1, 9, ARGB, 4)
-
-#define TESTPTOB(NAME, UYVYTOI420, UYVYTONV12)                                 \
-TEST_F(LibYUVConvertTest, NAME) {                                              \
-  const int kWidth = benchmark_width_;                                         \
-  const int kHeight = benchmark_height_;                                       \
-                                                                               \
-  align_buffer_page_end(orig_uyvy,                                             \
-                  4 * SUBSAMPLE(kWidth, 2) * kHeight);                         \
-  align_buffer_page_end(orig_y, kWidth * kHeight);                             \
-  align_buffer_page_end(orig_u,                                                \
-                  SUBSAMPLE(kWidth, 2) *                                       \
-                  SUBSAMPLE(kHeight, 2));                                      \
-  align_buffer_page_end(orig_v,                                                \
-                  SUBSAMPLE(kWidth, 2) *                                       \
-                  SUBSAMPLE(kHeight, 2));                                      \
-                                                                               \
-  align_buffer_page_end(dst_y_orig, kWidth * kHeight);                         \
-  align_buffer_page_end(dst_uv_orig, 2 *                                       \
-                  SUBSAMPLE(kWidth, 2) *                                       \
-                  SUBSAMPLE(kHeight, 2));                                      \
-                                                                               \
-  align_buffer_page_end(dst_y, kWidth * kHeight);                              \
-  align_buffer_page_end(dst_uv, 2 *                                            \
-                  SUBSAMPLE(kWidth, 2) *                                       \
-                  SUBSAMPLE(kHeight, 2));                                      \
-                                                                               \
-  MemRandomize(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2) * kHeight);                 \
-                                                                               \
-  /* Convert UYVY to NV12 in 2 steps for reference */                          \
-  libyuv::UYVYTOI420(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2),                      \
-                     orig_y, kWidth,                                           \
-                     orig_u, SUBSAMPLE(kWidth, 2),                             \
-                     orig_v, SUBSAMPLE(kWidth, 2),                             \
-                     kWidth, kHeight);                                         \
-  libyuv::I420ToNV12(orig_y, kWidth,                                           \
-                     orig_u, SUBSAMPLE(kWidth, 2),                             \
-                     orig_v, SUBSAMPLE(kWidth, 2),                             \
-                     dst_y_orig, kWidth,                                       \
-                     dst_uv_orig, 2 * SUBSAMPLE(kWidth, 2),                    \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      src_u[i + OFF] = (fastrand() & 0xff);                                    \
+      src_v[i + OFF] = (fastrand() & 0xff);                                    \
+    }                                                                          \
+    memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                           \
+    memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                       \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B##Dither(src_y + OFF, kWidth, src_u + OFF, kStrideUV, \
+                                  src_v + OFF, kStrideUV, dst_argb_c + OFF,    \
+                                  kStrideB, NULL, kWidth, NEG kHeight);        \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B##Dither(                                           \
+          src_y + OFF, kWidth, src_u + OFF, kStrideUV, src_v + OFF, kStrideUV, \
+          dst_argb_opt + OFF, kStrideB, NULL, kWidth, NEG kHeight);            \
+    }                                                                          \
+    /* Convert to ARGB so 565 is expanded to bytes that can be compared. */    \
+    align_buffer_page_end(dst_argb32_c, kWidth* BPP_C* kHeight);               \
+    align_buffer_page_end(dst_argb32_opt, kWidth* BPP_C* kHeight);             \
+    memset(dst_argb32_c, 2, kWidth* BPP_C* kHeight);                           \
+    memset(dst_argb32_opt, 102, kWidth* BPP_C* kHeight);                       \
+    FMT_B##To##FMT_C(dst_argb_c + OFF, kStrideB, dst_argb32_c, kWidth * BPP_C, \
                      kWidth, kHeight);                                         \
-                                                                               \
-  /* Convert to NV12 */                                                        \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    libyuv::UYVYTONV12(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2),                    \
-                       dst_y, kWidth,                                          \
-                       dst_uv, 2 * SUBSAMPLE(kWidth, 2),                       \
-                       kWidth, kHeight);                                       \
-  }                                                                            \
-                                                                               \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    EXPECT_EQ(orig_y[i], dst_y[i]);                                            \
-  }                                                                            \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    EXPECT_EQ(dst_y_orig[i], dst_y[i]);                                        \
-  }                                                                            \
-  for (int i = 0; i < 2 * SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2); ++i) { \
-    EXPECT_EQ(dst_uv_orig[i], dst_uv[i]);                                      \
-  }                                                                            \
-                                                                               \
-  free_aligned_buffer_page_end(orig_uyvy);                                     \
-  free_aligned_buffer_page_end(orig_y);                                        \
-  free_aligned_buffer_page_end(orig_u);                                        \
-  free_aligned_buffer_page_end(orig_v);                                        \
-  free_aligned_buffer_page_end(dst_y_orig);                                    \
-  free_aligned_buffer_page_end(dst_uv_orig);                                   \
-  free_aligned_buffer_page_end(dst_y);                                         \
-  free_aligned_buffer_page_end(dst_uv);                                        \
-}
+    FMT_B##To##FMT_C(dst_argb_opt + OFF, kStrideB, dst_argb32_opt,             \
+                     kWidth * BPP_C, kWidth, kHeight);                         \
+    for (int i = 0; i < kWidth * BPP_C * kHeight; ++i) {                       \
+      EXPECT_EQ(dst_argb32_c[i], dst_argb32_opt[i]);                           \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_u);                                       \
+    free_aligned_buffer_page_end(src_v);                                       \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+    free_aligned_buffer_page_end(dst_argb32_c);                                \
+    free_aligned_buffer_page_end(dst_argb32_opt);                              \
+  }
+
+#define TESTPLANARTOBD(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                       YALIGN, FMT_C, BPP_C)                                  \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C)     \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Unaligned, +, 2, FMT_C, BPP_C)   \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Invert, -, 0, FMT_C, BPP_C)      \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANARTOBD(I420, 2, 2, RGB565, 2, 2, 1, ARGB, 4)
+#endif
+
+#define TESTPTOB(NAME, UYVYTOI420, UYVYTONV12)                                \
+  TEST_F(LibYUVConvertTest, NAME) {                                           \
+    const int kWidth = benchmark_width_;                                      \
+    const int kHeight = benchmark_height_;                                    \
+                                                                              \
+    align_buffer_page_end(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2) * kHeight);     \
+    align_buffer_page_end(orig_y, kWidth* kHeight);                           \
+    align_buffer_page_end(orig_u,                                             \
+                          SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));      \
+    align_buffer_page_end(orig_v,                                             \
+                          SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));      \
+                                                                              \
+    align_buffer_page_end(dst_y_orig, kWidth* kHeight);                       \
+    align_buffer_page_end(dst_uv_orig,                                        \
+                          2 * SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));  \
+                                                                              \
+    align_buffer_page_end(dst_y, kWidth* kHeight);                            \
+    align_buffer_page_end(dst_uv,                                             \
+                          2 * SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));  \
+                                                                              \
+    MemRandomize(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2) * kHeight);              \
+                                                                              \
+    /* Convert UYVY to NV12 in 2 steps for reference */                       \
+    libyuv::UYVYTOI420(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2), orig_y, kWidth,   \
+                       orig_u, SUBSAMPLE(kWidth, 2), orig_v,                  \
+                       SUBSAMPLE(kWidth, 2), kWidth, kHeight);                \
+    libyuv::I420ToNV12(orig_y, kWidth, orig_u, SUBSAMPLE(kWidth, 2), orig_v,  \
+                       SUBSAMPLE(kWidth, 2), dst_y_orig, kWidth, dst_uv_orig, \
+                       2 * SUBSAMPLE(kWidth, 2), kWidth, kHeight);            \
+                                                                              \
+    /* Convert to NV12 */                                                     \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      libyuv::UYVYTONV12(orig_uyvy, 4 * SUBSAMPLE(kWidth, 2), dst_y, kWidth,  \
+                         dst_uv, 2 * SUBSAMPLE(kWidth, 2), kWidth, kHeight);  \
+    }                                                                         \
+                                                                              \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      EXPECT_EQ(orig_y[i], dst_y[i]);                                         \
+    }                                                                         \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      EXPECT_EQ(dst_y_orig[i], dst_y[i]);                                     \
+    }                                                                         \
+    for (int i = 0; i < 2 * SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2);     \
+         ++i) {                                                               \
+      EXPECT_EQ(dst_uv_orig[i], dst_uv[i]);                                   \
+    }                                                                         \
+                                                                              \
+    free_aligned_buffer_page_end(orig_uyvy);                                  \
+    free_aligned_buffer_page_end(orig_y);                                     \
+    free_aligned_buffer_page_end(orig_u);                                     \
+    free_aligned_buffer_page_end(orig_v);                                     \
+    free_aligned_buffer_page_end(dst_y_orig);                                 \
+    free_aligned_buffer_page_end(dst_uv_orig);                                \
+    free_aligned_buffer_page_end(dst_y);                                      \
+    free_aligned_buffer_page_end(dst_uv);                                     \
+  }
 
 TESTPTOB(TestYUY2ToNV12, YUY2ToI420, YUY2ToNV12)
 TESTPTOB(TestUYVYToNV12, UYVYToI420, UYVYToNV12)
 
-#define TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,  \
-                       W1280, N, NEG, OFF, FMT_C, BPP_C)                       \
-TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) {                \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                       \
-  const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                          \
-  const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);               \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_v, kSizeUV + OFF);                                 \
-  align_buffer_page_end(dst_argb_b, kStrideB * kHeight + OFF);                 \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    src_y[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  for (int i = 0; i < kSizeUV; ++i) {                                          \
-    src_u[i + OFF] = (fastrand() & 0xff);                                      \
-    src_v[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                             \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                 \
-                          src_u + OFF, kStrideUV,                              \
-                          src_v + OFF, kStrideUV,                              \
-                          dst_argb_b + OFF, kStrideB,                          \
-                          kWidth, NEG kHeight);                                \
-  }                                                                            \
-  /* Convert to a 3rd format in 1 step and 2 steps and compare  */             \
-  const int kStrideC = kWidth * BPP_C;                                         \
-  align_buffer_page_end(dst_argb_c, kStrideC * kHeight + OFF);                 \
-  align_buffer_page_end(dst_argb_bc, kStrideC * kHeight + OFF);                \
-  memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                             \
-  memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                            \
-  FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth,                                   \
-                        src_u + OFF, kStrideUV,                                \
-                        src_v + OFF, kStrideUV,                                \
-                        dst_argb_c + OFF, kStrideC,                            \
-                        kWidth, NEG kHeight);                                  \
-  /* Convert B to C */                                                         \
-  FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB,                                 \
-                   dst_argb_bc + OFF, kStrideC,                                \
-                   kWidth, kHeight);                                           \
-  for (int i = 0; i < kStrideC * kHeight; ++i) {                               \
-    EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]);                      \
-  }                                                                            \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
-  free_aligned_buffer_page_end(dst_argb_b);                                    \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_bc);                                   \
+TEST_F(LibYUVConvertTest, MM21ToYUY2) {
+  const int kWidth = (benchmark_width_ + 15) & (~15);
+  const int kHeight = (benchmark_height_ + 31) & (~31);
+
+  align_buffer_page_end(orig_y, kWidth * kHeight);
+  align_buffer_page_end(orig_uv,
+                        2 * SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));
+
+  align_buffer_page_end(tmp_y, kWidth * kHeight);
+  align_buffer_page_end(tmp_u, SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));
+  align_buffer_page_end(tmp_v, SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));
+
+  align_buffer_page_end(dst_yuyv, 4 * SUBSAMPLE(kWidth, 2) * kHeight);
+  align_buffer_page_end(golden_yuyv, 4 * SUBSAMPLE(kWidth, 2) * kHeight);
+
+  MemRandomize(orig_y, kWidth * kHeight);
+  MemRandomize(orig_uv, 2 * SUBSAMPLE(kWidth, 2) * SUBSAMPLE(kHeight, 2));
+
+  /* Convert MM21 to YUY2 in 2 steps for reference */
+  libyuv::MM21ToI420(orig_y, kWidth, orig_uv, 2 * SUBSAMPLE(kWidth, 2), tmp_y,
+                     kWidth, tmp_u, SUBSAMPLE(kWidth, 2), tmp_v,
+                     SUBSAMPLE(kWidth, 2), kWidth, kHeight);
+  libyuv::I420ToYUY2(tmp_y, kWidth, tmp_u, SUBSAMPLE(kWidth, 2), tmp_v,
+                     SUBSAMPLE(kWidth, 2), golden_yuyv,
+                     4 * SUBSAMPLE(kWidth, 2), kWidth, kHeight);
+
+  /* Convert to NV12 */
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    libyuv::MM21ToYUY2(orig_y, kWidth, orig_uv, 2 * SUBSAMPLE(kWidth, 2),
+                       dst_yuyv, 4 * SUBSAMPLE(kWidth, 2), kWidth, kHeight);
+  }
+
+  for (int i = 0; i < 4 * SUBSAMPLE(kWidth, 2) * kHeight; ++i) {
+    EXPECT_EQ(dst_yuyv[i], golden_yuyv[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  free_aligned_buffer_page_end(orig_uv);
+  free_aligned_buffer_page_end(tmp_y);
+  free_aligned_buffer_page_end(tmp_u);
+  free_aligned_buffer_page_end(tmp_v);
+  free_aligned_buffer_page_end(dst_yuyv);
+  free_aligned_buffer_page_end(golden_yuyv);
 }
 
-#define TESTPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,   \
-                      FMT_C, BPP_C)                                            \
-    TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-        benchmark_width_ - 4, _Any, +, 0, FMT_C, BPP_C)                        \
-    TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-        benchmark_width_, _Unaligned, +, 1, FMT_C, BPP_C)                      \
-    TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-        benchmark_width_, _Invert, -, 0, FMT_C, BPP_C)                         \
-    TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-        benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+// Transitive test.  A to B to C is same as A to C.
+// Benchmarks A To B to C for comparison to 1 step, benchmarked elsewhere.
+#define TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                       W1280, N, NEG, OFF, FMT_C, BPP_C)                      \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##To##FMT_C##N) {            \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                    \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);            \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                      \
+    align_buffer_page_end(src_u, kSizeUV + OFF);                              \
+    align_buffer_page_end(src_v, kSizeUV + OFF);                              \
+    align_buffer_page_end(dst_argb_b, kStrideB* kHeight + OFF);               \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      src_y[i + OFF] = (fastrand() & 0xff);                                   \
+    }                                                                         \
+    for (int i = 0; i < kSizeUV; ++i) {                                       \
+      src_u[i + OFF] = (fastrand() & 0xff);                                   \
+      src_v[i + OFF] = (fastrand() & 0xff);                                   \
+    }                                                                         \
+    memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                          \
+    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,        \
+                          src_v + OFF, kStrideUV, dst_argb_b + OFF, kStrideB, \
+                          kWidth, NEG kHeight);                               \
+    /* Convert to a 3rd format in 1 step and 2 steps and compare  */          \
+    const int kStrideC = kWidth * BPP_C;                                      \
+    align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);               \
+    align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);              \
+    memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                          \
+    memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth, src_u + OFF, kStrideUV,      \
+                            src_v + OFF, kStrideUV, dst_argb_c + OFF,         \
+                            kStrideC, kWidth, NEG kHeight);                   \
+      /* Convert B to C */                                                    \
+      FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF,         \
+                       kStrideC, kWidth, kHeight);                            \
+    }                                                                         \
+    for (int i = 0; i < kStrideC * kHeight; ++i) {                            \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]);                   \
+    }                                                                         \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_u);                                      \
+    free_aligned_buffer_page_end(src_v);                                      \
+    free_aligned_buffer_page_end(dst_argb_b);                                 \
+    free_aligned_buffer_page_end(dst_argb_c);                                 \
+    free_aligned_buffer_page_end(dst_argb_bc);                                \
+  }
 
+#if defined(ENABLE_FULL_TESTS)
+#define TESTPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                      FMT_C, BPP_C)                                          \
+  TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                 benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C)             \
+  TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                 benchmark_width_, _Unaligned, +, 2, FMT_C, BPP_C)           \
+  TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                 benchmark_width_, _Invert, -, 0, FMT_C, BPP_C)              \
+  TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                 benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+#else
+#define TESTPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                      FMT_C, BPP_C)                                          \
+  TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                 benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+#endif
+
+#if defined(ENABLE_FULL_TESTS)
+TESTPLANARTOE(I420, 2, 2, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ABGR, 4)
-TESTPLANARTOE(J420, 2, 2, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(J420, 2, 2, ABGR, 1, 4, ARGB, 4)
-TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(H420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB24, 3)
 TESTPLANARTOE(I420, 2, 2, BGRA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I420, 2, 2, ABGR, 1, 4, ARGB, 4)
-TESTPLANARTOE(I420, 2, 2, RGBA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, RGB24, 3)
+TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, RAW, 3)
-TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RAW, 3)
-TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RGBA, 1, 4, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, RAW, 3)
+TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, RGB24, 3)
+TESTPLANARTOE(H420, 2, 2, RAW, 1, 3, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, RAW, 1, 3, RGB24, 3)
+TESTPLANARTOE(H420, 2, 2, RGB24, 1, 3, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, RGB24, 1, 3, RAW, 3)
+TESTPLANARTOE(J420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(J420, 2, 2, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(U420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(U420, 2, 2, ARGB, 1, 4, ARGB, 4)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB565, 2)
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB1555, 2)
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB4444, 2)
-TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, RGB565, 2)
+#endif
+TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTPLANARTOE(I422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(J422, 2, 1, ARGB, 1, 4, ARGB, 4)
 TESTPLANARTOE(J422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(H422, 2, 1, ARGB, 1, 4, ARGB, 4)
 TESTPLANARTOE(H422, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(U422, 2, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(U422, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(V422, 2, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(V422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, BGRA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, RGBA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I411, 4, 1, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(I444, 1, 1, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(J444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(I444, 1, 1, ARGB, 1, 4, ABGR, 4)
 TESTPLANARTOE(I444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(J444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(J444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(H444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(H444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(U444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(U444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(V444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(V444, 1, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, YUY2, 2, 4, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, UYVY, 2, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, YUY2, 2, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
+#else
+TESTPLANARTOE(I420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB1555, 2)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB4444, 2)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB24, 3)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB565, 2)
+TESTPLANARTOE(I420, 2, 2, BGRA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, RGB24, 3)
+TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, RGBA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, UYVY, 2, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, YUY2, 2, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, RGB565, 2)
+TESTPLANARTOE(I422, 2, 1, BGRA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, RGBA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, YUY2, 2, 4, ARGB, 4)
+TESTPLANARTOE(I444, 1, 1, ABGR, 1, 4, ARGB, 4)
+#endif
 
+// Transitive test: Compare 1 step vs 2 step conversion for YUVA to ARGB.
+// Benchmark 2 step conversion for comparison to 1 step conversion.
 #define TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
-                       W1280, N, NEG, OFF, FMT_C, BPP_C, ATTEN)                \
-TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) {                \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                       \
-  const int kSizeUV =                                                          \
-    SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y);              \
-  align_buffer_page_end(src_y, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(src_u, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_v, kSizeUV + OFF);                                 \
-  align_buffer_page_end(src_a, kWidth * kHeight + OFF);                        \
-  align_buffer_page_end(dst_argb_b, kStrideB * kHeight + OFF);                 \
-  for (int i = 0; i < kWidth * kHeight; ++i) {                                 \
-    src_y[i + OFF] = (fastrand() & 0xff);                                      \
-    src_a[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  for (int i = 0; i < kSizeUV; ++i) {                                          \
-    src_u[i + OFF] = (fastrand() & 0xff);                                      \
-    src_v[i + OFF] = (fastrand() & 0xff);                                      \
-  }                                                                            \
-  memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                             \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth,                                 \
-                          src_u + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),           \
-                          src_v + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),           \
-                          src_a + OFF, kWidth,                                 \
-                          dst_argb_b + OFF, kStrideB,                          \
-                          kWidth, NEG kHeight, ATTEN);                         \
-  }                                                                            \
-  int max_diff = 0;                                                            \
-  /* Convert to a 3rd format in 1 step and 2 steps and compare  */             \
-  const int kStrideC = kWidth * BPP_C;                                         \
-  align_buffer_page_end(dst_argb_c, kStrideC * kHeight + OFF);                 \
-  align_buffer_page_end(dst_argb_bc, kStrideC * kHeight + OFF);                \
-  memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                             \
-  memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                            \
-  FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth,                                   \
-                        src_u + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),             \
-                        src_v + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),             \
-                        src_a + OFF, kWidth,                                   \
-                        dst_argb_c + OFF, kStrideC,                            \
-                        kWidth, NEG kHeight, ATTEN);                           \
-  /* Convert B to C */                                                         \
-  FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB,                                 \
-                   dst_argb_bc + OFF, kStrideC,                                \
-                   kWidth, kHeight);                                           \
-  for (int i = 0; i < kStrideC * kHeight; ++i) {                               \
-    EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]);                      \
-  }                                                                            \
-  free_aligned_buffer_page_end(src_y);                                         \
-  free_aligned_buffer_page_end(src_u);                                         \
-  free_aligned_buffer_page_end(src_v);                                         \
-  free_aligned_buffer_page_end(src_a);                                         \
-  free_aligned_buffer_page_end(dst_argb_b);                                    \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_bc);                                   \
-}
+                        W1280, N, NEG, OFF, FMT_C, BPP_C, ATTEN)               \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##To##FMT_C##N) {             \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
+    const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                     \
+    const int kSizeUV =                                                        \
+        SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y);          \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                       \
+    align_buffer_page_end(src_u, kSizeUV + OFF);                               \
+    align_buffer_page_end(src_v, kSizeUV + OFF);                               \
+    align_buffer_page_end(src_a, kWidth* kHeight + OFF);                       \
+    align_buffer_page_end(dst_argb_b, kStrideB* kHeight + OFF);                \
+    const int kStrideC = kWidth * BPP_C;                                       \
+    align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);               \
+    memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                           \
+    memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                           \
+    memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                          \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      src_y[i + OFF] = (fastrand() & 0xff);                                    \
+      src_a[i + OFF] = (fastrand() & 0xff);                                    \
+    }                                                                          \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      src_u[i + OFF] = (fastrand() & 0xff);                                    \
+      src_v[i + OFF] = (fastrand() & 0xff);                                    \
+    }                                                                          \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      /* Convert A to B */                                                     \
+      FMT_PLANAR##To##FMT_B(                                                   \
+          src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),      \
+          src_v + OFF, SUBSAMPLE(kWidth, SUBSAMP_X), src_a + OFF, kWidth,      \
+          dst_argb_b + OFF, kStrideB, kWidth, NEG kHeight, ATTEN);             \
+      /* Convert B to C */                                                     \
+      FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF,          \
+                       kStrideC, kWidth, kHeight);                             \
+    }                                                                          \
+    /* Convert A to C */                                                       \
+    FMT_PLANAR##To##FMT_C(                                                     \
+        src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),        \
+        src_v + OFF, SUBSAMPLE(kWidth, SUBSAMP_X), src_a + OFF, kWidth,        \
+        dst_argb_c + OFF, kStrideC, kWidth, NEG kHeight, ATTEN);               \
+    for (int i = 0; i < kStrideC * kHeight; ++i) {                             \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]);                    \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_u);                                       \
+    free_aligned_buffer_page_end(src_v);                                       \
+    free_aligned_buffer_page_end(src_a);                                       \
+    free_aligned_buffer_page_end(dst_argb_b);                                  \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_bc);                                 \
+  }
 
-#define TESTQPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,  \
-                      FMT_C, BPP_C)                                            \
-    TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,     \
-        benchmark_width_ - 4, _Any, +, 0, FMT_C, BPP_C, 0)                     \
-    TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,     \
-        benchmark_width_, _Unaligned, +, 1, FMT_C, BPP_C, 0)                   \
-    TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,     \
-        benchmark_width_, _Invert, -, 0, FMT_C, BPP_C, 0)                      \
-    TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,     \
-        benchmark_width_, _Opt, +, 0, FMT_C, BPP_C, 0)                         \
-      TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,   \
-          benchmark_width_, _Premult, +, 0, FMT_C, BPP_C, 1)
+#if defined(ENABLE_FULL_TESTS)
+#define TESTQPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                       FMT_C, BPP_C)                                          \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C, 0)          \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_, _Unaligned, +, 2, FMT_C, BPP_C, 0)        \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_, _Invert, -, 0, FMT_C, BPP_C, 0)           \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_, _Opt, +, 0, FMT_C, BPP_C, 0)              \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_, _Premult, +, 0, FMT_C, BPP_C, 1)
+#else
+#define TESTQPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                       FMT_C, BPP_C)                                          \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_, _Opt, +, 0, FMT_C, BPP_C, 0)
+#endif
 
+#if defined(ENABLE_FULL_TESTS)
 TESTQPLANARTOE(I420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
 TESTQPLANARTOE(I420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(J420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(J420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(H420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(H420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(F420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(F420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(U420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(U420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(V420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(V420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(I422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(J422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(J422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(F422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(F422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(H422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(H422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(U422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(U422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(V422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(V422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(I444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(J444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(J444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(H444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(H444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(U444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(U444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(V444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(V444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+#else
+TESTQPLANARTOE(I420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+#endif
+
+#define TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, W1280, N, NEG, \
+                      OFF, FMT_C, BPP_C)                                       \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##To##FMT_C##N) {                  \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
+    const int kStrideA = SUBSAMPLE(kWidth, SUB_A) * BPP_A;                     \
+    const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                     \
+    align_buffer_page_end(src_argb_a, kStrideA* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_b, kStrideB* kHeight + OFF);                \
+    MemRandomize(src_argb_a + OFF, kStrideA * kHeight);                        \
+    memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                           \
+    FMT_A##To##FMT_B(src_argb_a + OFF, kStrideA, dst_argb_b + OFF, kStrideB,   \
+                     kWidth, NEG kHeight);                                     \
+    /* Convert to a 3rd format in 1 step and 2 steps and compare  */           \
+    const int kStrideC = kWidth * BPP_C;                                       \
+    align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);               \
+    memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                           \
+    memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                          \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_A##To##FMT_C(src_argb_a + OFF, kStrideA, dst_argb_c + OFF, kStrideC, \
+                       kWidth, NEG kHeight);                                   \
+      /* Convert B to C */                                                     \
+      FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF,          \
+                       kStrideC, kWidth, kHeight);                             \
+    }                                                                          \
+    for (int i = 0; i < kStrideC * kHeight; i += 4) {                          \
+      EXPECT_EQ(dst_argb_c[i + OFF + 0], dst_argb_bc[i + OFF + 0]);            \
+      EXPECT_EQ(dst_argb_c[i + OFF + 1], dst_argb_bc[i + OFF + 1]);            \
+      EXPECT_EQ(dst_argb_c[i + OFF + 2], dst_argb_bc[i + OFF + 2]);            \
+      EXPECT_NEAR(dst_argb_c[i + OFF + 3], dst_argb_bc[i + OFF + 3], 64);      \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_argb_a);                                  \
+    free_aligned_buffer_page_end(dst_argb_b);                                  \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_bc);                                 \
+  }
+
+#define TESTPLANETOE(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, FMT_C, BPP_C) \
+  TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B,                    \
+                benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C)              \
+  TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, benchmark_width_,  \
+                _Unaligned, +, 4, FMT_C, BPP_C)                              \
+  TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, benchmark_width_,  \
+                _Invert, -, 0, FMT_C, BPP_C)                                 \
+  TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, benchmark_width_,  \
+                _Opt, +, 0, FMT_C, BPP_C)
+
+// Caveat: Destination needs to be 4 bytes
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANETOE(ARGB, 1, 4, AR30, 1, 4, ARGB, 4)
+TESTPLANETOE(ABGR, 1, 4, AR30, 1, 4, ABGR, 4)
+TESTPLANETOE(AR30, 1, 4, ARGB, 1, 4, ABGR, 4)
+TESTPLANETOE(AR30, 1, 4, ABGR, 1, 4, ARGB, 4)
+TESTPLANETOE(ARGB, 1, 4, AB30, 1, 4, ARGB, 4)
+TESTPLANETOE(ABGR, 1, 4, AB30, 1, 4, ABGR, 4)
+TESTPLANETOE(AB30, 1, 4, ARGB, 1, 4, ABGR, 4)
+TESTPLANETOE(AB30, 1, 4, ABGR, 1, 4, ARGB, 4)
+#endif
+
+TEST_F(LibYUVConvertTest, RotateWithARGBSource) {
+  // 2x2 frames
+  uint32_t src[4];
+  uint32_t dst[4];
+  // some random input
+  src[0] = 0x11000000;
+  src[1] = 0x00450000;
+  src[2] = 0x00009f00;
+  src[3] = 0x000000ff;
+  // zeros on destination
+  dst[0] = 0x00000000;
+  dst[1] = 0x00000000;
+  dst[2] = 0x00000000;
+  dst[3] = 0x00000000;
+
+  int r = ConvertToARGB(reinterpret_cast<uint8_t*>(src),
+                        16,  // input size
+                        reinterpret_cast<uint8_t*>(dst),
+                        8,  // destination stride
+                        0,  // crop_x
+                        0,  // crop_y
+                        2,  // width
+                        2,  // height
+                        2,  // crop width
+                        2,  // crop height
+                        kRotate90, FOURCC_ARGB);
+
+  EXPECT_EQ(r, 0);
+  // 90 degrees rotation, no conversion
+  EXPECT_EQ(dst[0], src[2]);
+  EXPECT_EQ(dst[1], src[0]);
+  EXPECT_EQ(dst[2], src[3]);
+  EXPECT_EQ(dst[3], src[1]);
+}
+
+#ifdef HAS_ARGBTOAR30ROW_AVX2
+TEST_F(LibYUVConvertTest, ARGBToAR30Row_Opt) {
+  // ARGBToAR30Row_AVX2 expects a multiple of 8 pixels.
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 7) & ~7;
+  align_buffer_page_end(src, kPixels * 4);
+  align_buffer_page_end(dst_opt, kPixels * 4);
+  align_buffer_page_end(dst_c, kPixels * 4);
+  MemRandomize(src, kPixels * 4);
+  memset(dst_opt, 0, kPixels * 4);
+  memset(dst_c, 1, kPixels * 4);
+
+  ARGBToAR30Row_C(src, dst_c, kPixels);
+
+  int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+  int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    if (has_avx2) {
+      ARGBToAR30Row_AVX2(src, dst_opt, kPixels);
+    } else if (has_ssse3) {
+      ARGBToAR30Row_SSSE3(src, dst_opt, kPixels);
+    } else {
+      ARGBToAR30Row_C(src, dst_opt, kPixels);
+    }
+  }
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_opt[i], dst_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src);
+  free_aligned_buffer_page_end(dst_opt);
+  free_aligned_buffer_page_end(dst_c);
+}
+#endif  // HAS_ARGBTOAR30ROW_AVX2
+
+#ifdef HAS_ABGRTOAR30ROW_AVX2
+TEST_F(LibYUVConvertTest, ABGRToAR30Row_Opt) {
+  // ABGRToAR30Row_AVX2 expects a multiple of 8 pixels.
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 7) & ~7;
+  align_buffer_page_end(src, kPixels * 4);
+  align_buffer_page_end(dst_opt, kPixels * 4);
+  align_buffer_page_end(dst_c, kPixels * 4);
+  MemRandomize(src, kPixels * 4);
+  memset(dst_opt, 0, kPixels * 4);
+  memset(dst_c, 1, kPixels * 4);
+
+  ABGRToAR30Row_C(src, dst_c, kPixels);
+
+  int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+  int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    if (has_avx2) {
+      ABGRToAR30Row_AVX2(src, dst_opt, kPixels);
+    } else if (has_ssse3) {
+      ABGRToAR30Row_SSSE3(src, dst_opt, kPixels);
+    } else {
+      ABGRToAR30Row_C(src, dst_opt, kPixels);
+    }
+  }
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_opt[i], dst_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src);
+  free_aligned_buffer_page_end(dst_opt);
+  free_aligned_buffer_page_end(dst_c);
+}
+#endif  // HAS_ABGRTOAR30ROW_AVX2
+
+// Provide matrix wrappers for 12 bit YUV
+#define I012ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I012ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define I012ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I012ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+
+#define I410ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I410ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define I410ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I410ToABGRMatrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define H410ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I410ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define H410ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I410ToABGRMatrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define U410ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I410ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+#define U410ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I410ToABGRMatrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+#define I410ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define I410ToAB30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAB30Matrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define H410ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define H410ToAB30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAB30Matrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define U410ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+#define U410ToAB30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAB30Matrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+
+#define I010ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I010ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I010ToAR30Filter(a, b, c, d, e, f, g, h, i, j)                     \
+  I010ToAR30MatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I210ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I210ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I210ToAR30Filter(a, b, c, d, e, f, g, h, i, j)                     \
+  I210ToAR30MatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+
+// TODO(fbarchard): Fix clamping issue affected by U channel.
+#define TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B,   \
+                         BPP_B, ALIGN, YALIGN, W1280, N, NEG, SOFF, DOFF)     \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                       \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                  \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                     \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);            \
+    const int kBpc = 2;                                                       \
+    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + SOFF);               \
+    align_buffer_page_end(src_u, kSizeUV* kBpc + SOFF);                       \
+    align_buffer_page_end(src_v, kSizeUV* kBpc + SOFF);                       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + DOFF);              \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + DOFF);            \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      reinterpret_cast<uint16_t*>(src_y + SOFF)[i] = (fastrand() & FMT_MASK); \
+    }                                                                         \
+    for (int i = 0; i < kSizeUV; ++i) {                                       \
+      reinterpret_cast<uint16_t*>(src_u + SOFF)[i] = (fastrand() & FMT_MASK); \
+      reinterpret_cast<uint16_t*>(src_v + SOFF)[i] = (fastrand() & FMT_MASK); \
+    }                                                                         \
+    memset(dst_argb_c + DOFF, 1, kStrideB * kHeight);                         \
+    memset(dst_argb_opt + DOFF, 101, kStrideB * kHeight);                     \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    FMT_PLANAR##To##FMT_B(                                                    \
+        reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,                    \
+        reinterpret_cast<uint16_t*>(src_u + SOFF), kStrideUV,                 \
+        reinterpret_cast<uint16_t*>(src_v + SOFF), kStrideUV,                 \
+        dst_argb_c + DOFF, kStrideB, kWidth, NEG kHeight);                    \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_PLANAR##To##FMT_B(                                                  \
+          reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,                  \
+          reinterpret_cast<uint16_t*>(src_u + SOFF), kStrideUV,               \
+          reinterpret_cast<uint16_t*>(src_v + SOFF), kStrideUV,               \
+          dst_argb_opt + DOFF, kStrideB, kWidth, NEG kHeight);                \
+    }                                                                         \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                      \
+      EXPECT_EQ(dst_argb_c[i + DOFF], dst_argb_opt[i + DOFF]);                \
+    }                                                                         \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_u);                                      \
+    free_aligned_buffer_page_end(src_v);                                      \
+    free_aligned_buffer_page_end(dst_argb_c);                                 \
+    free_aligned_buffer_page_end(dst_argb_opt);                               \
+  }
+
+#define TESTPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B,   \
+                        BPP_B, ALIGN, YALIGN)                                \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_ + 1, _Any, +, 0, 0)       \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_, _Unaligned, +, 4, 4)     \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_, _Invert, -, 0, 0)        \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_, _Opt, +, 0, 0)
+
+// These conversions are only optimized for x86
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(I012, 2, 2, 0xfff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, ARGBFilter, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, ARGBFilter, 4, 4, 1)
+
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(I012, 2, 2, 0xfff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, AR30Filter, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, AR30Filter, 4, 4, 1)
+#endif  // LITTLE_ENDIAN_ONLY_TEST
+#endif  // DISABLE_SLOW_TESTS
+
+#define TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,      \
+                          ALIGN, YALIGN, W1280, N, NEG, OFF, ATTEN, S_DEPTH)   \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);             \
+    const int kBpc = 2;                                                        \
+    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + OFF);                 \
+    align_buffer_page_end(src_u, kSizeUV* kBpc + OFF);                         \
+    align_buffer_page_end(src_v, kSizeUV* kBpc + OFF);                         \
+    align_buffer_page_end(src_a, kWidth* kHeight* kBpc + OFF);                 \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + OFF);              \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      reinterpret_cast<uint16_t*>(src_y + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+      reinterpret_cast<uint16_t*>(src_a + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+    }                                                                          \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      reinterpret_cast<uint16_t*>(src_u + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+      reinterpret_cast<uint16_t*>(src_v + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+    }                                                                          \
+    memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                           \
+    memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                       \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B(reinterpret_cast<uint16_t*>(src_y + OFF), kWidth,    \
+                          reinterpret_cast<uint16_t*>(src_u + OFF), kStrideUV, \
+                          reinterpret_cast<uint16_t*>(src_v + OFF), kStrideUV, \
+                          reinterpret_cast<uint16_t*>(src_a + OFF), kWidth,    \
+                          dst_argb_c + OFF, kStrideB, kWidth, NEG kHeight,     \
+                          ATTEN);                                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B(                                                   \
+          reinterpret_cast<uint16_t*>(src_y + OFF), kWidth,                    \
+          reinterpret_cast<uint16_t*>(src_u + OFF), kStrideUV,                 \
+          reinterpret_cast<uint16_t*>(src_v + OFF), kStrideUV,                 \
+          reinterpret_cast<uint16_t*>(src_a + OFF), kWidth,                    \
+          dst_argb_opt + OFF, kStrideB, kWidth, NEG kHeight, ATTEN);           \
+    }                                                                          \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                       \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_opt[i + OFF]);                   \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_u);                                       \
+    free_aligned_buffer_page_end(src_v);                                       \
+    free_aligned_buffer_page_end(src_a);                                       \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+#define TESTQPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,    \
+                         ALIGN, YALIGN, S_DEPTH)                            \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_ + 1, _Any, +, 0, 0, S_DEPTH)   \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Unaligned, +, 2, 0, S_DEPTH) \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Invert, -, 0, 0, S_DEPTH)    \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Opt, +, 0, 0, S_DEPTH)       \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Premult, +, 0, 1, S_DEPTH)
+#else
+#define TESTQPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,   \
+                         ALIGN, YALIGN, S_DEPTH)                           \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                    YALIGN, benchmark_width_, _Opt, +, 0, 0, S_DEPTH)
+#endif
+
+#define I010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define I010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define J010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define I210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define I210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define J210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define I410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define I410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define J410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define I010AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I010AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+#define I210AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I010AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+
+// These conversions are only optimized for x86
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+TESTQPLANAR16TOB(I010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I010Alpha, 2, 2, ARGBFilter, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I210Alpha, 2, 1, ARGBFilter, 4, 4, 1, 10)
+#endif  // DISABLE_SLOW_TESTS
+
+#define TESTBP16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,    \
+                     YALIGN, W1280, N, NEG, SOFF, DOFF, S_DEPTH)               \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X) * 2;                    \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2;         \
+    const int kBpc = 2;                                                        \
+    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + SOFF);                \
+    align_buffer_page_end(src_uv, kSizeUV* kBpc + SOFF);                       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + DOFF);               \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + DOFF);             \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      reinterpret_cast<uint16_t*>(src_y + SOFF)[i] =                           \
+          (fastrand() & (((uint16_t)(-1)) << (16 - S_DEPTH)));                 \
+    }                                                                          \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      reinterpret_cast<uint16_t*>(src_uv + SOFF)[i] =                          \
+          (fastrand() & (((uint16_t)(-1)) << (16 - S_DEPTH)));                 \
+    }                                                                          \
+    memset(dst_argb_c + DOFF, 1, kStrideB * kHeight);                          \
+    memset(dst_argb_opt + DOFF, 101, kStrideB * kHeight);                      \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B(reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,   \
+                          reinterpret_cast<uint16_t*>(src_uv + SOFF),          \
+                          kStrideUV, dst_argb_c + DOFF, kStrideB, kWidth,      \
+                          NEG kHeight);                                        \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B(reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth, \
+                            reinterpret_cast<uint16_t*>(src_uv + SOFF),        \
+                            kStrideUV, dst_argb_opt + DOFF, kStrideB, kWidth,  \
+                            NEG kHeight);                                      \
+    }                                                                          \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                       \
+      EXPECT_EQ(dst_argb_c[i + DOFF], dst_argb_opt[i + DOFF]);                 \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_uv);                                      \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+  }
+
+#define TESTBP16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,    \
+                    YALIGN, S_DEPTH)                                          \
+  TESTBP16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, YALIGN, \
+               benchmark_width_ + 1, _Any, +, 0, 0, S_DEPTH)                  \
+  TESTBP16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, YALIGN, \
+               benchmark_width_, _Unaligned, +, 4, 4, S_DEPTH)                \
+  TESTBP16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, YALIGN, \
+               benchmark_width_, _Invert, -, 0, 0, S_DEPTH)                   \
+  TESTBP16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, YALIGN, \
+               benchmark_width_, _Opt, +, 0, 0, S_DEPTH)
+
+#define P010ToARGB(a, b, c, d, e, f, g, h) \
+  P010ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P210ToARGB(a, b, c, d, e, f, g, h) \
+  P210ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P010ToAR30(a, b, c, d, e, f, g, h) \
+  P010ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P210ToAR30(a, b, c, d, e, f, g, h) \
+  P210ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+
+#define P012ToARGB(a, b, c, d, e, f, g, h) \
+  P012ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P212ToARGB(a, b, c, d, e, f, g, h) \
+  P212ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P012ToAR30(a, b, c, d, e, f, g, h) \
+  P012ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P212ToAR30(a, b, c, d, e, f, g, h) \
+  P212ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+
+#define P016ToARGB(a, b, c, d, e, f, g, h) \
+  P016ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P216ToARGB(a, b, c, d, e, f, g, h) \
+  P216ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P016ToAR30(a, b, c, d, e, f, g, h) \
+  P016ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P216ToAR30(a, b, c, d, e, f, g, h) \
+  P216ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+
+#define P010ToARGBFilter(a, b, c, d, e, f, g, h)                     \
+  P010ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+#define P210ToARGBFilter(a, b, c, d, e, f, g, h)                     \
+  P210ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+#define P010ToAR30Filter(a, b, c, d, e, f, g, h)                     \
+  P010ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+#define P210ToAR30Filter(a, b, c, d, e, f, g, h)                     \
+  P210ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+TESTBP16TOB(P010, 2, 2, ARGB, 4, 4, 1, 10)
+TESTBP16TOB(P210, 2, 1, ARGB, 4, 4, 1, 10)
+TESTBP16TOB(P012, 2, 2, ARGB, 4, 4, 1, 12)
+TESTBP16TOB(P212, 2, 1, ARGB, 4, 4, 1, 12)
+TESTBP16TOB(P016, 2, 2, ARGB, 4, 4, 1, 16)
+TESTBP16TOB(P216, 2, 1, ARGB, 4, 4, 1, 16)
+TESTBP16TOB(P010, 2, 2, ARGBFilter, 4, 4, 1, 10)
+TESTBP16TOB(P210, 2, 1, ARGBFilter, 4, 4, 1, 10)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTBP16TOB(P010, 2, 2, AR30, 4, 4, 1, 10)
+TESTBP16TOB(P210, 2, 1, AR30, 4, 4, 1, 10)
+TESTBP16TOB(P012, 2, 2, AR30, 4, 4, 1, 12)
+TESTBP16TOB(P212, 2, 1, AR30, 4, 4, 1, 12)
+TESTBP16TOB(P016, 2, 2, AR30, 4, 4, 1, 16)
+TESTBP16TOB(P216, 2, 1, AR30, 4, 4, 1, 16)
+TESTBP16TOB(P010, 2, 2, AR30Filter, 4, 4, 1, 10)
+TESTBP16TOB(P210, 2, 1, AR30Filter, 4, 4, 1, 10)
+#endif  // LITTLE_ENDIAN_ONLY_TEST
+#endif  // DISABLE_SLOW_TESTS
+
+static int Clamp(int y) {
+  if (y < 0) {
+    y = 0;
+  }
+  if (y > 255) {
+    y = 255;
+  }
+  return y;
+}
+
+static int Clamp10(int y) {
+  if (y < 0) {
+    y = 0;
+  }
+  if (y > 1023) {
+    y = 1023;
+  }
+  return y;
+}
+
+// Test 8 bit YUV to 8 bit RGB
+TEST_F(LibYUVConvertTest, TestH420ToARGB) {
+  const int kSize = 256;
+  int histogram_b[256];
+  int histogram_g[256];
+  int histogram_r[256];
+  memset(histogram_b, 0, sizeof(histogram_b));
+  memset(histogram_g, 0, sizeof(histogram_g));
+  memset(histogram_r, 0, sizeof(histogram_r));
+  align_buffer_page_end(orig_yuv, kSize + kSize / 2 * 2);
+  align_buffer_page_end(argb_pixels, kSize * 4);
+  uint8_t* orig_y = orig_yuv;
+  uint8_t* orig_u = orig_y + kSize;
+  uint8_t* orig_v = orig_u + kSize / 2;
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_y[i] = i;
+  }
+  for (int i = 0; i < kSize / 2; ++i) {
+    orig_u[i] = 128;  // 128 is 0.
+    orig_v[i] = 128;
+  }
+
+  H420ToARGB(orig_y, 0, orig_u, 0, orig_v, 0, argb_pixels, 0, kSize, 1);
+
+  for (int i = 0; i < kSize; ++i) {
+    int b = argb_pixels[i * 4 + 0];
+    int g = argb_pixels[i * 4 + 1];
+    int r = argb_pixels[i * 4 + 2];
+    int a = argb_pixels[i * 4 + 3];
+    ++histogram_b[b];
+    ++histogram_g[g];
+    ++histogram_r[r];
+    // Reference formula for Y channel contribution in YUV to RGB conversions:
+    int expected_y = Clamp(static_cast<int>((i - 16) * 1.164f + 0.5f));
+    EXPECT_EQ(b, expected_y);
+    EXPECT_EQ(g, expected_y);
+    EXPECT_EQ(r, expected_y);
+    EXPECT_EQ(a, 255);
+  }
+
+  int count_b = 0;
+  int count_g = 0;
+  int count_r = 0;
+  for (int i = 0; i < kSize; ++i) {
+    if (histogram_b[i]) {
+      ++count_b;
+    }
+    if (histogram_g[i]) {
+      ++count_g;
+    }
+    if (histogram_r[i]) {
+      ++count_r;
+    }
+  }
+  printf("uniques: B %d, G, %d, R %d\n", count_b, count_g, count_r);
+
+  free_aligned_buffer_page_end(orig_yuv);
+  free_aligned_buffer_page_end(argb_pixels);
+}
+
+// Test 10 bit YUV to 8 bit RGB
+TEST_F(LibYUVConvertTest, TestH010ToARGB) {
+  const int kSize = 1024;
+  int histogram_b[1024];
+  int histogram_g[1024];
+  int histogram_r[1024];
+  memset(histogram_b, 0, sizeof(histogram_b));
+  memset(histogram_g, 0, sizeof(histogram_g));
+  memset(histogram_r, 0, sizeof(histogram_r));
+  align_buffer_page_end(orig_yuv, kSize * 2 + kSize / 2 * 2 * 2);
+  align_buffer_page_end(argb_pixels, kSize * 4);
+  uint16_t* orig_y = reinterpret_cast<uint16_t*>(orig_yuv);
+  uint16_t* orig_u = orig_y + kSize;
+  uint16_t* orig_v = orig_u + kSize / 2;
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_y[i] = i;
+  }
+  for (int i = 0; i < kSize / 2; ++i) {
+    orig_u[i] = 512;  // 512 is 0.
+    orig_v[i] = 512;
+  }
+
+  H010ToARGB(orig_y, 0, orig_u, 0, orig_v, 0, argb_pixels, 0, kSize, 1);
+
+  for (int i = 0; i < kSize; ++i) {
+    int b = argb_pixels[i * 4 + 0];
+    int g = argb_pixels[i * 4 + 1];
+    int r = argb_pixels[i * 4 + 2];
+    int a = argb_pixels[i * 4 + 3];
+    ++histogram_b[b];
+    ++histogram_g[g];
+    ++histogram_r[r];
+    int expected_y = Clamp(static_cast<int>((i - 64) * 1.164f / 4));
+    EXPECT_NEAR(b, expected_y, 1);
+    EXPECT_NEAR(g, expected_y, 1);
+    EXPECT_NEAR(r, expected_y, 1);
+    EXPECT_EQ(a, 255);
+  }
+
+  int count_b = 0;
+  int count_g = 0;
+  int count_r = 0;
+  for (int i = 0; i < kSize; ++i) {
+    if (histogram_b[i]) {
+      ++count_b;
+    }
+    if (histogram_g[i]) {
+      ++count_g;
+    }
+    if (histogram_r[i]) {
+      ++count_r;
+    }
+  }
+  printf("uniques: B %d, G, %d, R %d\n", count_b, count_g, count_r);
+
+  free_aligned_buffer_page_end(orig_yuv);
+  free_aligned_buffer_page_end(argb_pixels);
+}
+
+// Test 10 bit YUV to 10 bit RGB
+// Caveat: Result is near due to float rounding in expected
+// result.
+TEST_F(LibYUVConvertTest, TestH010ToAR30) {
+  const int kSize = 1024;
+  int histogram_b[1024];
+  int histogram_g[1024];
+  int histogram_r[1024];
+  memset(histogram_b, 0, sizeof(histogram_b));
+  memset(histogram_g, 0, sizeof(histogram_g));
+  memset(histogram_r, 0, sizeof(histogram_r));
+
+  align_buffer_page_end(orig_yuv, kSize * 2 + kSize / 2 * 2 * 2);
+  align_buffer_page_end(ar30_pixels, kSize * 4);
+  uint16_t* orig_y = reinterpret_cast<uint16_t*>(orig_yuv);
+  uint16_t* orig_u = orig_y + kSize;
+  uint16_t* orig_v = orig_u + kSize / 2;
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_y[i] = i;
+  }
+  for (int i = 0; i < kSize / 2; ++i) {
+    orig_u[i] = 512;  // 512 is 0.
+    orig_v[i] = 512;
+  }
+
+  H010ToAR30(orig_y, 0, orig_u, 0, orig_v, 0, ar30_pixels, 0, kSize, 1);
+
+  for (int i = 0; i < kSize; ++i) {
+    int b10 = reinterpret_cast<uint32_t*>(ar30_pixels)[i] & 1023;
+    int g10 = (reinterpret_cast<uint32_t*>(ar30_pixels)[i] >> 10) & 1023;
+    int r10 = (reinterpret_cast<uint32_t*>(ar30_pixels)[i] >> 20) & 1023;
+    int a2 = (reinterpret_cast<uint32_t*>(ar30_pixels)[i] >> 30) & 3;
+    ++histogram_b[b10];
+    ++histogram_g[g10];
+    ++histogram_r[r10];
+    int expected_y = Clamp10(static_cast<int>((i - 64) * 1.164f + 0.5));
+    EXPECT_NEAR(b10, expected_y, 4);
+    EXPECT_NEAR(g10, expected_y, 4);
+    EXPECT_NEAR(r10, expected_y, 4);
+    EXPECT_EQ(a2, 3);
+  }
+
+  int count_b = 0;
+  int count_g = 0;
+  int count_r = 0;
+  for (int i = 0; i < kSize; ++i) {
+    if (histogram_b[i]) {
+      ++count_b;
+    }
+    if (histogram_g[i]) {
+      ++count_g;
+    }
+    if (histogram_r[i]) {
+      ++count_r;
+    }
+  }
+  printf("uniques: B %d, G, %d, R %d\n", count_b, count_g, count_r);
+
+  free_aligned_buffer_page_end(orig_yuv);
+  free_aligned_buffer_page_end(ar30_pixels);
+}
+
+// Test 10 bit YUV to 10 bit RGB
+// Caveat: Result is near due to float rounding in expected
+// result.
+TEST_F(LibYUVConvertTest, TestH010ToAB30) {
+  const int kSize = 1024;
+  int histogram_b[1024];
+  int histogram_g[1024];
+  int histogram_r[1024];
+  memset(histogram_b, 0, sizeof(histogram_b));
+  memset(histogram_g, 0, sizeof(histogram_g));
+  memset(histogram_r, 0, sizeof(histogram_r));
+
+  align_buffer_page_end(orig_yuv, kSize * 2 + kSize / 2 * 2 * 2);
+  align_buffer_page_end(ab30_pixels, kSize * 4);
+  uint16_t* orig_y = reinterpret_cast<uint16_t*>(orig_yuv);
+  uint16_t* orig_u = orig_y + kSize;
+  uint16_t* orig_v = orig_u + kSize / 2;
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_y[i] = i;
+  }
+  for (int i = 0; i < kSize / 2; ++i) {
+    orig_u[i] = 512;  // 512 is 0.
+    orig_v[i] = 512;
+  }
+
+  H010ToAB30(orig_y, 0, orig_u, 0, orig_v, 0, ab30_pixels, 0, kSize, 1);
+
+  for (int i = 0; i < kSize; ++i) {
+    int r10 = reinterpret_cast<uint32_t*>(ab30_pixels)[i] & 1023;
+    int g10 = (reinterpret_cast<uint32_t*>(ab30_pixels)[i] >> 10) & 1023;
+    int b10 = (reinterpret_cast<uint32_t*>(ab30_pixels)[i] >> 20) & 1023;
+    int a2 = (reinterpret_cast<uint32_t*>(ab30_pixels)[i] >> 30) & 3;
+    ++histogram_b[b10];
+    ++histogram_g[g10];
+    ++histogram_r[r10];
+    int expected_y = Clamp10(static_cast<int>((i - 64) * 1.164f));
+    EXPECT_NEAR(b10, expected_y, 4);
+    EXPECT_NEAR(g10, expected_y, 4);
+    EXPECT_NEAR(r10, expected_y, 4);
+    EXPECT_EQ(a2, 3);
+  }
+
+  int count_b = 0;
+  int count_g = 0;
+  int count_r = 0;
+  for (int i = 0; i < kSize; ++i) {
+    if (histogram_b[i]) {
+      ++count_b;
+    }
+    if (histogram_g[i]) {
+      ++count_g;
+    }
+    if (histogram_r[i]) {
+      ++count_r;
+    }
+  }
+  printf("uniques: B %d, G, %d, R %d\n", count_b, count_g, count_r);
+
+  free_aligned_buffer_page_end(orig_yuv);
+  free_aligned_buffer_page_end(ab30_pixels);
+}
+
+// Test 8 bit YUV to 10 bit RGB
+TEST_F(LibYUVConvertTest, TestH420ToAR30) {
+  const int kSize = 256;
+  const int kHistSize = 1024;
+  int histogram_b[kHistSize];
+  int histogram_g[kHistSize];
+  int histogram_r[kHistSize];
+  memset(histogram_b, 0, sizeof(histogram_b));
+  memset(histogram_g, 0, sizeof(histogram_g));
+  memset(histogram_r, 0, sizeof(histogram_r));
+  align_buffer_page_end(orig_yuv, kSize + kSize / 2 * 2);
+  align_buffer_page_end(ar30_pixels, kSize * 4);
+  uint8_t* orig_y = orig_yuv;
+  uint8_t* orig_u = orig_y + kSize;
+  uint8_t* orig_v = orig_u + kSize / 2;
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_y[i] = i;
+  }
+  for (int i = 0; i < kSize / 2; ++i) {
+    orig_u[i] = 128;  // 128 is 0.
+    orig_v[i] = 128;
+  }
+
+  H420ToAR30(orig_y, 0, orig_u, 0, orig_v, 0, ar30_pixels, 0, kSize, 1);
+
+  for (int i = 0; i < kSize; ++i) {
+    int b10 = reinterpret_cast<uint32_t*>(ar30_pixels)[i] & 1023;
+    int g10 = (reinterpret_cast<uint32_t*>(ar30_pixels)[i] >> 10) & 1023;
+    int r10 = (reinterpret_cast<uint32_t*>(ar30_pixels)[i] >> 20) & 1023;
+    int a2 = (reinterpret_cast<uint32_t*>(ar30_pixels)[i] >> 30) & 3;
+    ++histogram_b[b10];
+    ++histogram_g[g10];
+    ++histogram_r[r10];
+    int expected_y = Clamp10(static_cast<int>((i - 16) * 1.164f * 4.f));
+    EXPECT_NEAR(b10, expected_y, 4);
+    EXPECT_NEAR(g10, expected_y, 4);
+    EXPECT_NEAR(r10, expected_y, 4);
+    EXPECT_EQ(a2, 3);
+  }
+
+  int count_b = 0;
+  int count_g = 0;
+  int count_r = 0;
+  for (int i = 0; i < kHistSize; ++i) {
+    if (histogram_b[i]) {
+      ++count_b;
+    }
+    if (histogram_g[i]) {
+      ++count_g;
+    }
+    if (histogram_r[i]) {
+      ++count_r;
+    }
+  }
+  printf("uniques: B %d, G, %d, R %d\n", count_b, count_g, count_r);
+
+  free_aligned_buffer_page_end(orig_yuv);
+  free_aligned_buffer_page_end(ar30_pixels);
+}
+
+// Test I400 with jpeg matrix is same as J400
+TEST_F(LibYUVConvertTest, TestI400) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_i400, kSize);
+  align_buffer_page_end(argb_pixels_i400, kSize * 4);
+  align_buffer_page_end(argb_pixels_j400, kSize * 4);
+  align_buffer_page_end(argb_pixels_jpeg_i400, kSize * 4);
+  align_buffer_page_end(argb_pixels_h709_i400, kSize * 4);
+  align_buffer_page_end(argb_pixels_2020_i400, kSize * 4);
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_i400[i] = i;
+  }
+
+  J400ToARGB(orig_i400, 0, argb_pixels_j400, 0, kSize, 1);
+  I400ToARGB(orig_i400, 0, argb_pixels_i400, 0, kSize, 1);
+  I400ToARGBMatrix(orig_i400, 0, argb_pixels_jpeg_i400, 0, &kYuvJPEGConstants,
+                   kSize, 1);
+  I400ToARGBMatrix(orig_i400, 0, argb_pixels_h709_i400, 0, &kYuvH709Constants,
+                   kSize, 1);
+  I400ToARGBMatrix(orig_i400, 0, argb_pixels_2020_i400, 0, &kYuv2020Constants,
+                   kSize, 1);
+
+  EXPECT_EQ(0, argb_pixels_i400[0]);
+  EXPECT_EQ(0, argb_pixels_j400[0]);
+  EXPECT_EQ(0, argb_pixels_jpeg_i400[0]);
+  EXPECT_EQ(0, argb_pixels_h709_i400[0]);
+  EXPECT_EQ(0, argb_pixels_2020_i400[0]);
+  EXPECT_EQ(0, argb_pixels_i400[16 * 4]);
+  EXPECT_EQ(16, argb_pixels_j400[16 * 4]);
+  EXPECT_EQ(16, argb_pixels_jpeg_i400[16 * 4]);
+  EXPECT_EQ(0, argb_pixels_h709_i400[16 * 4]);
+  EXPECT_EQ(0, argb_pixels_2020_i400[16 * 4]);
+  EXPECT_EQ(130, argb_pixels_i400[128 * 4]);
+  EXPECT_EQ(128, argb_pixels_j400[128 * 4]);
+  EXPECT_EQ(128, argb_pixels_jpeg_i400[128 * 4]);
+  EXPECT_EQ(130, argb_pixels_h709_i400[128 * 4]);
+  EXPECT_EQ(130, argb_pixels_2020_i400[128 * 4]);
+  EXPECT_EQ(255, argb_pixels_i400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_j400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_jpeg_i400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_h709_i400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_2020_i400[255 * 4]);
+
+  for (int i = 0; i < kSize * 4; ++i) {
+    if ((i & 3) == 3) {
+      EXPECT_EQ(255, argb_pixels_j400[i]);
+    } else {
+      EXPECT_EQ(i / 4, argb_pixels_j400[i]);
+    }
+    EXPECT_EQ(argb_pixels_jpeg_i400[i], argb_pixels_j400[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_i400);
+  free_aligned_buffer_page_end(argb_pixels_i400);
+  free_aligned_buffer_page_end(argb_pixels_j400);
+  free_aligned_buffer_page_end(argb_pixels_jpeg_i400);
+  free_aligned_buffer_page_end(argb_pixels_h709_i400);
+  free_aligned_buffer_page_end(argb_pixels_2020_i400);
+}
+
+// Test RGB24 to ARGB and back to RGB24
+TEST_F(LibYUVConvertTest, TestARGBToRGB24) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_rgb24, kSize * 3);
+  align_buffer_page_end(argb_pixels, kSize * 4);
+  align_buffer_page_end(dest_rgb24, kSize * 3);
+
+  // Test grey scale
+  for (int i = 0; i < kSize * 3; ++i) {
+    orig_rgb24[i] = i;
+  }
+
+  RGB24ToARGB(orig_rgb24, 0, argb_pixels, 0, kSize, 1);
+  ARGBToRGB24(argb_pixels, 0, dest_rgb24, 0, kSize, 1);
+
+  for (int i = 0; i < kSize * 3; ++i) {
+    EXPECT_EQ(orig_rgb24[i], dest_rgb24[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_rgb24);
+  free_aligned_buffer_page_end(argb_pixels);
+  free_aligned_buffer_page_end(dest_rgb24);
+}
+
+TEST_F(LibYUVConvertTest, Test565) {
+  SIMD_ALIGNED(uint8_t orig_pixels[256][4]);
+  SIMD_ALIGNED(uint8_t pixels565[256][2]);
+
+  for (int i = 0; i < 256; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      orig_pixels[i][j] = i;
+    }
+  }
+  ARGBToRGB565(&orig_pixels[0][0], 0, &pixels565[0][0], 0, 256, 1);
+  uint32_t checksum = HashDjb2(&pixels565[0][0], sizeof(pixels565), 5381);
+  EXPECT_EQ(610919429u, checksum);
+}
+
+// Test RGB24 to J420 is exact
+#if defined(LIBYUV_BIT_EXACT)
+TEST_F(LibYUVConvertTest, TestRGB24ToJ420) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_rgb24, kSize * 3 * 2);  // 2 rows of RGB24
+  align_buffer_page_end(dest_j420, kSize * 3 / 2 * 2);
+  int iterations256 = (benchmark_width_ * benchmark_height_ + (kSize * 2 - 1)) /
+                      (kSize * 2) * benchmark_iterations_;
+
+  for (int i = 0; i < kSize * 3 * 2; ++i) {
+    orig_rgb24[i] = i;
+  }
+
+  for (int i = 0; i < iterations256; ++i) {
+    RGB24ToJ420(orig_rgb24, kSize * 3, dest_j420, kSize,  // Y plane
+                dest_j420 + kSize * 2, kSize / 2,         // U plane
+                dest_j420 + kSize * 5 / 2, kSize / 2,     // V plane
+                kSize, 2);
+  }
+
+  uint32_t checksum = HashDjb2(dest_j420, kSize * 3 / 2 * 2, 5381);
+  EXPECT_EQ(2755440272u, checksum);
+
+  free_aligned_buffer_page_end(orig_rgb24);
+  free_aligned_buffer_page_end(dest_j420);
+}
+#endif
+
+// Test RGB24 to I420 is exact
+#if defined(LIBYUV_BIT_EXACT)
+TEST_F(LibYUVConvertTest, TestRGB24ToI420) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_rgb24, kSize * 3 * 2);  // 2 rows of RGB24
+  align_buffer_page_end(dest_i420, kSize * 3 / 2 * 2);
+  int iterations256 = (benchmark_width_ * benchmark_height_ + (kSize * 2 - 1)) /
+                      (kSize * 2) * benchmark_iterations_;
+
+  for (int i = 0; i < kSize * 3 * 2; ++i) {
+    orig_rgb24[i] = i;
+  }
+
+  for (int i = 0; i < iterations256; ++i) {
+    RGB24ToI420(orig_rgb24, kSize * 3, dest_i420, kSize,  // Y plane
+                dest_i420 + kSize * 2, kSize / 2,         // U plane
+                dest_i420 + kSize * 5 / 2, kSize / 2,     // V plane
+                kSize, 2);
+  }
+
+  uint32_t checksum = HashDjb2(dest_i420, kSize * 3 / 2 * 2, 5381);
+  EXPECT_EQ(1526656597u, checksum);
+
+  free_aligned_buffer_page_end(orig_rgb24);
+  free_aligned_buffer_page_end(dest_i420);
+}
+#endif
 
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/cpu_test.cc b/media/libyuv/unit_test/cpu_test.cc
index 0cd06f9b75..080778f5ff 100644
--- a/media/libyuv/unit_test/cpu_test.cc
+++ b/media/libyuv/unit_test/cpu_test.cc
@@ -11,74 +11,180 @@
 #include <stdlib.h>
 #include <string.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/basic_types.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/version.h"
-#include "../unit_test/unit_test.h"
 
 namespace libyuv {
 
 TEST_F(LibYUVBaseTest, TestCpuHas) {
   int cpu_flags = TestCpuFlag(-1);
-  printf("Cpu Flags %x\n", cpu_flags);
+  printf("Cpu Flags %d\n", cpu_flags);
+#if defined(__arm__) || defined(__aarch64__)
   int has_arm = TestCpuFlag(kCpuHasARM);
-  printf("Has ARM %x\n", has_arm);
+  printf("Has ARM %d\n", has_arm);
   int has_neon = TestCpuFlag(kCpuHasNEON);
-  printf("Has NEON %x\n", has_neon);
+  printf("Has NEON %d\n", has_neon);
+#endif
   int has_x86 = TestCpuFlag(kCpuHasX86);
-  printf("Has X86 %x\n", has_x86);
   int has_sse2 = TestCpuFlag(kCpuHasSSE2);
-  printf("Has SSE2 %x\n", has_sse2);
   int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
-  printf("Has SSSE3 %x\n", has_ssse3);
   int has_sse41 = TestCpuFlag(kCpuHasSSE41);
-  printf("Has SSE4.1 %x\n", has_sse41);
   int has_sse42 = TestCpuFlag(kCpuHasSSE42);
-  printf("Has SSE4.2 %x\n", has_sse42);
   int has_avx = TestCpuFlag(kCpuHasAVX);
-  printf("Has AVX %x\n", has_avx);
   int has_avx2 = TestCpuFlag(kCpuHasAVX2);
-  printf("Has AVX2 %x\n", has_avx2);
   int has_erms = TestCpuFlag(kCpuHasERMS);
-  printf("Has ERMS %x\n", has_erms);
   int has_fma3 = TestCpuFlag(kCpuHasFMA3);
-  printf("Has FMA3 %x\n", has_fma3);
-  int has_avx3 = TestCpuFlag(kCpuHasAVX3);
-  printf("Has AVX3 %x\n", has_avx3);
+  int has_f16c = TestCpuFlag(kCpuHasF16C);
+  int has_gfni = TestCpuFlag(kCpuHasGFNI);
+  int has_avx512bw = TestCpuFlag(kCpuHasAVX512BW);
+  int has_avx512vl = TestCpuFlag(kCpuHasAVX512VL);
+  int has_avx512vnni = TestCpuFlag(kCpuHasAVX512VNNI);
+  int has_avx512vbmi = TestCpuFlag(kCpuHasAVX512VBMI);
+  int has_avx512vbmi2 = TestCpuFlag(kCpuHasAVX512VBMI2);
+  int has_avx512vbitalg = TestCpuFlag(kCpuHasAVX512VBITALG);
+  int has_avx512vpopcntdq = TestCpuFlag(kCpuHasAVX512VPOPCNTDQ);
+  printf("Has X86 %d\n", has_x86);
+  printf("Has SSE2 %d\n", has_sse2);
+  printf("Has SSSE3 %d\n", has_ssse3);
+  printf("Has SSE41 %d\n", has_sse41);
+  printf("Has SSE42 %d\n", has_sse42);
+  printf("Has AVX %d\n", has_avx);
+  printf("Has AVX2 %d\n", has_avx2);
+  printf("Has ERMS %d\n", has_erms);
+  printf("Has FMA3 %d\n", has_fma3);
+  printf("Has F16C %d\n", has_f16c);
+  printf("Has GFNI %d\n", has_gfni);
+  printf("Has AVX512BW %d\n", has_avx512bw);
+  printf("Has AVX512VL %d\n", has_avx512vl);
+  printf("Has AVX512VNNI %d\n", has_avx512vnni);
+  printf("Has AVX512VBMI %d\n", has_avx512vbmi);
+  printf("Has AVX512VBMI2 %d\n", has_avx512vbmi2);
+  printf("Has AVX512VBITALG %d\n", has_avx512vbitalg);
+  printf("Has AVX512VPOPCNTDQ %d\n", has_avx512vpopcntdq);
+
+#if defined(__mips__)
   int has_mips = TestCpuFlag(kCpuHasMIPS);
-  printf("Has MIPS %x\n", has_mips);
-  int has_dspr2 = TestCpuFlag(kCpuHasDSPR2);
-  printf("Has DSPR2 %x\n", has_dspr2);
+  printf("Has MIPS %d\n", has_mips);
+  int has_msa = TestCpuFlag(kCpuHasMSA);
+  printf("Has MSA %d\n", has_msa);
+#endif
+
+#if defined(__loongarch__)
+  int has_loongarch = TestCpuFlag(kCpuHasLOONGARCH);
+  printf("Has LOONGARCH %d\n", has_loongarch);
+  int has_lsx = TestCpuFlag(kCpuHasLSX);
+  printf("Has LSX %d\n", has_lsx);
+  int has_lasx = TestCpuFlag(kCpuHasLASX);
+  printf("Has LASX %d\n", has_lasx);
+#endif
 }
 
-TEST_F(LibYUVBaseTest, TestCpuCompilerEnabled) {
-#if defined(__aarch64__)
-  printf("Arm64 build\n");
+TEST_F(LibYUVBaseTest, TestCompilerMacros) {
+  // Tests all macros used in public headers.
+#ifdef __ATOMIC_RELAXED
+  printf("__ATOMIC_RELAXED %d\n", __ATOMIC_RELAXED);
+#endif
+#ifdef __cplusplus
+  printf("__cplusplus %ld\n", __cplusplus);
+#endif
+#ifdef __clang_major__
+  printf("__clang_major__ %d\n", __clang_major__);
+#endif
+#ifdef __clang_minor__
+  printf("__clang_minor__ %d\n", __clang_minor__);
+#endif
+#ifdef __GNUC__
+  printf("__GNUC__ %d\n", __GNUC__);
+#endif
+#ifdef __GNUC_MINOR__
+  printf("__GNUC_MINOR__ %d\n", __GNUC_MINOR__);
+#endif
+#ifdef __i386__
+  printf("__i386__ %d\n", __i386__);
 #endif
-#if defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON)
-  printf("Neon build enabled\n");
+#ifdef __mips
+  printf("__mips %d\n", __mips);
 #endif
-#if defined(__x86_64__) || defined(_M_X64)
-  printf("x64 build\n");
+#ifdef __mips_isa_rev
+  printf("__mips_isa_rev %d\n", __mips_isa_rev);
+#endif
+#ifdef __x86_64__
+  printf("__x86_64__ %d\n", __x86_64__);
 #endif
 #ifdef _MSC_VER
-printf("_MSC_VER %d\n", _MSC_VER);
+  printf("_MSC_VER %d\n", _MSC_VER);
+#endif
+#ifdef __aarch64__
+  printf("__aarch64__ %d\n", __aarch64__);
+#endif
+#ifdef __APPLE__
+  printf("__APPLE__ %d\n", __APPLE__);
+#endif
+#ifdef __arm__
+  printf("__arm__ %d\n", __arm__);
+#endif
+#ifdef __clang__
+  printf("__clang__ %d\n", __clang__);
+#endif
+#ifdef __CLR_VER
+  printf("__CLR_VER %d\n", __CLR_VER);
+#endif
+#ifdef __CYGWIN__
+  printf("__CYGWIN__ %d\n", __CYGWIN__);
+#endif
+#ifdef __llvm__
+  printf("__llvm__ %d\n", __llvm__);
+#endif
+#ifdef __mips_msa
+  printf("__mips_msa %d\n", __mips_msa);
+#endif
+#ifdef __native_client__
+  printf("__native_client__ %d\n", __native_client__);
+#endif
+#ifdef __pic__
+  printf("__pic__ %d\n", __pic__);
+#endif
+#ifdef __pnacl__
+  printf("__pnacl__ %d\n", __pnacl__);
+#endif
+#ifdef _M_IX86
+  printf("_M_IX86 %d\n", _M_IX86);
+#endif
+#ifdef _M_X64
+  printf("_M_X64 %d\n", _M_X64);
+#endif
+#ifdef _MIPS_ARCH_LOONGSON3A
+  printf("_MIPS_ARCH_LOONGSON3A %d\n", _MIPS_ARCH_LOONGSON3A);
+#endif
+#ifdef __loongarch__
+  printf("__loongarch__ %d\n", __loongarch__);
+#endif
+#ifdef _WIN32
+  printf("_WIN32 %d\n", _WIN32);
+#endif
+#ifdef GG_LONGLONG
+  printf("GG_LONGLONG %d\n", GG_LONGLONG);
+#endif
+#ifdef INT_TYPES_DEFINED
+  printf("INT_TYPES_DEFINED\n");
+#endif
+#ifdef __has_feature
+  printf("__has_feature\n");
+#if __has_feature(memory_sanitizer)
+  printf("__has_feature(memory_sanitizer) %d\n",
+         __has_feature(memory_sanitizer));
 #endif
-#if !defined(LIBYUV_DISABLE_X86) && (defined(GCC_HAS_AVX2) || \
-    defined(CLANG_HAS_AVX2) || defined(VISUALC_HAS_AVX2))
-  printf("Has AVX2 1\n");
-#else
-  printf("Has AVX2 0\n");
-  // If compiler does not support AVX2, the following function not expected:
 #endif
 }
 
-#if defined(__i386__) || defined(__x86_64__) || \
-    defined(_M_IX86) || defined(_M_X64)
+#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || \
+    defined(_M_X64)
 TEST_F(LibYUVBaseTest, TestCpuId) {
   int has_x86 = TestCpuFlag(kCpuHasX86);
   if (has_x86) {
-    uint32 cpu_info[4];
+    int cpu_info[4];
     // Vendor ID:
     // AuthenticAMD AMD processor
     // CentaurHauls Centaur processor
@@ -96,7 +202,7 @@ TEST_F(LibYUVBaseTest, TestCpuId) {
     cpu_info[3] = 0;
     printf("Cpu Vendor: %s %x %x %x\n", reinterpret_cast<char*>(&cpu_info[0]),
            cpu_info[0], cpu_info[1], cpu_info[2]);
-    EXPECT_EQ(12, strlen(reinterpret_cast<char*>(&cpu_info[0])));
+    EXPECT_EQ(12u, strlen(reinterpret_cast<char*>(&cpu_info[0])));
 
     // CPU Family and Model
     // 3:0 - Stepping
@@ -108,8 +214,8 @@ TEST_F(LibYUVBaseTest, TestCpuId) {
     CpuId(1, 0, cpu_info);
     int family = ((cpu_info[0] >> 8) & 0x0f) | ((cpu_info[0] >> 16) & 0xff0);
     int model = ((cpu_info[0] >> 4) & 0x0f) | ((cpu_info[0] >> 12) & 0xf0);
-    printf("Cpu Family %d (0x%x), Model %d (0x%x)\n", family, family,
-           model, model);
+    printf("Cpu Family %d (0x%x), Model %d (0x%x)\n", family, family, model,
+           model);
   }
 }
 #endif
@@ -125,6 +231,8 @@ static int FileExists(const char* file_name) {
 
 TEST_F(LibYUVBaseTest, TestLinuxNeon) {
   if (FileExists("../../unit_test/testdata/arm_v7.txt")) {
+    printf("Note: testing to load \"../../unit_test/testdata/arm_v7.txt\"\n");
+
     EXPECT_EQ(0, ArmCpuCaps("../../unit_test/testdata/arm_v7.txt"));
     EXPECT_EQ(kCpuHasNEON, ArmCpuCaps("../../unit_test/testdata/tegra3.txt"));
     EXPECT_EQ(kCpuHasNEON, ArmCpuCaps("../../unit_test/testdata/juno.txt"));
@@ -132,8 +240,56 @@ TEST_F(LibYUVBaseTest, TestLinuxNeon) {
     printf("WARNING: unable to load \"../../unit_test/testdata/arm_v7.txt\"\n");
   }
 #if defined(__linux__) && defined(__ARM_NEON__)
-  EXPECT_EQ(kCpuHasNEON, ArmCpuCaps("/proc/cpuinfo"));
+  if (FileExists("/proc/cpuinfo")) {
+    if (kCpuHasNEON != ArmCpuCaps("/proc/cpuinfo")) {
+      // This can happen on ARM emulator but /proc/cpuinfo is from host.
+      printf("WARNING: Neon build enabled but CPU does not have NEON\n");
+    }
+  } else {
+    printf("WARNING: unable to load \"/proc/cpuinfo\"\n");
+  }
+#endif
+}
+
+TEST_F(LibYUVBaseTest, TestLinuxMipsMsa) {
+  if (FileExists("../../unit_test/testdata/mips.txt")) {
+    printf("Note: testing to load \"../../unit_test/testdata/mips.txt\"\n");
+
+    EXPECT_EQ(0, MipsCpuCaps("../../unit_test/testdata/mips.txt"));
+    EXPECT_EQ(kCpuHasMSA, MipsCpuCaps("../../unit_test/testdata/mips_msa.txt"));
+    EXPECT_EQ(kCpuHasMSA,
+              MipsCpuCaps("../../unit_test/testdata/mips_loongson2k.txt"));
+  } else {
+    printf("WARNING: unable to load \"../../unit_test/testdata/mips.txt\"\n");
+  }
+}
+
+// TODO(fbarchard): Fix clangcl test of cpuflags.
+#ifdef _MSC_VER
+TEST_F(LibYUVBaseTest, DISABLED_TestSetCpuFlags) {
+#else
+TEST_F(LibYUVBaseTest, TestSetCpuFlags) {
 #endif
+  // Reset any masked flags that may have been set so auto init is enabled.
+  MaskCpuFlags(0);
+
+  int original_cpu_flags = TestCpuFlag(-1);
+
+  // Test setting different CPU configurations.
+  int cpu_flags = kCpuHasARM | kCpuHasNEON | kCpuInitialized;
+  SetCpuFlags(cpu_flags);
+  EXPECT_EQ(cpu_flags, TestCpuFlag(-1));
+
+  cpu_flags = kCpuHasX86 | kCpuInitialized;
+  SetCpuFlags(cpu_flags);
+  EXPECT_EQ(cpu_flags, TestCpuFlag(-1));
+
+  // Test that setting 0 turns auto-init back on.
+  SetCpuFlags(0);
+  EXPECT_EQ(original_cpu_flags, TestCpuFlag(-1));
+
+  // Restore the CPU flag mask.
+  MaskCpuFlags(benchmark_cpu_info_);
 }
 
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/cpu_thread_test.cc b/media/libyuv/unit_test/cpu_thread_test.cc
new file mode 100644
index 0000000000..69aab74e7c
--- /dev/null
+++ b/media/libyuv/unit_test/cpu_thread_test.cc
@@ -0,0 +1,63 @@
+/*
+ *  Copyright 2017 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <gtest/gtest.h>
+
+#include "libyuv/cpu_id.h"
+
+#if defined(__clang__) && !defined(__wasm__)
+#if __has_include(<pthread.h>)
+#define LIBYUV_HAVE_PTHREAD 1
+#endif
+#elif defined(__linux__)
+#define LIBYUV_HAVE_PTHREAD 1
+#endif
+
+#ifdef LIBYUV_HAVE_PTHREAD
+#include <pthread.h>
+#endif
+
+namespace libyuv {
+
+#ifdef LIBYUV_HAVE_PTHREAD
+void* ThreadMain(void* arg) {
+  int* flags = static_cast<int*>(arg);
+
+  *flags = TestCpuFlag(kCpuInitialized);
+  return nullptr;
+}
+#endif  // LIBYUV_HAVE_PTHREAD
+
+// Call TestCpuFlag() from two threads. ThreadSanitizer should not report any
+// data race.
+TEST(LibYUVCpuThreadTest, TestCpuFlagMultipleThreads) {
+#ifdef LIBYUV_HAVE_PTHREAD
+  int cpu_flags1;
+  int cpu_flags2;
+  int ret;
+  pthread_t thread1;
+  pthread_t thread2;
+
+  MaskCpuFlags(0);  // Reset to 0 to allow auto detect.
+  ret = pthread_create(&thread1, nullptr, ThreadMain, &cpu_flags1);
+  ASSERT_EQ(ret, 0);
+  ret = pthread_create(&thread2, nullptr, ThreadMain, &cpu_flags2);
+  ASSERT_EQ(ret, 0);
+  ret = pthread_join(thread1, nullptr);
+  EXPECT_EQ(ret, 0);
+  ret = pthread_join(thread2, nullptr);
+  EXPECT_EQ(ret, 0);
+  EXPECT_EQ(cpu_flags1, cpu_flags2);
+#else
+  printf("pthread unavailable; Test skipped.");
+#endif  // LIBYUV_HAVE_PTHREAD
+}
+
+}  // namespace libyuv
diff --git a/media/libyuv/unit_test/math_test.cc b/media/libyuv/unit_test/math_test.cc
index 19af9f6b07..a1544c122b 100644
--- a/media/libyuv/unit_test/math_test.cc
+++ b/media/libyuv/unit_test/math_test.cc
@@ -12,14 +12,18 @@
 #include <string.h>
 #include <time.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/basic_types.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/scale.h"
+
+#ifdef ENABLE_ROW_TESTS
 #include "libyuv/scale_row.h"
-#include "../unit_test/unit_test.h"
+#endif
 
 namespace libyuv {
 
+#ifdef ENABLE_ROW_TESTS
 TEST_F(LibYUVBaseTest, TestFixedDiv) {
   int num[1280];
   int div[1280];
@@ -65,8 +69,8 @@ TEST_F(LibYUVBaseTest, TestFixedDiv) {
   }
   EXPECT_EQ(123 * 65536, libyuv::FixedDiv(123, 1));
 
-  MemRandomize(reinterpret_cast<uint8*>(&num[0]), sizeof(num));
-  MemRandomize(reinterpret_cast<uint8*>(&div[0]), sizeof(div));
+  MemRandomize(reinterpret_cast<uint8_t*>(&num[0]), sizeof(num));
+  MemRandomize(reinterpret_cast<uint8_t*>(&div[0]), sizeof(div));
   for (int j = 0; j < 1280; ++j) {
     if (div[j] == 0) {
       div[j] = 1280;
@@ -90,8 +94,8 @@ TEST_F(LibYUVBaseTest, TestFixedDiv_Opt) {
   int result_opt[1280];
   int result_c[1280];
 
-  MemRandomize(reinterpret_cast<uint8*>(&num[0]), sizeof(num));
-  MemRandomize(reinterpret_cast<uint8*>(&div[0]), sizeof(div));
+  MemRandomize(reinterpret_cast<uint8_t*>(&num[0]), sizeof(num));
+  MemRandomize(reinterpret_cast<uint8_t*>(&div[0]), sizeof(div));
   for (int j = 0; j < 1280; ++j) {
     num[j] &= 4095;  // Make numerator smaller.
     div[j] &= 4095;  // Make divisor smaller.
@@ -124,8 +128,8 @@ TEST_F(LibYUVBaseTest, TestFixedDiv1_Opt) {
   int result_opt[1280];
   int result_c[1280];
 
-  MemRandomize(reinterpret_cast<uint8*>(&num[0]), sizeof(num));
-  MemRandomize(reinterpret_cast<uint8*>(&div[0]), sizeof(div));
+  MemRandomize(reinterpret_cast<uint8_t*>(&num[0]), sizeof(num));
+  MemRandomize(reinterpret_cast<uint8_t*>(&div[0]), sizeof(div));
   for (int j = 0; j < 1280; ++j) {
     num[j] &= 4095;  // Make numerator smaller.
     div[j] &= 4095;  // Make divisor smaller.
@@ -151,5 +155,6 @@ TEST_F(LibYUVBaseTest, TestFixedDiv1_Opt) {
     EXPECT_NEAR(result_c[j], result_opt[j], 1);
   }
 }
+#endif  // ENABLE_ROW_TESTS
 
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/planar_test.cc b/media/libyuv/unit_test/planar_test.cc
index bc0eebb528..ad97b87e21 100644
--- a/media/libyuv/unit_test/planar_test.cc
+++ b/media/libyuv/unit_test/planar_test.cc
@@ -8,9 +8,11 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <math.h>
 #include <stdlib.h>
 #include <time.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/compare.h"
 #include "libyuv/convert.h"
 #include "libyuv/convert_argb.h"
@@ -19,7 +21,19 @@
 #include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
-#include "../unit_test/unit_test.h"
+#include "libyuv/scale.h"
+
+#ifdef ENABLE_ROW_TESTS
+// row.h defines SIMD_ALIGNED, overriding unit_test.h
+// TODO(fbarchard): Remove row.h from unittests.  Test public functions.
+#include "libyuv/row.h" /* For ScaleSumSamples_Neon */
+#endif
+
+#if defined(LIBYUV_BIT_EXACT)
+#define EXPECTED_ATTENUATE_DIFF 0
+#else
+#define EXPECTED_ATTENUATE_DIFF 2
+#endif
 
 namespace libyuv {
 
@@ -90,11 +104,11 @@ TEST_F(LibYUVPlanarTest, TestAttenuate) {
   EXPECT_EQ(0, atten_pixels[0 * 4 + 3]);
   EXPECT_EQ(64, atten_pixels[128 * 4 + 0]);
   EXPECT_EQ(32, atten_pixels[128 * 4 + 1]);
-  EXPECT_EQ(21,  atten_pixels[128 * 4 + 2]);
+  EXPECT_EQ(21, atten_pixels[128 * 4 + 2]);
   EXPECT_EQ(128, atten_pixels[128 * 4 + 3]);
-  EXPECT_NEAR(255, atten_pixels[255 * 4 + 0], 1);
-  EXPECT_NEAR(127, atten_pixels[255 * 4 + 1], 1);
-  EXPECT_NEAR(85,  atten_pixels[255 * 4 + 2], 1);
+  EXPECT_NEAR(254, atten_pixels[255 * 4 + 0], EXPECTED_ATTENUATE_DIFF);
+  EXPECT_NEAR(127, atten_pixels[255 * 4 + 1], EXPECTED_ATTENUATE_DIFF);
+  EXPECT_NEAR(85, atten_pixels[255 * 4 + 2], EXPECTED_ATTENUATE_DIFF);
   EXPECT_EQ(255, atten_pixels[255 * 4 + 3]);
 
   free_aligned_buffer_page_end(atten2_pixels);
@@ -103,9 +117,13 @@ TEST_F(LibYUVPlanarTest, TestAttenuate) {
   free_aligned_buffer_page_end(orig_pixels);
 }
 
-static int TestAttenuateI(int width, int height, int benchmark_iterations,
-                          int disable_cpu_flags, int benchmark_cpu_info,
-                          int invert, int off) {
+static int TestAttenuateI(int width,
+                          int height,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info,
+                          int invert,
+                          int off) {
   if (width < 1) {
     width = 1;
   }
@@ -121,20 +139,17 @@ static int TestAttenuateI(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBAttenuate(src_argb + off, kStride,
-                dst_argb_c, kStride,
-                width, invert * height);
+  ARGBAttenuate(src_argb + off, kStride, dst_argb_c, kStride, width,
+                invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBAttenuate(src_argb + off, kStride,
-                  dst_argb_opt, kStride,
-                  width, invert * height);
+    ARGBAttenuate(src_argb + off, kStride, dst_argb_opt, kStride, width,
+                  invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -146,40 +161,41 @@ static int TestAttenuateI(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Any) {
-  int max_diff = TestAttenuateI(benchmark_width_ - 1, benchmark_height_,
-                                benchmark_iterations_,
-                                disable_cpu_flags_, benchmark_cpu_info_,
-                                +1, 0);
-  EXPECT_LE(max_diff, 2);
+  int max_diff = TestAttenuateI(benchmark_width_ + 1, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, +1, 0);
+
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Unaligned) {
-  int max_diff = TestAttenuateI(benchmark_width_, benchmark_height_,
-                                benchmark_iterations_,
-                                disable_cpu_flags_, benchmark_cpu_info_,
-                                +1, 1);
-  EXPECT_LE(max_diff, 2);
+  int max_diff =
+      TestAttenuateI(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Invert) {
-  int max_diff = TestAttenuateI(benchmark_width_, benchmark_height_,
-                                benchmark_iterations_,
-                                disable_cpu_flags_, benchmark_cpu_info_,
-                                -1, 0);
-  EXPECT_LE(max_diff, 2);
+  int max_diff =
+      TestAttenuateI(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Opt) {
-  int max_diff = TestAttenuateI(benchmark_width_, benchmark_height_,
-                                benchmark_iterations_,
-                                disable_cpu_flags_, benchmark_cpu_info_,
-                                +1, 0);
-  EXPECT_LE(max_diff, 2);
+  int max_diff =
+      TestAttenuateI(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
-static int TestUnattenuateI(int width, int height, int benchmark_iterations,
-                            int disable_cpu_flags, int benchmark_cpu_info,
-                            int invert, int off) {
+static int TestUnattenuateI(int width,
+                            int height,
+                            int benchmark_iterations,
+                            int disable_cpu_flags,
+                            int benchmark_cpu_info,
+                            int invert,
+                            int off) {
   if (width < 1) {
     width = 1;
   }
@@ -191,27 +207,23 @@ static int TestUnattenuateI(int width, int height, int benchmark_iterations,
   for (int i = 0; i < kStride * height; ++i) {
     src_argb[i + off] = (fastrand() & 0xff);
   }
-  ARGBAttenuate(src_argb + off, kStride,
-                src_argb + off, kStride,
-                width, height);
+  ARGBAttenuate(src_argb + off, kStride, src_argb + off, kStride, width,
+                height);
   memset(dst_argb_c, 0, kStride * height);
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBUnattenuate(src_argb + off, kStride,
-                  dst_argb_c, kStride,
-                  width, invert * height);
+  ARGBUnattenuate(src_argb + off, kStride, dst_argb_c, kStride, width,
+                  invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBUnattenuate(src_argb + off, kStride,
-                    dst_argb_opt, kStride,
-                    width, invert * height);
+    ARGBUnattenuate(src_argb + off, kStride, dst_argb_opt, kStride, width,
+                    invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -223,40 +235,36 @@ static int TestUnattenuateI(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Any) {
-  int max_diff = TestUnattenuateI(benchmark_width_ - 1, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  +1, 0);
-  EXPECT_LE(max_diff, 2);
+  int max_diff = TestUnattenuateI(benchmark_width_ + 1, benchmark_height_,
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, +1, 0);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Unaligned) {
   int max_diff = TestUnattenuateI(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  +1, 1);
-  EXPECT_LE(max_diff, 2);
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, +1, 1);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Invert) {
   int max_diff = TestUnattenuateI(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  -1, 0);
-  EXPECT_LE(max_diff, 2);
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, -1, 0);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Opt) {
   int max_diff = TestUnattenuateI(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  +1, 0);
-  EXPECT_LE(max_diff, 2);
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, +1, 0);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBComputeCumulativeSum) {
-  SIMD_ALIGNED(uint8 orig_pixels[16][16][4]);
-  SIMD_ALIGNED(int32 added_pixels[16][16][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16][16][4]);
+  SIMD_ALIGNED(int32_t added_pixels[16][16][4]);
 
   for (int y = 0; y < 16; ++y) {
     for (int x = 0; x < 16; ++x) {
@@ -268,8 +276,7 @@ TEST_F(LibYUVPlanarTest, TestARGBComputeCumulativeSum) {
   }
 
   ARGBComputeCumulativeSum(&orig_pixels[0][0][0], 16 * 4,
-                           &added_pixels[0][0][0], 16 * 4,
-                           16, 16);
+                           &added_pixels[0][0][0], 16 * 4, 16, 16);
 
   for (int y = 0; y < 16; ++y) {
     for (int x = 0; x < 16; ++x) {
@@ -281,8 +288,9 @@ TEST_F(LibYUVPlanarTest, TestARGBComputeCumulativeSum) {
   }
 }
 
+// near is for legacy platforms.
 TEST_F(LibYUVPlanarTest, TestARGBGray) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   // Test blue
@@ -317,17 +325,17 @@ TEST_F(LibYUVPlanarTest, TestARGBGray) {
   orig_pixels[5][3] = 224u;
   // Do 16 to test asm version.
   ARGBGray(&orig_pixels[0][0], 0, 0, 0, 16, 1);
-  EXPECT_EQ(30u, orig_pixels[0][0]);
-  EXPECT_EQ(30u, orig_pixels[0][1]);
-  EXPECT_EQ(30u, orig_pixels[0][2]);
+  EXPECT_NEAR(29u, orig_pixels[0][0], 1);
+  EXPECT_NEAR(29u, orig_pixels[0][1], 1);
+  EXPECT_NEAR(29u, orig_pixels[0][2], 1);
   EXPECT_EQ(128u, orig_pixels[0][3]);
   EXPECT_EQ(149u, orig_pixels[1][0]);
   EXPECT_EQ(149u, orig_pixels[1][1]);
   EXPECT_EQ(149u, orig_pixels[1][2]);
   EXPECT_EQ(0u, orig_pixels[1][3]);
-  EXPECT_EQ(76u, orig_pixels[2][0]);
-  EXPECT_EQ(76u, orig_pixels[2][1]);
-  EXPECT_EQ(76u, orig_pixels[2][2]);
+  EXPECT_NEAR(77u, orig_pixels[2][0], 1);
+  EXPECT_NEAR(77u, orig_pixels[2][1], 1);
+  EXPECT_NEAR(77u, orig_pixels[2][2], 1);
   EXPECT_EQ(255u, orig_pixels[2][3]);
   EXPECT_EQ(0u, orig_pixels[3][0]);
   EXPECT_EQ(0u, orig_pixels[3][1]);
@@ -337,9 +345,9 @@ TEST_F(LibYUVPlanarTest, TestARGBGray) {
   EXPECT_EQ(255u, orig_pixels[4][1]);
   EXPECT_EQ(255u, orig_pixels[4][2]);
   EXPECT_EQ(255u, orig_pixels[4][3]);
-  EXPECT_EQ(96u, orig_pixels[5][0]);
-  EXPECT_EQ(96u, orig_pixels[5][1]);
-  EXPECT_EQ(96u, orig_pixels[5][2]);
+  EXPECT_NEAR(97u, orig_pixels[5][0], 1);
+  EXPECT_NEAR(97u, orig_pixels[5][1], 1);
+  EXPECT_NEAR(97u, orig_pixels[5][2], 1);
   EXPECT_EQ(224u, orig_pixels[5][3]);
   for (int i = 0; i < 1280; ++i) {
     orig_pixels[i][0] = i;
@@ -353,8 +361,8 @@ TEST_F(LibYUVPlanarTest, TestARGBGray) {
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBGrayTo) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8 gray_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t gray_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   // Test blue
@@ -389,30 +397,30 @@ TEST_F(LibYUVPlanarTest, TestARGBGrayTo) {
   orig_pixels[5][3] = 224u;
   // Do 16 to test asm version.
   ARGBGrayTo(&orig_pixels[0][0], 0, &gray_pixels[0][0], 0, 16, 1);
-  EXPECT_EQ(30u, gray_pixels[0][0]);
-  EXPECT_EQ(30u, gray_pixels[0][1]);
-  EXPECT_EQ(30u, gray_pixels[0][2]);
-  EXPECT_EQ(128u, gray_pixels[0][3]);
-  EXPECT_EQ(149u, gray_pixels[1][0]);
-  EXPECT_EQ(149u, gray_pixels[1][1]);
-  EXPECT_EQ(149u, gray_pixels[1][2]);
-  EXPECT_EQ(0u, gray_pixels[1][3]);
-  EXPECT_EQ(76u, gray_pixels[2][0]);
-  EXPECT_EQ(76u, gray_pixels[2][1]);
-  EXPECT_EQ(76u, gray_pixels[2][2]);
-  EXPECT_EQ(255u, gray_pixels[2][3]);
-  EXPECT_EQ(0u, gray_pixels[3][0]);
-  EXPECT_EQ(0u, gray_pixels[3][1]);
-  EXPECT_EQ(0u, gray_pixels[3][2]);
-  EXPECT_EQ(255u, gray_pixels[3][3]);
-  EXPECT_EQ(255u, gray_pixels[4][0]);
-  EXPECT_EQ(255u, gray_pixels[4][1]);
-  EXPECT_EQ(255u, gray_pixels[4][2]);
-  EXPECT_EQ(255u, gray_pixels[4][3]);
-  EXPECT_EQ(96u, gray_pixels[5][0]);
-  EXPECT_EQ(96u, gray_pixels[5][1]);
-  EXPECT_EQ(96u, gray_pixels[5][2]);
-  EXPECT_EQ(224u, gray_pixels[5][3]);
+  EXPECT_NEAR(30u, gray_pixels[0][0], 1);
+  EXPECT_NEAR(30u, gray_pixels[0][1], 1);
+  EXPECT_NEAR(30u, gray_pixels[0][2], 1);
+  EXPECT_NEAR(128u, gray_pixels[0][3], 1);
+  EXPECT_NEAR(149u, gray_pixels[1][0], 1);
+  EXPECT_NEAR(149u, gray_pixels[1][1], 1);
+  EXPECT_NEAR(149u, gray_pixels[1][2], 1);
+  EXPECT_NEAR(0u, gray_pixels[1][3], 1);
+  EXPECT_NEAR(76u, gray_pixels[2][0], 1);
+  EXPECT_NEAR(76u, gray_pixels[2][1], 1);
+  EXPECT_NEAR(76u, gray_pixels[2][2], 1);
+  EXPECT_NEAR(255u, gray_pixels[2][3], 1);
+  EXPECT_NEAR(0u, gray_pixels[3][0], 1);
+  EXPECT_NEAR(0u, gray_pixels[3][1], 1);
+  EXPECT_NEAR(0u, gray_pixels[3][2], 1);
+  EXPECT_NEAR(255u, gray_pixels[3][3], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][0], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][1], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][2], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][3], 1);
+  EXPECT_NEAR(96u, gray_pixels[5][0], 1);
+  EXPECT_NEAR(96u, gray_pixels[5][1], 1);
+  EXPECT_NEAR(96u, gray_pixels[5][2], 1);
+  EXPECT_NEAR(224u, gray_pixels[5][3], 1);
   for (int i = 0; i < 1280; ++i) {
     orig_pixels[i][0] = i;
     orig_pixels[i][1] = i / 2;
@@ -422,10 +430,24 @@ TEST_F(LibYUVPlanarTest, TestARGBGrayTo) {
   for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
     ARGBGrayTo(&orig_pixels[0][0], 0, &gray_pixels[0][0], 0, 1280, 1);
   }
+
+  for (int i = 0; i < 256; ++i) {
+    orig_pixels[i][0] = i;
+    orig_pixels[i][1] = i;
+    orig_pixels[i][2] = i;
+    orig_pixels[i][3] = i;
+  }
+  ARGBGray(&orig_pixels[0][0], 0, 0, 0, 256, 1);
+  for (int i = 0; i < 256; ++i) {
+    EXPECT_EQ(i, orig_pixels[i][0]);
+    EXPECT_EQ(i, orig_pixels[i][1]);
+    EXPECT_EQ(i, orig_pixels[i][2]);
+    EXPECT_EQ(i, orig_pixels[i][3]);
+  }
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBSepia) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   // Test blue
@@ -497,16 +519,14 @@ TEST_F(LibYUVPlanarTest, TestARGBSepia) {
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBColorMatrix) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels_opt[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels_c[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t dst_pixels_opt[1280][4]);
+  SIMD_ALIGNED(uint8_t dst_pixels_c[1280][4]);
 
   // Matrix for Sepia.
-  SIMD_ALIGNED(static const int8 kRGBToSepia[]) = {
-    17 / 2, 68 / 2, 35 / 2, 0,
-    22 / 2, 88 / 2, 45 / 2, 0,
-    24 / 2, 98 / 2, 50 / 2, 0,
-    0, 0, 0, 64,  // Copy alpha.
+  SIMD_ALIGNED(static const int8_t kRGBToSepia[]) = {
+      17 / 2, 68 / 2, 35 / 2, 0, 22 / 2, 88 / 2, 45 / 2, 0,
+      24 / 2, 98 / 2, 50 / 2, 0, 0,      0,      0,      64,  // Copy alpha.
   };
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
@@ -575,14 +595,12 @@ TEST_F(LibYUVPlanarTest, TestARGBColorMatrix) {
 }
 
 TEST_F(LibYUVPlanarTest, TestRGBColorMatrix) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
 
   // Matrix for Sepia.
-  SIMD_ALIGNED(static const int8 kRGBToSepia[]) = {
-    17, 68, 35, 0,
-    22, 88, 45, 0,
-    24, 98, 50, 0,
-    0, 0, 0, 0,  // Unused but makes matrix 16 bytes.
+  SIMD_ALIGNED(static const int8_t kRGBToSepia[]) = {
+      17, 68, 35, 0, 22, 88, 45, 0,
+      24, 98, 50, 0, 0,  0,  0,  0,  // Unused but makes matrix 16 bytes.
   };
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
@@ -637,15 +655,12 @@ TEST_F(LibYUVPlanarTest, TestRGBColorMatrix) {
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBColorTable) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   // Matrix for Sepia.
-  static const uint8 kARGBTable[256 * 4] = {
-    1u, 2u, 3u, 4u,
-    5u, 6u, 7u, 8u,
-    9u, 10u, 11u, 12u,
-    13u, 14u, 15u, 16u,
+  static const uint8_t kARGBTable[256 * 4] = {
+      1u, 2u, 3u, 4u, 5u, 6u, 7u, 8u, 9u, 10u, 11u, 12u, 13u, 14u, 15u, 16u,
   };
 
   orig_pixels[0][0] = 0u;
@@ -696,15 +711,12 @@ TEST_F(LibYUVPlanarTest, TestARGBColorTable) {
 
 // Same as TestARGBColorTable except alpha does not change.
 TEST_F(LibYUVPlanarTest, TestRGBColorTable) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   // Matrix for Sepia.
-  static const uint8 kARGBTable[256 * 4] = {
-    1u, 2u, 3u, 4u,
-    5u, 6u, 7u, 8u,
-    9u, 10u, 11u, 12u,
-    13u, 14u, 15u, 16u,
+  static const uint8_t kARGBTable[256 * 4] = {
+      1u, 2u, 3u, 4u, 5u, 6u, 7u, 8u, 9u, 10u, 11u, 12u, 13u, 14u, 15u, 16u,
   };
 
   orig_pixels[0][0] = 0u;
@@ -754,7 +766,7 @@ TEST_F(LibYUVPlanarTest, TestRGBColorTable) {
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBQuantize) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
 
   for (int i = 0; i < 1280; ++i) {
     orig_pixels[i][0] = i;
@@ -762,8 +774,8 @@ TEST_F(LibYUVPlanarTest, TestARGBQuantize) {
     orig_pixels[i][2] = i / 3;
     orig_pixels[i][3] = i;
   }
-  ARGBQuantize(&orig_pixels[0][0], 0,
-               (65536 + (8 / 2)) / 8, 8, 8 / 2, 0, 0, 1280, 1);
+  ARGBQuantize(&orig_pixels[0][0], 0, (65536 + (8 / 2)) / 8, 8, 8 / 2, 0, 0,
+               1280, 1);
 
   for (int i = 0; i < 1280; ++i) {
     EXPECT_EQ((i / 8 * 8 + 8 / 2) & 255, orig_pixels[i][0]);
@@ -772,37 +784,85 @@ TEST_F(LibYUVPlanarTest, TestARGBQuantize) {
     EXPECT_EQ(i & 255, orig_pixels[i][3]);
   }
   for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
-    ARGBQuantize(&orig_pixels[0][0], 0,
-                 (65536 + (8 / 2)) / 8, 8, 8 / 2, 0, 0, 1280, 1);
+    ARGBQuantize(&orig_pixels[0][0], 0, (65536 + (8 / 2)) / 8, 8, 8 / 2, 0, 0,
+                 1280, 1);
   }
 }
 
-TEST_F(LibYUVPlanarTest, TestARGBMirror) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels[1280][4]);
+TEST_F(LibYUVPlanarTest, ARGBMirror_Opt) {
+  align_buffer_page_end(src_pixels, benchmark_width_ * benchmark_height_ * 4);
+  align_buffer_page_end(dst_pixels_opt,
+                        benchmark_width_ * benchmark_height_ * 4);
+  align_buffer_page_end(dst_pixels_c, benchmark_width_ * benchmark_height_ * 4);
 
-  for (int i = 0; i < 1280; ++i) {
-    orig_pixels[i][0] = i;
-    orig_pixels[i][1] = i / 2;
-    orig_pixels[i][2] = i / 3;
-    orig_pixels[i][3] = i / 4;
+  MemRandomize(src_pixels, benchmark_width_ * benchmark_height_ * 4);
+  MaskCpuFlags(disable_cpu_flags_);
+  ARGBMirror(src_pixels, benchmark_width_ * 4, dst_pixels_c,
+             benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ARGBMirror(src_pixels, benchmark_width_ * 4, dst_pixels_opt,
+               benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+  }
+  for (int i = 0; i < benchmark_width_ * benchmark_height_ * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
-  ARGBMirror(&orig_pixels[0][0], 0, &dst_pixels[0][0], 0, 1280, 1);
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
 
-  for (int i = 0; i < 1280; ++i) {
-    EXPECT_EQ(i & 255, dst_pixels[1280 - 1 - i][0]);
-    EXPECT_EQ((i / 2) & 255, dst_pixels[1280 - 1 - i][1]);
-    EXPECT_EQ((i / 3) & 255, dst_pixels[1280 - 1 - i][2]);
-    EXPECT_EQ((i / 4) & 255, dst_pixels[1280 - 1 - i][3]);
+TEST_F(LibYUVPlanarTest, MirrorPlane_Opt) {
+  align_buffer_page_end(src_pixels, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_pixels_opt, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_pixels_c, benchmark_width_ * benchmark_height_);
+
+  MemRandomize(src_pixels, benchmark_width_ * benchmark_height_);
+  MaskCpuFlags(disable_cpu_flags_);
+  MirrorPlane(src_pixels, benchmark_width_, dst_pixels_c, benchmark_width_,
+              benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MirrorPlane(src_pixels, benchmark_width_, dst_pixels_opt, benchmark_width_,
+                benchmark_width_, benchmark_height_);
   }
-  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
-    ARGBMirror(&orig_pixels[0][0], 0, &dst_pixels[0][0], 0, 1280, 1);
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, MirrorUVPlane_Opt) {
+  align_buffer_page_end(src_pixels, benchmark_width_ * benchmark_height_ * 2);
+  align_buffer_page_end(dst_pixels_opt,
+                        benchmark_width_ * benchmark_height_ * 2);
+  align_buffer_page_end(dst_pixels_c, benchmark_width_ * benchmark_height_ * 2);
+
+  MemRandomize(src_pixels, benchmark_width_ * benchmark_height_ * 2);
+  MaskCpuFlags(disable_cpu_flags_);
+  MirrorUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_c,
+                benchmark_width_ * 2, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MirrorUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_opt,
+                  benchmark_width_ * 2, benchmark_width_, benchmark_height_);
+  }
+  for (int i = 0; i < benchmark_width_ * benchmark_height_ * 2; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
 }
 
 TEST_F(LibYUVPlanarTest, TestShade) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8 shade_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t shade_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   orig_pixels[0][0] = 10u;
@@ -859,9 +919,9 @@ TEST_F(LibYUVPlanarTest, TestShade) {
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBInterpolate) {
-  SIMD_ALIGNED(uint8 orig_pixels_0[1280][4]);
-  SIMD_ALIGNED(uint8 orig_pixels_1[1280][4]);
-  SIMD_ALIGNED(uint8 interpolate_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels_0[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels_1[1280][4]);
+  SIMD_ALIGNED(uint8_t interpolate_pixels[1280][4]);
   memset(orig_pixels_0, 0, sizeof(orig_pixels_0));
   memset(orig_pixels_1, 0, sizeof(orig_pixels_1));
 
@@ -940,9 +1000,9 @@ TEST_F(LibYUVPlanarTest, TestARGBInterpolate) {
 }
 
 TEST_F(LibYUVPlanarTest, TestInterpolatePlane) {
-  SIMD_ALIGNED(uint8 orig_pixels_0[1280]);
-  SIMD_ALIGNED(uint8 orig_pixels_1[1280]);
-  SIMD_ALIGNED(uint8 interpolate_pixels[1280]);
+  SIMD_ALIGNED(uint8_t orig_pixels_0[1280]);
+  SIMD_ALIGNED(uint8_t orig_pixels_1[1280]);
+  SIMD_ALIGNED(uint8_t interpolate_pixels[1280]);
   memset(orig_pixels_0, 0, sizeof(orig_pixels_0));
   memset(orig_pixels_1, 0, sizeof(orig_pixels_1));
 
@@ -1020,48 +1080,126 @@ TEST_F(LibYUVPlanarTest, TestInterpolatePlane) {
   }
 }
 
-#define TESTTERP(FMT_A, BPP_A, STRIDE_A,                                       \
-                 FMT_B, BPP_B, STRIDE_B,                                       \
-                 W1280, TERP, N, NEG, OFF)                                     \
-TEST_F(LibYUVPlanarTest, ARGBInterpolate##TERP##N) {                           \
-  const int kWidth = ((W1280) > 0) ? (W1280) : 1;                              \
-  const int kHeight = benchmark_height_;                                       \
-  const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;  \
-  const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;  \
-  align_buffer_page_end(src_argb_a, kStrideA * kHeight + OFF);                 \
-  align_buffer_page_end(src_argb_b, kStrideA * kHeight + OFF);                 \
-  align_buffer_page_end(dst_argb_c, kStrideB * kHeight);                       \
-  align_buffer_page_end(dst_argb_opt, kStrideB * kHeight);                     \
-  for (int i = 0; i < kStrideA * kHeight; ++i) {                               \
-    src_argb_a[i + OFF] = (fastrand() & 0xff);                                 \
-    src_argb_b[i + OFF] = (fastrand() & 0xff);                                 \
-  }                                                                            \
-  MaskCpuFlags(disable_cpu_flags_);                                            \
-  ARGBInterpolate(src_argb_a + OFF, kStrideA,                                  \
-                  src_argb_b + OFF, kStrideA,                                  \
-                  dst_argb_c, kStrideB,                                        \
-                  kWidth, NEG kHeight, TERP);                                  \
-  MaskCpuFlags(benchmark_cpu_info_);                                           \
-  for (int i = 0; i < benchmark_iterations_; ++i) {                            \
-    ARGBInterpolate(src_argb_a + OFF, kStrideA,                                \
-                    src_argb_b + OFF, kStrideA,                                \
-                    dst_argb_opt, kStrideB,                                    \
-                    kWidth, NEG kHeight, TERP);                                \
-  }                                                                            \
-  for (int i = 0; i < kStrideB * kHeight; ++i) {                               \
-    EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                                 \
-  }                                                                            \
-  free_aligned_buffer_page_end(src_argb_a);                                    \
-  free_aligned_buffer_page_end(src_argb_b);                                    \
-  free_aligned_buffer_page_end(dst_argb_c);                                    \
-  free_aligned_buffer_page_end(dst_argb_opt);                                  \
-}
-
-#define TESTINTERPOLATE(TERP)                                                  \
-    TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_ - 1, TERP, _Any, +, 0)   \
-    TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Unaligned, +, 1) \
-    TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Invert, -, 0)    \
-    TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Opt, +, 0)
+TEST_F(LibYUVPlanarTest, TestInterpolatePlane_16) {
+  SIMD_ALIGNED(uint16_t orig_pixels_0[1280]);
+  SIMD_ALIGNED(uint16_t orig_pixels_1[1280]);
+  SIMD_ALIGNED(uint16_t interpolate_pixels[1280]);
+  memset(orig_pixels_0, 0, sizeof(orig_pixels_0));
+  memset(orig_pixels_1, 0, sizeof(orig_pixels_1));
+
+  orig_pixels_0[0] = 16u;
+  orig_pixels_0[1] = 32u;
+  orig_pixels_0[2] = 64u;
+  orig_pixels_0[3] = 128u;
+  orig_pixels_0[4] = 0u;
+  orig_pixels_0[5] = 0u;
+  orig_pixels_0[6] = 0u;
+  orig_pixels_0[7] = 255u;
+  orig_pixels_0[8] = 0u;
+  orig_pixels_0[9] = 0u;
+  orig_pixels_0[10] = 0u;
+  orig_pixels_0[11] = 0u;
+  orig_pixels_0[12] = 0u;
+  orig_pixels_0[13] = 0u;
+  orig_pixels_0[14] = 0u;
+  orig_pixels_0[15] = 0u;
+
+  orig_pixels_1[0] = 0u;
+  orig_pixels_1[1] = 0u;
+  orig_pixels_1[2] = 0u;
+  orig_pixels_1[3] = 0u;
+  orig_pixels_1[4] = 0u;
+  orig_pixels_1[5] = 0u;
+  orig_pixels_1[6] = 0u;
+  orig_pixels_1[7] = 0u;
+  orig_pixels_1[8] = 0u;
+  orig_pixels_1[9] = 0u;
+  orig_pixels_1[10] = 0u;
+  orig_pixels_1[11] = 0u;
+  orig_pixels_1[12] = 255u;
+  orig_pixels_1[13] = 255u;
+  orig_pixels_1[14] = 255u;
+  orig_pixels_1[15] = 255u;
+
+  InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                      &interpolate_pixels[0], 0, 16, 1, 128);
+  EXPECT_EQ(8u, interpolate_pixels[0]);
+  EXPECT_EQ(16u, interpolate_pixels[1]);
+  EXPECT_EQ(32u, interpolate_pixels[2]);
+  EXPECT_EQ(64u, interpolate_pixels[3]);
+  EXPECT_EQ(0u, interpolate_pixels[4]);
+  EXPECT_EQ(0u, interpolate_pixels[5]);
+  EXPECT_EQ(0u, interpolate_pixels[6]);
+  EXPECT_EQ(128u, interpolate_pixels[7]);
+  EXPECT_EQ(0u, interpolate_pixels[8]);
+  EXPECT_EQ(0u, interpolate_pixels[9]);
+  EXPECT_EQ(0u, interpolate_pixels[10]);
+  EXPECT_EQ(0u, interpolate_pixels[11]);
+  EXPECT_EQ(128u, interpolate_pixels[12]);
+  EXPECT_EQ(128u, interpolate_pixels[13]);
+  EXPECT_EQ(128u, interpolate_pixels[14]);
+  EXPECT_EQ(128u, interpolate_pixels[15]);
+
+  InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                      &interpolate_pixels[0], 0, 16, 1, 0);
+  EXPECT_EQ(16u, interpolate_pixels[0]);
+  EXPECT_EQ(32u, interpolate_pixels[1]);
+  EXPECT_EQ(64u, interpolate_pixels[2]);
+  EXPECT_EQ(128u, interpolate_pixels[3]);
+
+  InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                      &interpolate_pixels[0], 0, 16, 1, 192);
+
+  EXPECT_EQ(4u, interpolate_pixels[0]);
+  EXPECT_EQ(8u, interpolate_pixels[1]);
+  EXPECT_EQ(16u, interpolate_pixels[2]);
+  EXPECT_EQ(32u, interpolate_pixels[3]);
+
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+    InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                        &interpolate_pixels[0], 0, 1280, 1, 123);
+  }
+}
+
+#define TESTTERP(FMT_A, BPP_A, STRIDE_A, FMT_B, BPP_B, STRIDE_B, W1280, TERP, \
+                 N, NEG, OFF)                                                 \
+  TEST_F(LibYUVPlanarTest, ARGBInterpolate##TERP##N) {                        \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kStrideA =                                                      \
+        (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                \
+    const int kStrideB =                                                      \
+        (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;                \
+    align_buffer_page_end(src_argb_a, kStrideA* kHeight + OFF);               \
+    align_buffer_page_end(src_argb_b, kStrideA* kHeight + OFF);               \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight);                     \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight);                   \
+    for (int i = 0; i < kStrideA * kHeight; ++i) {                            \
+      src_argb_a[i + OFF] = (fastrand() & 0xff);                              \
+      src_argb_b[i + OFF] = (fastrand() & 0xff);                              \
+    }                                                                         \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    ARGBInterpolate(src_argb_a + OFF, kStrideA, src_argb_b + OFF, kStrideA,   \
+                    dst_argb_c, kStrideB, kWidth, NEG kHeight, TERP);         \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      ARGBInterpolate(src_argb_a + OFF, kStrideA, src_argb_b + OFF, kStrideA, \
+                      dst_argb_opt, kStrideB, kWidth, NEG kHeight, TERP);     \
+    }                                                                         \
+    for (int i = 0; i < kStrideB * kHeight; ++i) {                            \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                              \
+    }                                                                         \
+    free_aligned_buffer_page_end(src_argb_a);                                 \
+    free_aligned_buffer_page_end(src_argb_b);                                 \
+    free_aligned_buffer_page_end(dst_argb_c);                                 \
+    free_aligned_buffer_page_end(dst_argb_opt);                               \
+  }
+
+#define TESTINTERPOLATE(TERP)                                                \
+  TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_ + 1, TERP, _Any, +, 0)   \
+  TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Unaligned, +, 1) \
+  TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Invert, -, 0)    \
+  TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Opt, +, 0)
 
 TESTINTERPOLATE(0)
 TESTINTERPOLATE(64)
@@ -1069,9 +1207,14 @@ TESTINTERPOLATE(128)
 TESTINTERPOLATE(192)
 TESTINTERPOLATE(255)
 
-static int TestBlend(int width, int height, int benchmark_iterations,
-                     int disable_cpu_flags, int benchmark_cpu_info,
-                     int invert, int off) {
+static int TestBlend(int width,
+                     int height,
+                     int benchmark_iterations,
+                     int disable_cpu_flags,
+                     int benchmark_cpu_info,
+                     int invert,
+                     int off,
+                     int attenuate) {
   if (width < 1) {
     width = 1;
   }
@@ -1085,30 +1228,27 @@ static int TestBlend(int width, int height, int benchmark_iterations,
     src_argb_a[i + off] = (fastrand() & 0xff);
     src_argb_b[i + off] = (fastrand() & 0xff);
   }
-  ARGBAttenuate(src_argb_a + off, kStride, src_argb_a + off, kStride, width,
-                height);
-  ARGBAttenuate(src_argb_b + off, kStride, src_argb_b + off, kStride, width,
-                height);
+  MemRandomize(src_argb_a, kStride * height + off);
+  MemRandomize(src_argb_b, kStride * height + off);
+  if (attenuate) {
+    ARGBAttenuate(src_argb_a + off, kStride, src_argb_a + off, kStride, width,
+                  height);
+  }
   memset(dst_argb_c, 255, kStride * height);
   memset(dst_argb_opt, 255, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBBlend(src_argb_a + off, kStride,
-            src_argb_b + off, kStride,
-            dst_argb_c, kStride,
-            width, invert * height);
+  ARGBBlend(src_argb_a + off, kStride, src_argb_b + off, kStride, dst_argb_c,
+            kStride, width, invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBBlend(src_argb_a + off, kStride,
-              src_argb_b + off, kStride,
-              dst_argb_opt, kStride,
-              width, invert * height);
+    ARGBBlend(src_argb_a + off, kStride, src_argb_b + off, kStride,
+              dst_argb_opt, kStride, width, invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1121,36 +1261,47 @@ static int TestBlend(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Any) {
-  int max_diff = TestBlend(benchmark_width_ - 4, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_,  benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestBlend(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Unaligned) {
-  int max_diff = TestBlend(benchmark_width_, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  int max_diff =
+      TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 1, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Invert) {
-  int max_diff = TestBlend(benchmark_width_, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  int max_diff =
+      TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, -1, 0, 1);
+  EXPECT_LE(max_diff, 1);
+}
+
+TEST_F(LibYUVPlanarTest, ARGBBlend_Unattenuated) {
+  int max_diff =
+      TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Opt) {
-  int max_diff = TestBlend(benchmark_width_, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 1);
   EXPECT_LE(max_diff, 1);
 }
 
-static void TestBlendPlane(int width, int height, int benchmark_iterations,
-                           int disable_cpu_flags, int benchmark_cpu_info,
-                           int invert, int off) {
+static void TestBlendPlane(int width,
+                           int height,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info,
+                           int invert,
+                           int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1170,21 +1321,15 @@ static void TestBlendPlane(int width, int height, int benchmark_iterations,
     src_argb_b[i + off] = 255 - (i & 255);
   }
   memset(src_argb_alpha + off, 255, width);
-  BlendPlane(src_argb_a + off, width,
-             src_argb_b + off, width,
-             src_argb_alpha + off, width,
-             dst_argb_opt + off, width,
-             width, 1);
+  BlendPlane(src_argb_a + off, width, src_argb_b + off, width,
+             src_argb_alpha + off, width, dst_argb_opt + off, width, width, 1);
   for (int i = 0; i < width; ++i) {
     EXPECT_EQ(src_argb_a[i + off], dst_argb_opt[i + off]);
   }
   // Test destination is maintained exactly if alpha is 0.
   memset(src_argb_alpha + off, 0, width);
-  BlendPlane(src_argb_a + off, width,
-             src_argb_b + off, width,
-             src_argb_alpha + off, width,
-             dst_argb_opt + off, width,
-             width, 1);
+  BlendPlane(src_argb_a + off, width, src_argb_b + off, width,
+             src_argb_alpha + off, width, dst_argb_opt + off, width, width, 1);
   for (int i = 0; i < width; ++i) {
     EXPECT_EQ(src_argb_b[i + off], dst_argb_opt[i + off]);
   }
@@ -1195,18 +1340,14 @@ static void TestBlendPlane(int width, int height, int benchmark_iterations,
   }
 
   MaskCpuFlags(disable_cpu_flags);
-  BlendPlane(src_argb_a + off, width,
-             src_argb_b + off, width,
-             src_argb_alpha + off, width,
-             dst_argb_c + off, width,
-             width, height);
+  BlendPlane(src_argb_a + off, width, src_argb_b + off, width,
+             src_argb_alpha + off, width, dst_argb_c + off, width, width,
+             invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    BlendPlane(src_argb_a + off, width,
-               src_argb_b + off, width,
-               src_argb_alpha + off, width,
-               dst_argb_opt + off, width,
-               width, height);
+    BlendPlane(src_argb_a + off, width, src_argb_b + off, width,
+               src_argb_alpha + off, width, dst_argb_opt + off, width, width,
+               invert * height);
   }
   for (int i = 0; i < kStride * height; ++i) {
     EXPECT_EQ(dst_argb_c[i + off], dst_argb_opt[i + off]);
@@ -1216,7 +1357,6 @@ static void TestBlendPlane(int width, int height, int benchmark_iterations,
   free_aligned_buffer_page_end(src_argb_alpha);
   free_aligned_buffer_page_end(dst_argb_c);
   free_aligned_buffer_page_end(dst_argb_opt);
-  return;
 }
 
 TEST_F(LibYUVPlanarTest, BlendPlane_Opt) {
@@ -1228,7 +1368,7 @@ TEST_F(LibYUVPlanarTest, BlendPlane_Unaligned) {
                  disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
 }
 TEST_F(LibYUVPlanarTest, BlendPlane_Any) {
-  TestBlendPlane(benchmark_width_ - 4, benchmark_height_, benchmark_iterations_,
+  TestBlendPlane(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
 }
 TEST_F(LibYUVPlanarTest, BlendPlane_Invert) {
@@ -1236,11 +1376,15 @@ TEST_F(LibYUVPlanarTest, BlendPlane_Invert) {
                  disable_cpu_flags_, benchmark_cpu_info_, -1, 1);
 }
 
-#define SUBSAMPLE(v, a) ((((v) + (a) - 1)) / (a))
+#define SUBSAMPLE(v, a) ((((v) + (a)-1)) / (a))
 
-static void TestI420Blend(int width, int height, int benchmark_iterations,
-                          int disable_cpu_flags, int benchmark_cpu_info,
-                          int invert, int off) {
+static void TestI420Blend(int width,
+                          int height,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info,
+                          int invert,
+                          int off) {
   width = ((width) > 0) ? (width) : 1;
   const int kStrideUV = SUBSAMPLE(width, 2);
   const int kSizeUV = kStrideUV * SUBSAMPLE(height, 2);
@@ -1273,30 +1417,18 @@ static void TestI420Blend(int width, int height, int benchmark_iterations,
   memset(dst_v_opt, 255, kSizeUV + off);
 
   MaskCpuFlags(disable_cpu_flags);
-  I420Blend(src_y0 + off, width,
-            src_u0 + off, kStrideUV,
-            src_v0 + off, kStrideUV,
-            src_y1 + off, width,
-            src_u1 + off, kStrideUV,
-            src_v1 + off, kStrideUV,
-            src_a + off, width,
-            dst_y_c + off, width,
-            dst_u_c + off, kStrideUV,
-            dst_v_c + off, kStrideUV,
-            width, height);
+  I420Blend(src_y0 + off, width, src_u0 + off, kStrideUV, src_v0 + off,
+            kStrideUV, src_y1 + off, width, src_u1 + off, kStrideUV,
+            src_v1 + off, kStrideUV, src_a + off, width, dst_y_c + off, width,
+            dst_u_c + off, kStrideUV, dst_v_c + off, kStrideUV, width,
+            invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    I420Blend(src_y0 + off, width,
-              src_u0 + off, kStrideUV,
-              src_v0 + off, kStrideUV,
-              src_y1 + off, width,
-              src_u1 + off, kStrideUV,
-              src_v1 + off, kStrideUV,
-              src_a + off, width,
-              dst_y_opt + off, width,
-              dst_u_opt + off, kStrideUV,
-              dst_v_opt + off, kStrideUV,
-              width, height);
+    I420Blend(src_y0 + off, width, src_u0 + off, kStrideUV, src_v0 + off,
+              kStrideUV, src_y1 + off, width, src_u1 + off, kStrideUV,
+              src_v1 + off, kStrideUV, src_a + off, width, dst_y_opt + off,
+              width, dst_u_opt + off, kStrideUV, dst_v_opt + off, kStrideUV,
+              width, invert * height);
   }
   for (int i = 0; i < width * height; ++i) {
     EXPECT_EQ(dst_y_c[i + off], dst_y_opt[i + off]);
@@ -1318,31 +1450,30 @@ static void TestI420Blend(int width, int height, int benchmark_iterations,
   free_aligned_buffer_page_end(dst_y_opt);
   free_aligned_buffer_page_end(dst_u_opt);
   free_aligned_buffer_page_end(dst_v_opt);
-  return;
 }
 
 TEST_F(LibYUVPlanarTest, I420Blend_Opt) {
   TestI420Blend(benchmark_width_, benchmark_height_, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
 }
 TEST_F(LibYUVPlanarTest, I420Blend_Unaligned) {
   TestI420Blend(benchmark_width_, benchmark_height_, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
 }
 
 // TODO(fbarchard): DISABLED because _Any uses C.  Avoid C and re-enable.
 TEST_F(LibYUVPlanarTest, DISABLED_I420Blend_Any) {
-  TestI420Blend(benchmark_width_ - 4, benchmark_height_, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  TestI420Blend(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
 }
 TEST_F(LibYUVPlanarTest, I420Blend_Invert) {
   TestI420Blend(benchmark_width_, benchmark_height_, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+                disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
 }
 
 TEST_F(LibYUVPlanarTest, TestAffine) {
-  SIMD_ALIGNED(uint8 orig_pixels_0[1280][4]);
-  SIMD_ALIGNED(uint8 interpolate_pixels_C[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels_0[1280][4]);
+  SIMD_ALIGNED(uint8_t interpolate_pixels_C[1280][4]);
 
   for (int i = 0; i < 1280; ++i) {
     for (int j = 0; j < 4; ++j) {
@@ -1350,16 +1481,16 @@ TEST_F(LibYUVPlanarTest, TestAffine) {
     }
   }
 
-  float uv_step[4] = { 0.f, 0.f, 0.75f, 0.f };
+  float uv_step[4] = {0.f, 0.f, 0.75f, 0.f};
 
-  ARGBAffineRow_C(&orig_pixels_0[0][0], 0, &interpolate_pixels_C[0][0],
-                  uv_step, 1280);
+  ARGBAffineRow_C(&orig_pixels_0[0][0], 0, &interpolate_pixels_C[0][0], uv_step,
+                  1280);
   EXPECT_EQ(0u, interpolate_pixels_C[0][0]);
   EXPECT_EQ(96u, interpolate_pixels_C[128][0]);
   EXPECT_EQ(191u, interpolate_pixels_C[255][3]);
 
 #if defined(HAS_ARGBAFFINEROW_SSE2)
-  SIMD_ALIGNED(uint8 interpolate_pixels_Opt[1280][4]);
+  SIMD_ALIGNED(uint8_t interpolate_pixels_Opt[1280][4]);
   ARGBAffineRow_SSE2(&orig_pixels_0[0][0], 0, &interpolate_pixels_Opt[0][0],
                      uv_step, 1280);
   EXPECT_EQ(0, memcmp(interpolate_pixels_Opt, interpolate_pixels_C, 1280 * 4));
@@ -1399,7 +1530,7 @@ TEST_F(LibYUVPlanarTest, TestCopyPlane) {
 
   // Fill destination buffers with random data.
   for (i = 0; i < y_plane_size; ++i) {
-    uint8 random_number = fastrand() & 0x7f;
+    uint8_t random_number = fastrand() & 0x7f;
     dst_c[i] = random_number;
     dst_opt[i] = dst_c[i];
   }
@@ -1411,23 +1542,20 @@ TEST_F(LibYUVPlanarTest, TestCopyPlane) {
 
   // Disable all optimizations.
   MaskCpuFlags(disable_cpu_flags_);
-  double c_time = get_time();
   for (j = 0; j < benchmark_iterations_; j++) {
     CopyPlane(orig_y + y_off, y_st, dst_c + y_off, stride, yw, yh);
   }
-  c_time = (get_time() - c_time) / benchmark_iterations_;
 
   // Enable optimizations.
   MaskCpuFlags(benchmark_cpu_info_);
-  double opt_time = get_time();
   for (j = 0; j < benchmark_iterations_; j++) {
     CopyPlane(orig_y + y_off, y_st, dst_opt + y_off, stride, yw, yh);
   }
-  opt_time = (get_time() - opt_time) / benchmark_iterations_;
 
   for (i = 0; i < y_plane_size; ++i) {
-    if (dst_c[i] != dst_opt[i])
+    if (dst_c[i] != dst_opt[i]) {
       ++err;
+    }
   }
 
   free_aligned_buffer_page_end(orig_y);
@@ -1437,9 +1565,258 @@ TEST_F(LibYUVPlanarTest, TestCopyPlane) {
   EXPECT_EQ(0, err);
 }
 
-static int TestMultiply(int width, int height, int benchmark_iterations,
-                        int disable_cpu_flags, int benchmark_cpu_info,
-                        int invert, int off) {
+TEST_F(LibYUVPlanarTest, CopyPlane_Opt) {
+  int i;
+  int y_plane_size = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_y, y_plane_size);
+  align_buffer_page_end(dst_c, y_plane_size);
+  align_buffer_page_end(dst_opt, y_plane_size);
+
+  MemRandomize(orig_y, y_plane_size);
+  memset(dst_c, 1, y_plane_size);
+  memset(dst_opt, 2, y_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  for (i = 0; i < benchmark_iterations_; i++) {
+    CopyPlane(orig_y, benchmark_width_, dst_c, benchmark_width_,
+              benchmark_width_, benchmark_height_);
+  }
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (i = 0; i < benchmark_iterations_; i++) {
+    CopyPlane(orig_y, benchmark_width_, dst_opt, benchmark_width_,
+              benchmark_width_, benchmark_height_);
+  }
+
+  for (i = 0; i < y_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  free_aligned_buffer_page_end(dst_c);
+  free_aligned_buffer_page_end(dst_opt);
+}
+
+TEST_F(LibYUVPlanarTest, TestCopyPlaneZero) {
+  // Test to verify copying a rect with a zero height or width does
+  // not touch destination memory.
+  uint8_t src = 42;
+  uint8_t dst = 0;
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  CopyPlane(&src, 0, &dst, 0, 0, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 1, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 0, 1);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  CopyPlane(&src, 0, &dst, 0, 0, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 1, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 0, 1);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+}
+
+TEST_F(LibYUVPlanarTest, TestDetilePlane) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int tile_width = (benchmark_width_ + 15) & ~15;
+  int tile_height = (benchmark_height_ + 15) & ~15;
+  int tile_plane_size = tile_width * tile_height;
+  int y_plane_size = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(tile_y, tile_plane_size);
+  align_buffer_page_end(dst_c, y_plane_size);
+  align_buffer_page_end(dst_opt, y_plane_size);
+
+  MemRandomize(tile_y, tile_plane_size);
+  memset(dst_c, 0, y_plane_size);
+  memset(dst_opt, 0, y_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane(tile_y, tile_width, dst_c, benchmark_width_, benchmark_width_,
+                benchmark_height_, 16);
+  }
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane(tile_y, tile_width, dst_opt, benchmark_width_, benchmark_width_,
+                benchmark_height_, 16);
+  }
+
+  for (i = 0; i < y_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(tile_y);
+  free_aligned_buffer_page_end(dst_c);
+  free_aligned_buffer_page_end(dst_opt);
+}
+
+TEST_F(LibYUVPlanarTest, TestDetilePlane_16) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int tile_width = (benchmark_width_ + 15) & ~15;
+  int tile_height = (benchmark_height_ + 15) & ~15;
+  int tile_plane_size = tile_width * tile_height * 2;
+  int y_plane_size = benchmark_width_ * benchmark_height_ * 2;
+  align_buffer_page_end(tile_y, tile_plane_size);
+  align_buffer_page_end(dst_c, y_plane_size);
+  align_buffer_page_end(dst_opt, y_plane_size);
+
+  MemRandomize(tile_y, tile_plane_size);
+  memset(dst_c, 0, y_plane_size);
+  memset(dst_opt, 0, y_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane_16((const uint16_t*)tile_y, tile_width, (uint16_t*)dst_c,
+                   benchmark_width_, benchmark_width_, benchmark_height_, 16);
+  }
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane_16((const uint16_t*)tile_y, tile_width, (uint16_t*)dst_opt,
+                   benchmark_width_, benchmark_width_, benchmark_height_, 16);
+  }
+
+  for (i = 0; i < y_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(tile_y);
+  free_aligned_buffer_page_end(dst_c);
+  free_aligned_buffer_page_end(dst_opt);
+}
+
+// Compares DetileSplitUV to 2 step Detile + SplitUV
+TEST_F(LibYUVPlanarTest, TestDetileSplitUVPlane_Correctness) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int tile_width = (benchmark_width_ + 15) & ~15;
+  int tile_height = (benchmark_height_ + 15) & ~15;
+  int tile_plane_size = tile_width * tile_height;
+  int uv_plane_size = ((benchmark_width_ + 1) / 2) * benchmark_height_;
+  align_buffer_page_end(tile_uv, tile_plane_size);
+  align_buffer_page_end(detiled_uv, tile_plane_size);
+  align_buffer_page_end(dst_u_two_stage, uv_plane_size);
+  align_buffer_page_end(dst_u_opt, uv_plane_size);
+  align_buffer_page_end(dst_v_two_stage, uv_plane_size);
+  align_buffer_page_end(dst_v_opt, uv_plane_size);
+
+  MemRandomize(tile_uv, tile_plane_size);
+  memset(detiled_uv, 0, tile_plane_size);
+  memset(dst_u_two_stage, 0, uv_plane_size);
+  memset(dst_u_opt, 0, uv_plane_size);
+  memset(dst_v_two_stage, 0, uv_plane_size);
+  memset(dst_v_opt, 0, uv_plane_size);
+
+  DetileSplitUVPlane(tile_uv, tile_width, dst_u_opt, (benchmark_width_ + 1) / 2,
+                     dst_v_opt, (benchmark_width_ + 1) / 2, benchmark_width_,
+                     benchmark_height_, 16);
+
+  // Benchmark 2 step conversion for comparison.
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane(tile_uv, tile_width, detiled_uv, benchmark_width_,
+                benchmark_width_, benchmark_height_, 16);
+    SplitUVPlane(detiled_uv, tile_width, dst_u_two_stage,
+                 (benchmark_width_ + 1) / 2, dst_v_two_stage,
+                 (benchmark_width_ + 1) / 2, (benchmark_width_ + 1) / 2,
+                 benchmark_height_);
+  }
+
+  for (i = 0; i < uv_plane_size; ++i) {
+    EXPECT_EQ(dst_u_two_stage[i], dst_u_opt[i]);
+    EXPECT_EQ(dst_v_two_stage[i], dst_v_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(tile_uv);
+  free_aligned_buffer_page_end(detiled_uv);
+  free_aligned_buffer_page_end(dst_u_two_stage);
+  free_aligned_buffer_page_end(dst_u_opt);
+  free_aligned_buffer_page_end(dst_v_two_stage);
+  free_aligned_buffer_page_end(dst_v_opt);
+}
+
+TEST_F(LibYUVPlanarTest, TestDetileSplitUVPlane_Benchmark) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int tile_width = (benchmark_width_ + 15) & ~15;
+  int tile_height = (benchmark_height_ + 15) & ~15;
+  int tile_plane_size = tile_width * tile_height;
+  int uv_plane_size = ((benchmark_width_ + 1) / 2) * benchmark_height_;
+  align_buffer_page_end(tile_uv, tile_plane_size);
+  align_buffer_page_end(dst_u_c, uv_plane_size);
+  align_buffer_page_end(dst_u_opt, uv_plane_size);
+  align_buffer_page_end(dst_v_c, uv_plane_size);
+  align_buffer_page_end(dst_v_opt, uv_plane_size);
+
+  MemRandomize(tile_uv, tile_plane_size);
+  memset(dst_u_c, 0, uv_plane_size);
+  memset(dst_u_opt, 0, uv_plane_size);
+  memset(dst_v_c, 0, uv_plane_size);
+  memset(dst_v_opt, 0, uv_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+
+  DetileSplitUVPlane(tile_uv, tile_width, dst_u_c, (benchmark_width_ + 1) / 2,
+                     dst_v_c, (benchmark_width_ + 1) / 2, benchmark_width_,
+                     benchmark_height_, 16);
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetileSplitUVPlane(
+        tile_uv, tile_width, dst_u_opt, (benchmark_width_ + 1) / 2, dst_v_opt,
+        (benchmark_width_ + 1) / 2, benchmark_width_, benchmark_height_, 16);
+  }
+
+  for (i = 0; i < uv_plane_size; ++i) {
+    EXPECT_EQ(dst_u_c[i], dst_u_opt[i]);
+    EXPECT_EQ(dst_v_c[i], dst_v_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(tile_uv);
+  free_aligned_buffer_page_end(dst_u_c);
+  free_aligned_buffer_page_end(dst_u_opt);
+  free_aligned_buffer_page_end(dst_v_c);
+  free_aligned_buffer_page_end(dst_v_opt);
+}
+
+static int TestMultiply(int width,
+                        int height,
+                        int benchmark_iterations,
+                        int disable_cpu_flags,
+                        int benchmark_cpu_info,
+                        int invert,
+                        int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1457,22 +1834,17 @@ static int TestMultiply(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBMultiply(src_argb_a + off, kStride,
-               src_argb_b + off, kStride,
-               dst_argb_c, kStride,
-               width, invert * height);
+  ARGBMultiply(src_argb_a + off, kStride, src_argb_b + off, kStride, dst_argb_c,
+               kStride, width, invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBMultiply(src_argb_a + off, kStride,
-                 src_argb_b + off, kStride,
-                 dst_argb_opt, kStride,
-                 width, invert * height);
+    ARGBMultiply(src_argb_a + off, kStride, src_argb_b + off, kStride,
+                 dst_argb_opt, kStride, width, invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1485,36 +1857,40 @@ static int TestMultiply(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBMultiply_Any) {
-  int max_diff = TestMultiply(benchmark_width_ - 1, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff = TestMultiply(benchmark_width_ + 1, benchmark_height_,
+                              benchmark_iterations_, disable_cpu_flags_,
+                              benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBMultiply_Unaligned) {
-  int max_diff = TestMultiply(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  int max_diff =
+      TestMultiply(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBMultiply_Invert) {
-  int max_diff = TestMultiply(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  int max_diff =
+      TestMultiply(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBMultiply_Opt) {
-  int max_diff = TestMultiply(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestMultiply(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
-static int TestAdd(int width, int height, int benchmark_iterations,
-                   int disable_cpu_flags,  int benchmark_cpu_info,
-                   int invert, int off) {
+static int TestAdd(int width,
+                   int height,
+                   int benchmark_iterations,
+                   int disable_cpu_flags,
+                   int benchmark_cpu_info,
+                   int invert,
+                   int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1532,22 +1908,17 @@ static int TestAdd(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBAdd(src_argb_a + off, kStride,
-          src_argb_b + off, kStride,
-          dst_argb_c, kStride,
-          width, invert * height);
+  ARGBAdd(src_argb_a + off, kStride, src_argb_b + off, kStride, dst_argb_c,
+          kStride, width, invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBAdd(src_argb_a + off, kStride,
-            src_argb_b + off, kStride,
-            dst_argb_opt, kStride,
-            width, invert * height);
+    ARGBAdd(src_argb_a + off, kStride, src_argb_b + off, kStride, dst_argb_opt,
+            kStride, width, invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1560,36 +1931,40 @@ static int TestAdd(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAdd_Any) {
-  int max_diff = TestAdd(benchmark_width_ - 1, benchmark_height_,
-                         benchmark_iterations_,
-                         disable_cpu_flags_,  benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestAdd(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+              disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAdd_Unaligned) {
-  int max_diff = TestAdd(benchmark_width_, benchmark_height_,
-                         benchmark_iterations_,
-                         disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  int max_diff =
+      TestAdd(benchmark_width_, benchmark_height_, benchmark_iterations_,
+              disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAdd_Invert) {
-  int max_diff = TestAdd(benchmark_width_, benchmark_height_,
-                         benchmark_iterations_,
-                         disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  int max_diff =
+      TestAdd(benchmark_width_, benchmark_height_, benchmark_iterations_,
+              disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAdd_Opt) {
-  int max_diff = TestAdd(benchmark_width_, benchmark_height_,
-                         benchmark_iterations_,
-                         disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestAdd(benchmark_width_, benchmark_height_, benchmark_iterations_,
+              disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
-static int TestSubtract(int width, int height, int benchmark_iterations,
-                        int disable_cpu_flags, int benchmark_cpu_info,
-                        int invert, int off) {
+static int TestSubtract(int width,
+                        int height,
+                        int benchmark_iterations,
+                        int disable_cpu_flags,
+                        int benchmark_cpu_info,
+                        int invert,
+                        int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1607,22 +1982,17 @@ static int TestSubtract(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBSubtract(src_argb_a + off, kStride,
-               src_argb_b + off, kStride,
-               dst_argb_c, kStride,
-               width, invert * height);
+  ARGBSubtract(src_argb_a + off, kStride, src_argb_b + off, kStride, dst_argb_c,
+               kStride, width, invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBSubtract(src_argb_a + off, kStride,
-                 src_argb_b + off, kStride,
-                 dst_argb_opt, kStride,
-                 width, invert * height);
+    ARGBSubtract(src_argb_a + off, kStride, src_argb_b + off, kStride,
+                 dst_argb_opt, kStride, width, invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1635,36 +2005,40 @@ static int TestSubtract(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSubtract_Any) {
-  int max_diff = TestSubtract(benchmark_width_ - 1, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff = TestSubtract(benchmark_width_ + 1, benchmark_height_,
+                              benchmark_iterations_, disable_cpu_flags_,
+                              benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSubtract_Unaligned) {
-  int max_diff = TestSubtract(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  int max_diff =
+      TestSubtract(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSubtract_Invert) {
-  int max_diff = TestSubtract(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  int max_diff =
+      TestSubtract(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSubtract_Opt) {
-  int max_diff = TestSubtract(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestSubtract(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
 
-static int TestSobel(int width, int height, int benchmark_iterations,
-                     int disable_cpu_flags, int benchmark_cpu_info,
-                     int invert, int off) {
+static int TestSobel(int width,
+                     int height,
+                     int benchmark_iterations,
+                     int disable_cpu_flags,
+                     int benchmark_cpu_info,
+                     int invert,
+                     int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1681,20 +2055,17 @@ static int TestSobel(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBSobel(src_argb_a + off, kStride,
-            dst_argb_c, kStride,
-            width, invert * height);
+  ARGBSobel(src_argb_a + off, kStride, dst_argb_c, kStride, width,
+            invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBSobel(src_argb_a + off, kStride,
-              dst_argb_opt, kStride,
-              width, invert * height);
+    ARGBSobel(src_argb_a + off, kStride, dst_argb_opt, kStride, width,
+              invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1706,36 +2077,40 @@ static int TestSobel(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobel_Any) {
-  int max_diff = TestSobel(benchmark_width_ - 1, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestSobel(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobel_Unaligned) {
-  int max_diff = TestSobel(benchmark_width_, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  int max_diff =
+      TestSobel(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobel_Invert) {
-  int max_diff = TestSobel(benchmark_width_, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  int max_diff =
+      TestSobel(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobel_Opt) {
-  int max_diff = TestSobel(benchmark_width_, benchmark_height_,
-                           benchmark_iterations_,
-                           disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestSobel(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
-static int TestSobelToPlane(int width, int height, int benchmark_iterations,
-                            int disable_cpu_flags, int benchmark_cpu_info,
-                            int invert, int off) {
+static int TestSobelToPlane(int width,
+                            int height,
+                            int benchmark_iterations,
+                            int disable_cpu_flags,
+                            int benchmark_cpu_info,
+                            int invert,
+                            int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1754,20 +2129,17 @@ static int TestSobelToPlane(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kDstStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBSobelToPlane(src_argb_a + off, kSrcStride,
-                   dst_argb_c, kDstStride,
-                   width, invert * height);
+  ARGBSobelToPlane(src_argb_a + off, kSrcStride, dst_argb_c, kDstStride, width,
+                   invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBSobelToPlane(src_argb_a + off, kSrcStride,
-                     dst_argb_opt, kDstStride,
+    ARGBSobelToPlane(src_argb_a + off, kSrcStride, dst_argb_opt, kDstStride,
                      width, invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kDstStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1779,40 +2151,40 @@ static int TestSobelToPlane(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelToPlane_Any) {
-  int max_diff = TestSobelToPlane(benchmark_width_ - 1, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  +1, 0);
+  int max_diff = TestSobelToPlane(benchmark_width_ + 1, benchmark_height_,
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelToPlane_Unaligned) {
   int max_diff = TestSobelToPlane(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  +1, 1);
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, +1, 1);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelToPlane_Invert) {
   int max_diff = TestSobelToPlane(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  -1, 0);
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, -1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelToPlane_Opt) {
   int max_diff = TestSobelToPlane(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_,
-                                  disable_cpu_flags_, benchmark_cpu_info_,
-                                  +1, 0);
+                                  benchmark_iterations_, disable_cpu_flags_,
+                                  benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
-static int TestSobelXY(int width, int height, int benchmark_iterations,
-                       int disable_cpu_flags, int benchmark_cpu_info,
-                       int invert, int off) {
+static int TestSobelXY(int width,
+                       int height,
+                       int benchmark_iterations,
+                       int disable_cpu_flags,
+                       int benchmark_cpu_info,
+                       int invert,
+                       int off) {
   if (width < 1) {
     width = 1;
   }
@@ -1829,20 +2201,17 @@ static int TestSobelXY(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBSobelXY(src_argb_a + off, kStride,
-            dst_argb_c, kStride,
-            width, invert * height);
+  ARGBSobelXY(src_argb_a + off, kStride, dst_argb_c, kStride, width,
+              invert * height);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBSobelXY(src_argb_a + off, kStride,
-              dst_argb_opt, kStride,
-              width, invert * height);
+    ARGBSobelXY(src_argb_a + off, kStride, dst_argb_opt, kStride, width,
+                invert * height);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1854,36 +2223,41 @@ static int TestSobelXY(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelXY_Any) {
-  int max_diff = TestSobelXY(benchmark_width_ - 1, benchmark_height_,
-                             benchmark_iterations_,
-                             disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff = TestSobelXY(benchmark_width_ + 1, benchmark_height_,
+                             benchmark_iterations_, disable_cpu_flags_,
+                             benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelXY_Unaligned) {
-  int max_diff = TestSobelXY(benchmark_width_, benchmark_height_,
-                             benchmark_iterations_,
-                             disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+  int max_diff =
+      TestSobelXY(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelXY_Invert) {
-  int max_diff = TestSobelXY(benchmark_width_, benchmark_height_,
-                             benchmark_iterations_,
-                             disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+  int max_diff =
+      TestSobelXY(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelXY_Opt) {
-  int max_diff = TestSobelXY(benchmark_width_, benchmark_height_,
-                             benchmark_iterations_,
-                             disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+  int max_diff =
+      TestSobelXY(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
 
-static int TestBlur(int width, int height, int benchmark_iterations,
-                    int disable_cpu_flags, int benchmark_cpu_info,
-                    int invert, int off, int radius) {
+static int TestBlur(int width,
+                    int height,
+                    int benchmark_iterations,
+                    int disable_cpu_flags,
+                    int benchmark_cpu_info,
+                    int invert,
+                    int off,
+                    int radius) {
   if (width < 1) {
     width = 1;
   }
@@ -1901,22 +2275,19 @@ static int TestBlur(int width, int height, int benchmark_iterations,
   memset(dst_argb_opt, 0, kStride * height);
 
   MaskCpuFlags(disable_cpu_flags);
-  ARGBBlur(src_argb_a + off, kStride,
-           dst_argb_c, kStride,
-           reinterpret_cast<int32*>(dst_cumsum), width * 4,
-           width, invert * height, radius);
+  ARGBBlur(src_argb_a + off, kStride, dst_argb_c, kStride,
+           reinterpret_cast<int32_t*>(dst_cumsum), width * 4, width,
+           invert * height, radius);
   MaskCpuFlags(benchmark_cpu_info);
   for (int i = 0; i < benchmark_iterations; ++i) {
-    ARGBBlur(src_argb_a + off, kStride,
-             dst_argb_opt, kStride,
-             reinterpret_cast<int32*>(dst_cumsum), width * 4,
-             width, invert * height, radius);
+    ARGBBlur(src_argb_a + off, kStride, dst_argb_opt, kStride,
+             reinterpret_cast<int32_t*>(dst_cumsum), width * 4, width,
+             invert * height, radius);
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i]) -
-            static_cast<int>(dst_argb_opt[i]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -
+                       static_cast<int>(dst_argb_opt[i]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -1928,83 +2299,81 @@ static int TestBlur(int width, int height, int benchmark_iterations,
   return max_diff;
 }
 
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+#define DISABLED_ARM(name) name
+#else
+#define DISABLED_ARM(name) DISABLED_##name
+#endif
+
 static const int kBlurSize = 55;
-TEST_F(LibYUVPlanarTest, ARGBBlur_Any) {
-  int max_diff = TestBlur(benchmark_width_ - 1, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          +1, 0, kBlurSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Any)) {
+  int max_diff =
+      TestBlur(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlur_Unaligned) {
-  int max_diff = TestBlur(benchmark_width_, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          +1, 1, kBlurSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Unaligned)) {
+  int max_diff =
+      TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, +1, 1, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlur_Invert) {
-  int max_diff = TestBlur(benchmark_width_, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          -1, 0, kBlurSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Invert)) {
+  int max_diff =
+      TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, -1, 0, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlur_Opt) {
-  int max_diff = TestBlur(benchmark_width_, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          +1, 0, kBlurSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Opt)) {
+  int max_diff =
+      TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
 static const int kBlurSmallSize = 5;
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Any) {
-  int max_diff = TestBlur(benchmark_width_ - 1, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          +1, 0, kBlurSmallSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Any)) {
+  int max_diff =
+      TestBlur(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Unaligned) {
-  int max_diff = TestBlur(benchmark_width_, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          +1, 1, kBlurSmallSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Unaligned)) {
+  int max_diff =
+      TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, +1, 1, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Invert) {
-  int max_diff = TestBlur(benchmark_width_, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          -1, 0, kBlurSmallSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Invert)) {
+  int max_diff =
+      TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, -1, 0, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Opt) {
-  int max_diff = TestBlur(benchmark_width_, benchmark_height_,
-                          benchmark_iterations_,
-                          disable_cpu_flags_, benchmark_cpu_info_,
-                          +1, 0, kBlurSmallSize);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Opt)) {
+  int max_diff =
+      TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
+               disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, TestARGBPolynomial) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels_opt[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels_c[1280][4]);
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(TestARGBPolynomial)) {
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t dst_pixels_opt[1280][4]);
+  SIMD_ALIGNED(uint8_t dst_pixels_c[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   SIMD_ALIGNED(static const float kWarmifyPolynomial[16]) = {
-    0.94230f,  -3.03300f,    -2.92500f,  0.f,  // C0
-    0.584500f,  1.112000f,    1.535000f, 1.f,  // C1 x
-    0.001313f, -0.002503f,   -0.004496f, 0.f,  // C2 x * x
-    0.0f,       0.000006965f, 0.000008781f, 0.f,  // C3 x * x * x
+      0.94230f,  -3.03300f,    -2.92500f,    0.f,  // C0
+      0.584500f, 1.112000f,    1.535000f,    1.f,  // C1 x
+      0.001313f, -0.002503f,   -0.004496f,   0.f,  // C2 x * x
+      0.0f,      0.000006965f, 0.000008781f, 0.f,  // C3 x * x * x
   };
 
   // Test blue
@@ -2081,10 +2450,190 @@ TEST_F(LibYUVPlanarTest, TestARGBPolynomial) {
   }
 }
 
+int TestHalfFloatPlane(int benchmark_width,
+                       int benchmark_height,
+                       int benchmark_iterations,
+                       int disable_cpu_flags,
+                       int benchmark_cpu_info,
+                       float scale,
+                       int mask) {
+  int i, j;
+  const int y_plane_size = benchmark_width * benchmark_height * 2;
+
+  align_buffer_page_end(orig_y, y_plane_size * 3);
+  uint8_t* dst_opt = orig_y + y_plane_size;
+  uint8_t* dst_c = orig_y + y_plane_size * 2;
+
+  MemRandomize(orig_y, y_plane_size);
+  memset(dst_c, 0, y_plane_size);
+  memset(dst_opt, 1, y_plane_size);
+
+  for (i = 0; i < y_plane_size / 2; ++i) {
+    reinterpret_cast<uint16_t*>(orig_y)[i] &= mask;
+  }
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags);
+  for (j = 0; j < benchmark_iterations; j++) {
+    HalfFloatPlane(reinterpret_cast<uint16_t*>(orig_y), benchmark_width * 2,
+                   reinterpret_cast<uint16_t*>(dst_c), benchmark_width * 2,
+                   scale, benchmark_width, benchmark_height);
+  }
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info);
+  for (j = 0; j < benchmark_iterations; j++) {
+    HalfFloatPlane(reinterpret_cast<uint16_t*>(orig_y), benchmark_width * 2,
+                   reinterpret_cast<uint16_t*>(dst_opt), benchmark_width * 2,
+                   scale, benchmark_width, benchmark_height);
+  }
+
+  int max_diff = 0;
+  for (i = 0; i < y_plane_size / 2; ++i) {
+    int abs_diff =
+        abs(static_cast<int>(reinterpret_cast<uint16_t*>(dst_c)[i]) -
+            static_cast<int>(reinterpret_cast<uint16_t*>(dst_opt)[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  return max_diff;
+}
+
+#if defined(__arm__)
+static void EnableFlushDenormalToZero(void) {
+  uint32_t cw;
+  __asm__ __volatile__(
+      "vmrs   %0, fpscr         \n"
+      "orr    %0, %0, #0x1000000        \n"
+      "vmsr   fpscr, %0         \n"
+      : "=r"(cw)::"memory");
+}
+#endif
+
+// 5 bit exponent with bias of 15 will underflow to a denormal if scale causes
+// exponent to be less than 0.  15 - log2(65536) = -1/  This shouldnt normally
+// happen since scale is 1/(1<<bits) where bits is 9, 10 or 12.
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_denormal) {
+// 32 bit arm rounding on denormal case is off by 1 compared to C.
+#if defined(__arm__)
+  EnableFlushDenormalToZero();
+#endif
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f / 65536.0f, 65535);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_One) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f, 65535);
+  EXPECT_LE(diff, 1);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_Opt) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f / 4096.0f, 65535);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_10bit_Opt) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f / 1024.0f, 1023);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_9bit_Opt) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f / 512.0f, 511);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Opt) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f / 4096.0f, 4095);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Offby1) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f / 4095.0f, 4095);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_One) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f, 2047);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_12bit_One) {
+  int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, disable_cpu_flags_,
+                                benchmark_cpu_info_, 1.0f, 4095);
+  EXPECT_LE(diff, 1);
+}
+
+float TestByteToFloat(int benchmark_width,
+                      int benchmark_height,
+                      int benchmark_iterations,
+                      int disable_cpu_flags,
+                      int benchmark_cpu_info,
+                      float scale) {
+  int i, j;
+  const int y_plane_size = benchmark_width * benchmark_height;
+
+  align_buffer_page_end(orig_y, y_plane_size * (1 + 4 + 4));
+  float* dst_opt = reinterpret_cast<float*>(orig_y + y_plane_size);
+  float* dst_c = reinterpret_cast<float*>(orig_y + y_plane_size * 5);
+
+  MemRandomize(orig_y, y_plane_size);
+  memset(dst_c, 0, y_plane_size * 4);
+  memset(dst_opt, 1, y_plane_size * 4);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags);
+  ByteToFloat(orig_y, dst_c, scale, y_plane_size);
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info);
+  for (j = 0; j < benchmark_iterations; j++) {
+    ByteToFloat(orig_y, dst_opt, scale, y_plane_size);
+  }
+
+  float max_diff = 0;
+  for (i = 0; i < y_plane_size; ++i) {
+    float abs_diff = fabs(dst_c[i] - dst_opt[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  return max_diff;
+}
+
+TEST_F(LibYUVPlanarTest, TestByteToFloat) {
+  float diff = TestByteToFloat(benchmark_width_, benchmark_height_,
+                               benchmark_iterations_, disable_cpu_flags_,
+                               benchmark_cpu_info_, 1.0f);
+  EXPECT_EQ(0.f, diff);
+}
+
 TEST_F(LibYUVPlanarTest, TestARGBLumaColorTable) {
-  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels_opt[1280][4]);
-  SIMD_ALIGNED(uint8 dst_pixels_c[1280][4]);
+  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8_t dst_pixels_opt[1280][4]);
+  SIMD_ALIGNED(uint8_t dst_pixels_c[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
 
   align_buffer_page_end(lumacolortable, 32768);
@@ -2170,15 +2719,13 @@ TEST_F(LibYUVPlanarTest, TestARGBCopyAlpha) {
   memcpy(dst_pixels_c, dst_pixels_opt, kSize);
 
   MaskCpuFlags(disable_cpu_flags_);
-  ARGBCopyAlpha(orig_pixels, benchmark_width_ * 4,
-                dst_pixels_c, benchmark_width_ * 4,
-                benchmark_width_, benchmark_height_);
+  ARGBCopyAlpha(orig_pixels, benchmark_width_ * 4, dst_pixels_c,
+                benchmark_width_ * 4, benchmark_width_, benchmark_height_);
   MaskCpuFlags(benchmark_cpu_info_);
 
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    ARGBCopyAlpha(orig_pixels, benchmark_width_ * 4,
-                  dst_pixels_opt, benchmark_width_ * 4,
-                  benchmark_width_, benchmark_height_);
+    ARGBCopyAlpha(orig_pixels, benchmark_width_ * 4, dst_pixels_opt,
+                  benchmark_width_ * 4, benchmark_width_, benchmark_height_);
   }
   for (int i = 0; i < kSize; ++i) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
@@ -2200,15 +2747,13 @@ TEST_F(LibYUVPlanarTest, TestARGBExtractAlpha) {
   memcpy(dst_pixels_c, dst_pixels_opt, kPixels);
 
   MaskCpuFlags(disable_cpu_flags_);
-  ARGBExtractAlpha(src_pixels, benchmark_width_ * 4,
-                   dst_pixels_c, benchmark_width_,
-                   benchmark_width_, benchmark_height_);
+  ARGBExtractAlpha(src_pixels, benchmark_width_ * 4, dst_pixels_c,
+                   benchmark_width_, benchmark_width_, benchmark_height_);
   MaskCpuFlags(benchmark_cpu_info_);
 
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    ARGBExtractAlpha(src_pixels, benchmark_width_ * 4,
-                     dst_pixels_opt, benchmark_width_,
-                     benchmark_width_, benchmark_height_);
+    ARGBExtractAlpha(src_pixels, benchmark_width_ * 4, dst_pixels_opt,
+                     benchmark_width_, benchmark_width_, benchmark_height_);
   }
   for (int i = 0; i < kPixels; ++i) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
@@ -2230,15 +2775,13 @@ TEST_F(LibYUVPlanarTest, TestARGBCopyYToAlpha) {
   memcpy(dst_pixels_c, dst_pixels_opt, kPixels * 4);
 
   MaskCpuFlags(disable_cpu_flags_);
-  ARGBCopyYToAlpha(orig_pixels, benchmark_width_,
-                   dst_pixels_c, benchmark_width_ * 4,
-                   benchmark_width_, benchmark_height_);
+  ARGBCopyYToAlpha(orig_pixels, benchmark_width_, dst_pixels_c,
+                   benchmark_width_ * 4, benchmark_width_, benchmark_height_);
   MaskCpuFlags(benchmark_cpu_info_);
 
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    ARGBCopyYToAlpha(orig_pixels, benchmark_width_,
-                     dst_pixels_opt, benchmark_width_ * 4,
-                     benchmark_width_, benchmark_height_);
+    ARGBCopyYToAlpha(orig_pixels, benchmark_width_, dst_pixels_opt,
+                     benchmark_width_ * 4, benchmark_width_, benchmark_height_);
   }
   for (int i = 0; i < kPixels * 4; ++i) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
@@ -2249,15 +2792,20 @@ TEST_F(LibYUVPlanarTest, TestARGBCopyYToAlpha) {
   free_aligned_buffer_page_end(orig_pixels);
 }
 
-static int TestARGBRect(int width, int height, int benchmark_iterations,
-                        int disable_cpu_flags, int benchmark_cpu_info,
-                        int invert, int off, int bpp) {
+static int TestARGBRect(int width,
+                        int height,
+                        int benchmark_iterations,
+                        int disable_cpu_flags,
+                        int benchmark_cpu_info,
+                        int invert,
+                        int off,
+                        int bpp) {
   if (width < 1) {
     width = 1;
   }
   const int kStride = width * bpp;
   const int kSize = kStride * height;
-  const uint32 v32 = fastrand() & (bpp == 4 ? 0xffffffff : 0xff);
+  const uint32_t v32 = fastrand() & (bpp == 4 ? 0xffffffff : 0xff);
 
   align_buffer_page_end(dst_argb_c, kSize + off);
   align_buffer_page_end(dst_argb_opt, kSize + off);
@@ -2282,9 +2830,8 @@ static int TestARGBRect(int width, int height, int benchmark_iterations,
   }
   int max_diff = 0;
   for (int i = 0; i < kStride * height; ++i) {
-    int abs_diff =
-        abs(static_cast<int>(dst_argb_c[i + off]) -
-            static_cast<int>(dst_argb_opt[i + off]));
+    int abs_diff = abs(static_cast<int>(dst_argb_c[i + off]) -
+                       static_cast<int>(dst_argb_opt[i + off]));
     if (abs_diff > max_diff) {
       max_diff = abs_diff;
     }
@@ -2295,67 +2842,1630 @@ static int TestARGBRect(int width, int height, int benchmark_iterations,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBRect_Any) {
-  int max_diff = TestARGBRect(benchmark_width_ - 1, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              +1, 0, 4);
+  int max_diff = TestARGBRect(benchmark_width_ + 1, benchmark_height_,
+                              benchmark_iterations_, disable_cpu_flags_,
+                              benchmark_cpu_info_, +1, 0, 4);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBRect_Unaligned) {
-  int max_diff = TestARGBRect(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              +1, 1, 4);
+  int max_diff =
+      TestARGBRect(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 1, 4);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBRect_Invert) {
-  int max_diff = TestARGBRect(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              -1, 0, 4);
+  int max_diff =
+      TestARGBRect(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, -1, 0, 4);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBRect_Opt) {
-  int max_diff = TestARGBRect(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              +1, 0, 4);
+  int max_diff =
+      TestARGBRect(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 4);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, SetPlane_Any) {
-  int max_diff = TestARGBRect(benchmark_width_ - 1, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              +1, 0, 1);
+  int max_diff = TestARGBRect(benchmark_width_ + 1, benchmark_height_,
+                              benchmark_iterations_, disable_cpu_flags_,
+                              benchmark_cpu_info_, +1, 0, 1);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, SetPlane_Unaligned) {
-  int max_diff = TestARGBRect(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              +1, 1, 1);
+  int max_diff =
+      TestARGBRect(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 1, 1);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, SetPlane_Invert) {
-  int max_diff = TestARGBRect(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              -1, 0, 1);
+  int max_diff =
+      TestARGBRect(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, -1, 0, 1);
   EXPECT_EQ(0, max_diff);
 }
 
 TEST_F(LibYUVPlanarTest, SetPlane_Opt) {
-  int max_diff = TestARGBRect(benchmark_width_, benchmark_height_,
-                              benchmark_iterations_,
-                              disable_cpu_flags_, benchmark_cpu_info_,
-                              +1, 0, 1);
+  int max_diff =
+      TestARGBRect(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                   disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 1);
   EXPECT_EQ(0, max_diff);
 }
 
+TEST_F(LibYUVPlanarTest, MergeUVPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_u, kPixels);
+  align_buffer_page_end(src_pixels_v, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_c, kPixels * 2);
+
+  MemRandomize(src_pixels_u, kPixels);
+  MemRandomize(src_pixels_v, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 2);
+  MemRandomize(dst_pixels_c, kPixels * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  MergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v, benchmark_width_,
+               dst_pixels_c, benchmark_width_ * 2, benchmark_width_,
+               benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v, benchmark_width_,
+                 dst_pixels_opt, benchmark_width_ * 2, benchmark_width_,
+                 benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+// 16 bit channel split and merge
+TEST_F(LibYUVPlanarTest, MergeUVPlane_16_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_u, kPixels * 2);
+  align_buffer_page_end(src_pixels_v, kPixels * 2);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 2 * 2);
+  align_buffer_page_end(dst_pixels_c, kPixels * 2 * 2);
+  MemRandomize(src_pixels_u, kPixels * 2);
+  MemRandomize(src_pixels_v, kPixels * 2);
+  MemRandomize(dst_pixels_opt, kPixels * 2 * 2);
+  MemRandomize(dst_pixels_c, kPixels * 2 * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  MergeUVPlane_16((const uint16_t*)src_pixels_u, benchmark_width_,
+                  (const uint16_t*)src_pixels_v, benchmark_width_,
+                  (uint16_t*)dst_pixels_c, benchmark_width_ * 2,
+                  benchmark_width_, benchmark_height_, 12);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeUVPlane_16((const uint16_t*)src_pixels_u, benchmark_width_,
+                    (const uint16_t*)src_pixels_v, benchmark_width_,
+                    (uint16_t*)dst_pixels_opt, benchmark_width_ * 2,
+                    benchmark_width_, benchmark_height_, 12);
+  }
+
+  for (int i = 0; i < kPixels * 2 * 2; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, SplitUVPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 2);
+  align_buffer_page_end(dst_pixels_u_c, kPixels);
+  align_buffer_page_end(dst_pixels_v_c, kPixels);
+  align_buffer_page_end(dst_pixels_u_opt, kPixels);
+  align_buffer_page_end(dst_pixels_v_opt, kPixels);
+
+  MemRandomize(src_pixels, kPixels * 2);
+  MemRandomize(dst_pixels_u_c, kPixels);
+  MemRandomize(dst_pixels_v_c, kPixels);
+  MemRandomize(dst_pixels_u_opt, kPixels);
+  MemRandomize(dst_pixels_v_opt, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_u_c,
+               benchmark_width_, dst_pixels_v_c, benchmark_width_,
+               benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_u_opt,
+                 benchmark_width_, dst_pixels_v_opt, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_u_c[i], dst_pixels_u_opt[i]);
+    EXPECT_EQ(dst_pixels_v_c[i], dst_pixels_v_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_u_c);
+  free_aligned_buffer_page_end(dst_pixels_v_c);
+  free_aligned_buffer_page_end(dst_pixels_u_opt);
+  free_aligned_buffer_page_end(dst_pixels_v_opt);
+}
+
+// 16 bit channel split
+TEST_F(LibYUVPlanarTest, SplitUVPlane_16_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 2 * 2);
+  align_buffer_page_end(dst_pixels_u_c, kPixels * 2);
+  align_buffer_page_end(dst_pixels_v_c, kPixels * 2);
+  align_buffer_page_end(dst_pixels_u_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_v_opt, kPixels * 2);
+  MemRandomize(src_pixels, kPixels * 2 * 2);
+  MemRandomize(dst_pixels_u_c, kPixels * 2);
+  MemRandomize(dst_pixels_v_c, kPixels * 2);
+  MemRandomize(dst_pixels_u_opt, kPixels * 2);
+  MemRandomize(dst_pixels_v_opt, kPixels * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitUVPlane_16((const uint16_t*)src_pixels, benchmark_width_ * 2,
+                  (uint16_t*)dst_pixels_u_c, benchmark_width_,
+                  (uint16_t*)dst_pixels_v_c, benchmark_width_, benchmark_width_,
+                  benchmark_height_, 10);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitUVPlane_16((const uint16_t*)src_pixels, benchmark_width_ * 2,
+                    (uint16_t*)dst_pixels_u_opt, benchmark_width_,
+                    (uint16_t*)dst_pixels_v_opt, benchmark_width_,
+                    benchmark_width_, benchmark_height_, 10);
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_u_c[i], dst_pixels_u_opt[i]);
+    EXPECT_EQ(dst_pixels_v_c[i], dst_pixels_v_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_u_c);
+  free_aligned_buffer_page_end(dst_pixels_v_c);
+  free_aligned_buffer_page_end(dst_pixels_u_opt);
+  free_aligned_buffer_page_end(dst_pixels_v_opt);
+}
+
+TEST_F(LibYUVPlanarTest, SwapUVPlane_Opt) {
+  // Round count up to multiple of 16
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 2);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_c, kPixels * 2);
+
+  MemRandomize(src_pixels, kPixels * 2);
+  MemRandomize(dst_pixels_opt, kPixels * 2);
+  MemRandomize(dst_pixels_c, kPixels * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SwapUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_c,
+              benchmark_width_ * 2, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SwapUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_opt,
+                benchmark_width_ * 2, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, MergeRGBPlane_Opt) {
+  // Round count up to multiple of 16
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 3);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 3);
+  align_buffer_page_end(dst_pixels_c, kPixels * 3);
+
+  MemRandomize(src_pixels, kPixels * 3);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 3);
+  MemRandomize(dst_pixels_c, kPixels * 3);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitRGBPlane(src_pixels, benchmark_width_ * 3, tmp_pixels_r,
+                benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                benchmark_width_, benchmark_width_, benchmark_height_);
+  MergeRGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                tmp_pixels_b, benchmark_width_, dst_pixels_c,
+                benchmark_width_ * 3, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  SplitRGBPlane(src_pixels, benchmark_width_ * 3, tmp_pixels_r,
+                benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                benchmark_width_, benchmark_width_, benchmark_height_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeRGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g,
+                  benchmark_width_, tmp_pixels_b, benchmark_width_,
+                  dst_pixels_opt, benchmark_width_ * 3, benchmark_width_,
+                  benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 3; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, SplitRGBPlane_Opt) {
+  // Round count up to multiple of 16
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 3);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 3);
+  align_buffer_page_end(dst_pixels_c, kPixels * 3);
+
+  MemRandomize(src_pixels, kPixels * 3);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 3);
+  MemRandomize(dst_pixels_c, kPixels * 3);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitRGBPlane(src_pixels, benchmark_width_ * 3, tmp_pixels_r,
+                benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                benchmark_width_, benchmark_width_, benchmark_height_);
+  MergeRGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                tmp_pixels_b, benchmark_width_, dst_pixels_c,
+                benchmark_width_ * 3, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitRGBPlane(src_pixels, benchmark_width_ * 3, tmp_pixels_r,
+                  benchmark_width_, tmp_pixels_g, benchmark_width_,
+                  tmp_pixels_b, benchmark_width_, benchmark_width_,
+                  benchmark_height_);
+  }
+  MergeRGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                tmp_pixels_b, benchmark_width_, dst_pixels_opt,
+                benchmark_width_ * 3, benchmark_width_, benchmark_height_);
+
+  for (int i = 0; i < kPixels * 3; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, MergeARGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(tmp_pixels_a, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(tmp_pixels_a, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 dst_pixels_c, benchmark_width_ * 4, benchmark_width_,
+                 benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g,
+                   benchmark_width_, tmp_pixels_b, benchmark_width_,
+                   tmp_pixels_a, benchmark_width_, dst_pixels_opt,
+                   benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(tmp_pixels_a);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, SplitARGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(tmp_pixels_a, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(tmp_pixels_a, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 dst_pixels_c, benchmark_width_ * 4, benchmark_width_,
+                 benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                   benchmark_width_, tmp_pixels_g, benchmark_width_,
+                   tmp_pixels_b, benchmark_width_, tmp_pixels_a,
+                   benchmark_width_, benchmark_width_, benchmark_height_);
+  }
+
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 dst_pixels_opt, benchmark_width_ * 4, benchmark_width_,
+                 benchmark_height_);
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(tmp_pixels_a);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, MergeXRGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, NULL, 0, benchmark_width_,
+                 benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, NULL, 0, dst_pixels_c,
+                 benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, NULL, 0, benchmark_width_,
+                 benchmark_height_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g,
+                   benchmark_width_, tmp_pixels_b, benchmark_width_, NULL, 0,
+                   dst_pixels_opt, benchmark_width_ * 4, benchmark_width_,
+                   benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, SplitXRGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, NULL, 0, benchmark_width_,
+                 benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, NULL, 0, dst_pixels_c,
+                 benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                   benchmark_width_, tmp_pixels_g, benchmark_width_,
+                   tmp_pixels_b, benchmark_width_, NULL, 0, benchmark_width_,
+                   benchmark_height_);
+  }
+
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, NULL, 0, dst_pixels_opt,
+                 benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+// Merge 4 channels
+#define TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, W1280, N, NEG, OFF)      \
+  TEST_F(LibYUVPlanarTest, FUNC##Plane_##DEPTH##N) {                        \
+    const int kWidth = W1280;                                               \
+    const int kPixels = kWidth * benchmark_height_;                         \
+    align_buffer_page_end(src_memory_r, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_g, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_b, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_a, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(dst_memory_c, kPixels * 4 * sizeof(DTYPE));       \
+    align_buffer_page_end(dst_memory_opt, kPixels * 4 * sizeof(DTYPE));     \
+    MemRandomize(src_memory_r, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_g, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_b, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_a, kPixels * sizeof(STYPE) + OFF);              \
+    memset(dst_memory_c, 0, kPixels * 4 * sizeof(DTYPE));                   \
+    memset(dst_memory_opt, 0, kPixels * 4 * sizeof(DTYPE));                 \
+    STYPE* src_pixels_r = reinterpret_cast<STYPE*>(src_memory_r + OFF);     \
+    STYPE* src_pixels_g = reinterpret_cast<STYPE*>(src_memory_g + OFF);     \
+    STYPE* src_pixels_b = reinterpret_cast<STYPE*>(src_memory_b + OFF);     \
+    STYPE* src_pixels_a = reinterpret_cast<STYPE*>(src_memory_a + OFF);     \
+    DTYPE* dst_pixels_c = reinterpret_cast<DTYPE*>(dst_memory_c);           \
+    DTYPE* dst_pixels_opt = reinterpret_cast<DTYPE*>(dst_memory_opt);       \
+    MaskCpuFlags(disable_cpu_flags_);                                       \
+    FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b,   \
+                kWidth, src_pixels_a, kWidth, dst_pixels_c, kWidth * 4,     \
+                kWidth, NEG benchmark_height_, DEPTH);                      \
+    MaskCpuFlags(benchmark_cpu_info_);                                      \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                       \
+      FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b, \
+                  kWidth, src_pixels_a, kWidth, dst_pixels_opt, kWidth * 4, \
+                  kWidth, NEG benchmark_height_, DEPTH);                    \
+    }                                                                       \
+    for (int i = 0; i < kPixels * 4; ++i) {                                 \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                        \
+    }                                                                       \
+    free_aligned_buffer_page_end(src_memory_r);                             \
+    free_aligned_buffer_page_end(src_memory_g);                             \
+    free_aligned_buffer_page_end(src_memory_b);                             \
+    free_aligned_buffer_page_end(src_memory_a);                             \
+    free_aligned_buffer_page_end(dst_memory_c);                             \
+    free_aligned_buffer_page_end(dst_memory_opt);                           \
+  }
+
+// Merge 3 channel RGB into 4 channel XRGB with opaque alpha
+#define TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, W1280, N, NEG, OFF)     \
+  TEST_F(LibYUVPlanarTest, FUNC##Plane_Opaque_##DEPTH##N) {                 \
+    const int kWidth = W1280;                                               \
+    const int kPixels = kWidth * benchmark_height_;                         \
+    align_buffer_page_end(src_memory_r, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_g, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_b, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(dst_memory_c, kPixels * 4 * sizeof(DTYPE));       \
+    align_buffer_page_end(dst_memory_opt, kPixels * 4 * sizeof(DTYPE));     \
+    MemRandomize(src_memory_r, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_g, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_b, kPixels * sizeof(STYPE) + OFF);              \
+    memset(dst_memory_c, 0, kPixels * 4 * sizeof(DTYPE));                   \
+    memset(dst_memory_opt, 0, kPixels * 4 * sizeof(DTYPE));                 \
+    STYPE* src_pixels_r = reinterpret_cast<STYPE*>(src_memory_r + OFF);     \
+    STYPE* src_pixels_g = reinterpret_cast<STYPE*>(src_memory_g + OFF);     \
+    STYPE* src_pixels_b = reinterpret_cast<STYPE*>(src_memory_b + OFF);     \
+    DTYPE* dst_pixels_c = reinterpret_cast<DTYPE*>(dst_memory_c);           \
+    DTYPE* dst_pixels_opt = reinterpret_cast<DTYPE*>(dst_memory_opt);       \
+    MaskCpuFlags(disable_cpu_flags_);                                       \
+    FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b,   \
+                kWidth, NULL, 0, dst_pixels_c, kWidth * 4, kWidth,          \
+                NEG benchmark_height_, DEPTH);                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                      \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                       \
+      FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b, \
+                  kWidth, NULL, 0, dst_pixels_opt, kWidth * 4, kWidth,      \
+                  NEG benchmark_height_, DEPTH);                            \
+    }                                                                       \
+    for (int i = 0; i < kPixels * 4; ++i) {                                 \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                        \
+    }                                                                       \
+    free_aligned_buffer_page_end(src_memory_r);                             \
+    free_aligned_buffer_page_end(src_memory_g);                             \
+    free_aligned_buffer_page_end(src_memory_b);                             \
+    free_aligned_buffer_page_end(dst_memory_c);                             \
+    free_aligned_buffer_page_end(dst_memory_opt);                           \
+  }
+
+#define TESTQPLANARTOP(FUNC, STYPE, DTYPE, DEPTH)                              \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_ + 1, _Any, +, 0) \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Unaligned, +,  \
+                  2)                                                           \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Invert, -, 0)  \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Opt, +, 0)     \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_ + 1, _Any, +,   \
+                   0)                                                          \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Unaligned, +, \
+                   2)                                                          \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Invert, -, 0) \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Opt, +, 0)
+
+TESTQPLANARTOP(MergeAR64, uint16_t, uint16_t, 10)
+TESTQPLANARTOP(MergeAR64, uint16_t, uint16_t, 12)
+TESTQPLANARTOP(MergeAR64, uint16_t, uint16_t, 16)
+TESTQPLANARTOP(MergeARGB16To8, uint16_t, uint8_t, 10)
+TESTQPLANARTOP(MergeARGB16To8, uint16_t, uint8_t, 12)
+TESTQPLANARTOP(MergeARGB16To8, uint16_t, uint8_t, 16)
+
+#define TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, W1280, N, NEG, OFF)      \
+  TEST_F(LibYUVPlanarTest, FUNC##Plane_##DEPTH##N) {                        \
+    const int kWidth = W1280;                                               \
+    const int kPixels = kWidth * benchmark_height_;                         \
+    align_buffer_page_end(src_memory_r, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_g, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_b, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(dst_memory_c, kPixels * 4 * sizeof(DTYPE));       \
+    align_buffer_page_end(dst_memory_opt, kPixels * 4 * sizeof(DTYPE));     \
+    MemRandomize(src_memory_r, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_g, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_b, kPixels * sizeof(STYPE) + OFF);              \
+    STYPE* src_pixels_r = reinterpret_cast<STYPE*>(src_memory_r + OFF);     \
+    STYPE* src_pixels_g = reinterpret_cast<STYPE*>(src_memory_g + OFF);     \
+    STYPE* src_pixels_b = reinterpret_cast<STYPE*>(src_memory_b + OFF);     \
+    DTYPE* dst_pixels_c = reinterpret_cast<DTYPE*>(dst_memory_c);           \
+    DTYPE* dst_pixels_opt = reinterpret_cast<DTYPE*>(dst_memory_opt);       \
+    memset(dst_pixels_c, 1, kPixels * 4 * sizeof(DTYPE));                   \
+    memset(dst_pixels_opt, 2, kPixels * 4 * sizeof(DTYPE));                 \
+    MaskCpuFlags(disable_cpu_flags_);                                       \
+    FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b,   \
+                kWidth, dst_pixels_c, kWidth * 4, kWidth,                   \
+                NEG benchmark_height_, DEPTH);                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                      \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                       \
+      FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b, \
+                  kWidth, dst_pixels_opt, kWidth * 4, kWidth,               \
+                  NEG benchmark_height_, DEPTH);                            \
+    }                                                                       \
+    for (int i = 0; i < kPixels * 4; ++i) {                                 \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                        \
+    }                                                                       \
+    free_aligned_buffer_page_end(src_memory_r);                             \
+    free_aligned_buffer_page_end(src_memory_g);                             \
+    free_aligned_buffer_page_end(src_memory_b);                             \
+    free_aligned_buffer_page_end(dst_memory_c);                             \
+    free_aligned_buffer_page_end(dst_memory_opt);                           \
+  }
+
+#define TESTTPLANARTOP(FUNC, STYPE, DTYPE, DEPTH)                              \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_ + 1, _Any, +, 0) \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Unaligned, +,  \
+                  2)                                                           \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Invert, -, 0)  \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Opt, +, 0)
+
+TESTTPLANARTOP(MergeXR30, uint16_t, uint8_t, 10)
+TESTTPLANARTOP(MergeXR30, uint16_t, uint8_t, 12)
+TESTTPLANARTOP(MergeXR30, uint16_t, uint8_t, 16)
+
+// TODO(fbarchard): improve test for platforms and cpu detect
+#ifdef HAS_MERGEUVROW_16_AVX2
+TEST_F(LibYUVPlanarTest, MergeUVRow_16_Opt) {
+  // Round count up to multiple of 8
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 7) & ~7;
+
+  align_buffer_page_end(src_pixels_u, kPixels * 2);
+  align_buffer_page_end(src_pixels_v, kPixels * 2);
+  align_buffer_page_end(dst_pixels_uv_opt, kPixels * 2 * 2);
+  align_buffer_page_end(dst_pixels_uv_c, kPixels * 2 * 2);
+
+  MemRandomize(src_pixels_u, kPixels * 2);
+  MemRandomize(src_pixels_v, kPixels * 2);
+  memset(dst_pixels_uv_opt, 0, kPixels * 2 * 2);
+  memset(dst_pixels_uv_c, 1, kPixels * 2 * 2);
+
+  MergeUVRow_16_C(reinterpret_cast<const uint16_t*>(src_pixels_u),
+                  reinterpret_cast<const uint16_t*>(src_pixels_v),
+                  reinterpret_cast<uint16_t*>(dst_pixels_uv_c), 16, kPixels);
+
+  int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    if (has_avx2) {
+      MergeUVRow_16_AVX2(reinterpret_cast<const uint16_t*>(src_pixels_u),
+                         reinterpret_cast<const uint16_t*>(src_pixels_v),
+                         reinterpret_cast<uint16_t*>(dst_pixels_uv_opt), 16,
+                         kPixels);
+    } else {
+      MergeUVRow_16_C(reinterpret_cast<const uint16_t*>(src_pixels_u),
+                      reinterpret_cast<const uint16_t*>(src_pixels_v),
+                      reinterpret_cast<uint16_t*>(dst_pixels_uv_opt), 16,
+                      kPixels);
+    }
+  }
+
+  for (int i = 0; i < kPixels * 2 * 2; ++i) {
+    EXPECT_EQ(dst_pixels_uv_opt[i], dst_pixels_uv_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
+  free_aligned_buffer_page_end(dst_pixels_uv_opt);
+  free_aligned_buffer_page_end(dst_pixels_uv_c);
+}
+#endif
+
+// TODO(fbarchard): Improve test for more platforms.
+#ifdef HAS_MULTIPLYROW_16_AVX2
+TEST_F(LibYUVPlanarTest, MultiplyRow_16_Opt) {
+  // Round count up to multiple of 32
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 31) & ~31;
+
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_c, kPixels * 2);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels * 2);
+  memset(dst_pixels_y_c, 1, kPixels * 2);
+
+  MultiplyRow_16_C(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                   reinterpret_cast<uint16_t*>(dst_pixels_y_c), 64, kPixels);
+
+  int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    if (has_avx2) {
+      MultiplyRow_16_AVX2(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                          reinterpret_cast<uint16_t*>(dst_pixels_y_opt), 64,
+                          kPixels);
+    } else {
+      MultiplyRow_16_C(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                       reinterpret_cast<uint16_t*>(dst_pixels_y_opt), 64,
+                       kPixels);
+    }
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+#endif  // HAS_MULTIPLYROW_16_AVX2
+
+TEST_F(LibYUVPlanarTest, Convert16To8Plane) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  Convert16To8Plane(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                    benchmark_width_, dst_pixels_y_c, benchmark_width_, 16384,
+                    benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    Convert16To8Plane(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                      benchmark_width_, dst_pixels_y_opt, benchmark_width_,
+                      16384, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+
+TEST_F(LibYUVPlanarTest, YUY2ToY) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  YUY2ToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_c, benchmark_width_,
+          benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    YUY2ToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_opt,
+            benchmark_width_, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+
+TEST_F(LibYUVPlanarTest, UYVYToY) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  UYVYToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_c, benchmark_width_,
+          benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    UYVYToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_opt,
+            benchmark_width_, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+
+#ifdef ENABLE_ROW_TESTS
+// TODO(fbarchard): Improve test for more platforms.
+#ifdef HAS_CONVERT16TO8ROW_AVX2
+TEST_F(LibYUVPlanarTest, Convert16To8Row_Opt) {
+  // AVX2 does multiple of 32, so round count up
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 31) & ~31;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  // clamp source range to 10 bits.
+  for (int i = 0; i < kPixels; ++i) {
+    reinterpret_cast<uint16_t*>(src_pixels_y)[i] &= 1023;
+  }
+
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  Convert16To8Row_C(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                    dst_pixels_y_c, 16384, kPixels);
+
+  int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+  int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    if (has_avx2) {
+      Convert16To8Row_AVX2(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                           dst_pixels_y_opt, 16384, kPixels);
+    } else if (has_ssse3) {
+      Convert16To8Row_SSSE3(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                            dst_pixels_y_opt, 16384, kPixels);
+    } else {
+      Convert16To8Row_C(reinterpret_cast<const uint16_t*>(src_pixels_y),
+                        dst_pixels_y_opt, 16384, kPixels);
+    }
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+#endif  // HAS_CONVERT16TO8ROW_AVX2
+
+#ifdef HAS_UYVYTOYROW_NEON
+TEST_F(LibYUVPlanarTest, UYVYToYRow_Opt) {
+  // NEON does multiple of 16, so round count up
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 15) & ~15;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  UYVYToYRow_C(src_pixels_y, dst_pixels_y_c, kPixels);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    UYVYToYRow_NEON(src_pixels_y, dst_pixels_y_opt, kPixels);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+#endif  // HAS_UYVYTOYROW_NEON
+
+#endif  // ENABLE_ROW_TESTS
+
+TEST_F(LibYUVPlanarTest, Convert8To16Plane) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_y, kPixels);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_c, kPixels * 2);
+
+  MemRandomize(src_pixels_y, kPixels);
+  memset(dst_pixels_y_opt, 0, kPixels * 2);
+  memset(dst_pixels_y_c, 1, kPixels * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  Convert8To16Plane(src_pixels_y, benchmark_width_,
+                    reinterpret_cast<uint16_t*>(dst_pixels_y_c),
+                    benchmark_width_, 1024, benchmark_width_,
+                    benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    Convert8To16Plane(src_pixels_y, benchmark_width_,
+                      reinterpret_cast<uint16_t*>(dst_pixels_y_opt),
+                      benchmark_width_, 1024, benchmark_width_,
+                      benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+
+#ifdef ENABLE_ROW_TESTS
+// TODO(fbarchard): Improve test for more platforms.
+#ifdef HAS_CONVERT8TO16ROW_AVX2
+TEST_F(LibYUVPlanarTest, Convert8To16Row_Opt) {
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 31) & ~31;
+  align_buffer_page_end(src_pixels_y, kPixels);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_c, kPixels * 2);
+
+  MemRandomize(src_pixels_y, kPixels);
+  memset(dst_pixels_y_opt, 0, kPixels * 2);
+  memset(dst_pixels_y_c, 1, kPixels * 2);
+
+  Convert8To16Row_C(src_pixels_y, reinterpret_cast<uint16_t*>(dst_pixels_y_c),
+                    1024, kPixels);
+
+  int has_avx2 = TestCpuFlag(kCpuHasAVX2);
+  int has_sse2 = TestCpuFlag(kCpuHasSSE2);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    if (has_avx2) {
+      Convert8To16Row_AVX2(src_pixels_y,
+                           reinterpret_cast<uint16_t*>(dst_pixels_y_opt), 1024,
+                           kPixels);
+    } else if (has_sse2) {
+      Convert8To16Row_SSE2(src_pixels_y,
+                           reinterpret_cast<uint16_t*>(dst_pixels_y_opt), 1024,
+                           kPixels);
+    } else {
+      Convert8To16Row_C(src_pixels_y,
+                        reinterpret_cast<uint16_t*>(dst_pixels_y_opt), 1024,
+                        kPixels);
+    }
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+#endif  // HAS_CONVERT8TO16ROW_AVX2
+
+float TestScaleMaxSamples(int benchmark_width,
+                          int benchmark_height,
+                          int benchmark_iterations,
+                          float scale,
+                          bool opt) {
+  int i, j;
+  float max_c, max_opt = 0.f;
+  // NEON does multiple of 8, so round count up
+  const int kPixels = (benchmark_width * benchmark_height + 7) & ~7;
+  align_buffer_page_end(orig_y, kPixels * 4 * 3 + 48);
+  uint8_t* dst_c = orig_y + kPixels * 4 + 16;
+  uint8_t* dst_opt = orig_y + kPixels * 4 * 2 + 32;
+
+  // Randomize works but may contain some denormals affecting performance.
+  // MemRandomize(orig_y, kPixels * 4);
+  // large values are problematic.  audio is really -1 to 1.
+  for (i = 0; i < kPixels; ++i) {
+    (reinterpret_cast<float*>(orig_y))[i] = sinf(static_cast<float>(i) * 0.1f);
+  }
+  memset(dst_c, 0, kPixels * 4);
+  memset(dst_opt, 1, kPixels * 4);
+
+  max_c = ScaleMaxSamples_C(reinterpret_cast<float*>(orig_y),
+                            reinterpret_cast<float*>(dst_c), scale, kPixels);
+
+  for (j = 0; j < benchmark_iterations; j++) {
+    if (opt) {
+#ifdef HAS_SCALESUMSAMPLES_NEON
+      max_opt = ScaleMaxSamples_NEON(reinterpret_cast<float*>(orig_y),
+                                     reinterpret_cast<float*>(dst_opt), scale,
+                                     kPixels);
+#else
+      max_opt =
+          ScaleMaxSamples_C(reinterpret_cast<float*>(orig_y),
+                            reinterpret_cast<float*>(dst_opt), scale, kPixels);
+#endif
+    } else {
+      max_opt =
+          ScaleMaxSamples_C(reinterpret_cast<float*>(orig_y),
+                            reinterpret_cast<float*>(dst_opt), scale, kPixels);
+    }
+  }
+
+  float max_diff = FAbs(max_opt - max_c);
+  for (i = 0; i < kPixels; ++i) {
+    float abs_diff = FAbs((reinterpret_cast<float*>(dst_c)[i]) -
+                          (reinterpret_cast<float*>(dst_opt)[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  return max_diff;
+}
+
+TEST_F(LibYUVPlanarTest, TestScaleMaxSamples_C) {
+  float diff = TestScaleMaxSamples(benchmark_width_, benchmark_height_,
+                                   benchmark_iterations_, 1.2f, false);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestScaleMaxSamples_Opt) {
+  float diff = TestScaleMaxSamples(benchmark_width_, benchmark_height_,
+                                   benchmark_iterations_, 1.2f, true);
+  EXPECT_EQ(0, diff);
+}
+
+float TestScaleSumSamples(int benchmark_width,
+                          int benchmark_height,
+                          int benchmark_iterations,
+                          float scale,
+                          bool opt) {
+  int i, j;
+  float sum_c, sum_opt = 0.f;
+  // NEON does multiple of 8, so round count up
+  const int kPixels = (benchmark_width * benchmark_height + 7) & ~7;
+  align_buffer_page_end(orig_y, kPixels * 4 * 3);
+  uint8_t* dst_c = orig_y + kPixels * 4;
+  uint8_t* dst_opt = orig_y + kPixels * 4 * 2;
+
+  // Randomize works but may contain some denormals affecting performance.
+  // MemRandomize(orig_y, kPixels * 4);
+  // large values are problematic.  audio is really -1 to 1.
+  for (i = 0; i < kPixels; ++i) {
+    (reinterpret_cast<float*>(orig_y))[i] = sinf(static_cast<float>(i) * 0.1f);
+  }
+  memset(dst_c, 0, kPixels * 4);
+  memset(dst_opt, 1, kPixels * 4);
+
+  sum_c = ScaleSumSamples_C(reinterpret_cast<float*>(orig_y),
+                            reinterpret_cast<float*>(dst_c), scale, kPixels);
+
+  for (j = 0; j < benchmark_iterations; j++) {
+    if (opt) {
+#ifdef HAS_SCALESUMSAMPLES_NEON
+      sum_opt = ScaleSumSamples_NEON(reinterpret_cast<float*>(orig_y),
+                                     reinterpret_cast<float*>(dst_opt), scale,
+                                     kPixels);
+#else
+      sum_opt =
+          ScaleSumSamples_C(reinterpret_cast<float*>(orig_y),
+                            reinterpret_cast<float*>(dst_opt), scale, kPixels);
+#endif
+    } else {
+      sum_opt =
+          ScaleSumSamples_C(reinterpret_cast<float*>(orig_y),
+                            reinterpret_cast<float*>(dst_opt), scale, kPixels);
+    }
+  }
+
+  float mse_opt = sum_opt / kPixels * 4;
+  float mse_c = sum_c / kPixels * 4;
+  float mse_error = FAbs(mse_opt - mse_c) / mse_c;
+
+  // If the sum of a float is more than 4 million, small adds are round down on
+  // float and produce different results with vectorized sum vs scalar sum.
+  // Ignore the difference if the sum is large.
+  float max_diff = 0.f;
+  if (mse_error > 0.0001 && sum_c < 4000000) {  // allow .01% difference of mse
+    max_diff = mse_error;
+  }
+
+  for (i = 0; i < kPixels; ++i) {
+    float abs_diff = FAbs((reinterpret_cast<float*>(dst_c)[i]) -
+                          (reinterpret_cast<float*>(dst_opt)[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  return max_diff;
+}
+
+TEST_F(LibYUVPlanarTest, TestScaleSumSamples_C) {
+  float diff = TestScaleSumSamples(benchmark_width_, benchmark_height_,
+                                   benchmark_iterations_, 1.2f, false);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestScaleSumSamples_Opt) {
+  float diff = TestScaleSumSamples(benchmark_width_, benchmark_height_,
+                                   benchmark_iterations_, 1.2f, true);
+  EXPECT_EQ(0, diff);
+}
+
+float TestScaleSamples(int benchmark_width,
+                       int benchmark_height,
+                       int benchmark_iterations,
+                       float scale,
+                       bool opt) {
+  int i, j;
+  // NEON does multiple of 8, so round count up
+  const int kPixels = (benchmark_width * benchmark_height + 7) & ~7;
+  align_buffer_page_end(orig_y, kPixels * 4 * 3);
+  uint8_t* dst_c = orig_y + kPixels * 4;
+  uint8_t* dst_opt = orig_y + kPixels * 4 * 2;
+
+  // Randomize works but may contain some denormals affecting performance.
+  // MemRandomize(orig_y, kPixels * 4);
+  // large values are problematic.  audio is really -1 to 1.
+  for (i = 0; i < kPixels; ++i) {
+    (reinterpret_cast<float*>(orig_y))[i] = sinf(static_cast<float>(i) * 0.1f);
+  }
+  memset(dst_c, 0, kPixels * 4);
+  memset(dst_opt, 1, kPixels * 4);
+
+  ScaleSamples_C(reinterpret_cast<float*>(orig_y),
+                 reinterpret_cast<float*>(dst_c), scale, kPixels);
+
+  for (j = 0; j < benchmark_iterations; j++) {
+    if (opt) {
+#ifdef HAS_SCALESUMSAMPLES_NEON
+      ScaleSamples_NEON(reinterpret_cast<float*>(orig_y),
+                        reinterpret_cast<float*>(dst_opt), scale, kPixels);
+#else
+      ScaleSamples_C(reinterpret_cast<float*>(orig_y),
+                     reinterpret_cast<float*>(dst_opt), scale, kPixels);
+#endif
+    } else {
+      ScaleSamples_C(reinterpret_cast<float*>(orig_y),
+                     reinterpret_cast<float*>(dst_opt), scale, kPixels);
+    }
+  }
+
+  float max_diff = 0.f;
+  for (i = 0; i < kPixels; ++i) {
+    float abs_diff = FAbs((reinterpret_cast<float*>(dst_c)[i]) -
+                          (reinterpret_cast<float*>(dst_opt)[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  return max_diff;
+}
+
+TEST_F(LibYUVPlanarTest, TestScaleSamples_C) {
+  float diff = TestScaleSamples(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, 1.2f, false);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestScaleSamples_Opt) {
+  float diff = TestScaleSamples(benchmark_width_, benchmark_height_,
+                                benchmark_iterations_, 1.2f, true);
+  EXPECT_EQ(0, diff);
+}
+
+float TestCopySamples(int benchmark_width,
+                      int benchmark_height,
+                      int benchmark_iterations,
+                      bool opt) {
+  int i, j;
+  // NEON does multiple of 16 floats, so round count up
+  const int kPixels = (benchmark_width * benchmark_height + 15) & ~15;
+  align_buffer_page_end(orig_y, kPixels * 4 * 3);
+  uint8_t* dst_c = orig_y + kPixels * 4;
+  uint8_t* dst_opt = orig_y + kPixels * 4 * 2;
+
+  // Randomize works but may contain some denormals affecting performance.
+  // MemRandomize(orig_y, kPixels * 4);
+  // large values are problematic.  audio is really -1 to 1.
+  for (i = 0; i < kPixels; ++i) {
+    (reinterpret_cast<float*>(orig_y))[i] = sinf(static_cast<float>(i) * 0.1f);
+  }
+  memset(dst_c, 0, kPixels * 4);
+  memset(dst_opt, 1, kPixels * 4);
+
+  memcpy(reinterpret_cast<void*>(dst_c), reinterpret_cast<void*>(orig_y),
+         kPixels * 4);
+
+  for (j = 0; j < benchmark_iterations; j++) {
+    if (opt) {
+#ifdef HAS_COPYROW_NEON
+      CopyRow_NEON(orig_y, dst_opt, kPixels * 4);
+#else
+      CopyRow_C(orig_y, dst_opt, kPixels * 4);
+#endif
+    } else {
+      CopyRow_C(orig_y, dst_opt, kPixels * 4);
+    }
+  }
+
+  float max_diff = 0.f;
+  for (i = 0; i < kPixels; ++i) {
+    float abs_diff = FAbs((reinterpret_cast<float*>(dst_c)[i]) -
+                          (reinterpret_cast<float*>(dst_opt)[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  return max_diff;
+}
+
+TEST_F(LibYUVPlanarTest, TestCopySamples_C) {
+  float diff = TestCopySamples(benchmark_width_, benchmark_height_,
+                               benchmark_iterations_, false);
+  EXPECT_EQ(0, diff);
+}
+
+TEST_F(LibYUVPlanarTest, TestCopySamples_Opt) {
+  float diff = TestCopySamples(benchmark_width_, benchmark_height_,
+                               benchmark_iterations_, true);
+  EXPECT_EQ(0, diff);
+}
+
+extern "C" void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width);
+extern "C" void GaussRow_C(const uint32_t* src, uint16_t* dst, int width);
+
+TEST_F(LibYUVPlanarTest, TestGaussRow_Opt) {
+  SIMD_ALIGNED(uint32_t orig_pixels[1280 + 8]);
+  SIMD_ALIGNED(uint16_t dst_pixels_c[1280]);
+  SIMD_ALIGNED(uint16_t dst_pixels_opt[1280]);
+
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
+
+  for (int i = 0; i < 1280 + 8; ++i) {
+    orig_pixels[i] = i * 256;
+  }
+  GaussRow_C(&orig_pixels[0], &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      GaussRow_NEON(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+    } else {
+      GaussRow_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+    }
+#else
+    GaussRow_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+#endif
+  }
+
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  EXPECT_EQ(dst_pixels_c[0],
+            static_cast<uint16_t>(0 * 1 + 1 * 4 + 2 * 6 + 3 * 4 + 4 * 1));
+  EXPECT_EQ(dst_pixels_c[639], static_cast<uint16_t>(10256));
+}
+
+extern "C" void GaussCol_NEON(const uint16_t* src0,
+                              const uint16_t* src1,
+                              const uint16_t* src2,
+                              const uint16_t* src3,
+                              const uint16_t* src4,
+                              uint32_t* dst,
+                              int width);
+
+extern "C" void GaussCol_C(const uint16_t* src0,
+                           const uint16_t* src1,
+                           const uint16_t* src2,
+                           const uint16_t* src3,
+                           const uint16_t* src4,
+                           uint32_t* dst,
+                           int width);
+
+TEST_F(LibYUVPlanarTest, TestGaussCol_Opt) {
+  SIMD_ALIGNED(uint16_t orig_pixels[1280 * 5]);
+  SIMD_ALIGNED(uint32_t dst_pixels_c[1280]);
+  SIMD_ALIGNED(uint32_t dst_pixels_opt[1280]);
+
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
+
+  for (int i = 0; i < 1280 * 5; ++i) {
+    orig_pixels[i] = static_cast<float>(i);
+  }
+  GaussCol_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+             &orig_pixels[1280 * 3], &orig_pixels[1280 * 4], &dst_pixels_c[0],
+             1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      GaussCol_NEON(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                    &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                    &dst_pixels_opt[0], 1280);
+    } else {
+      GaussCol_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                 &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                 &dst_pixels_opt[0], 1280);
+    }
+#else
+    GaussCol_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+               &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+               &dst_pixels_opt[0], 1280);
+#endif
+  }
+
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+}
+
+TEST_F(LibYUVPlanarTest, TestGaussRow_F32_Opt) {
+  SIMD_ALIGNED(float orig_pixels[1280 + 4]);
+  SIMD_ALIGNED(float dst_pixels_c[1280]);
+  SIMD_ALIGNED(float dst_pixels_opt[1280]);
+
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
+
+  for (int i = 0; i < 1280 + 4; ++i) {
+    orig_pixels[i] = static_cast<float>(i);
+  }
+  GaussRow_F32_C(&orig_pixels[0], &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      GaussRow_F32_NEON(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+    } else {
+      GaussRow_F32_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+    }
+#else
+    GaussRow_F32_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+#endif
+  }
+
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+}
+
+TEST_F(LibYUVPlanarTest, TestGaussCol_F32_Opt) {
+  SIMD_ALIGNED(float dst_pixels_c[1280]);
+  SIMD_ALIGNED(float dst_pixels_opt[1280]);
+  align_buffer_page_end(orig_pixels_buf, 1280 * 5 * 4);  // 5 rows
+  float* orig_pixels = reinterpret_cast<float*>(orig_pixels_buf);
+
+  memset(orig_pixels, 0, 1280 * 5 * 4);
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
+
+  for (int i = 0; i < 1280 * 5; ++i) {
+    orig_pixels[i] = static_cast<float>(i);
+  }
+  GaussCol_F32_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                 &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                 &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      GaussCol_F32_NEON(&orig_pixels[0], &orig_pixels[1280],
+                        &orig_pixels[1280 * 2], &orig_pixels[1280 * 3],
+                        &orig_pixels[1280 * 4], &dst_pixels_opt[0], 1280);
+    } else {
+      GaussCol_F32_C(&orig_pixels[0], &orig_pixels[1280],
+                     &orig_pixels[1280 * 2], &orig_pixels[1280 * 3],
+                     &orig_pixels[1280 * 4], &dst_pixels_opt[0], 1280);
+    }
+#else
+    GaussCol_F32_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                   &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                   &dst_pixels_opt[0], 1280);
+#endif
+  }
+
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(orig_pixels_buf);
+}
+
+TEST_F(LibYUVPlanarTest, SwapUVRow) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  void (*SwapUVRow)(const uint8_t* src_uv, uint8_t* dst_vu, int width) =
+      SwapUVRow_C;
+
+  align_buffer_page_end(src_pixels_vu, kPixels * 2);
+  align_buffer_page_end(dst_pixels_uv, kPixels * 2);
+  MemRandomize(src_pixels_vu, kPixels * 2);
+  memset(dst_pixels_uv, 1, kPixels * 2);
+
+#if defined(HAS_SWAPUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SwapUVRow = SwapUVRow_Any_NEON;
+    if (IS_ALIGNED(kPixels, 16)) {
+      SwapUVRow = SwapUVRow_NEON;
+    }
+  }
+#endif
+
+  for (int j = 0; j < benchmark_iterations_; j++) {
+    SwapUVRow(src_pixels_vu, dst_pixels_uv, kPixels);
+  }
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_uv[i * 2 + 0], src_pixels_vu[i * 2 + 1]);
+    EXPECT_EQ(dst_pixels_uv[i * 2 + 1], src_pixels_vu[i * 2 + 0]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_vu);
+  free_aligned_buffer_page_end(dst_pixels_uv);
+}
+#endif  // ENABLE_ROW_TESTS
+
+TEST_F(LibYUVPlanarTest, TestGaussPlane_F32) {
+  const int kSize = benchmark_width_ * benchmark_height_ * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  align_buffer_page_end(dst_pixels_opt, kSize);
+  align_buffer_page_end(dst_pixels_c, kSize);
+
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    ((float*)(orig_pixels))[i] = (i & 1023) * 3.14f;
+  }
+  memset(dst_pixels_opt, 1, kSize);
+  memset(dst_pixels_c, 2, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  GaussPlane_F32((const float*)(orig_pixels), benchmark_width_,
+                 (float*)(dst_pixels_c), benchmark_width_, benchmark_width_,
+                 benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    GaussPlane_F32((const float*)(orig_pixels), benchmark_width_,
+                   (float*)(dst_pixels_opt), benchmark_width_, benchmark_width_,
+                   benchmark_height_);
+  }
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_NEAR(((float*)(dst_pixels_c))[i], ((float*)(dst_pixels_opt))[i], 1.f)
+        << i;
+  }
+
+  free_aligned_buffer_page_end(dst_pixels_c);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVPlanarTest, HalfMergeUVPlane_Opt) {
+  int dst_width = (benchmark_width_ + 1) / 2;
+  int dst_height = (benchmark_height_ + 1) / 2;
+  align_buffer_page_end(src_pixels_u, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_pixels_v, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(tmp_pixels_u, dst_width * dst_height);
+  align_buffer_page_end(tmp_pixels_v, dst_width * dst_height);
+  align_buffer_page_end(dst_pixels_uv_opt, dst_width * 2 * dst_height);
+  align_buffer_page_end(dst_pixels_uv_c, dst_width * 2 * dst_height);
+
+  MemRandomize(src_pixels_u, benchmark_width_ * benchmark_height_);
+  MemRandomize(src_pixels_v, benchmark_width_ * benchmark_height_);
+  MemRandomize(tmp_pixels_u, dst_width * dst_height);
+  MemRandomize(tmp_pixels_v, dst_width * dst_height);
+  MemRandomize(dst_pixels_uv_opt, dst_width * 2 * dst_height);
+  MemRandomize(dst_pixels_uv_c, dst_width * 2 * dst_height);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  HalfMergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v,
+                   benchmark_width_, dst_pixels_uv_c, dst_width * 2,
+                   benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    HalfMergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v,
+                     benchmark_width_, dst_pixels_uv_opt, dst_width * 2,
+                     benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < dst_width * 2 * dst_height; ++i) {
+    EXPECT_EQ(dst_pixels_uv_c[i], dst_pixels_uv_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
+  free_aligned_buffer_page_end(tmp_pixels_u);
+  free_aligned_buffer_page_end(tmp_pixels_v);
+  free_aligned_buffer_page_end(dst_pixels_uv_opt);
+  free_aligned_buffer_page_end(dst_pixels_uv_c);
+}
+
+TEST_F(LibYUVPlanarTest, NV12Copy) {
+  const int halfwidth = (benchmark_width_ + 1) >> 1;
+  const int halfheight = (benchmark_height_ + 1) >> 1;
+  align_buffer_page_end(src_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_uv, halfwidth * 2 * halfheight);
+  align_buffer_page_end(dst_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_uv, halfwidth * 2 * halfheight);
+
+  MemRandomize(src_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(src_uv, halfwidth * 2 * halfheight);
+  MemRandomize(dst_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(dst_uv, halfwidth * 2 * halfheight);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    NV12Copy(src_y, benchmark_width_, src_uv, halfwidth * 2, dst_y,
+             benchmark_width_, dst_uv, halfwidth * 2, benchmark_width_,
+             benchmark_height_);
+  }
+
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_EQ(src_y[i], dst_y[i]);
+  }
+  for (int i = 0; i < halfwidth * 2 * halfheight; ++i) {
+    EXPECT_EQ(src_uv[i], dst_uv[i]);
+  }
+
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_uv);
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVPlanarTest, NV21Copy) {
+  const int halfwidth = (benchmark_width_ + 1) >> 1;
+  const int halfheight = (benchmark_height_ + 1) >> 1;
+  align_buffer_page_end(src_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_vu, halfwidth * 2 * halfheight);
+  align_buffer_page_end(dst_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_vu, halfwidth * 2 * halfheight);
+
+  MemRandomize(src_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(src_vu, halfwidth * 2 * halfheight);
+  MemRandomize(dst_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(dst_vu, halfwidth * 2 * halfheight);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    NV21Copy(src_y, benchmark_width_, src_vu, halfwidth * 2, dst_y,
+             benchmark_width_, dst_vu, halfwidth * 2, benchmark_width_,
+             benchmark_height_);
+  }
+
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_EQ(src_y[i], dst_y[i]);
+  }
+  for (int i = 0; i < halfwidth * 2 * halfheight; ++i) {
+    EXPECT_EQ(src_vu[i], dst_vu[i]);
+  }
+
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_vu);
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/rotate_argb_test.cc b/media/libyuv/unit_test/rotate_argb_test.cc
index 9c83c35630..74952c4e60 100644
--- a/media/libyuv/unit_test/rotate_argb_test.cc
+++ b/media/libyuv/unit_test/rotate_argb_test.cc
@@ -10,14 +10,16 @@
 
 #include <stdlib.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/rotate_argb.h"
-#include "../unit_test/unit_test.h"
 
 namespace libyuv {
 
-void TestRotateBpp(int src_width, int src_height,
-                   int dst_width, int dst_height,
+void TestRotateBpp(int src_width,
+                   int src_height,
+                   int dst_width,
+                   int dst_height,
                    libyuv::RotationMode mode,
                    int benchmark_iterations,
                    int disable_cpu_flags,
@@ -51,26 +53,22 @@ void TestRotateBpp(int src_width, int src_height,
 
   if (kBpp == 1) {
     MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
-    RotatePlane(src_argb, src_stride_argb,
-                dst_argb_c, dst_stride_argb,
+    RotatePlane(src_argb, src_stride_argb, dst_argb_c, dst_stride_argb,
                 src_width, src_height, mode);
 
     MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
     for (int i = 0; i < benchmark_iterations; ++i) {
-      RotatePlane(src_argb, src_stride_argb,
-                  dst_argb_opt, dst_stride_argb,
+      RotatePlane(src_argb, src_stride_argb, dst_argb_opt, dst_stride_argb,
                   src_width, src_height, mode);
     }
   } else if (kBpp == 4) {
     MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
-    ARGBRotate(src_argb, src_stride_argb,
-               dst_argb_c, dst_stride_argb,
+    ARGBRotate(src_argb, src_stride_argb, dst_argb_c, dst_stride_argb,
                src_width, src_height, mode);
 
     MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
     for (int i = 0; i < benchmark_iterations; ++i) {
-      ARGBRotate(src_argb, src_stride_argb,
-                 dst_argb_opt, dst_stride_argb,
+      ARGBRotate(src_argb, src_stride_argb, dst_argb_opt, dst_stride_argb,
                  src_width, src_height, mode);
     }
   }
@@ -85,112 +83,252 @@ void TestRotateBpp(int src_width, int src_height,
   free_aligned_buffer_page_end(src_argb);
 }
 
-static void ARGBTestRotate(int src_width, int src_height,
-                           int dst_width, int dst_height,
+static void ARGBTestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
                            libyuv::RotationMode mode,
                            int benchmark_iterations,
                            int disable_cpu_flags,
                            int benchmark_cpu_info) {
-  TestRotateBpp(src_width, src_height,
-                dst_width, dst_height,
-                mode, benchmark_iterations,
-                disable_cpu_flags, benchmark_cpu_info, 4);
+  TestRotateBpp(src_width, src_height, dst_width, dst_height, mode,
+                benchmark_iterations, disable_cpu_flags, benchmark_cpu_info, 4);
 }
 
 TEST_F(LibYUVRotateTest, ARGBRotate0_Opt) {
-  ARGBTestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate0, benchmark_iterations_,
+  ARGBTestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, ARGBRotate90_Opt) {
-  ARGBTestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate90, benchmark_iterations_,
+  ARGBTestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, ARGBRotate180_Opt) {
-  ARGBTestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate180, benchmark_iterations_,
+  ARGBTestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, ARGBRotate270_Opt) {
-  ARGBTestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate270, benchmark_iterations_,
+  ARGBTestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
-static void TestRotatePlane(int src_width, int src_height,
-                            int dst_width, int dst_height,
+static void TestRotatePlane(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
                             libyuv::RotationMode mode,
                             int benchmark_iterations,
                             int disable_cpu_flags,
                             int benchmark_cpu_info) {
-  TestRotateBpp(src_width, src_height,
-                dst_width, dst_height,
-                mode, benchmark_iterations,
-                disable_cpu_flags, benchmark_cpu_info, 1);
+  TestRotateBpp(src_width, src_height, dst_width, dst_height, mode,
+                benchmark_iterations, disable_cpu_flags, benchmark_cpu_info, 1);
 }
 
 TEST_F(LibYUVRotateTest, RotatePlane0_Opt) {
-  TestRotatePlane(benchmark_width_, benchmark_height_,
-                  benchmark_width_, benchmark_height_,
-                  kRotate0, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_, benchmark_height_, benchmark_width_,
+                  benchmark_height_, kRotate0, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, RotatePlane90_Opt) {
-  TestRotatePlane(benchmark_width_, benchmark_height_,
-                  benchmark_height_, benchmark_width_,
-                  kRotate90, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_, benchmark_height_, benchmark_height_,
+                  benchmark_width_, kRotate90, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, RotatePlane180_Opt) {
-  TestRotatePlane(benchmark_width_, benchmark_height_,
-                  benchmark_width_, benchmark_height_,
-                  kRotate180, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_, benchmark_height_, benchmark_width_,
+                  benchmark_height_, kRotate180, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, RotatePlane270_Opt) {
-  TestRotatePlane(benchmark_width_, benchmark_height_,
-                  benchmark_height_, benchmark_width_,
-                  kRotate270, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_, benchmark_height_, benchmark_height_,
+                  benchmark_width_, kRotate270, benchmark_iterations_,
+                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane0_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_width_ - 3, benchmark_height_ - 1,
-                  kRotate0, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
+                  benchmark_iterations_, disable_cpu_flags_,
+                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane90_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_height_ - 1, benchmark_width_ - 3,
-                  kRotate90, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
+                  benchmark_iterations_, disable_cpu_flags_,
+                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane180_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_width_ - 3, benchmark_height_ - 1,
-                  kRotate180, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
+                  benchmark_iterations_, disable_cpu_flags_,
+                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane270_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_height_ - 1, benchmark_width_ - 3,
-                  kRotate270, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
+                  benchmark_iterations_, disable_cpu_flags_,
+                  benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane90_TestStride) {
+  int argb_plane_size = benchmark_width_ * 4 * abs(benchmark_height_);
+
+  align_buffer_page_end(src_argb, argb_plane_size);
+  align_buffer_page_end(dst_argb, argb_plane_size);
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          benchmark_width_ * 4, benchmark_width_,
+                          benchmark_height_, kRotate0));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                          benchmark_width_ * 4 - 1, benchmark_width_ - 1,
+                          benchmark_height_, kRotate0));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          benchmark_width_ * 4, benchmark_width_,
+                          benchmark_height_, kRotate180));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                          benchmark_width_ * 4 - 1, benchmark_width_ - 1,
+                          benchmark_height_, kRotate180));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          abs(benchmark_height_) * 4, benchmark_width_,
+                          benchmark_height_, kRotate90));
+
+  EXPECT_EQ(-1, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                           abs(benchmark_height_) * 4, benchmark_width_ - 1,
+                           benchmark_height_, kRotate90));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          abs(benchmark_height_) * 4, benchmark_width_,
+                          benchmark_height_, kRotate270));
+
+  EXPECT_EQ(-1, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                           abs(benchmark_height_) * 4, benchmark_width_ - 1,
+                           benchmark_height_, kRotate270));
+
+  free_aligned_buffer_page_end(dst_argb);
+  free_aligned_buffer_page_end(src_argb);
+}
+
+static void TestRotatePlane_16(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               libyuv::RotationMode mode,
+                               int benchmark_iterations,
+                               int disable_cpu_flags,
+                               int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height < 1) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_stride = src_width;
+  int src_plane_size = src_stride * abs(src_height);
+  align_buffer_page_end_16(src, src_plane_size);
+  for (int i = 0; i < src_plane_size; ++i) {
+    src[i] = fastrand() & 0xff;
+  }
+
+  int dst_stride = dst_width;
+  int dst_plane_size = dst_stride * dst_height;
+  align_buffer_page_end_16(dst_c, dst_plane_size);
+  align_buffer_page_end_16(dst_opt, dst_plane_size);
+  memset(dst_c, 2, dst_plane_size);
+  memset(dst_opt, 3, dst_plane_size);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  RotatePlane_16(src, src_stride, dst_c, dst_stride, src_width, src_height,
+                 mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    RotatePlane_16(src, src_stride, dst_opt, dst_stride, src_width, src_height,
+                   mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_c);
+  free_aligned_buffer_page_end_16(dst_opt);
+  free_aligned_buffer_page_end_16(src);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane0_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_width_,
+                     benchmark_height_, kRotate0, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane90_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_height_,
+                     benchmark_width_, kRotate90, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane180_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_width_,
+                     benchmark_height_, kRotate180, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane270_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_height_,
+                     benchmark_width_, kRotate270, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane0_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane90_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane180_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane270_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
 }
 
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/rotate_test.cc b/media/libyuv/unit_test/rotate_test.cc
index 07e2f73a03..e8d2ca164e 100644
--- a/media/libyuv/unit_test/rotate_test.cc
+++ b/media/libyuv/unit_test/rotate_test.cc
@@ -10,17 +10,26 @@
 
 #include <stdlib.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/rotate.h"
-#include "../unit_test/unit_test.h"
+
+#ifdef ENABLE_ROW_TESTS
+#include "libyuv/rotate_row.h"
+#endif
 
 namespace libyuv {
 
-static void I420TestRotate(int src_width, int src_height,
-                           int dst_width, int dst_height,
+#define SUBSAMPLE(v, a) ((((v) + (a)-1)) / (a))
+
+static void I420TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
                            libyuv::RotationMode mode,
                            int benchmark_iterations,
-                           int disable_cpu_flags, int benchmark_cpu_info) {
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
   if (src_width < 1) {
     src_width = 1;
   }
@@ -50,26 +59,21 @@ static void I420TestRotate(int src_width, int src_height,
   memset(dst_i420_opt, 3, dst_i420_size);
 
   MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
-  I420Rotate(src_i420, src_width,
-             src_i420 + src_i420_y_size, (src_width + 1) / 2,
-             src_i420 + src_i420_y_size + src_i420_uv_size, (src_width + 1) / 2,
-             dst_i420_c, dst_width,
+  I420Rotate(src_i420, src_width, src_i420 + src_i420_y_size,
+             (src_width + 1) / 2, src_i420 + src_i420_y_size + src_i420_uv_size,
+             (src_width + 1) / 2, dst_i420_c, dst_width,
              dst_i420_c + dst_i420_y_size, (dst_width + 1) / 2,
              dst_i420_c + dst_i420_y_size + dst_i420_uv_size,
-               (dst_width + 1) / 2,
-             src_width, src_height, mode);
+             (dst_width + 1) / 2, src_width, src_height, mode);
 
   MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
   for (int i = 0; i < benchmark_iterations; ++i) {
-    I420Rotate(src_i420, src_width,
-               src_i420 + src_i420_y_size, (src_width + 1) / 2,
-               src_i420 + src_i420_y_size + src_i420_uv_size,
-                 (src_width + 1) / 2,
-               dst_i420_opt, dst_width,
-               dst_i420_opt + dst_i420_y_size, (dst_width + 1) / 2,
-               dst_i420_opt + dst_i420_y_size + dst_i420_uv_size,
-                 (dst_width + 1) / 2,
-               src_width, src_height, mode);
+    I420Rotate(
+        src_i420, src_width, src_i420 + src_i420_y_size, (src_width + 1) / 2,
+        src_i420 + src_i420_y_size + src_i420_uv_size, (src_width + 1) / 2,
+        dst_i420_opt, dst_width, dst_i420_opt + dst_i420_y_size,
+        (dst_width + 1) / 2, dst_i420_opt + dst_i420_y_size + dst_i420_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
   }
 
   // Rotation should be exact.
@@ -83,30 +87,26 @@ static void I420TestRotate(int src_width, int src_height,
 }
 
 TEST_F(LibYUVRotateTest, I420Rotate0_Opt) {
-  I420TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate0, benchmark_iterations_,
+  I420TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, I420Rotate90_Opt) {
-  I420TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate90, benchmark_iterations_,
+  I420TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, I420Rotate180_Opt) {
-  I420TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate180, benchmark_iterations_,
+  I420TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, I420Rotate270_Opt) {
-  I420TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate270, benchmark_iterations_,
+  I420TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
@@ -114,38 +114,246 @@ TEST_F(LibYUVRotateTest, I420Rotate270_Opt) {
 // Odd width tests work but disabled because they use C code and can be
 // tested by passing an odd width command line or environment variable.
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate0_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1,
-                 kRotate0, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate90_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3,
-                 kRotate90, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate180_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1,
-                 kRotate180, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate270_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3,
-                 kRotate270, benchmark_iterations_,
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
+}
+
+static void I422TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i422_y_size = src_width * Abs(src_height);
+  int src_i422_uv_size = ((src_width + 1) / 2) * Abs(src_height);
+  int src_i422_size = src_i422_y_size + src_i422_uv_size * 2;
+  align_buffer_page_end(src_i422, src_i422_size);
+  for (int i = 0; i < src_i422_size; ++i) {
+    src_i422[i] = fastrand() & 0xff;
+  }
+
+  int dst_i422_y_size = dst_width * dst_height;
+  int dst_i422_uv_size = ((dst_width + 1) / 2) * dst_height;
+  int dst_i422_size = dst_i422_y_size + dst_i422_uv_size * 2;
+  align_buffer_page_end(dst_i422_c, dst_i422_size);
+  align_buffer_page_end(dst_i422_opt, dst_i422_size);
+  memset(dst_i422_c, 2, dst_i422_size);
+  memset(dst_i422_opt, 3, dst_i422_size);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I422Rotate(src_i422, src_width, src_i422 + src_i422_y_size,
+             (src_width + 1) / 2, src_i422 + src_i422_y_size + src_i422_uv_size,
+             (src_width + 1) / 2, dst_i422_c, dst_width,
+             dst_i422_c + dst_i422_y_size, (dst_width + 1) / 2,
+             dst_i422_c + dst_i422_y_size + dst_i422_uv_size,
+             (dst_width + 1) / 2, src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I422Rotate(
+        src_i422, src_width, src_i422 + src_i422_y_size, (src_width + 1) / 2,
+        src_i422 + src_i422_y_size + src_i422_uv_size, (src_width + 1) / 2,
+        dst_i422_opt, dst_width, dst_i422_opt + dst_i422_y_size,
+        (dst_width + 1) / 2, dst_i422_opt + dst_i422_y_size + dst_i422_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i422_size; ++i) {
+    EXPECT_EQ(dst_i422_c[i], dst_i422_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(dst_i422_c);
+  free_aligned_buffer_page_end(dst_i422_opt);
+  free_aligned_buffer_page_end(src_i422);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate0_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate90_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate180_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate270_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+static void I444TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i444_y_size = src_width * Abs(src_height);
+  int src_i444_uv_size = src_width * Abs(src_height);
+  int src_i444_size = src_i444_y_size + src_i444_uv_size * 2;
+  align_buffer_page_end(src_i444, src_i444_size);
+  for (int i = 0; i < src_i444_size; ++i) {
+    src_i444[i] = fastrand() & 0xff;
+  }
+
+  int dst_i444_y_size = dst_width * dst_height;
+  int dst_i444_uv_size = dst_width * dst_height;
+  int dst_i444_size = dst_i444_y_size + dst_i444_uv_size * 2;
+  align_buffer_page_end(dst_i444_c, dst_i444_size);
+  align_buffer_page_end(dst_i444_opt, dst_i444_size);
+  memset(dst_i444_c, 2, dst_i444_size);
+  memset(dst_i444_opt, 3, dst_i444_size);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I444Rotate(src_i444, src_width, src_i444 + src_i444_y_size, src_width,
+             src_i444 + src_i444_y_size + src_i444_uv_size, src_width,
+             dst_i444_c, dst_width, dst_i444_c + dst_i444_y_size, dst_width,
+             dst_i444_c + dst_i444_y_size + dst_i444_uv_size, dst_width,
+             src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I444Rotate(src_i444, src_width, src_i444 + src_i444_y_size, src_width,
+               src_i444 + src_i444_y_size + src_i444_uv_size, src_width,
+               dst_i444_opt, dst_width, dst_i444_opt + dst_i444_y_size,
+               dst_width, dst_i444_opt + dst_i444_y_size + dst_i444_uv_size,
+               dst_width, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i444_size; ++i) {
+    EXPECT_EQ(dst_i444_c[i], dst_i444_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(dst_i444_c);
+  free_aligned_buffer_page_end(dst_i444_opt);
+  free_aligned_buffer_page_end(src_i444);
+}
+
+TEST_F(LibYUVRotateTest, I444Rotate0_Opt) {
+  I444TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
-static void NV12TestRotate(int src_width, int src_height,
-                           int dst_width, int dst_height,
+TEST_F(LibYUVRotateTest, I444Rotate90_Opt) {
+  I444TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I444Rotate180_Opt) {
+  I444TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I444Rotate270_Opt) {
+  I444TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+// TODO(fbarchard): Remove odd width tests.
+// Odd width tests work but disabled because they use C code and can be
+// tested by passing an odd width command line or environment variable.
+TEST_F(LibYUVRotateTest, DISABLED_I444Rotate0_Odd) {
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, DISABLED_I444Rotate90_Odd) {
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, DISABLED_I444Rotate180_Odd) {
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, DISABLED_I444Rotate270_Odd) {
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
+}
+
+static void NV12TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
                            libyuv::RotationMode mode,
                            int benchmark_iterations,
-                           int disable_cpu_flags, int benchmark_cpu_info) {
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
   if (src_width < 1) {
     src_width = 1;
   }
@@ -176,23 +384,19 @@ static void NV12TestRotate(int src_width, int src_height,
   memset(dst_i420_opt, 3, dst_i420_size);
 
   MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
-  NV12ToI420Rotate(src_nv12, src_width,
-                   src_nv12 + src_nv12_y_size, (src_width + 1) & ~1,
-                   dst_i420_c, dst_width,
+  NV12ToI420Rotate(src_nv12, src_width, src_nv12 + src_nv12_y_size,
+                   (src_width + 1) & ~1, dst_i420_c, dst_width,
                    dst_i420_c + dst_i420_y_size, (dst_width + 1) / 2,
                    dst_i420_c + dst_i420_y_size + dst_i420_uv_size,
-                     (dst_width + 1) / 2,
-                   src_width, src_height, mode);
+                   (dst_width + 1) / 2, src_width, src_height, mode);
 
   MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
   for (int i = 0; i < benchmark_iterations; ++i) {
-    NV12ToI420Rotate(src_nv12, src_width,
-                     src_nv12 + src_nv12_y_size, (src_width + 1) & ~1,
-                     dst_i420_opt, dst_width,
+    NV12ToI420Rotate(src_nv12, src_width, src_nv12 + src_nv12_y_size,
+                     (src_width + 1) & ~1, dst_i420_opt, dst_width,
                      dst_i420_opt + dst_i420_y_size, (dst_width + 1) / 2,
                      dst_i420_opt + dst_i420_y_size + dst_i420_uv_size,
-                       (dst_width + 1) / 2,
-                     src_width, src_height, mode);
+                     (dst_width + 1) / 2, src_width, src_height, mode);
   }
 
   // Rotation should be exact.
@@ -206,91 +410,499 @@ static void NV12TestRotate(int src_width, int src_height,
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate0_Opt) {
-  NV12TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate0, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate90_Opt) {
-  NV12TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate90, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate180_Opt) {
-  NV12TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate180, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate270_Opt) {
-  NV12TestRotate(benchmark_width_, benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate270, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate0_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1,
-                 kRotate0, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate90_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3,
-                 kRotate90, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate180_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1,
-                 kRotate180, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate270_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3,
-                 kRotate270, benchmark_iterations_,
-                 disable_cpu_flags_, benchmark_cpu_info_);
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
+                 benchmark_iterations_, disable_cpu_flags_,
+                 benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate0_Invert) {
-  NV12TestRotate(benchmark_width_, -benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate0, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, -benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate90_Invert) {
-  NV12TestRotate(benchmark_width_, -benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate90, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, -benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate180_Invert) {
-  NV12TestRotate(benchmark_width_, -benchmark_height_,
-                 benchmark_width_, benchmark_height_,
-                 kRotate180, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, -benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, NV12Rotate270_Invert) {
-  NV12TestRotate(benchmark_width_, -benchmark_height_,
-                 benchmark_height_, benchmark_width_,
-                 kRotate270, benchmark_iterations_,
+  NV12TestRotate(benchmark_width_, -benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+// Test Android 420 to I420 Rotate
+#define TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X,          \
+                        SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                        W1280, N, NEG, OFF, PN, OFF_U, OFF_V, ROT)            \
+  TEST_F(LibYUVRotateTest,                                                    \
+         SRC_FMT_PLANAR##To##FMT_PLANAR##Rotate##ROT##To##PN##N) {            \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kSizeUV =                                                       \
+        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                      \
+    align_buffer_page_end(src_uv,                                             \
+                          kSizeUV*((PIXEL_STRIDE == 3) ? 3 : 2) + OFF);       \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight);                          \
+    align_buffer_page_end(dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X) *             \
+                                       SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    align_buffer_page_end(dst_v_c, SUBSAMPLE(kWidth, SUBSAMP_X) *             \
+                                       SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                        \
+    align_buffer_page_end(dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *           \
+                                         SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    align_buffer_page_end(dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *           \
+                                         SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    uint8_t* src_u = src_uv + OFF_U;                                          \
+    uint8_t* src_v = src_uv + (PIXEL_STRIDE == 1 ? kSizeUV : OFF_V);          \
+    int src_stride_uv = SUBSAMPLE(kWidth, SUBSAMP_X) * PIXEL_STRIDE;          \
+    for (int i = 0; i < kHeight; ++i)                                         \
+      for (int j = 0; j < kWidth; ++j)                                        \
+        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                    \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {             \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {            \
+        src_u[(i * src_stride_uv) + j * PIXEL_STRIDE + OFF] =                 \
+            (fastrand() & 0xff);                                              \
+        src_v[(i * src_stride_uv) + j * PIXEL_STRIDE + OFF] =                 \
+            (fastrand() & 0xff);                                              \
+      }                                                                       \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight);                                      \
+    memset(dst_u_c, 2,                                                        \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_v_c, 3,                                                        \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_y_opt, 101, kWidth* kHeight);                                  \
+    memset(dst_u_opt, 102,                                                    \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_v_opt, 103,                                                    \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    SRC_FMT_PLANAR##To##FMT_PLANAR##Rotate(                                   \
+        src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X),   \
+        src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), PIXEL_STRIDE, dst_y_c, \
+        kWidth, dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X), dst_v_c,               \
+        SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight,                    \
+        (libyuv::RotationMode)ROT);                                           \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      SRC_FMT_PLANAR##To##FMT_PLANAR##Rotate(                                 \
+          src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), \
+          src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), PIXEL_STRIDE,        \
+          dst_y_opt, kWidth, dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X),         \
+          dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight,       \
+          (libyuv::RotationMode)ROT);                                         \
+    }                                                                         \
+    for (int i = 0; i < kHeight; ++i) {                                       \
+      for (int j = 0; j < kWidth; ++j) {                                      \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);        \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
+        EXPECT_EQ(dst_u_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_u_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
+        EXPECT_EQ(dst_v_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_v_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
+      }                                                                       \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_u_c);                                    \
+    free_aligned_buffer_page_end(dst_v_c);                                    \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_u_opt);                                  \
+    free_aligned_buffer_page_end(dst_v_opt);                                  \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_uv);                                     \
+  }
+
+#define TESTAPLANARTOP(SRC_FMT_PLANAR, PN, PIXEL_STRIDE, OFF_U, OFF_V,         \
+                       SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X,    \
+                       SUBSAMP_Y)                                              \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_ + 1,      \
+                  _Any, +, 0, PN, OFF_U, OFF_V, 0)                             \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_,          \
+                  _Unaligned, +, 2, PN, OFF_U, OFF_V, 0)                       \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, \
+                  -, 0, PN, OFF_U, OFF_V, 0)                                   \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, \
+                  0, PN, OFF_U, OFF_V, 0)                                      \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, \
+                  0, PN, OFF_U, OFF_V, 180)
+
+TESTAPLANARTOP(Android420, I420, 1, 0, 0, 2, 2, I420, 2, 2)
+TESTAPLANARTOP(Android420, NV12, 2, 0, 1, 2, 2, I420, 2, 2)
+TESTAPLANARTOP(Android420, NV21, 2, 1, 0, 2, 2, I420, 2, 2)
+#undef TESTAPLANARTOP
+#undef TESTAPLANARTOPI
+
+static void I010TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i010_y_size = src_width * Abs(src_height);
+  int src_i010_uv_size = ((src_width + 1) / 2) * ((Abs(src_height) + 1) / 2);
+  int src_i010_size = src_i010_y_size + src_i010_uv_size * 2;
+  align_buffer_page_end_16(src_i010, src_i010_size);
+  for (int i = 0; i < src_i010_size; ++i) {
+    src_i010[i] = fastrand() & 0x3ff;
+  }
+
+  int dst_i010_y_size = dst_width * dst_height;
+  int dst_i010_uv_size = ((dst_width + 1) / 2) * ((dst_height + 1) / 2);
+  int dst_i010_size = dst_i010_y_size + dst_i010_uv_size * 2;
+  align_buffer_page_end_16(dst_i010_c, dst_i010_size);
+  align_buffer_page_end_16(dst_i010_opt, dst_i010_size);
+  memset(dst_i010_c, 2, dst_i010_size * 2);
+  memset(dst_i010_opt, 3, dst_i010_size * 2);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I010Rotate(src_i010, src_width, src_i010 + src_i010_y_size,
+             (src_width + 1) / 2, src_i010 + src_i010_y_size + src_i010_uv_size,
+             (src_width + 1) / 2, dst_i010_c, dst_width,
+             dst_i010_c + dst_i010_y_size, (dst_width + 1) / 2,
+             dst_i010_c + dst_i010_y_size + dst_i010_uv_size,
+             (dst_width + 1) / 2, src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I010Rotate(
+        src_i010, src_width, src_i010 + src_i010_y_size, (src_width + 1) / 2,
+        src_i010 + src_i010_y_size + src_i010_uv_size, (src_width + 1) / 2,
+        dst_i010_opt, dst_width, dst_i010_opt + dst_i010_y_size,
+        (dst_width + 1) / 2, dst_i010_opt + dst_i010_y_size + dst_i010_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i010_size; ++i) {
+    EXPECT_EQ(dst_i010_c[i], dst_i010_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_i010_c);
+  free_aligned_buffer_page_end_16(dst_i010_opt);
+  free_aligned_buffer_page_end_16(src_i010);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate0_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate90_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate180_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate270_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+static void I210TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i210_y_size = src_width * Abs(src_height);
+  int src_i210_uv_size = ((src_width + 1) / 2) * Abs(src_height);
+  int src_i210_size = src_i210_y_size + src_i210_uv_size * 2;
+  align_buffer_page_end_16(src_i210, src_i210_size);
+  for (int i = 0; i < src_i210_size; ++i) {
+    src_i210[i] = fastrand() & 0x3ff;
+  }
+
+  int dst_i210_y_size = dst_width * dst_height;
+  int dst_i210_uv_size = ((dst_width + 1) / 2) * dst_height;
+  int dst_i210_size = dst_i210_y_size + dst_i210_uv_size * 2;
+  align_buffer_page_end_16(dst_i210_c, dst_i210_size);
+  align_buffer_page_end_16(dst_i210_opt, dst_i210_size);
+  memset(dst_i210_c, 2, dst_i210_size * 2);
+  memset(dst_i210_opt, 3, dst_i210_size * 2);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I210Rotate(src_i210, src_width, src_i210 + src_i210_y_size,
+             (src_width + 1) / 2, src_i210 + src_i210_y_size + src_i210_uv_size,
+             (src_width + 1) / 2, dst_i210_c, dst_width,
+             dst_i210_c + dst_i210_y_size, (dst_width + 1) / 2,
+             dst_i210_c + dst_i210_y_size + dst_i210_uv_size,
+             (dst_width + 1) / 2, src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I210Rotate(
+        src_i210, src_width, src_i210 + src_i210_y_size, (src_width + 1) / 2,
+        src_i210 + src_i210_y_size + src_i210_uv_size, (src_width + 1) / 2,
+        dst_i210_opt, dst_width, dst_i210_opt + dst_i210_y_size,
+        (dst_width + 1) / 2, dst_i210_opt + dst_i210_y_size + dst_i210_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i210_size; ++i) {
+    EXPECT_EQ(dst_i210_c[i], dst_i210_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_i210_c);
+  free_aligned_buffer_page_end_16(dst_i210_opt);
+  free_aligned_buffer_page_end_16(src_i210);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate0_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate90_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
+TEST_F(LibYUVRotateTest, I210Rotate180_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate270_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+static void I410TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i410_y_size = src_width * Abs(src_height);
+  int src_i410_uv_size = src_width * Abs(src_height);
+  int src_i410_size = src_i410_y_size + src_i410_uv_size * 2;
+  align_buffer_page_end_16(src_i410, src_i410_size);
+  for (int i = 0; i < src_i410_size; ++i) {
+    src_i410[i] = fastrand() & 0x3ff;
+  }
 
+  int dst_i410_y_size = dst_width * dst_height;
+  int dst_i410_uv_size = dst_width * dst_height;
+  int dst_i410_size = dst_i410_y_size + dst_i410_uv_size * 2;
+  align_buffer_page_end_16(dst_i410_c, dst_i410_size);
+  align_buffer_page_end_16(dst_i410_opt, dst_i410_size);
+  memset(dst_i410_c, 2, dst_i410_size * 2);
+  memset(dst_i410_opt, 3, dst_i410_size * 2);
 
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I410Rotate(src_i410, src_width, src_i410 + src_i410_y_size, src_width,
+             src_i410 + src_i410_y_size + src_i410_uv_size, src_width,
+             dst_i410_c, dst_width, dst_i410_c + dst_i410_y_size, dst_width,
+             dst_i410_c + dst_i410_y_size + dst_i410_uv_size, dst_width,
+             src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I410Rotate(src_i410, src_width, src_i410 + src_i410_y_size, src_width,
+               src_i410 + src_i410_y_size + src_i410_uv_size, src_width,
+               dst_i410_opt, dst_width, dst_i410_opt + dst_i410_y_size,
+               dst_width, dst_i410_opt + dst_i410_y_size + dst_i410_uv_size,
+               dst_width, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i410_size; ++i) {
+    EXPECT_EQ(dst_i410_c[i], dst_i410_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_i410_c);
+  free_aligned_buffer_page_end_16(dst_i410_opt);
+  free_aligned_buffer_page_end_16(src_i410);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate0_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate90_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate180_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate270_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+#if defined(ENABLE_ROW_TESTS)
+
+TEST_F(LibYUVRotateTest, Transpose4x4) {
+  // dst width and height
+  const int width = ((benchmark_width_ * benchmark_height_ + 3) / 4 + 3) & ~3;
+  const int height = 4;
+  align_buffer_page_end(src_pixels, height * width * 4);
+  align_buffer_page_end(dst_pixels_c, width * height * 4);
+  align_buffer_page_end(dst_pixels_opt, width * height * 4);
+
+  MemRandomize(src_pixels, height * width * 4);
+  memset(dst_pixels_c, 1, width * height * 4);
+  memset(dst_pixels_opt, 1, width * height * 4);
+
+  Transpose4x4_32_C((const uint8_t*)src_pixels, height * 4,
+                    (uint8_t*)dst_pixels_c, width * 4, width);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+#if defined(__aarch64__)
+    if (TestCpuFlag(kCpuHasNEON)) {
+      Transpose4x4_32_NEON((const uint8_t*)src_pixels, height * 4,
+                           (uint8_t*)dst_pixels_opt, width * 4, width);
+    } else {
+      Transpose4x4_32_C((const uint8_t*)src_pixels, height * 4,
+                        (uint8_t*)dst_pixels_opt, width * 4, width);
+    }
+#else
+    Transpose4x4_32_C((const uint8_t*)src_pixels, height * 4,
+                      (uint8_t*)dst_pixels_opt, width * 4, width);
+#endif
+  }
+
+  //  for (int i = 0; i < width * height; ++i) {
+  //    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  //  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_c);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+}
 
+#endif  // ENABLE_ROW_TESTS
 
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/scale_argb_test.cc b/media/libyuv/unit_test/scale_argb_test.cc
index f99782f731..f54a68f11f 100644
--- a/media/libyuv/unit_test/scale_argb_test.cc
+++ b/media/libyuv/unit_test/scale_argb_test.cc
@@ -11,30 +11,40 @@
 #include <stdlib.h>
 #include <time.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/convert_argb.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/scale_argb.h"
 #include "libyuv/video_common.h"
-#include "../unit_test/unit_test.h"
 
 namespace libyuv {
 
 #define STRINGIZE(line) #line
 #define FILELINESTR(file, line) file ":" STRINGIZE(line)
 
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
-static int ARGBTestFilter(int src_width, int src_height,
-                          int dst_width, int dst_height,
-                          FilterMode f, int benchmark_iterations,
-                          int disable_cpu_flags, int benchmark_cpu_info) {
+static int ARGBTestFilter(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          FilterMode f,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info) {
   if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
     return 0;
   }
 
   int i, j;
   const int b = 0;  // 128 to test for padding/stride.
-  int64 src_argb_plane_size = (Abs(src_width) + b * 2) *
-      (Abs(src_height) + b * 2) * 4LL;
+  int64_t src_argb_plane_size =
+      (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2) * 4LL;
   int src_stride_argb = (b * 2 + Abs(src_width)) * 4;
 
   align_buffer_page_end(src_argb, src_argb_plane_size);
@@ -44,7 +54,8 @@ static int ARGBTestFilter(int src_width, int src_height,
   }
   MemRandomize(src_argb, src_argb_plane_size);
 
-  int64 dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4LL;
+  int64_t dst_argb_plane_size =
+      (dst_width + b * 2) * (dst_height + b * 2) * 4LL;
   int dst_stride_argb = (b * 2 + dst_width) * 4;
 
   align_buffer_page_end(dst_argb_c, dst_argb_plane_size);
@@ -59,21 +70,18 @@ static int ARGBTestFilter(int src_width, int src_height,
   // Warm up both versions for consistent benchmarks.
   MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
   ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
-            src_width, src_height,
-            dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb,
-            dst_width, dst_height, f);
+            src_width, src_height, dst_argb_c + (dst_stride_argb * b) + b * 4,
+            dst_stride_argb, dst_width, dst_height, f);
   MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
   ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
-            src_width, src_height,
-            dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb,
-            dst_width, dst_height, f);
+            src_width, src_height, dst_argb_opt + (dst_stride_argb * b) + b * 4,
+            dst_stride_argb, dst_width, dst_height, f);
 
   MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
   double c_time = get_time();
   ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
-            src_width, src_height,
-            dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb,
-            dst_width, dst_height, f);
+            src_width, src_height, dst_argb_c + (dst_stride_argb * b) + b * 4,
+            dst_stride_argb, dst_width, dst_height, f);
 
   c_time = (get_time() - c_time);
 
@@ -88,8 +96,8 @@ static int ARGBTestFilter(int src_width, int src_height,
   opt_time = (get_time() - opt_time) / benchmark_iterations;
 
   // Report performance of C vs OPT
-  printf("filter %d - %8d us C - %8d us OPT\n",
-         f, static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
 
   // C version may be a little off from the optimized. Order of
   //  operations may introduce rounding somewhere. So do a difference
@@ -112,13 +120,17 @@ static int ARGBTestFilter(int src_width, int src_height,
   return max_diff;
 }
 
-static const int kTileX = 8;
-static const int kTileY = 8;
-
-static int TileARGBScale(const uint8* src_argb, int src_stride_argb,
-                         int src_width, int src_height,
-                         uint8* dst_argb, int dst_stride_argb,
-                         int dst_width, int dst_height,
+static const int kTileX = 64;
+static const int kTileY = 64;
+
+static int TileARGBScale(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         int src_width,
+                         int src_height,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int dst_width,
+                         int dst_height,
                          FilterMode filtering) {
   for (int y = 0; y < dst_height; y += kTileY) {
     for (int x = 0; x < dst_width; x += kTileX) {
@@ -130,11 +142,9 @@ static int TileARGBScale(const uint8* src_argb, int src_stride_argb,
       if (y + clip_height > dst_height) {
         clip_height = dst_height - y;
       }
-      int r = ARGBScaleClip(src_argb, src_stride_argb,
-                            src_width, src_height,
-                            dst_argb, dst_stride_argb,
-                            dst_width, dst_height,
-                            x, y, clip_width, clip_height, filtering);
+      int r = ARGBScaleClip(src_argb, src_stride_argb, src_width, src_height,
+                            dst_argb, dst_stride_argb, dst_width, dst_height, x,
+                            y, clip_width, clip_height, filtering);
       if (r) {
         return r;
       }
@@ -143,16 +153,19 @@ static int TileARGBScale(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
-static int ARGBClipTestFilter(int src_width, int src_height,
-                              int dst_width, int dst_height,
-                              FilterMode f, int benchmark_iterations) {
+static int ARGBClipTestFilter(int src_width,
+                              int src_height,
+                              int dst_width,
+                              int dst_height,
+                              FilterMode f,
+                              int benchmark_iterations) {
   if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
     return 0;
   }
 
   const int b = 128;
-  int64 src_argb_plane_size = (Abs(src_width) + b * 2) *
-      (Abs(src_height) + b * 2) * 4;
+  int64_t src_argb_plane_size =
+      (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2) * 4;
   int src_stride_argb = (b * 2 + Abs(src_width)) * 4;
 
   align_buffer_page_end(src_argb, src_argb_plane_size);
@@ -162,7 +175,7 @@ static int ARGBClipTestFilter(int src_width, int src_height,
   }
   memset(src_argb, 1, src_argb_plane_size);
 
-  int64 dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4;
+  int64_t dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4;
   int dst_stride_argb = (b * 2 + dst_width) * 4;
 
   int i, j;
@@ -184,9 +197,8 @@ static int ARGBClipTestFilter(int src_width, int src_height,
   // Do full image, no clipping.
   double c_time = get_time();
   ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
-            src_width, src_height,
-            dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb,
-            dst_width, dst_height, f);
+            src_width, src_height, dst_argb_c + (dst_stride_argb * b) + b * 4,
+            dst_stride_argb, dst_width, dst_height, f);
   c_time = (get_time() - c_time);
 
   // Do tiled image, clipping scale to a tile at a time.
@@ -200,8 +212,8 @@ static int ARGBClipTestFilter(int src_width, int src_height,
   opt_time = (get_time() - opt_time) / benchmark_iterations;
 
   // Report performance of Full vs Tiled.
-  printf("filter %d - %8d us Full - %8d us Tiled\n",
-         f, static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+  printf("filter %d - %8d us Full - %8d us Tiled\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
 
   // Compare full scaled image vs tiled image.
   int max_diff = 0;
@@ -226,36 +238,49 @@ static int ARGBClipTestFilter(int src_width, int src_height,
 #define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
 #define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
 
-#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                       \
-    TEST_F(LibYUVScaleTest, ARGBScaleDownBy##name##_##filter) {                \
-      int diff = ARGBTestFilter(SX(benchmark_width_, nom, denom),              \
-                                SX(benchmark_height_, nom, denom),             \
-                                DX(benchmark_width_, nom, denom),              \
-                                DX(benchmark_height_, nom, denom),             \
-                                kFilter##filter, benchmark_iterations_,        \
-                                disable_cpu_flags_, benchmark_cpu_info_);      \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest, ARGBScaleDownClipBy##name##_##filter) {            \
-      int diff = ARGBClipTestFilter(SX(benchmark_width_, nom, denom),          \
-                                    SX(benchmark_height_, nom, denom),         \
-                                    DX(benchmark_width_, nom, denom),          \
-                                    DX(benchmark_height_, nom, denom),         \
-                                    kFilter##filter, benchmark_iterations_);   \
-      EXPECT_LE(diff, max_diff);                                               \
-    }
+#define TEST_FACTOR1(DISABLED_, name, filter, nom, denom, max_diff)          \
+  TEST_F(LibYUVScaleTest, ARGBScaleDownBy##name##_##filter) {                \
+    int diff = ARGBTestFilter(                                               \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
+        benchmark_cpu_info_);                                                \
+    EXPECT_LE(diff, max_diff);                                               \
+  }                                                                          \
+  TEST_F(LibYUVScaleTest, DISABLED_##ARGBScaleDownClipBy##name##_##filter) { \
+    int diff = ARGBClipTestFilter(                                           \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
+        kFilter##filter, benchmark_iterations_);                             \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
 
 // Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
 // filtering is different fixed point implementations for SSSE3, Neon and C.
-#define TEST_FACTOR(name, nom, denom)                                          \
-    TEST_FACTOR1(name, None, nom, denom, 0)                                    \
-    TEST_FACTOR1(name, Linear, nom, denom, 3)                                  \
-    TEST_FACTOR1(name, Bilinear, nom, denom, 3)                                \
-    TEST_FACTOR1(name, Box, nom, denom, 3)
+#ifndef DISABLE_SLOW_TESTS
+#define TEST_FACTOR(name, nom, denom)           \
+  TEST_FACTOR1(, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(, name, Box, nom, denom, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_FACTOR(name, nom, denom)                    \
+  TEST_FACTOR1(DISABLED_, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(DISABLED_, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(DISABLED_, name, Box, nom, denom, 3)
+#else
+#define TEST_FACTOR(name, nom, denom) \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3)
+#endif
+#endif
 
 TEST_FACTOR(2, 1, 2)
 TEST_FACTOR(4, 1, 4)
+#ifndef DISABLE_SLOW_TESTS
 TEST_FACTOR(8, 1, 8)
+#endif
 TEST_FACTOR(3by4, 3, 4)
 TEST_FACTOR(3by8, 3, 8)
 TEST_FACTOR(3, 1, 3)
@@ -264,82 +289,125 @@ TEST_FACTOR(3, 1, 3)
 #undef SX
 #undef DX
 
-#define TEST_SCALETO1(name, width, height, filter, max_diff)                   \
-    TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {           \
-      int diff = ARGBTestFilter(benchmark_width_, benchmark_height_,           \
-                                width, height,                                 \
-                                kFilter##filter, benchmark_iterations_,        \
-                                disable_cpu_flags_, benchmark_cpu_info_);      \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {         \
-      int diff = ARGBTestFilter(width, height,                                 \
-                                Abs(benchmark_width_), Abs(benchmark_height_), \
-                                kFilter##filter, benchmark_iterations_,        \
-                                disable_cpu_flags_, benchmark_cpu_info_);      \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest, name##ClipTo##width##x##height##_##filter) {       \
-      int diff = ARGBClipTestFilter(benchmark_width_, benchmark_height_,       \
-                                    width, height,                             \
-                                    kFilter##filter, benchmark_iterations_);   \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest, name##ClipFrom##width##x##height##_##filter) {     \
-      int diff = ARGBClipTestFilter(width, height,                             \
-                                    Abs(benchmark_width_),                     \
-                                    Abs(benchmark_height_),                    \
-                                    kFilter##filter, benchmark_iterations_);   \
-      EXPECT_LE(diff, max_diff);                                               \
-    }
+#define TEST_SCALETO1(DISABLED_, name, width, height, filter, max_diff)        \
+  TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {             \
+    int diff = ARGBTestFilter(benchmark_width_, benchmark_height_, width,      \
+                              height, kFilter##filter, benchmark_iterations_,  \
+                              disable_cpu_flags_, benchmark_cpu_info_);        \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }                                                                            \
+  TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {           \
+    int diff = ARGBTestFilter(width, height, Abs(benchmark_width_),            \
+                              Abs(benchmark_height_), kFilter##filter,         \
+                              benchmark_iterations_, disable_cpu_flags_,       \
+                              benchmark_cpu_info_);                            \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }                                                                            \
+  TEST_F(LibYUVScaleTest,                                                      \
+         DISABLED_##name##ClipTo##width##x##height##_##filter) {               \
+    int diff =                                                                 \
+        ARGBClipTestFilter(benchmark_width_, benchmark_height_, width, height, \
+                           kFilter##filter, benchmark_iterations_);            \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }                                                                            \
+  TEST_F(LibYUVScaleTest,                                                      \
+         DISABLED_##name##ClipFrom##width##x##height##_##filter) {             \
+    int diff = ARGBClipTestFilter(width, height, Abs(benchmark_width_),        \
+                                  Abs(benchmark_height_), kFilter##filter,     \
+                                  benchmark_iterations_);                      \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }
 
-/// Test scale to a specified size with all 4 filters.
-#define TEST_SCALETO(name, width, height)                                      \
-    TEST_SCALETO1(name, width, height, None, 0)                                \
-    TEST_SCALETO1(name, width, height, Linear, 3)                              \
-    TEST_SCALETO1(name, width, height, Bilinear, 3)
+#ifndef DISABLE_SLOW_TESTS
+// Test scale to a specified size with all 4 filters.
+#define TEST_SCALETO(name, width, height)         \
+  TEST_SCALETO1(, name, width, height, None, 0)   \
+  TEST_SCALETO1(, name, width, height, Linear, 3) \
+  TEST_SCALETO1(, name, width, height, Bilinear, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_SCALETO(name, width, height)                  \
+  TEST_SCALETO1(DISABLED_, name, width, height, None, 0)   \
+  TEST_SCALETO1(DISABLED_, name, width, height, Linear, 3) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3)
+#else
+#define TEST_SCALETO(name, width, height) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3)
+#endif
+#endif
 
 TEST_SCALETO(ARGBScale, 1, 1)
-TEST_SCALETO(ARGBScale, 320, 240)
-TEST_SCALETO(ARGBScale, 352, 288)
 TEST_SCALETO(ARGBScale, 569, 480)
 TEST_SCALETO(ARGBScale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(ARGBScale, 256, 144) /* 128x72 * 2 */
+TEST_SCALETO(ARGBScale, 320, 240)
 TEST_SCALETO(ARGBScale, 1280, 720)
+TEST_SCALETO(ARGBScale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
 #undef TEST_SCALETO1
 #undef TEST_SCALETO
 
+#define TEST_SCALESWAPXY1(name, filter, max_diff)                       \
+  TEST_F(LibYUVScaleTest, name##SwapXY_##filter) {                      \
+    int diff = ARGBTestFilter(benchmark_width_, benchmark_height_,      \
+                              benchmark_height_, benchmark_width_,      \
+                              kFilter##filter, benchmark_iterations_,   \
+                              disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                          \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test scale with swapped width and height with all 3 filters.
+TEST_SCALESWAPXY1(ARGBScale, None, 0)
+TEST_SCALESWAPXY1(ARGBScale, Linear, 0)
+TEST_SCALESWAPXY1(ARGBScale, Bilinear, 0)
+#else
+TEST_SCALESWAPXY1(ARGBScale, Bilinear, 0)
+#endif
+#undef TEST_SCALESWAPXY1
+
 // Scale with YUV conversion to ARGB and clipping.
+// TODO(fbarchard): Add fourcc support.  All 4 ARGB formats is easy to support.
 LIBYUV_API
-int YUVToARGBScaleReference2(const uint8* src_y, int src_stride_y,
-                             const uint8* src_u, int src_stride_u,
-                             const uint8* src_v, int src_stride_v,
-                             uint32 src_fourcc,
-                             int src_width, int src_height,
-                             uint8* dst_argb, int dst_stride_argb,
-                             uint32 dst_fourcc,
-                             int dst_width, int dst_height,
-                             int clip_x, int clip_y,
-                             int clip_width, int clip_height,
+int YUVToARGBScaleReference2(const uint8_t* src_y,
+                             int src_stride_y,
+                             const uint8_t* src_u,
+                             int src_stride_u,
+                             const uint8_t* src_v,
+                             int src_stride_v,
+                             uint32_t /* src_fourcc */,
+                             int src_width,
+                             int src_height,
+                             uint8_t* dst_argb,
+                             int dst_stride_argb,
+                             uint32_t /* dst_fourcc */,
+                             int dst_width,
+                             int dst_height,
+                             int clip_x,
+                             int clip_y,
+                             int clip_width,
+                             int clip_height,
                              enum FilterMode filtering) {
-  uint8* argb_buffer = static_cast<uint8*>(malloc(src_width * src_height * 4));
+  uint8_t* argb_buffer =
+      static_cast<uint8_t*>(malloc(src_width * src_height * 4));
   int r;
-  I420ToARGB(src_y, src_stride_y,
-             src_u, src_stride_u,
-             src_v, src_stride_v,
-             argb_buffer, src_width * 4,
-             src_width, src_height);
-
-  r = ARGBScaleClip(argb_buffer, src_width * 4,
-                    src_width, src_height,
-                    dst_argb, dst_stride_argb,
-                    dst_width, dst_height,
-                    clip_x, clip_y, clip_width, clip_height,
-                    filtering);
+  I420ToARGB(src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+             argb_buffer, src_width * 4, src_width, src_height);
+
+  r = ARGBScaleClip(argb_buffer, src_width * 4, src_width, src_height, dst_argb,
+                    dst_stride_argb, dst_width, dst_height, clip_x, clip_y,
+                    clip_width, clip_height, filtering);
   free(argb_buffer);
   return r;
 }
 
-static void FillRamp(uint8* buf, int width, int height, int v, int dx, int dy) {
+static void FillRamp(uint8_t* buf,
+                     int width,
+                     int height,
+                     int v,
+                     int dx,
+                     int dy) {
   int rv = v;
   for (int y = 0; y < height; ++y) {
     for (int x = 0; x < width; ++x) {
@@ -360,13 +428,15 @@ static void FillRamp(uint8* buf, int width, int height, int v, int dx, int dy) {
 }
 
 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
-static int YUVToARGBTestFilter(int src_width, int src_height,
-                               int dst_width, int dst_height,
-                               FilterMode f, int benchmark_iterations,
-                               int disable_cpu_flags, int benchmark_cpu_info) {
-  int64 src_y_plane_size = Abs(src_width) * Abs(src_height);
-  int64 src_uv_plane_size = ((Abs(src_width) + 1) / 2) *
-      ((Abs(src_height) + 1) / 2);
+static int YUVToARGBTestFilter(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               FilterMode f,
+                               int benchmark_iterations) {
+  int64_t src_y_plane_size = Abs(src_width) * Abs(src_height);
+  int64_t src_uv_plane_size =
+      ((Abs(src_width) + 1) / 2) * ((Abs(src_height) + 1) / 2);
   int src_stride_y = Abs(src_width);
   int src_stride_uv = (Abs(src_width) + 1) / 2;
 
@@ -374,8 +444,8 @@ static int YUVToARGBTestFilter(int src_width, int src_height,
   align_buffer_page_end(src_u, src_uv_plane_size);
   align_buffer_page_end(src_v, src_uv_plane_size);
 
-  int64 dst_argb_plane_size = (dst_width) * (dst_height) * 4LL;
-  int dst_stride_argb = (dst_width) * 4;
+  int64_t dst_argb_plane_size = (dst_width) * (dst_height)*4LL;
+  int dst_stride_argb = (dst_width)*4;
   align_buffer_page_end(dst_argb_c, dst_argb_plane_size);
   align_buffer_page_end(dst_argb_opt, dst_argb_plane_size);
   if (!dst_argb_c || !dst_argb_opt || !src_y || !src_u || !src_v) {
@@ -390,28 +460,18 @@ static int YUVToARGBTestFilter(int src_width, int src_height,
   memset(dst_argb_c, 2, dst_argb_plane_size);
   memset(dst_argb_opt, 3, dst_argb_plane_size);
 
-  YUVToARGBScaleReference2(src_y, src_stride_y,
-                           src_u, src_stride_uv,
-                           src_v, src_stride_uv,
-                           libyuv::FOURCC_I420,
-                           src_width, src_height,
-                           dst_argb_c, dst_stride_argb,
-                           libyuv::FOURCC_I420,
-                           dst_width, dst_height,
-                           0, 0, dst_width, dst_height,
-                           f);
+  YUVToARGBScaleReference2(src_y, src_stride_y, src_u, src_stride_uv, src_v,
+                           src_stride_uv, libyuv::FOURCC_I420, src_width,
+                           src_height, dst_argb_c, dst_stride_argb,
+                           libyuv::FOURCC_I420, dst_width, dst_height, 0, 0,
+                           dst_width, dst_height, f);
 
   for (int i = 0; i < benchmark_iterations; ++i) {
-    YUVToARGBScaleClip(src_y, src_stride_y,
-                       src_u, src_stride_uv,
-                       src_v, src_stride_uv,
-                       libyuv::FOURCC_I420,
-                       src_width, src_height,
-                       dst_argb_opt, dst_stride_argb,
-                       libyuv::FOURCC_I420,
-                       dst_width, dst_height,
-                       0, 0, dst_width, dst_height,
-                       f);
+    YUVToARGBScaleClip(src_y, src_stride_y, src_u, src_stride_uv, src_v,
+                       src_stride_uv, libyuv::FOURCC_I420, src_width,
+                       src_height, dst_argb_opt, dst_stride_argb,
+                       libyuv::FOURCC_I420, dst_width, dst_height, 0, 0,
+                       dst_width, dst_height, f);
   }
   int max_diff = 0;
   for (int i = 0; i < dst_height; ++i) {
@@ -419,9 +479,7 @@ static int YUVToARGBTestFilter(int src_width, int src_height,
       int abs_diff = Abs(dst_argb_c[(i * dst_stride_argb) + j] -
                          dst_argb_opt[(i * dst_stride_argb) + j]);
       if (abs_diff > max_diff) {
-        printf("error %d at %d,%d c %d opt %d",
-               abs_diff,
-               j, i,
+        printf("error %d at %d,%d c %d opt %d", abs_diff, j, i,
                dst_argb_c[(i * dst_stride_argb) + j],
                dst_argb_opt[(i * dst_stride_argb) + j]);
         EXPECT_LE(abs_diff, 40);
@@ -439,24 +497,92 @@ static int YUVToARGBTestFilter(int src_width, int src_height,
 }
 
 TEST_F(LibYUVScaleTest, YUVToRGBScaleUp) {
-  int diff = YUVToARGBTestFilter(benchmark_width_, benchmark_height_,
-                                 benchmark_width_ * 3 / 2,
-                                 benchmark_height_ * 3 / 2,
-                                 libyuv::kFilterBilinear,
-                                 benchmark_iterations_,
-                                 disable_cpu_flags_, benchmark_cpu_info_);
+  int diff =
+      YUVToARGBTestFilter(benchmark_width_, benchmark_height_,
+                          benchmark_width_ * 3 / 2, benchmark_height_ * 3 / 2,
+                          libyuv::kFilterBilinear, benchmark_iterations_);
   EXPECT_LE(diff, 10);
 }
 
 TEST_F(LibYUVScaleTest, YUVToRGBScaleDown) {
-  int diff = YUVToARGBTestFilter(benchmark_width_ * 3 / 2,
-                                 benchmark_height_ * 3 / 2,
-                                 benchmark_width_, benchmark_height_,
-                                 libyuv::kFilterBilinear,
-                                 benchmark_iterations_,
-                                 disable_cpu_flags_, benchmark_cpu_info_);
+  int diff = YUVToARGBTestFilter(
+      benchmark_width_ * 3 / 2, benchmark_height_ * 3 / 2, benchmark_width_,
+      benchmark_height_, libyuv::kFilterBilinear, benchmark_iterations_);
   EXPECT_LE(diff, 10);
 }
 
+TEST_F(LibYUVScaleTest, ARGBTest3x) {
+  const int kSrcStride = 480 * 4;
+  const int kDstStride = 160 * 4;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i * 4 + 0] = i;
+    orig_pixels[i * 4 + 1] = 255 - i;
+    orig_pixels[i * 4 + 2] = i + 1;
+    orig_pixels[i * 4 + 3] = i + 10;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ARGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+              kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+  EXPECT_EQ(226, dest_pixels[2]);
+  EXPECT_EQ(235, dest_pixels[3]);
+
+  ARGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+            kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+  EXPECT_EQ(226, dest_pixels[2]);
+  EXPECT_EQ(235, dest_pixels[3]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, ARGBTest4x) {
+  const int kSrcStride = 640 * 4;
+  const int kDstStride = 160 * 4;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i * 4 + 0] = i;
+    orig_pixels[i * 4 + 1] = 255 - i;
+    orig_pixels[i * 4 + 2] = i + 1;
+    orig_pixels[i * 4 + 3] = i + 10;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ARGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+              kFilterBilinear);
+  }
+
+  EXPECT_NEAR(66, dest_pixels[0], 4);
+  EXPECT_NEAR(255 - 66, dest_pixels[1], 4);
+  EXPECT_NEAR(67, dest_pixels[2], 4);
+  EXPECT_NEAR(76, dest_pixels[3], 4);
+
+  ARGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+            kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);
+  EXPECT_EQ(255 - 2, dest_pixels[1]);
+  EXPECT_EQ(3, dest_pixels[2]);
+  EXPECT_EQ(12, dest_pixels[3]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
 
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/scale_rgb_test.cc b/media/libyuv/unit_test/scale_rgb_test.cc
new file mode 100644
index 0000000000..8296abe31d
--- /dev/null
+++ b/media/libyuv/unit_test/scale_rgb_test.cc
@@ -0,0 +1,280 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <time.h>
+
+#include "../unit_test/unit_test.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/scale_rgb.h"
+
+namespace libyuv {
+
+#define STRINGIZE(line) #line
+#define FILELINESTR(file, line) file ":" STRINGIZE(line)
+
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int RGBTestFilter(int src_width,
+                         int src_height,
+                         int dst_width,
+                         int dst_height,
+                         FilterMode f,
+                         int benchmark_iterations,
+                         int disable_cpu_flags,
+                         int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  const int b = 0;  // 128 to test for padding/stride.
+  int64_t src_rgb_plane_size =
+      (Abs(src_width) + b * 3) * (Abs(src_height) + b * 3) * 3LL;
+  int src_stride_rgb = (b * 3 + Abs(src_width)) * 3;
+
+  align_buffer_page_end(src_rgb, src_rgb_plane_size);
+  if (!src_rgb) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_rgb, src_rgb_plane_size);
+
+  int64_t dst_rgb_plane_size = (dst_width + b * 3) * (dst_height + b * 3) * 3LL;
+  int dst_stride_rgb = (b * 3 + dst_width) * 3;
+
+  align_buffer_page_end(dst_rgb_c, dst_rgb_plane_size);
+  align_buffer_page_end(dst_rgb_opt, dst_rgb_plane_size);
+  if (!dst_rgb_c || !dst_rgb_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  memset(dst_rgb_c, 2, dst_rgb_plane_size);
+  memset(dst_rgb_opt, 3, dst_rgb_plane_size);
+
+  // Warm up both versions for consistent benchmarks.
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+           src_height, dst_rgb_c + (dst_stride_rgb * b) + b * 3, dst_stride_rgb,
+           dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+           src_height, dst_rgb_opt + (dst_stride_rgb * b) + b * 3,
+           dst_stride_rgb, dst_width, dst_height, f);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+           src_height, dst_rgb_c + (dst_stride_rgb * b) + b * 3, dst_stride_rgb,
+           dst_width, dst_height, f);
+
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
+  for (i = 0; i < benchmark_iterations; ++i) {
+    RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+             src_height, dst_rgb_opt + (dst_stride_rgb * b) + b * 3,
+             dst_stride_rgb, dst_width, dst_height, f);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+
+  // Report performance of C vs OPT
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference isn't
+  //  over 2.
+  int max_diff = 0;
+  for (i = b; i < (dst_height + b); ++i) {
+    for (j = b * 3; j < (dst_width + b) * 3; ++j) {
+      int abs_diff = Abs(dst_rgb_c[(i * dst_stride_rgb) + j] -
+                         dst_rgb_opt[(i * dst_stride_rgb) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_rgb_c);
+  free_aligned_buffer_page_end(dst_rgb_opt);
+  free_aligned_buffer_page_end(src_rgb);
+  return max_diff;
+}
+
+// The following adjustments in dimensions ensure the scale factor will be
+// exactly achieved.
+#define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
+#define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
+
+#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
+  TEST_F(LibYUVScaleTest, RGBScaleDownBy##name##_##filter) {                 \
+    int diff = RGBTestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
+        benchmark_cpu_info_);                                                \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
+// filtering is different fixed point implementations for SSSE3, Neon and C.
+#define TEST_FACTOR(name, nom, denom)         \
+  TEST_FACTOR1(name, None, nom, denom, 0)     \
+  TEST_FACTOR1(name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(name, Box, nom, denom, 3)
+#else
+// Test a scale factor with Bilinear.
+#define TEST_FACTOR(name, nom, denom) \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3)
+#endif
+
+TEST_FACTOR(2, 1, 2)
+#ifndef DISABLE_SLOW_TESTS
+TEST_FACTOR(4, 1, 4)
+// TEST_FACTOR(8, 1, 8)  Disable for benchmark performance.
+TEST_FACTOR(3by4, 3, 4)
+TEST_FACTOR(3by8, 3, 8)
+TEST_FACTOR(3, 1, 3)
+#endif
+#undef TEST_FACTOR1
+#undef TEST_FACTOR
+#undef SX
+#undef DX
+
+#define TEST_SCALETO1(name, width, height, filter, max_diff)                 \
+  TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {           \
+    int diff = RGBTestFilter(benchmark_width_, benchmark_height_, width,     \
+                             height, kFilter##filter, benchmark_iterations_, \
+                             disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                               \
+  }                                                                          \
+  TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {         \
+    int diff = RGBTestFilter(width, height, Abs(benchmark_width_),           \
+                             Abs(benchmark_height_), kFilter##filter,        \
+                             benchmark_iterations_, disable_cpu_flags_,      \
+                             benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+/// Test scale to a specified size with all 4 filters.
+#define TEST_SCALETO(name, width, height)       \
+  TEST_SCALETO1(name, width, height, None, 0)   \
+  TEST_SCALETO1(name, width, height, Linear, 3) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#else
+#define TEST_SCALETO(name, width, height) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#endif
+
+TEST_SCALETO(RGBScale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(RGBScale, 1, 1)
+TEST_SCALETO(RGBScale, 256, 144) /* 128x72 * 3 */
+TEST_SCALETO(RGBScale, 320, 240)
+TEST_SCALETO(RGBScale, 569, 480)
+TEST_SCALETO(RGBScale, 1280, 720)
+TEST_SCALETO(RGBScale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
+#undef TEST_SCALETO1
+#undef TEST_SCALETO
+
+#define TEST_SCALESWAPXY1(name, filter, max_diff)                      \
+  TEST_F(LibYUVScaleTest, name##SwapXY_##filter) {                     \
+    int diff = RGBTestFilter(benchmark_width_, benchmark_height_,      \
+                             benchmark_height_, benchmark_width_,      \
+                             kFilter##filter, benchmark_iterations_,   \
+                             disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                         \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test scale with swapped width and height with all 3 filters.
+TEST_SCALESWAPXY1(RGBScale, None, 0)
+TEST_SCALESWAPXY1(RGBScale, Linear, 0)
+TEST_SCALESWAPXY1(RGBScale, Bilinear, 0)
+#else
+TEST_SCALESWAPXY1(RGBScale, Bilinear, 0)
+#endif
+#undef TEST_SCALESWAPXY1
+
+TEST_F(LibYUVScaleTest, RGBTest3x) {
+  const int kSrcStride = 480 * 3;
+  const int kDstStride = 160 * 3;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i * 3 + 0] = i;
+    orig_pixels[i * 3 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    RGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+             kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  RGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+           kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, RGBTest4x) {
+  const int kSrcStride = 640 * 3;
+  const int kDstStride = 160 * 3;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i * 3 + 0] = i;
+    orig_pixels[i * 3 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    RGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+             kFilterBilinear);
+  }
+
+  EXPECT_EQ(66, dest_pixels[0]);
+  EXPECT_EQ(190, dest_pixels[1]);
+
+  RGBScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
+           kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);  // expect the 3rd pixel of the 3rd row
+  EXPECT_EQ(255 - 2, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+}  // namespace libyuv
diff --git a/media/libyuv/unit_test/scale_test.cc b/media/libyuv/unit_test/scale_test.cc
index f40443e290..a8c95268dc 100644
--- a/media/libyuv/unit_test/scale_test.cc
+++ b/media/libyuv/unit_test/scale_test.cc
@@ -11,38 +11,51 @@
 #include <stdlib.h>
 #include <time.h>
 
+#include "../unit_test/unit_test.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/scale.h"
-#include "../unit_test/unit_test.h"
+
+#ifdef ENABLE_ROW_TESTS
+#include "libyuv/scale_row.h"  // For ScaleRowDown2Box_Odd_C
+#endif
 
 #define STRINGIZE(line) #line
 #define FILELINESTR(file, line) file ":" STRINGIZE(line)
 
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
 namespace libyuv {
 
 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
-static int TestFilter(int src_width, int src_height,
-                      int dst_width, int dst_height,
-                      FilterMode f, int benchmark_iterations,
-                      int disable_cpu_flags, int benchmark_cpu_info) {
+static int I420TestFilter(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          FilterMode f,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info) {
   if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
     return 0;
   }
 
   int i, j;
-  const int b = 0;  // 128 to test for padding/stride.
   int src_width_uv = (Abs(src_width) + 1) >> 1;
   int src_height_uv = (Abs(src_height) + 1) >> 1;
 
-  int64 src_y_plane_size = (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2);
-  int64 src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2);
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
 
-  int src_stride_y = b * 2 + Abs(src_width);
-  int src_stride_uv = b * 2 + src_width_uv;
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
 
-  align_buffer_page_end(src_y, src_y_plane_size)
-  align_buffer_page_end(src_u, src_uv_plane_size)
-  align_buffer_page_end(src_v, src_uv_plane_size)
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
   if (!src_y || !src_u || !src_v) {
     printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
     return 0;
@@ -54,62 +67,51 @@ static int TestFilter(int src_width, int src_height,
   int dst_width_uv = (dst_width + 1) >> 1;
   int dst_height_uv = (dst_height + 1) >> 1;
 
-  int64 dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2);
-  int64 dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2);
+  int64_t dst_y_plane_size = (dst_width) * (dst_height);
+  int64_t dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
 
-  int dst_stride_y = b * 2 + dst_width;
-  int dst_stride_uv = b * 2 + dst_width_uv;
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
 
-  align_buffer_page_end(dst_y_c, dst_y_plane_size)
-  align_buffer_page_end(dst_u_c, dst_uv_plane_size)
-  align_buffer_page_end(dst_v_c, dst_uv_plane_size)
-  align_buffer_page_end(dst_y_opt, dst_y_plane_size)
-  align_buffer_page_end(dst_u_opt, dst_uv_plane_size)
-  align_buffer_page_end(dst_v_opt, dst_uv_plane_size)
-  if (!dst_y_c || !dst_u_c || !dst_v_c ||
-      !dst_y_opt|| !dst_u_opt|| !dst_v_opt) {
+  align_buffer_page_end(dst_y_c, dst_y_plane_size);
+  align_buffer_page_end(dst_u_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_opt, dst_y_plane_size);
+  align_buffer_page_end(dst_u_opt, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_opt, dst_uv_plane_size);
+  if (!dst_y_c || !dst_u_c || !dst_v_c || !dst_y_opt || !dst_u_opt ||
+      !dst_v_opt) {
     printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
     return 0;
   }
 
   MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
   double c_time = get_time();
-  I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
-            src_u + (src_stride_uv * b) + b, src_stride_uv,
-            src_v + (src_stride_uv * b) + b, src_stride_uv,
-            src_width, src_height,
-            dst_y_c + (dst_stride_y * b) + b, dst_stride_y,
-            dst_u_c + (dst_stride_uv * b) + b, dst_stride_uv,
-            dst_v_c + (dst_stride_uv * b) + b, dst_stride_uv,
-            dst_width, dst_height, f);
+  I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_c, dst_stride_y, dst_u_c,
+            dst_stride_uv, dst_v_c, dst_stride_uv, dst_width, dst_height, f);
   c_time = (get_time() - c_time);
 
   MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
   double opt_time = get_time();
   for (i = 0; i < benchmark_iterations; ++i) {
-    I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
-              src_u + (src_stride_uv * b) + b, src_stride_uv,
-              src_v + (src_stride_uv * b) + b, src_stride_uv,
-              src_width, src_height,
-              dst_y_opt + (dst_stride_y * b) + b, dst_stride_y,
-              dst_u_opt + (dst_stride_uv * b) + b, dst_stride_uv,
-              dst_v_opt + (dst_stride_uv * b) + b, dst_stride_uv,
-              dst_width, dst_height, f);
+    I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+              src_width, src_height, dst_y_opt, dst_stride_y, dst_u_opt,
+              dst_stride_uv, dst_v_opt, dst_stride_uv, dst_width, dst_height,
+              f);
   }
   opt_time = (get_time() - opt_time) / benchmark_iterations;
-  // Report performance of C vs OPT
-  printf("filter %d - %8d us C - %8d us OPT\n",
-         f,
-         static_cast<int>(c_time * 1e6),
-         static_cast<int>(opt_time * 1e6));
+  // Report performance of C vs OPT.
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
 
   // C version may be a little off from the optimized. Order of
   //  operations may introduce rounding somewhere. So do a difference
-  //  of the buffers and look to see that the max difference isn't
-  //  over 2.
+  //  of the buffers and look to see that the max difference is not
+  //  over 3.
   int max_diff = 0;
-  for (i = b; i < (dst_height + b); ++i) {
-    for (j = b; j < (dst_width + b); ++j) {
+  for (i = 0; i < (dst_height); ++i) {
+    for (j = 0; j < (dst_width); ++j) {
       int abs_diff = Abs(dst_y_c[(i * dst_stride_y) + j] -
                          dst_y_opt[(i * dst_stride_y) + j]);
       if (abs_diff > max_diff) {
@@ -118,8 +120,8 @@ static int TestFilter(int src_width, int src_height,
     }
   }
 
-  for (i = b; i < (dst_height_uv + b); ++i) {
-    for (j = b; j < (dst_width_uv + b); ++j) {
+  for (i = 0; i < (dst_height_uv); ++i) {
+    for (j = 0; j < (dst_width_uv); ++j) {
       int abs_diff = Abs(dst_u_c[(i * dst_stride_uv) + j] -
                          dst_u_opt[(i * dst_stride_uv) + j]);
       if (abs_diff > max_diff) {
@@ -133,188 +135,784 @@ static int TestFilter(int src_width, int src_height,
     }
   }
 
-  free_aligned_buffer_page_end(dst_y_c)
-  free_aligned_buffer_page_end(dst_u_c)
-  free_aligned_buffer_page_end(dst_v_c)
-  free_aligned_buffer_page_end(dst_y_opt)
-  free_aligned_buffer_page_end(dst_u_opt)
-  free_aligned_buffer_page_end(dst_v_opt)
-
-  free_aligned_buffer_page_end(src_y)
-  free_aligned_buffer_page_end(src_u)
-  free_aligned_buffer_page_end(src_v)
+  free_aligned_buffer_page_end(dst_y_c);
+  free_aligned_buffer_page_end(dst_u_c);
+  free_aligned_buffer_page_end(dst_v_c);
+  free_aligned_buffer_page_end(dst_y_opt);
+  free_aligned_buffer_page_end(dst_u_opt);
+  free_aligned_buffer_page_end(dst_v_opt);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
 
   return max_diff;
 }
 
-// Test scaling with 8 bit C vs 16 bit C and return maximum pixel difference.
+// Test scaling with 8 bit C vs 12 bit C and return maximum pixel difference.
 // 0 = exact.
-static int TestFilter_16(int src_width, int src_height,
-                         int dst_width, int dst_height,
-                         FilterMode f, int benchmark_iterations) {
+static int I420TestFilter_12(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             FilterMode f,
+                             int benchmark_iterations,
+                             int disable_cpu_flags,
+                             int benchmark_cpu_info) {
   if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
     return 0;
   }
 
-  int i, j;
-  const int b = 0;  // 128 to test for padding/stride.
+  int i;
   int src_width_uv = (Abs(src_width) + 1) >> 1;
   int src_height_uv = (Abs(src_height) + 1) >> 1;
 
-  int64 src_y_plane_size = (Abs(src_width) + b * 2) *
-      (Abs(src_height) + b * 2);
-  int64 src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2);
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
 
-  int src_stride_y = b * 2 + Abs(src_width);
-  int src_stride_uv = b * 2 + src_width_uv;
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
 
-  align_buffer_page_end(src_y, src_y_plane_size)
-  align_buffer_page_end(src_u, src_uv_plane_size)
-  align_buffer_page_end(src_v, src_uv_plane_size)
-  align_buffer_page_end(src_y_16, src_y_plane_size * 2)
-  align_buffer_page_end(src_u_16, src_uv_plane_size * 2)
-  align_buffer_page_end(src_v_16, src_uv_plane_size * 2)
-  uint16* p_src_y_16 = reinterpret_cast<uint16*>(src_y_16);
-  uint16* p_src_u_16 = reinterpret_cast<uint16*>(src_u_16);
-  uint16* p_src_v_16 = reinterpret_cast<uint16*>(src_v_16);
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  align_buffer_page_end(src_y_12, src_y_plane_size * 2);
+  align_buffer_page_end(src_u_12, src_uv_plane_size * 2);
+  align_buffer_page_end(src_v_12, src_uv_plane_size * 2);
+  if (!src_y || !src_u || !src_v || !src_y_12 || !src_u_12 || !src_v_12) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  uint16_t* p_src_y_12 = reinterpret_cast<uint16_t*>(src_y_12);
+  uint16_t* p_src_u_12 = reinterpret_cast<uint16_t*>(src_u_12);
+  uint16_t* p_src_v_12 = reinterpret_cast<uint16_t*>(src_v_12);
 
   MemRandomize(src_y, src_y_plane_size);
   MemRandomize(src_u, src_uv_plane_size);
   MemRandomize(src_v, src_uv_plane_size);
 
-  for (i = b; i < src_height + b; ++i) {
-    for (j = b; j < src_width + b; ++j) {
-      p_src_y_16[(i * src_stride_y) + j] = src_y[(i * src_stride_y) + j];
-    }
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_12[i] = src_y[i];
+  }
+  for (i = 0; i < src_uv_plane_size; ++i) {
+    p_src_u_12[i] = src_u[i];
+    p_src_v_12[i] = src_v[i];
   }
 
-  for (i = b; i < (src_height_uv + b); ++i) {
-    for (j = b; j < (src_width_uv + b); ++j) {
-      p_src_u_16[(i * src_stride_uv) + j] = src_u[(i * src_stride_uv) + j];
-      p_src_v_16[(i * src_stride_uv) + j] = src_v[(i * src_stride_uv) + j];
+  int dst_width_uv = (dst_width + 1) >> 1;
+  int dst_height_uv = (dst_height + 1) >> 1;
+
+  int dst_y_plane_size = (dst_width) * (dst_height);
+  int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
+
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_u_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_12, dst_y_plane_size * 2);
+  align_buffer_page_end(dst_u_12, dst_uv_plane_size * 2);
+  align_buffer_page_end(dst_v_12, dst_uv_plane_size * 2);
+
+  uint16_t* p_dst_y_12 = reinterpret_cast<uint16_t*>(dst_y_12);
+  uint16_t* p_dst_u_12 = reinterpret_cast<uint16_t*>(dst_u_12);
+  uint16_t* p_dst_v_12 = reinterpret_cast<uint16_t*>(dst_v_12);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_8, dst_stride_y, dst_u_8,
+            dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (i = 0; i < benchmark_iterations; ++i) {
+    I420Scale_12(p_src_y_12, src_stride_y, p_src_u_12, src_stride_uv,
+                 p_src_v_12, src_stride_uv, src_width, src_height, p_dst_y_12,
+                 dst_stride_y, p_dst_u_12, dst_stride_uv, p_dst_v_12,
+                 dst_stride_uv, dst_width, dst_height, f);
+  }
+
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
     }
   }
+  for (i = 0; i < dst_uv_plane_size; ++i) {
+    int abs_diff = Abs(dst_u_8[i] - p_dst_u_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+    abs_diff = Abs(dst_v_8[i] - p_dst_v_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_u_8);
+  free_aligned_buffer_page_end(dst_v_8);
+  free_aligned_buffer_page_end(dst_y_12);
+  free_aligned_buffer_page_end(dst_u_12);
+  free_aligned_buffer_page_end(dst_v_12);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+  free_aligned_buffer_page_end(src_y_12);
+  free_aligned_buffer_page_end(src_u_12);
+  free_aligned_buffer_page_end(src_v_12);
+
+  return max_diff;
+}
+
+// Test scaling with 8 bit C vs 16 bit C and return maximum pixel difference.
+// 0 = exact.
+static int I420TestFilter_16(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             FilterMode f,
+                             int benchmark_iterations,
+                             int disable_cpu_flags,
+                             int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i;
+  int src_width_uv = (Abs(src_width) + 1) >> 1;
+  int src_height_uv = (Abs(src_height) + 1) >> 1;
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  align_buffer_page_end(src_y_16, src_y_plane_size * 2);
+  align_buffer_page_end(src_u_16, src_uv_plane_size * 2);
+  align_buffer_page_end(src_v_16, src_uv_plane_size * 2);
+  if (!src_y || !src_u || !src_v || !src_y_16 || !src_u_16 || !src_v_16) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  uint16_t* p_src_y_16 = reinterpret_cast<uint16_t*>(src_y_16);
+  uint16_t* p_src_u_16 = reinterpret_cast<uint16_t*>(src_u_16);
+  uint16_t* p_src_v_16 = reinterpret_cast<uint16_t*>(src_v_16);
+
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_u, src_uv_plane_size);
+  MemRandomize(src_v, src_uv_plane_size);
+
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_16[i] = src_y[i];
+  }
+  for (i = 0; i < src_uv_plane_size; ++i) {
+    p_src_u_16[i] = src_u[i];
+    p_src_v_16[i] = src_v[i];
+  }
 
   int dst_width_uv = (dst_width + 1) >> 1;
   int dst_height_uv = (dst_height + 1) >> 1;
 
-  int dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2);
-  int dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2);
+  int dst_y_plane_size = (dst_width) * (dst_height);
+  int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
 
-  int dst_stride_y = b * 2 + dst_width;
-  int dst_stride_uv = b * 2 + dst_width_uv;
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
 
-  align_buffer_page_end(dst_y_8, dst_y_plane_size)
-  align_buffer_page_end(dst_u_8, dst_uv_plane_size)
-  align_buffer_page_end(dst_v_8, dst_uv_plane_size)
-  align_buffer_page_end(dst_y_16, dst_y_plane_size * 2)
-  align_buffer_page_end(dst_u_16, dst_uv_plane_size * 2)
-  align_buffer_page_end(dst_v_16, dst_uv_plane_size * 2)
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_u_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_16, dst_y_plane_size * 2);
+  align_buffer_page_end(dst_u_16, dst_uv_plane_size * 2);
+  align_buffer_page_end(dst_v_16, dst_uv_plane_size * 2);
 
-  uint16* p_dst_y_16 = reinterpret_cast<uint16*>(dst_y_16);
-  uint16* p_dst_u_16 = reinterpret_cast<uint16*>(dst_u_16);
-  uint16* p_dst_v_16 = reinterpret_cast<uint16*>(dst_v_16);
+  uint16_t* p_dst_y_16 = reinterpret_cast<uint16_t*>(dst_y_16);
+  uint16_t* p_dst_u_16 = reinterpret_cast<uint16_t*>(dst_u_16);
+  uint16_t* p_dst_v_16 = reinterpret_cast<uint16_t*>(dst_v_16);
 
-  I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
-            src_u + (src_stride_uv * b) + b, src_stride_uv,
-            src_v + (src_stride_uv * b) + b, src_stride_uv,
-            src_width, src_height,
-            dst_y_8 + (dst_stride_y * b) + b, dst_stride_y,
-            dst_u_8 + (dst_stride_uv * b) + b, dst_stride_uv,
-            dst_v_8 + (dst_stride_uv * b) + b, dst_stride_uv,
-            dst_width, dst_height, f);
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_8, dst_stride_y, dst_u_8,
+            dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (i = 0; i < benchmark_iterations; ++i) {
+    I420Scale_16(p_src_y_16, src_stride_y, p_src_u_16, src_stride_uv,
+                 p_src_v_16, src_stride_uv, src_width, src_height, p_dst_y_16,
+                 dst_stride_y, p_dst_u_16, dst_stride_uv, p_dst_v_16,
+                 dst_stride_uv, dst_width, dst_height, f);
+  }
 
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  for (i = 0; i < dst_uv_plane_size; ++i) {
+    int abs_diff = Abs(dst_u_8[i] - p_dst_u_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+    abs_diff = Abs(dst_v_8[i] - p_dst_v_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_u_8);
+  free_aligned_buffer_page_end(dst_v_8);
+  free_aligned_buffer_page_end(dst_y_16);
+  free_aligned_buffer_page_end(dst_u_16);
+  free_aligned_buffer_page_end(dst_v_16);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+  free_aligned_buffer_page_end(src_y_16);
+  free_aligned_buffer_page_end(src_u_16);
+  free_aligned_buffer_page_end(src_v_16);
+
+  return max_diff;
+}
+
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int I444TestFilter(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          FilterMode f,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  int src_width_uv = Abs(src_width);
+  int src_height_uv = Abs(src_height);
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  if (!src_y || !src_u || !src_v) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_u, src_uv_plane_size);
+  MemRandomize(src_v, src_uv_plane_size);
+
+  int dst_width_uv = dst_width;
+  int dst_height_uv = dst_height;
+
+  int64_t dst_y_plane_size = (dst_width) * (dst_height);
+  int64_t dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
+
+  align_buffer_page_end(dst_y_c, dst_y_plane_size);
+  align_buffer_page_end(dst_u_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_opt, dst_y_plane_size);
+  align_buffer_page_end(dst_u_opt, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_opt, dst_uv_plane_size);
+  if (!dst_y_c || !dst_u_c || !dst_v_c || !dst_y_opt || !dst_u_opt ||
+      !dst_v_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  I444Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_c, dst_stride_y, dst_u_c,
+            dst_stride_uv, dst_v_c, dst_stride_uv, dst_width, dst_height, f);
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
   for (i = 0; i < benchmark_iterations; ++i) {
-    I420Scale_16(p_src_y_16 + (src_stride_y * b) + b, src_stride_y,
-                 p_src_u_16 + (src_stride_uv * b) + b, src_stride_uv,
-                 p_src_v_16 + (src_stride_uv * b) + b, src_stride_uv,
-                 src_width, src_height,
-                 p_dst_y_16 + (dst_stride_y * b) + b, dst_stride_y,
-                 p_dst_u_16 + (dst_stride_uv * b) + b, dst_stride_uv,
-                 p_dst_v_16 + (dst_stride_uv * b) + b, dst_stride_uv,
-                 dst_width, dst_height, f);
+    I444Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+              src_width, src_height, dst_y_opt, dst_stride_y, dst_u_opt,
+              dst_stride_uv, dst_v_opt, dst_stride_uv, dst_width, dst_height,
+              f);
   }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+  // Report performance of C vs OPT.
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
 
-  // Expect an exact match
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference is not
+  //  over 3.
   int max_diff = 0;
-  for (i = b; i < (dst_height + b); ++i) {
-    for (j = b; j < (dst_width + b); ++j) {
-      int abs_diff = Abs(dst_y_8[(i * dst_stride_y) + j] -
-                         p_dst_y_16[(i * dst_stride_y) + j]);
+  for (i = 0; i < (dst_height); ++i) {
+    for (j = 0; j < (dst_width); ++j) {
+      int abs_diff = Abs(dst_y_c[(i * dst_stride_y) + j] -
+                         dst_y_opt[(i * dst_stride_y) + j]);
       if (abs_diff > max_diff) {
         max_diff = abs_diff;
       }
     }
   }
 
-  for (i = b; i < (dst_height_uv + b); ++i) {
-    for (j = b; j < (dst_width_uv + b); ++j) {
-      int abs_diff = Abs(dst_u_8[(i * dst_stride_uv) + j] -
-                         p_dst_u_16[(i * dst_stride_uv) + j]);
+  for (i = 0; i < (dst_height_uv); ++i) {
+    for (j = 0; j < (dst_width_uv); ++j) {
+      int abs_diff = Abs(dst_u_c[(i * dst_stride_uv) + j] -
+                         dst_u_opt[(i * dst_stride_uv) + j]);
       if (abs_diff > max_diff) {
         max_diff = abs_diff;
       }
-      abs_diff = Abs(dst_v_8[(i * dst_stride_uv) + j] -
-                     p_dst_v_16[(i * dst_stride_uv) + j]);
+      abs_diff = Abs(dst_v_c[(i * dst_stride_uv) + j] -
+                     dst_v_opt[(i * dst_stride_uv) + j]);
       if (abs_diff > max_diff) {
         max_diff = abs_diff;
       }
     }
   }
 
-  free_aligned_buffer_page_end(dst_y_8)
-  free_aligned_buffer_page_end(dst_u_8)
-  free_aligned_buffer_page_end(dst_v_8)
-  free_aligned_buffer_page_end(dst_y_16)
-  free_aligned_buffer_page_end(dst_u_16)
-  free_aligned_buffer_page_end(dst_v_16)
+  free_aligned_buffer_page_end(dst_y_c);
+  free_aligned_buffer_page_end(dst_u_c);
+  free_aligned_buffer_page_end(dst_v_c);
+  free_aligned_buffer_page_end(dst_y_opt);
+  free_aligned_buffer_page_end(dst_u_opt);
+  free_aligned_buffer_page_end(dst_v_opt);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+
+  return max_diff;
+}
+
+// Test scaling with 8 bit C vs 12 bit C and return maximum pixel difference.
+// 0 = exact.
+static int I444TestFilter_12(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             FilterMode f,
+                             int benchmark_iterations,
+                             int disable_cpu_flags,
+                             int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i;
+  int src_width_uv = Abs(src_width);
+  int src_height_uv = Abs(src_height);
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  align_buffer_page_end(src_y_12, src_y_plane_size * 2);
+  align_buffer_page_end(src_u_12, src_uv_plane_size * 2);
+  align_buffer_page_end(src_v_12, src_uv_plane_size * 2);
+  if (!src_y || !src_u || !src_v || !src_y_12 || !src_u_12 || !src_v_12) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  uint16_t* p_src_y_12 = reinterpret_cast<uint16_t*>(src_y_12);
+  uint16_t* p_src_u_12 = reinterpret_cast<uint16_t*>(src_u_12);
+  uint16_t* p_src_v_12 = reinterpret_cast<uint16_t*>(src_v_12);
+
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_u, src_uv_plane_size);
+  MemRandomize(src_v, src_uv_plane_size);
+
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_12[i] = src_y[i];
+  }
+  for (i = 0; i < src_uv_plane_size; ++i) {
+    p_src_u_12[i] = src_u[i];
+    p_src_v_12[i] = src_v[i];
+  }
+
+  int dst_width_uv = dst_width;
+  int dst_height_uv = dst_height;
+
+  int dst_y_plane_size = (dst_width) * (dst_height);
+  int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
+
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_u_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_12, dst_y_plane_size * 2);
+  align_buffer_page_end(dst_u_12, dst_uv_plane_size * 2);
+  align_buffer_page_end(dst_v_12, dst_uv_plane_size * 2);
+
+  uint16_t* p_dst_y_12 = reinterpret_cast<uint16_t*>(dst_y_12);
+  uint16_t* p_dst_u_12 = reinterpret_cast<uint16_t*>(dst_u_12);
+  uint16_t* p_dst_v_12 = reinterpret_cast<uint16_t*>(dst_v_12);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I444Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_8, dst_stride_y, dst_u_8,
+            dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (i = 0; i < benchmark_iterations; ++i) {
+    I444Scale_12(p_src_y_12, src_stride_y, p_src_u_12, src_stride_uv,
+                 p_src_v_12, src_stride_uv, src_width, src_height, p_dst_y_12,
+                 dst_stride_y, p_dst_u_12, dst_stride_uv, p_dst_v_12,
+                 dst_stride_uv, dst_width, dst_height, f);
+  }
+
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  for (i = 0; i < dst_uv_plane_size; ++i) {
+    int abs_diff = Abs(dst_u_8[i] - p_dst_u_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+    abs_diff = Abs(dst_v_8[i] - p_dst_v_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_u_8);
+  free_aligned_buffer_page_end(dst_v_8);
+  free_aligned_buffer_page_end(dst_y_12);
+  free_aligned_buffer_page_end(dst_u_12);
+  free_aligned_buffer_page_end(dst_v_12);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+  free_aligned_buffer_page_end(src_y_12);
+  free_aligned_buffer_page_end(src_u_12);
+  free_aligned_buffer_page_end(src_v_12);
 
-  free_aligned_buffer_page_end(src_y)
-  free_aligned_buffer_page_end(src_u)
-  free_aligned_buffer_page_end(src_v)
-  free_aligned_buffer_page_end(src_y_16)
-  free_aligned_buffer_page_end(src_u_16)
-  free_aligned_buffer_page_end(src_v_16)
+  return max_diff;
+}
+
+// Test scaling with 8 bit C vs 16 bit C and return maximum pixel difference.
+// 0 = exact.
+static int I444TestFilter_16(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             FilterMode f,
+                             int benchmark_iterations,
+                             int disable_cpu_flags,
+                             int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i;
+  int src_width_uv = Abs(src_width);
+  int src_height_uv = Abs(src_height);
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  align_buffer_page_end(src_y_16, src_y_plane_size * 2);
+  align_buffer_page_end(src_u_16, src_uv_plane_size * 2);
+  align_buffer_page_end(src_v_16, src_uv_plane_size * 2);
+  if (!src_y || !src_u || !src_v || !src_y_16 || !src_u_16 || !src_v_16) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  uint16_t* p_src_y_16 = reinterpret_cast<uint16_t*>(src_y_16);
+  uint16_t* p_src_u_16 = reinterpret_cast<uint16_t*>(src_u_16);
+  uint16_t* p_src_v_16 = reinterpret_cast<uint16_t*>(src_v_16);
+
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_u, src_uv_plane_size);
+  MemRandomize(src_v, src_uv_plane_size);
+
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_16[i] = src_y[i];
+  }
+  for (i = 0; i < src_uv_plane_size; ++i) {
+    p_src_u_16[i] = src_u[i];
+    p_src_v_16[i] = src_v[i];
+  }
+
+  int dst_width_uv = dst_width;
+  int dst_height_uv = dst_height;
+
+  int dst_y_plane_size = (dst_width) * (dst_height);
+  int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
+
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_u_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_16, dst_y_plane_size * 2);
+  align_buffer_page_end(dst_u_16, dst_uv_plane_size * 2);
+  align_buffer_page_end(dst_v_16, dst_uv_plane_size * 2);
+
+  uint16_t* p_dst_y_16 = reinterpret_cast<uint16_t*>(dst_y_16);
+  uint16_t* p_dst_u_16 = reinterpret_cast<uint16_t*>(dst_u_16);
+  uint16_t* p_dst_v_16 = reinterpret_cast<uint16_t*>(dst_v_16);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I444Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_8, dst_stride_y, dst_u_8,
+            dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (i = 0; i < benchmark_iterations; ++i) {
+    I444Scale_16(p_src_y_16, src_stride_y, p_src_u_16, src_stride_uv,
+                 p_src_v_16, src_stride_uv, src_width, src_height, p_dst_y_16,
+                 dst_stride_y, p_dst_u_16, dst_stride_uv, p_dst_v_16,
+                 dst_stride_uv, dst_width, dst_height, f);
+  }
+
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  for (i = 0; i < dst_uv_plane_size; ++i) {
+    int abs_diff = Abs(dst_u_8[i] - p_dst_u_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+    abs_diff = Abs(dst_v_8[i] - p_dst_v_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_u_8);
+  free_aligned_buffer_page_end(dst_v_8);
+  free_aligned_buffer_page_end(dst_y_16);
+  free_aligned_buffer_page_end(dst_u_16);
+  free_aligned_buffer_page_end(dst_v_16);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+  free_aligned_buffer_page_end(src_y_16);
+  free_aligned_buffer_page_end(src_u_16);
+  free_aligned_buffer_page_end(src_v_16);
+
+  return max_diff;
+}
+
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int NV12TestFilter(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          FilterMode f,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  int src_width_uv = (Abs(src_width) + 1) >> 1;
+  int src_height_uv = (Abs(src_height) + 1) >> 1;
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv)*2;
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv * 2;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_uv, src_uv_plane_size);
+  if (!src_y || !src_uv) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_uv, src_uv_plane_size);
+
+  int dst_width_uv = (dst_width + 1) >> 1;
+  int dst_height_uv = (dst_height + 1) >> 1;
+
+  int64_t dst_y_plane_size = (dst_width) * (dst_height);
+  int64_t dst_uv_plane_size = (dst_width_uv) * (dst_height_uv)*2;
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv * 2;
+
+  align_buffer_page_end(dst_y_c, dst_y_plane_size);
+  align_buffer_page_end(dst_uv_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_opt, dst_y_plane_size);
+  align_buffer_page_end(dst_uv_opt, dst_uv_plane_size);
+  if (!dst_y_c || !dst_uv_c || !dst_y_opt || !dst_uv_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  NV12Scale(src_y, src_stride_y, src_uv, src_stride_uv, src_width, src_height,
+            dst_y_c, dst_stride_y, dst_uv_c, dst_stride_uv, dst_width,
+            dst_height, f);
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
+  for (i = 0; i < benchmark_iterations; ++i) {
+    NV12Scale(src_y, src_stride_y, src_uv, src_stride_uv, src_width, src_height,
+              dst_y_opt, dst_stride_y, dst_uv_opt, dst_stride_uv, dst_width,
+              dst_height, f);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+  // Report performance of C vs OPT.
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference is not
+  //  over 3.
+  int max_diff = 0;
+  for (i = 0; i < (dst_height); ++i) {
+    for (j = 0; j < (dst_width); ++j) {
+      int abs_diff = Abs(dst_y_c[(i * dst_stride_y) + j] -
+                         dst_y_opt[(i * dst_stride_y) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  for (i = 0; i < (dst_height_uv); ++i) {
+    for (j = 0; j < (dst_width_uv * 2); ++j) {
+      int abs_diff = Abs(dst_uv_c[(i * dst_stride_uv) + j] -
+                         dst_uv_opt[(i * dst_stride_uv) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_c);
+  free_aligned_buffer_page_end(dst_uv_c);
+  free_aligned_buffer_page_end(dst_y_opt);
+  free_aligned_buffer_page_end(dst_uv_opt);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_uv);
 
   return max_diff;
 }
 
 // The following adjustments in dimensions ensure the scale factor will be
 // exactly achieved.
-// 2 is chroma subsample
+// 2 is chroma subsample.
 #define DX(x, nom, denom) static_cast<int>(((Abs(x) / nom + 1) / 2) * nom * 2)
 #define SX(x, nom, denom) static_cast<int>(((x / nom + 1) / 2) * denom * 2)
 
-#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                       \
-    TEST_F(LibYUVScaleTest, ScaleDownBy##name##_##filter) {                    \
-      int diff = TestFilter(SX(benchmark_width_, nom, denom),                  \
-                            SX(benchmark_height_, nom, denom),                 \
-                            DX(benchmark_width_, nom, denom),                  \
-                            DX(benchmark_height_, nom, denom),                 \
-                            kFilter##filter, benchmark_iterations_,            \
-                            disable_cpu_flags_, benchmark_cpu_info_);          \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest, DISABLED_ScaleDownBy##name##_##filter##_16) {      \
-      int diff = TestFilter_16(SX(benchmark_width_, nom, denom),               \
-                               SX(benchmark_height_, nom, denom),              \
-                               DX(benchmark_width_, nom, denom),               \
-                               DX(benchmark_height_, nom, denom),              \
-                               kFilter##filter, benchmark_iterations_);        \
-      EXPECT_LE(diff, max_diff);                                               \
-    }
+#define TEST_FACTOR1(DISABLED_, name, filter, nom, denom, max_diff)           \
+  TEST_F(LibYUVScaleTest, I420ScaleDownBy##name##_##filter) {                 \
+    int diff = I420TestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, I444ScaleDownBy##name##_##filter) {                 \
+    int diff = I444TestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, DISABLED_##I420ScaleDownBy##name##_##filter##_12) { \
+    int diff = I420TestFilter_12(                                             \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, DISABLED_##I444ScaleDownBy##name##_##filter##_12) { \
+    int diff = I444TestFilter_12(                                             \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, NV12ScaleDownBy##name##_##filter) {                 \
+    int diff = NV12TestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }
 
 // Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
 // filtering is different fixed point implementations for SSSE3, Neon and C.
-#define TEST_FACTOR(name, nom, denom, boxdiff)                                 \
-    TEST_FACTOR1(name, None, nom, denom, 0)                                    \
-    TEST_FACTOR1(name, Linear, nom, denom, 3)                                  \
-    TEST_FACTOR1(name, Bilinear, nom, denom, 3)                                \
-    TEST_FACTOR1(name, Box, nom, denom, boxdiff)
+#ifndef DISABLE_SLOW_TESTS
+#define TEST_FACTOR(name, nom, denom, boxdiff)  \
+  TEST_FACTOR1(, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(, name, Box, nom, denom, boxdiff)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_FACTOR(name, nom, denom, boxdiff)           \
+  TEST_FACTOR1(DISABLED_, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(DISABLED_, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(DISABLED_, name, Box, nom, denom, boxdiff)
+#else
+#define TEST_FACTOR(name, nom, denom, boxdiff)           \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(DISABLED_, name, Box, nom, denom, boxdiff)
+#endif
+#endif
 
 TEST_FACTOR(2, 1, 2, 0)
 TEST_FACTOR(4, 1, 4, 0)
+#ifndef DISABLE_SLOW_TESTS
 TEST_FACTOR(8, 1, 8, 0)
+#endif
 TEST_FACTOR(3by4, 3, 4, 1)
 TEST_FACTOR(3by8, 3, 8, 1)
 TEST_FACTOR(3, 1, 3, 0)
@@ -323,50 +921,681 @@ TEST_FACTOR(3, 1, 3, 0)
 #undef SX
 #undef DX
 
-#define TEST_SCALETO1(name, width, height, filter, max_diff)                   \
-    TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {           \
-      int diff = TestFilter(benchmark_width_, benchmark_height_,               \
-                            width, height,                                     \
-                            kFilter##filter, benchmark_iterations_,            \
-                            disable_cpu_flags_, benchmark_cpu_info_);          \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {         \
-      int diff = TestFilter(width, height,                                     \
-                            Abs(benchmark_width_), Abs(benchmark_height_),     \
-                            kFilter##filter, benchmark_iterations_,            \
-                            disable_cpu_flags_, benchmark_cpu_info_);          \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest,                                                    \
-        DISABLED_##name##To##width##x##height##_##filter##_16) {               \
-      int diff = TestFilter_16(benchmark_width_, benchmark_height_,            \
-                               width, height,                                  \
-                               kFilter##filter, benchmark_iterations_);        \
-      EXPECT_LE(diff, max_diff);                                               \
-    }                                                                          \
-    TEST_F(LibYUVScaleTest,                                                    \
-        DISABLED_##name##From##width##x##height##_##filter##_16) {             \
-      int diff = TestFilter_16(width, height,                                  \
-                               Abs(benchmark_width_), Abs(benchmark_height_),  \
-                               kFilter##filter, benchmark_iterations_);        \
-      EXPECT_LE(diff, max_diff);                                               \
-    }
+#define TEST_SCALETO1(DISABLED_, name, width, height, filter, max_diff)       \
+  TEST_F(LibYUVScaleTest, I420##name##To##width##x##height##_##filter) {      \
+    int diff = I420TestFilter(benchmark_width_, benchmark_height_, width,     \
+                              height, kFilter##filter, benchmark_iterations_, \
+                              disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, I444##name##To##width##x##height##_##filter) {      \
+    int diff = I444TestFilter(benchmark_width_, benchmark_height_, width,     \
+                              height, kFilter##filter, benchmark_iterations_, \
+                              disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##To##width##x##height##_##filter##_12) {       \
+    int diff = I420TestFilter_12(                                             \
+        benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
+        benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##To##width##x##height##_##filter##_12) {       \
+    int diff = I444TestFilter_12(                                             \
+        benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
+        benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##To##width##x##height##_##filter##_16) {       \
+    int diff = I420TestFilter_16(                                             \
+        benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
+        benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##To##width##x##height##_##filter##_16) {       \
+    int diff = I444TestFilter_16(                                             \
+        benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
+        benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, NV12##name##To##width##x##height##_##filter) {      \
+    int diff = NV12TestFilter(benchmark_width_, benchmark_height_, width,     \
+                              height, kFilter##filter, benchmark_iterations_, \
+                              disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, I420##name##From##width##x##height##_##filter) {    \
+    int diff = I420TestFilter(width, height, Abs(benchmark_width_),           \
+                              Abs(benchmark_height_), kFilter##filter,        \
+                              benchmark_iterations_, disable_cpu_flags_,      \
+                              benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, I444##name##From##width##x##height##_##filter) {    \
+    int diff = I444TestFilter(width, height, Abs(benchmark_width_),           \
+                              Abs(benchmark_height_), kFilter##filter,        \
+                              benchmark_iterations_, disable_cpu_flags_,      \
+                              benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##From##width##x##height##_##filter##_12) {     \
+    int diff = I420TestFilter_12(width, height, Abs(benchmark_width_),        \
+                                 Abs(benchmark_height_), kFilter##filter,     \
+                                 benchmark_iterations_, disable_cpu_flags_,   \
+                                 benchmark_cpu_info_);                        \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##From##width##x##height##_##filter##_12) {     \
+    int diff = I444TestFilter_12(width, height, Abs(benchmark_width_),        \
+                                 Abs(benchmark_height_), kFilter##filter,     \
+                                 benchmark_iterations_, disable_cpu_flags_,   \
+                                 benchmark_cpu_info_);                        \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##From##width##x##height##_##filter##_16) {     \
+    int diff = I420TestFilter_16(width, height, Abs(benchmark_width_),        \
+                                 Abs(benchmark_height_), kFilter##filter,     \
+                                 benchmark_iterations_, disable_cpu_flags_,   \
+                                 benchmark_cpu_info_);                        \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##From##width##x##height##_##filter##_16) {     \
+    int diff = I444TestFilter_16(width, height, Abs(benchmark_width_),        \
+                                 Abs(benchmark_height_), kFilter##filter,     \
+                                 benchmark_iterations_, disable_cpu_flags_,   \
+                                 benchmark_cpu_info_);                        \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, NV12##name##From##width##x##height##_##filter) {    \
+    int diff = NV12TestFilter(width, height, Abs(benchmark_width_),           \
+                              Abs(benchmark_height_), kFilter##filter,        \
+                              benchmark_iterations_, disable_cpu_flags_,      \
+                              benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                                \
+  }
 
+#ifndef DISABLE_SLOW_TESTS
 // Test scale to a specified size with all 4 filters.
-#define TEST_SCALETO(name, width, height)                                      \
-    TEST_SCALETO1(name, width, height, None, 0)                                \
-    TEST_SCALETO1(name, width, height, Linear, 0)                              \
-    TEST_SCALETO1(name, width, height, Bilinear, 0)                            \
-    TEST_SCALETO1(name, width, height, Box, 0)
+#define TEST_SCALETO(name, width, height)           \
+  TEST_SCALETO1(, name, width, height, None, 0)     \
+  TEST_SCALETO1(, name, width, height, Linear, 3)   \
+  TEST_SCALETO1(, name, width, height, Bilinear, 3) \
+  TEST_SCALETO1(, name, width, height, Box, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_SCALETO(name, width, height)                    \
+  TEST_SCALETO1(DISABLED_, name, width, height, None, 0)     \
+  TEST_SCALETO1(DISABLED_, name, width, height, Linear, 3)   \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Box, 3)
+#else
+#define TEST_SCALETO(name, width, height)                    \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Box, 3)
+#endif
+#endif
 
 TEST_SCALETO(Scale, 1, 1)
-TEST_SCALETO(Scale, 320, 240)
-TEST_SCALETO(Scale, 352, 288)
 TEST_SCALETO(Scale, 569, 480)
 TEST_SCALETO(Scale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(Scale, 256, 144) /* 128x72 * 2 */
+TEST_SCALETO(Scale, 320, 240)
 TEST_SCALETO(Scale, 1280, 720)
+TEST_SCALETO(Scale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
 #undef TEST_SCALETO1
 #undef TEST_SCALETO
 
+#define TEST_SCALESWAPXY1(DISABLED_, name, filter, max_diff)               \
+  TEST_F(LibYUVScaleTest, I420##name##SwapXY_##filter) {                   \
+    int diff = I420TestFilter(benchmark_width_, benchmark_height_,         \
+                              benchmark_height_, benchmark_width_,         \
+                              kFilter##filter, benchmark_iterations_,      \
+                              disable_cpu_flags_, benchmark_cpu_info_);    \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, I444##name##SwapXY_##filter) {                   \
+    int diff = I444TestFilter(benchmark_width_, benchmark_height_,         \
+                              benchmark_height_, benchmark_width_,         \
+                              kFilter##filter, benchmark_iterations_,      \
+                              disable_cpu_flags_, benchmark_cpu_info_);    \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I420##name##SwapXY_##filter##_12) {   \
+    int diff = I420TestFilter_12(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I444##name##SwapXY_##filter##_12) {   \
+    int diff = I444TestFilter_12(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I420##name##SwapXY_##filter##_16) {   \
+    int diff = I420TestFilter_16(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I444##name##SwapXY_##filter##_16) {   \
+    int diff = I444TestFilter_16(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, NV12##name##SwapXY_##filter) {                   \
+    int diff = NV12TestFilter(benchmark_width_, benchmark_height_,         \
+                              benchmark_height_, benchmark_width_,         \
+                              kFilter##filter, benchmark_iterations_,      \
+                              disable_cpu_flags_, benchmark_cpu_info_);    \
+    EXPECT_LE(diff, max_diff);                                             \
+  }
+
+// Test scale to a specified size with all 4 filters.
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALESWAPXY1(, Scale, None, 0)
+TEST_SCALESWAPXY1(, Scale, Linear, 3)
+TEST_SCALESWAPXY1(, Scale, Bilinear, 3)
+TEST_SCALESWAPXY1(, Scale, Box, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+TEST_SCALESWAPXY1(DISABLED_, Scale, None, 0)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Linear, 3)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Bilinear, 3)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Box, 3)
+#else
+TEST_SCALESWAPXY1(DISABLED_, Scale, Bilinear, 3)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Box, 3)
+#endif
+#endif
+
+#undef TEST_SCALESWAPXY1
+
+#ifdef ENABLE_ROW_TESTS
+#ifdef HAS_SCALEROWDOWN2_SSSE3
+TEST_F(LibYUVScaleTest, TestScaleRowDown2Box_Odd_SSSE3) {
+  SIMD_ALIGNED(uint8_t orig_pixels[128 * 2]);
+  SIMD_ALIGNED(uint8_t dst_pixels_opt[64]);
+  SIMD_ALIGNED(uint8_t dst_pixels_c[64]);
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_opt, 0, sizeof(dst_pixels_opt));
+  memset(dst_pixels_c, 0, sizeof(dst_pixels_c));
+
+  int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+  if (!has_ssse3) {
+    printf("Warning SSSE3 not detected; Skipping test.\n");
+  } else {
+    // TL.
+    orig_pixels[0] = 255u;
+    orig_pixels[1] = 0u;
+    orig_pixels[128 + 0] = 0u;
+    orig_pixels[128 + 1] = 0u;
+    // TR.
+    orig_pixels[2] = 0u;
+    orig_pixels[3] = 100u;
+    orig_pixels[128 + 2] = 0u;
+    orig_pixels[128 + 3] = 0u;
+    // BL.
+    orig_pixels[4] = 0u;
+    orig_pixels[5] = 0u;
+    orig_pixels[128 + 4] = 50u;
+    orig_pixels[128 + 5] = 0u;
+    // BR.
+    orig_pixels[6] = 0u;
+    orig_pixels[7] = 0u;
+    orig_pixels[128 + 6] = 0u;
+    orig_pixels[128 + 7] = 20u;
+    // Odd.
+    orig_pixels[126] = 4u;
+    orig_pixels[127] = 255u;
+    orig_pixels[128 + 126] = 16u;
+    orig_pixels[128 + 127] = 255u;
+
+    // Test regular half size.
+    ScaleRowDown2Box_C(orig_pixels, 128, dst_pixels_c, 64);
+
+    EXPECT_EQ(64u, dst_pixels_c[0]);
+    EXPECT_EQ(25u, dst_pixels_c[1]);
+    EXPECT_EQ(13u, dst_pixels_c[2]);
+    EXPECT_EQ(5u, dst_pixels_c[3]);
+    EXPECT_EQ(0u, dst_pixels_c[4]);
+    EXPECT_EQ(133u, dst_pixels_c[63]);
+
+    // Test Odd width version - Last pixel is just 1 horizontal pixel.
+    ScaleRowDown2Box_Odd_C(orig_pixels, 128, dst_pixels_c, 64);
+
+    EXPECT_EQ(64u, dst_pixels_c[0]);
+    EXPECT_EQ(25u, dst_pixels_c[1]);
+    EXPECT_EQ(13u, dst_pixels_c[2]);
+    EXPECT_EQ(5u, dst_pixels_c[3]);
+    EXPECT_EQ(0u, dst_pixels_c[4]);
+    EXPECT_EQ(10u, dst_pixels_c[63]);
+
+    // Test one pixel less, should skip the last pixel.
+    memset(dst_pixels_c, 0, sizeof(dst_pixels_c));
+    ScaleRowDown2Box_Odd_C(orig_pixels, 128, dst_pixels_c, 63);
+
+    EXPECT_EQ(64u, dst_pixels_c[0]);
+    EXPECT_EQ(25u, dst_pixels_c[1]);
+    EXPECT_EQ(13u, dst_pixels_c[2]);
+    EXPECT_EQ(5u, dst_pixels_c[3]);
+    EXPECT_EQ(0u, dst_pixels_c[4]);
+    EXPECT_EQ(0u, dst_pixels_c[63]);
+
+    // Test regular half size SSSE3.
+    ScaleRowDown2Box_SSSE3(orig_pixels, 128, dst_pixels_opt, 64);
+
+    EXPECT_EQ(64u, dst_pixels_opt[0]);
+    EXPECT_EQ(25u, dst_pixels_opt[1]);
+    EXPECT_EQ(13u, dst_pixels_opt[2]);
+    EXPECT_EQ(5u, dst_pixels_opt[3]);
+    EXPECT_EQ(0u, dst_pixels_opt[4]);
+    EXPECT_EQ(133u, dst_pixels_opt[63]);
+
+    // Compare C and SSSE3 match.
+    ScaleRowDown2Box_Odd_C(orig_pixels, 128, dst_pixels_c, 64);
+    ScaleRowDown2Box_Odd_SSSE3(orig_pixels, 128, dst_pixels_opt, 64);
+    for (int i = 0; i < 64; ++i) {
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+    }
+  }
+}
+#endif  // HAS_SCALEROWDOWN2_SSSE3
+
+extern "C" void ScaleRowUp2_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst,
+                                    int dst_width);
+extern "C" void ScaleRowUp2_16_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint16_t* dst,
+                                 int dst_width);
+
+TEST_F(LibYUVScaleTest, TestScaleRowUp2_16) {
+  SIMD_ALIGNED(uint16_t orig_pixels[640 * 2 + 1]);  // 2 rows + 1 pixel overrun.
+  SIMD_ALIGNED(uint16_t dst_pixels_opt[1280]);
+  SIMD_ALIGNED(uint16_t dst_pixels_c[1280]);
+
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_opt, 1, sizeof(dst_pixels_opt));
+  memset(dst_pixels_c, 2, sizeof(dst_pixels_c));
+
+  for (int i = 0; i < 640 * 2 + 1; ++i) {
+    orig_pixels[i] = i;
+  }
+  ScaleRowUp2_16_C(&orig_pixels[0], 640, &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      ScaleRowUp2_16_NEON(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
+    } else {
+      ScaleRowUp2_16_C(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
+    }
+#else
+    ScaleRowUp2_16_C(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
+#endif
+  }
+
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  EXPECT_EQ(dst_pixels_c[0], (0 * 9 + 1 * 3 + 640 * 3 + 641 * 1 + 8) / 16);
+  EXPECT_EQ(dst_pixels_c[1279], 800);
+}
+
+extern "C" void ScaleRowDown2Box_16_NEON(const uint16_t* src_ptr,
+                                         ptrdiff_t src_stride,
+                                         uint16_t* dst,
+                                         int dst_width);
+
+TEST_F(LibYUVScaleTest, TestScaleRowDown2Box_16) {
+  SIMD_ALIGNED(uint16_t orig_pixels[2560 * 2]);
+  SIMD_ALIGNED(uint16_t dst_pixels_c[1280]);
+  SIMD_ALIGNED(uint16_t dst_pixels_opt[1280]);
+
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
+
+  for (int i = 0; i < 2560 * 2; ++i) {
+    orig_pixels[i] = i;
+  }
+  ScaleRowDown2Box_16_C(&orig_pixels[0], 2560, &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      ScaleRowDown2Box_16_NEON(&orig_pixels[0], 2560, &dst_pixels_opt[0], 1280);
+    } else {
+      ScaleRowDown2Box_16_C(&orig_pixels[0], 2560, &dst_pixels_opt[0], 1280);
+    }
+#else
+    ScaleRowDown2Box_16_C(&orig_pixels[0], 2560, &dst_pixels_opt[0], 1280);
+#endif
+  }
+
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  EXPECT_EQ(dst_pixels_c[0], (0 + 1 + 2560 + 2561 + 2) / 4);
+  EXPECT_EQ(dst_pixels_c[1279], 3839);
+}
+#endif  // ENABLE_ROW_TESTS
+
+// Test scaling plane with 8 bit C vs 12 bit C and return maximum pixel
+// difference.
+// 0 = exact.
+static int TestPlaneFilter_16(int src_width,
+                              int src_height,
+                              int dst_width,
+                              int dst_height,
+                              FilterMode f,
+                              int benchmark_iterations,
+                              int disable_cpu_flags,
+                              int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i;
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int src_stride_y = Abs(src_width);
+  int dst_y_plane_size = dst_width * dst_height;
+  int dst_stride_y = dst_width;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_y_16, src_y_plane_size * 2);
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_y_16, dst_y_plane_size * 2);
+  uint16_t* p_src_y_16 = reinterpret_cast<uint16_t*>(src_y_16);
+  uint16_t* p_dst_y_16 = reinterpret_cast<uint16_t*>(dst_y_16);
+
+  MemRandomize(src_y, src_y_plane_size);
+  memset(dst_y_8, 0, dst_y_plane_size);
+  memset(dst_y_16, 1, dst_y_plane_size * 2);
+
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_16[i] = src_y[i] & 255;
+  }
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y_8, dst_stride_y,
+             dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+
+  for (i = 0; i < benchmark_iterations; ++i) {
+    ScalePlane_16(p_src_y_16, src_stride_y, src_width, src_height, p_dst_y_16,
+                  dst_stride_y, dst_width, dst_height, f);
+  }
+
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_16[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_y_16);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_y_16);
+
+  return max_diff;
+}
+
+// The following adjustments in dimensions ensure the scale factor will be
+// exactly achieved.
+// 2 is chroma subsample.
+#define DX(x, nom, denom) static_cast<int>(((Abs(x) / nom + 1) / 2) * nom * 2)
+#define SX(x, nom, denom) static_cast<int>(((x / nom + 1) / 2) * denom * 2)
+
+#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                       \
+  TEST_F(LibYUVScaleTest, DISABLED_##ScalePlaneDownBy##name##_##filter##_16) { \
+    int diff = TestPlaneFilter_16(                                             \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),   \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),   \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,            \
+        benchmark_cpu_info_);                                                  \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }
+
+// Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
+// filtering is different fixed point implementations for SSSE3, Neon and C.
+#define TEST_FACTOR(name, nom, denom, boxdiff)      \
+  TEST_FACTOR1(name, None, nom, denom, 0)           \
+  TEST_FACTOR1(name, Linear, nom, denom, boxdiff)   \
+  TEST_FACTOR1(name, Bilinear, nom, denom, boxdiff) \
+  TEST_FACTOR1(name, Box, nom, denom, boxdiff)
+
+TEST_FACTOR(2, 1, 2, 0)
+TEST_FACTOR(4, 1, 4, 0)
+// TEST_FACTOR(8, 1, 8, 0) Disable for benchmark performance.  Takes 90 seconds.
+TEST_FACTOR(3by4, 3, 4, 1)
+TEST_FACTOR(3by8, 3, 8, 1)
+TEST_FACTOR(3, 1, 3, 0)
+#undef TEST_FACTOR1
+#undef TEST_FACTOR
+#undef SX
+#undef DX
+
+TEST_F(LibYUVScaleTest, PlaneTest3x) {
+  const int kSrcStride = 480;
+  const int kDstStride = 160;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ScalePlane(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+               kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+
+  ScalePlane(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+             kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTest4x) {
+  const int kSrcStride = 640;
+  const int kDstStride = 160;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ScalePlane(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+               kFilterBilinear);
+  }
+
+  EXPECT_EQ(66, dest_pixels[0]);
+
+  ScalePlane(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+             kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);  // expect the 3rd pixel of the 3rd row
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+// Intent is to test 200x50 to 50x200 but width and height can be parameters.
+TEST_F(LibYUVScaleTest, PlaneTestRotate_None) {
+  const int kSize = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < kSize; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_opt_pixels, kSize);
+  align_buffer_page_end(dest_c_pixels, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);  // Disable all CPU optimization.
+  ScalePlane(orig_pixels, benchmark_width_, benchmark_width_, benchmark_height_,
+             dest_c_pixels, benchmark_height_, benchmark_height_,
+             benchmark_width_, kFilterNone);
+  MaskCpuFlags(benchmark_cpu_info_);  // Enable all CPU optimization.
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ScalePlane(orig_pixels, benchmark_width_, benchmark_width_,
+               benchmark_height_, dest_opt_pixels, benchmark_height_,
+               benchmark_height_, benchmark_width_, kFilterNone);
+  }
+
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dest_c_pixels[i], dest_opt_pixels[i]);
+  }
+
+  free_aligned_buffer_page_end(dest_c_pixels);
+  free_aligned_buffer_page_end(dest_opt_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTestRotate_Bilinear) {
+  const int kSize = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < kSize; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_opt_pixels, kSize);
+  align_buffer_page_end(dest_c_pixels, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);  // Disable all CPU optimization.
+  ScalePlane(orig_pixels, benchmark_width_, benchmark_width_, benchmark_height_,
+             dest_c_pixels, benchmark_height_, benchmark_height_,
+             benchmark_width_, kFilterBilinear);
+  MaskCpuFlags(benchmark_cpu_info_);  // Enable all CPU optimization.
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ScalePlane(orig_pixels, benchmark_width_, benchmark_width_,
+               benchmark_height_, dest_opt_pixels, benchmark_height_,
+               benchmark_height_, benchmark_width_, kFilterBilinear);
+  }
+
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dest_c_pixels[i], dest_opt_pixels[i]);
+  }
+
+  free_aligned_buffer_page_end(dest_c_pixels);
+  free_aligned_buffer_page_end(dest_opt_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+// Intent is to test 200x50 to 50x200 but width and height can be parameters.
+TEST_F(LibYUVScaleTest, PlaneTestRotate_Box) {
+  const int kSize = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < kSize; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_opt_pixels, kSize);
+  align_buffer_page_end(dest_c_pixels, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);  // Disable all CPU optimization.
+  ScalePlane(orig_pixels, benchmark_width_, benchmark_width_, benchmark_height_,
+             dest_c_pixels, benchmark_height_, benchmark_height_,
+             benchmark_width_, kFilterBox);
+  MaskCpuFlags(benchmark_cpu_info_);  // Enable all CPU optimization.
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ScalePlane(orig_pixels, benchmark_width_, benchmark_width_,
+               benchmark_height_, dest_opt_pixels, benchmark_height_,
+               benchmark_height_, benchmark_width_, kFilterBox);
+  }
+
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dest_c_pixels[i], dest_opt_pixels[i]);
+  }
+
+  free_aligned_buffer_page_end(dest_c_pixels);
+  free_aligned_buffer_page_end(dest_opt_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTest1_Box) {
+  align_buffer_page_end(orig_pixels, 3);
+  align_buffer_page_end(dst_pixels, 3);
+
+  // Pad the 1x1 byte image with invalid values before and after in case libyuv
+  // reads outside the memory boundaries.
+  orig_pixels[0] = 0;
+  orig_pixels[1] = 1;  // scale this pixel
+  orig_pixels[2] = 2;
+  dst_pixels[0] = 3;
+  dst_pixels[1] = 3;
+  dst_pixels[2] = 3;
+
+  libyuv::ScalePlane(orig_pixels + 1, /* src_stride= */ 1, /* src_width= */ 1,
+                     /* src_height= */ 1, dst_pixels, /* dst_stride= */ 1,
+                     /* dst_width= */ 1, /* dst_height= */ 2,
+                     libyuv::kFilterBox);
+
+  EXPECT_EQ(dst_pixels[0], 1);
+  EXPECT_EQ(dst_pixels[1], 1);
+  EXPECT_EQ(dst_pixels[2], 3);
+
+  free_aligned_buffer_page_end(dst_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTest1_16_Box) {
+  align_buffer_page_end(orig_pixels_alloc, 3 * 2);
+  align_buffer_page_end(dst_pixels_alloc, 3 * 2);
+  uint16_t* orig_pixels = (uint16_t*)orig_pixels_alloc;
+  uint16_t* dst_pixels = (uint16_t*)dst_pixels_alloc;
+
+  // Pad the 1x1 byte image with invalid values before and after in case libyuv
+  // reads outside the memory boundaries.
+  orig_pixels[0] = 0;
+  orig_pixels[1] = 1;  // scale this pixel
+  orig_pixels[2] = 2;
+  dst_pixels[0] = 3;
+  dst_pixels[1] = 3;
+  dst_pixels[2] = 3;
+
+  libyuv::ScalePlane_16(
+      orig_pixels + 1, /* src_stride= */ 1, /* src_width= */ 1,
+      /* src_height= */ 1, dst_pixels, /* dst_stride= */ 1,
+      /* dst_width= */ 1, /* dst_height= */ 2, libyuv::kFilterNone);
+
+  EXPECT_EQ(dst_pixels[0], 1);
+  EXPECT_EQ(dst_pixels[1], 1);
+  EXPECT_EQ(dst_pixels[2], 3);
+
+  free_aligned_buffer_page_end(dst_pixels_alloc);
+  free_aligned_buffer_page_end(orig_pixels_alloc);
+}
 }  // namespace libyuv
diff --git a/media/libyuv/unit_test/scale_uv_test.cc b/media/libyuv/unit_test/scale_uv_test.cc
new file mode 100644
index 0000000000..3d524bef1f
--- /dev/null
+++ b/media/libyuv/unit_test/scale_uv_test.cc
@@ -0,0 +1,278 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <time.h>
+
+#include "../unit_test/unit_test.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/scale_uv.h"
+
+namespace libyuv {
+
+#define STRINGIZE(line) #line
+#define FILELINESTR(file, line) file ":" STRINGIZE(line)
+
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int UVTestFilter(int src_width,
+                        int src_height,
+                        int dst_width,
+                        int dst_height,
+                        FilterMode f,
+                        int benchmark_iterations,
+                        int disable_cpu_flags,
+                        int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  const int b = 0;  // 128 to test for padding/stride.
+  int64_t src_uv_plane_size =
+      (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2) * 2LL;
+  int src_stride_uv = (b * 2 + Abs(src_width)) * 2;
+
+  align_buffer_page_end(src_uv, src_uv_plane_size);
+  if (!src_uv) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_uv, src_uv_plane_size);
+
+  int64_t dst_uv_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 2LL;
+  int dst_stride_uv = (b * 2 + dst_width) * 2;
+
+  align_buffer_page_end(dst_uv_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_uv_opt, dst_uv_plane_size);
+  if (!dst_uv_c || !dst_uv_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  memset(dst_uv_c, 2, dst_uv_plane_size);
+  memset(dst_uv_opt, 3, dst_uv_plane_size);
+
+  // Warm up both versions for consistent benchmarks.
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+          src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+          dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+          src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+          dst_width, dst_height, f);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+          src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+          dst_width, dst_height, f);
+
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
+  for (i = 0; i < benchmark_iterations; ++i) {
+    UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+            src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+            dst_width, dst_height, f);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+
+  // Report performance of C vs OPT
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference isn't
+  //  over 2.
+  int max_diff = 0;
+  for (i = b; i < (dst_height + b); ++i) {
+    for (j = b * 2; j < (dst_width + b) * 2; ++j) {
+      int abs_diff = Abs(dst_uv_c[(i * dst_stride_uv) + j] -
+                         dst_uv_opt[(i * dst_stride_uv) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_uv_c);
+  free_aligned_buffer_page_end(dst_uv_opt);
+  free_aligned_buffer_page_end(src_uv);
+  return max_diff;
+}
+
+// The following adjustments in dimensions ensure the scale factor will be
+// exactly achieved.
+#define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
+#define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
+
+#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
+  TEST_F(LibYUVScaleTest, UVScaleDownBy##name##_##filter) {                  \
+    int diff = UVTestFilter(                                                 \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
+        benchmark_cpu_info_);                                                \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
+// filtering is different fixed point implementations for SSSE3, Neon and C.
+#define TEST_FACTOR(name, nom, denom)         \
+  TEST_FACTOR1(name, None, nom, denom, 0)     \
+  TEST_FACTOR1(name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(name, Box, nom, denom, 3)
+#else
+// Test a scale factor with Bilinear.
+#define TEST_FACTOR(name, nom, denom) \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3)
+#endif
+
+TEST_FACTOR(2, 1, 2)
+TEST_FACTOR(4, 1, 4)
+// TEST_FACTOR(8, 1, 8)  Disable for benchmark performance.
+TEST_FACTOR(3by4, 3, 4)
+TEST_FACTOR(3by8, 3, 8)
+TEST_FACTOR(3, 1, 3)
+#undef TEST_FACTOR1
+#undef TEST_FACTOR
+#undef SX
+#undef DX
+
+#define TEST_SCALETO1(name, width, height, filter, max_diff)                \
+  TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {          \
+    int diff = UVTestFilter(benchmark_width_, benchmark_height_, width,     \
+                            height, kFilter##filter, benchmark_iterations_, \
+                            disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                              \
+  }                                                                         \
+  TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {        \
+    int diff = UVTestFilter(width, height, Abs(benchmark_width_),           \
+                            Abs(benchmark_height_), kFilter##filter,        \
+                            benchmark_iterations_, disable_cpu_flags_,      \
+                            benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                              \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+/// Test scale to a specified size with all 4 filters.
+#define TEST_SCALETO(name, width, height)       \
+  TEST_SCALETO1(name, width, height, None, 0)   \
+  TEST_SCALETO1(name, width, height, Linear, 3) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#else
+#define TEST_SCALETO(name, width, height) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#endif
+
+TEST_SCALETO(UVScale, 1, 1)
+TEST_SCALETO(UVScale, 569, 480)
+TEST_SCALETO(UVScale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(UVScale, 256, 144) /* 128x72 * 2 */
+TEST_SCALETO(UVScale, 320, 240)
+TEST_SCALETO(UVScale, 1280, 720)
+TEST_SCALETO(UVScale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
+#undef TEST_SCALETO1
+#undef TEST_SCALETO
+
+#define TEST_SCALESWAPXY1(name, filter, max_diff)                              \
+  TEST_F(LibYUVScaleTest, name##SwapXY_##filter) {                             \
+    int diff =                                                                 \
+        UVTestFilter(benchmark_width_, benchmark_height_, benchmark_height_,   \
+                     benchmark_width_, kFilter##filter, benchmark_iterations_, \
+                     disable_cpu_flags_, benchmark_cpu_info_);                 \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test scale with swapped width and height with all 3 filters.
+TEST_SCALESWAPXY1(UVScale, None, 0)
+TEST_SCALESWAPXY1(UVScale, Linear, 0)
+TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
+#else
+TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
+#endif
+#undef TEST_SCALESWAPXY1
+
+TEST_F(LibYUVScaleTest, UVTest3x) {
+  const int kSrcStride = 480 * 2;
+  const int kDstStride = 160 * 2;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i * 2 + 0] = i;
+    orig_pixels[i * 2 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    UVScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+            kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  UVScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+          kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, UVTest4x) {
+  const int kSrcStride = 640 * 2;
+  const int kDstStride = 160 * 2;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i * 2 + 0] = i;
+    orig_pixels[i * 2 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    UVScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+            kFilterBilinear);
+  }
+
+  EXPECT_EQ(66, dest_pixels[0]);
+  EXPECT_EQ(190, dest_pixels[1]);
+
+  UVScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
+          kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);  // expect the 3rd pixel of the 3rd row
+  EXPECT_EQ(255 - 2, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+}  // namespace libyuv
diff --git a/media/libyuv/unit_test/testdata/juno.txt b/media/libyuv/unit_test/testdata/juno.txt
index c275be7425..dd465272b8 100644
--- a/media/libyuv/unit_test/testdata/juno.txt
+++ b/media/libyuv/unit_test/testdata/juno.txt
@@ -1,15 +1,15 @@
-Processor       : AArch64 Processor rev 0 (aarch64)

-processor       : 0

-processor       : 1

-processor       : 2

-processor       : 3

-processor       : 4

-processor       : 5

-Features        : fp asimd evtstrm aes pmull sha1 sha2 crc32

-CPU implementer : 0x41

-CPU architecture: AArch64

-CPU variant     : 0x0

-CPU part        : 0xd07

-CPU revision    : 0

-

-Hardware        : Juno

+Processor       : AArch64 Processor rev 0 (aarch64)
+processor       : 0
+processor       : 1
+processor       : 2
+processor       : 3
+processor       : 4
+processor       : 5
+Features        : fp asimd evtstrm aes pmull sha1 sha2 crc32
+CPU implementer : 0x41
+CPU architecture: AArch64
+CPU variant     : 0x0
+CPU part        : 0xd07
+CPU revision    : 0
+
+Hardware        : Juno
diff --git a/media/libyuv/unit_test/testdata/mips.txt b/media/libyuv/unit_test/testdata/mips.txt
new file mode 100644
index 0000000000..d9f28cbfcb
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/mips.txt
@@ -0,0 +1,7 @@
+system type             : generic-loongson-machine
+machine                 : loongson,generic
+processor               : 0
+
+isa                     : mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented        : vz
+shadow register sets    : 1
diff --git a/media/libyuv/unit_test/testdata/mips_loongson2k.txt b/media/libyuv/unit_test/testdata/mips_loongson2k.txt
new file mode 100644
index 0000000000..8a88d38f2b
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/mips_loongson2k.txt
@@ -0,0 +1,5 @@
+system type		: Loongson2K-SBC
+machine			: loongson,LS2k1000-EVP
+processor		: 0
+cpu model		: Loongson-2K V0.3  FPU V0.1
+BogoMIPS		: 1980.41
diff --git a/media/libyuv/unit_test/testdata/mips_loongson3.txt b/media/libyuv/unit_test/testdata/mips_loongson3.txt
new file mode 100644
index 0000000000..1f540b1281
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/mips_loongson3.txt
@@ -0,0 +1,10 @@
+system type		: generic-loongson-machine
+machine			: Unknown
+processor		: 0
+cpu model		: ICT Loongson-3 V0.9  FPU V0.1
+model name		: ICT Loongson-3A R3 (Loongson-3A3000) @ 1500MHz
+BogoMIPS		: 2990.15
+
+isa			: mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented	: dsp dsp2 vz
+shadow register sets	: 1
diff --git a/media/libyuv/unit_test/testdata/mips_loongson_mmi.txt b/media/libyuv/unit_test/testdata/mips_loongson_mmi.txt
new file mode 100644
index 0000000000..0f10b8bb2b
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/mips_loongson_mmi.txt
@@ -0,0 +1,7 @@
+system type             : generic-loongson-machine
+machine                 : loongson,generic
+processor               : 0
+
+isa                     : mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented        : vz loongson-mmi loongson-ext
+shadow register sets    : 1
diff --git a/media/libyuv/unit_test/testdata/mips_msa.txt b/media/libyuv/unit_test/testdata/mips_msa.txt
new file mode 100644
index 0000000000..ac93061527
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/mips_msa.txt
@@ -0,0 +1,7 @@
+system type             : generic-loongson-machine
+machine                 : loongson,generic
+processor               : 0
+
+isa                     : mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented        : vz msa
+shadow register sets    : 1
diff --git a/media/libyuv/unit_test/testdata/test0.jpg b/media/libyuv/unit_test/testdata/test0.jpg
new file mode 100644
index 0000000000..f4461a813f
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/test0.jpg
diff --git a/media/libyuv/unit_test/testdata/test1.jpg b/media/libyuv/unit_test/testdata/test1.jpg
new file mode 100644
index 0000000000..a0210e9d4d
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/test1.jpg
diff --git a/media/libyuv/unit_test/testdata/test2.jpg b/media/libyuv/unit_test/testdata/test2.jpg
new file mode 100644
index 0000000000..816ca767b1
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/test2.jpg
diff --git a/media/libyuv/unit_test/testdata/test3.jpg b/media/libyuv/unit_test/testdata/test3.jpg
new file mode 100644
index 0000000000..792d91dc8d
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/test3.jpg
diff --git a/media/libyuv/unit_test/testdata/test4.jpg b/media/libyuv/unit_test/testdata/test4.jpg
new file mode 100644
index 0000000000..1ef41668fa
--- /dev/null
+++ b/media/libyuv/unit_test/testdata/test4.jpg
diff --git a/media/libyuv/unit_test/unit_test.cc b/media/libyuv/unit_test/unit_test.cc
index e75510fd74..61145a4628 100644
--- a/media/libyuv/unit_test/unit_test.cc
+++ b/media/libyuv/unit_test/unit_test.cc
@@ -14,35 +14,216 @@
 
 #include <cstring>
 
-#include "gflags/gflags.h"
-
-// Change this to 1000 for benchmarking.
-// TODO(fbarchard): Add command line parsing to pass this as option.
-#define BENCHMARK_ITERATIONS 1
+#ifdef LIBYUV_USE_ABSL_FLAGS
+#include "absl/flags/flag.h"
+#include "absl/flags/parse.h"
+#endif
+#include "libyuv/cpu_id.h"
 
 unsigned int fastrand_seed = 0xfb;
 
-DEFINE_int32(libyuv_width, 0, "width of test image.");
-DEFINE_int32(libyuv_height, 0, "height of test image.");
-DEFINE_int32(libyuv_repeat, 0, "number of times to repeat test.");
-DEFINE_int32(libyuv_flags, 0,
-             "cpu flags for reference code. 1 = C, -1 = SIMD");
-DEFINE_int32(libyuv_cpu_info, 0,
-             "cpu flags for benchmark code. 1 = C, -1 = SIMD");
+#ifdef LIBYUV_USE_ABSL_FLAGS
+ABSL_FLAG(int32_t, libyuv_width, 0, "width of test image.");
+ABSL_FLAG(int32_t, libyuv_height, 0, "height of test image.");
+ABSL_FLAG(int32_t, libyuv_repeat, 0, "number of times to repeat test.");
+ABSL_FLAG(int32_t,
+          libyuv_flags,
+          0,
+          "cpu flags for reference code. 1 = C, -1 = SIMD");
+ABSL_FLAG(int32_t,
+          libyuv_cpu_info,
+          0,
+          "cpu flags for benchmark code. 1 = C, -1 = SIMD");
+#else
+// Disable command line parameters if absl/flags disabled.
+static const int32_t FLAGS_libyuv_width = 0;
+static const int32_t FLAGS_libyuv_height = 0;
+static const int32_t FLAGS_libyuv_repeat = 0;
+static const int32_t FLAGS_libyuv_flags = 0;
+static const int32_t FLAGS_libyuv_cpu_info = 0;
+#endif
+
+#ifdef LIBYUV_USE_ABSL_FLAGS
+#define LIBYUV_GET_FLAG(f) absl::GetFlag(f)
+#else
+#define LIBYUV_GET_FLAG(f) f
+#endif
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+static LIBYUV_BOOL TestEnv(const char* name) {
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+int TestCpuEnv(int cpu_info) {
+#if defined(__arm__) || defined(__aarch64__)
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info &= ~libyuv::kCpuHasNEON;
+  }
+#endif
+#if defined(__mips__) && defined(__linux__)
+  if (TestEnv("LIBYUV_DISABLE_MSA")) {
+    cpu_info &= ~libyuv::kCpuHasMSA;
+  }
+#endif
+#if defined(__longarch__) && defined(__linux__)
+  if (TestEnv("LIBYUV_DISABLE_LSX")) {
+    cpu_info &= ~libyuv::kCpuHasLSX;
+  }
+#endif
+#if defined(__longarch__) && defined(__linux__)
+  if (TestEnv("LIBYUV_DISABLE_LASX")) {
+    cpu_info &= ~libyuv::kCpuHasLASX;
+  }
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) &&                   \
+    (defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \
+     defined(_M_IX86))
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info &= ~libyuv::kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info &= ~libyuv::kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info &= ~libyuv::kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info &= ~libyuv::kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info &= ~libyuv::kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info &= ~libyuv::kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info &= ~libyuv::kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info &= ~libyuv::kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info &= ~libyuv::kCpuHasFMA3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_F16C")) {
+    cpu_info &= ~libyuv::kCpuHasF16C;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512BW")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512BW;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VL")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VL;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VNNI")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VNNI;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VBMI")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VBMI;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VBMI2")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VBMI2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VBITALG")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VBITALG;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VPOPCNTDQ")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VPOPCNTDQ;
+  }
+  if (TestEnv("LIBYUV_DISABLE_GFNI")) {
+    cpu_info &= ~libyuv::kCpuHasGFNI;
+  }
+#endif
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info = libyuv::kCpuInitialized;
+  }
+  return cpu_info;
+}
 
 // For quicker unittests, default is 128 x 72.  But when benchmarking,
 // default to 720p.  Allow size to specify.
 // Set flags to -1 for benchmarking to avoid slower C code.
 
-LibYUVConvertTest::LibYUVConvertTest() :
-    benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
-    benchmark_height_(72), disable_cpu_flags_(1), benchmark_cpu_info_(-1) {
+LibYUVConvertTest::LibYUVConvertTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
+  const char* repeat = getenv("LIBYUV_REPEAT");
+  if (repeat) {
+    benchmark_iterations_ = atoi(repeat);  // NOLINT
+  }
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
+  }
+  if (benchmark_iterations_ > 1) {
+    benchmark_width_ = 1280;
+    benchmark_height_ = 720;
+  }
+  const char* width = getenv("LIBYUV_WIDTH");
+  if (width) {
+    benchmark_width_ = atoi(width);  // NOLINT
+  }
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
+  }
+  const char* height = getenv("LIBYUV_HEIGHT");
+  if (height) {
+    benchmark_height_ = atoi(height);  // NOLINT
+  }
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
+  }
+  const char* cpu_flags = getenv("LIBYUV_FLAGS");
+  if (cpu_flags) {
+    disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
+  }
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
+  }
+  const char* cpu_info = getenv("LIBYUV_CPU_INFO");
+  if (cpu_info) {
+    benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
+  }
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
+}
+
+LibYUVColorTest::LibYUVColorTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
   const char* repeat = getenv("LIBYUV_REPEAT");
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -52,49 +233,53 @@ LibYUVConvertTest::LibYUVConvertTest() :
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
-  }
-  benchmark_pixels_div256_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 255.0) / 256.0);
-  benchmark_pixels_div1280_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 1279.0) / 1280.0);
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
 }
 
-LibYUVColorTest::LibYUVColorTest() :
-    benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
-    benchmark_height_(72), disable_cpu_flags_(1), benchmark_cpu_info_(-1) {
+LibYUVScaleTest::LibYUVScaleTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
   const char* repeat = getenv("LIBYUV_REPEAT");
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -104,49 +289,53 @@ LibYUVColorTest::LibYUVColorTest() :
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
-  }
-  benchmark_pixels_div256_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 255.0) / 256.0);
-  benchmark_pixels_div1280_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 1279.0) / 1280.0);
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
 }
 
-LibYUVScaleTest::LibYUVScaleTest() :
-    benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
-    benchmark_height_(72), disable_cpu_flags_(1), benchmark_cpu_info_(-1) {
+LibYUVRotateTest::LibYUVRotateTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
   const char* repeat = getenv("LIBYUV_REPEAT");
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -156,49 +345,53 @@ LibYUVScaleTest::LibYUVScaleTest() :
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
-  }
-  benchmark_pixels_div256_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 255.0) / 256.0);
-  benchmark_pixels_div1280_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 1279.0) / 1280.0);
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
 }
 
-LibYUVRotateTest::LibYUVRotateTest() :
-    benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
-    benchmark_height_(72), disable_cpu_flags_(1), benchmark_cpu_info_(-1) {
+LibYUVPlanarTest::LibYUVPlanarTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
   const char* repeat = getenv("LIBYUV_REPEAT");
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -208,49 +401,53 @@ LibYUVRotateTest::LibYUVRotateTest() :
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
-  }
-  benchmark_pixels_div256_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 255.0) / 256.0);
-  benchmark_pixels_div1280_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 1279.0) / 1280.0);
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
 }
 
-LibYUVPlanarTest::LibYUVPlanarTest() :
-    benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
-    benchmark_height_(72), disable_cpu_flags_(1), benchmark_cpu_info_(-1) {
+LibYUVBaseTest::LibYUVBaseTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
   const char* repeat = getenv("LIBYUV_REPEAT");
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -260,49 +457,53 @@ LibYUVPlanarTest::LibYUVPlanarTest() :
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
-  }
-  benchmark_pixels_div256_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 255.0) / 256.0);
-  benchmark_pixels_div1280_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 1279.0) / 1280.0);
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
 }
 
-LibYUVBaseTest::LibYUVBaseTest() :
-    benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
-    benchmark_height_(72), disable_cpu_flags_(1), benchmark_cpu_info_(-1) {
+LibYUVCompareTest::LibYUVCompareTest()
+    : benchmark_iterations_(1),
+      benchmark_width_(128),
+      benchmark_height_(72),
+      disable_cpu_flags_(1),
+      benchmark_cpu_info_(-1) {
   const char* repeat = getenv("LIBYUV_REPEAT");
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -312,45 +513,45 @@ LibYUVBaseTest::LibYUVBaseTest() :
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
-  }
-  benchmark_pixels_div256_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 255.0) / 256.0);
-  benchmark_pixels_div1280_ = static_cast<int>((
-      static_cast<double>(Abs(benchmark_width_)) *
-      static_cast<double>(Abs(benchmark_height_)) *
-      static_cast<double>(benchmark_iterations_)  + 1279.0) / 1280.0);
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
+  }
+  disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
+  benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
+  libyuv::MaskCpuFlags(benchmark_cpu_info_);
+  benchmark_pixels_div1280_ =
+      static_cast<int>((static_cast<double>(Abs(benchmark_width_)) *
+                            static_cast<double>(Abs(benchmark_height_)) *
+                            static_cast<double>(benchmark_iterations_) +
+                        1279.0) /
+                       1280.0);
 }
 
 int main(int argc, char** argv) {
   ::testing::InitGoogleTest(&argc, argv);
-  // AllowCommandLineParsing allows us to ignore flags passed on to us by
-  // Chromium build bots without having to explicitly disable them.
-  google::AllowCommandLineReparsing();
-  google::ParseCommandLineFlags(&argc, &argv, true);
+#ifdef LIBYUV_USE_ABSL_FLAGS
+  absl::ParseCommandLine(argc, argv);
+#endif
   return RUN_ALL_TESTS();
 }
diff --git a/media/libyuv/unit_test/unit_test.h b/media/libyuv/unit_test/unit_test.h
index f2c4bef001..99cc8d193e 100644
--- a/media/libyuv/unit_test/unit_test.h
+++ b/media/libyuv/unit_test/unit_test.h
@@ -11,11 +11,11 @@
 #ifndef UNIT_TEST_UNIT_TEST_H_  // NOLINT
 #define UNIT_TEST_UNIT_TEST_H_
 
-#ifdef WIN32
+#include <stddef.h>  // For NULL
+#ifdef _WIN32
 #include <windows.h>
 #else
 #include <sys/time.h>
-#include <sys/resource.h>
 #endif
 
 #include <gtest/gtest.h>
@@ -36,6 +36,9 @@ static __inline int Abs(int v) {
   return v >= 0 ? v : -v;
 }
 
+static __inline float FAbs(float v) {
+  return v >= 0 ? v : -v;
+}
 #define OFFBY 0
 
 // Scaling uses 16.16 fixed point to step thru the source image, so a
@@ -54,8 +57,10 @@ static __inline int Abs(int v) {
 static const int kMaxWidth = 32768;
 static const int kMaxHeight = 32768;
 
-static inline bool SizeValid(int src_width, int src_height,
-                             int dst_width, int dst_height) {
+static inline bool SizeValid(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height) {
   if (src_width > kMaxWidth || src_height > kMaxHeight ||
       dst_width > kMaxWidth || dst_height > kMaxHeight) {
     printf("Warning - size too large to test.  Skipping\n");
@@ -64,16 +69,26 @@ static inline bool SizeValid(int src_width, int src_height,
   return true;
 }
 
-#define align_buffer_page_end(var, size)                                       \
-  uint8* var;                                                                  \
-  uint8* var##_mem;                                                            \
-  var##_mem = reinterpret_cast<uint8*>(malloc(((size) + 4095 + 63) & ~4095));  \
-  var = (uint8*)((intptr_t)(var##_mem + (((size) + 4095 + 63) & ~4095) -       \
-      (size)) & ~63);
+#define align_buffer_page_end(var, size)                                \
+  uint8_t* var##_mem =                                                  \
+      reinterpret_cast<uint8_t*>(malloc(((size) + 4095 + 63) & ~4095)); \
+  uint8_t* var = reinterpret_cast<uint8_t*>(                            \
+      (intptr_t)(var##_mem + (((size) + 4095 + 63) & ~4095) - (size)) & ~63)
 
 #define free_aligned_buffer_page_end(var) \
-  free(var##_mem);  \
-  var = 0;
+  free(var##_mem);                        \
+  var = NULL
+
+#define align_buffer_page_end_16(var, size)                                 \
+  uint8_t* var##_mem =                                                      \
+      reinterpret_cast<uint8_t*>(malloc(((size)*2 + 4095 + 63) & ~4095));   \
+  uint16_t* var = reinterpret_cast<uint16_t*>(                              \
+      (intptr_t)(var##_mem + (((size)*2 + 4095 + 63) & ~4095) - (size)*2) & \
+      ~63)
+
+#define free_aligned_buffer_page_end_16(var) \
+  free(var##_mem);                           \
+  var = NULL
 
 #ifdef WIN32
 static inline double get_time() {
@@ -107,10 +122,13 @@ inline int fastrand() {
   return static_cast<int>((fastrand_seed >> 16) & 0xffff);
 }
 
-static inline void MemRandomize(uint8* dst, int64 len) {
-  int64 i;
+// ubsan fails if dst is unaligned unless we use uint8
+static inline void MemRandomize(uint8_t* dst, int64_t len) {
+  int64_t i;
   for (i = 0; i < len - 1; i += 2) {
-    *reinterpret_cast<uint16*>(dst) = fastrand();
+    int r = fastrand();
+    dst[0] = static_cast<uint8_t>(r);
+    dst[1] = static_cast<uint8_t>(r >> 8);
     dst += 2;
   }
   for (; i < len; ++i) {
@@ -123,12 +141,11 @@ class LibYUVColorTest : public ::testing::Test {
   LibYUVColorTest();
 
   int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
-  int benchmark_width_;  // Default 1280.  Use 640 for benchmarking VGA.
-  int benchmark_height_;  // Default 720.  Use 360 for benchmarking VGA.
-  int benchmark_pixels_div256_;  // Total pixels to benchmark / 256.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
   int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
-  int disable_cpu_flags_;  // Default 1.  Use -1 for benchmarking.
-  int benchmark_cpu_info_;  // Default -1.  Use 1 to disable SIMD.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
 };
 
 class LibYUVConvertTest : public ::testing::Test {
@@ -136,12 +153,11 @@ class LibYUVConvertTest : public ::testing::Test {
   LibYUVConvertTest();
 
   int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
-  int benchmark_width_;  // Default 1280.  Use 640 for benchmarking VGA.
-  int benchmark_height_;  // Default 720.  Use 360 for benchmarking VGA.
-  int benchmark_pixels_div256_;  // Total pixels to benchmark / 256.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
   int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
-  int disable_cpu_flags_;  // Default 1.  Use -1 for benchmarking.
-  int benchmark_cpu_info_;  // Default -1.  Use 1 to disable SIMD.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
 };
 
 class LibYUVScaleTest : public ::testing::Test {
@@ -149,12 +165,11 @@ class LibYUVScaleTest : public ::testing::Test {
   LibYUVScaleTest();
 
   int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
-  int benchmark_width_;  // Default 1280.  Use 640 for benchmarking VGA.
-  int benchmark_height_;  // Default 720.  Use 360 for benchmarking VGA.
-  int benchmark_pixels_div256_;  // Total pixels to benchmark / 256.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
   int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
-  int disable_cpu_flags_;  // Default 1.  Use -1 for benchmarking.
-  int benchmark_cpu_info_;  // Default -1.  Use 1 to disable SIMD.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
 };
 
 class LibYUVRotateTest : public ::testing::Test {
@@ -162,12 +177,11 @@ class LibYUVRotateTest : public ::testing::Test {
   LibYUVRotateTest();
 
   int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
-  int benchmark_width_;  // Default 1280.  Use 640 for benchmarking VGA.
-  int benchmark_height_;  // Default 720.  Use 360 for benchmarking VGA.
-  int benchmark_pixels_div256_;  // Total pixels to benchmark / 256.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
   int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
-  int disable_cpu_flags_;  // Default 1.  Use -1 for benchmarking.
-  int benchmark_cpu_info_;  // Default -1.  Use 1 to disable SIMD.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
 };
 
 class LibYUVPlanarTest : public ::testing::Test {
@@ -175,12 +189,11 @@ class LibYUVPlanarTest : public ::testing::Test {
   LibYUVPlanarTest();
 
   int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
-  int benchmark_width_;  // Default 1280.  Use 640 for benchmarking VGA.
-  int benchmark_height_;  // Default 720.  Use 360 for benchmarking VGA.
-  int benchmark_pixels_div256_;  // Total pixels to benchmark / 256.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
   int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
-  int disable_cpu_flags_;  // Default 1.  Use -1 for benchmarking.
-  int benchmark_cpu_info_;  // Default -1.  Use 1 to disable SIMD.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
 };
 
 class LibYUVBaseTest : public ::testing::Test {
@@ -188,12 +201,23 @@ class LibYUVBaseTest : public ::testing::Test {
   LibYUVBaseTest();
 
   int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
-  int benchmark_width_;  // Default 1280.  Use 640 for benchmarking VGA.
-  int benchmark_height_;  // Default 720.  Use 360 for benchmarking VGA.
-  int benchmark_pixels_div256_;  // Total pixels to benchmark / 256.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
+  int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
+};
+
+class LibYUVCompareTest : public ::testing::Test {
+ protected:
+  LibYUVCompareTest();
+
+  int benchmark_iterations_;  // Default 1. Use 1000 for benchmarking.
+  int benchmark_width_;       // Default 1280.  Use 640 for benchmarking VGA.
+  int benchmark_height_;      // Default 720.  Use 360 for benchmarking VGA.
   int benchmark_pixels_div1280_;  // Total pixels to benchmark / 1280.
-  int disable_cpu_flags_;  // Default 1.  Use -1 for benchmarking.
-  int benchmark_cpu_info_;  // Default -1.  Use 1 to disable SIMD.
+  int disable_cpu_flags_;         // Default 1.  Use -1 for benchmarking.
+  int benchmark_cpu_info_;        // Default -1.  Use 1 to disable SIMD.
 };
 
 #endif  // UNIT_TEST_UNIT_TEST_H_  NOLINT
diff --git a/media/libyuv/unit_test/video_common_test.cc b/media/libyuv/unit_test/video_common_test.cc
index ac97d0f34f..36728ea900 100644
--- a/media/libyuv/unit_test/video_common_test.cc
+++ b/media/libyuv/unit_test/video_common_test.cc
@@ -11,54 +11,48 @@
 #include <stdlib.h>
 #include <string.h>
 
-#include "libyuv/video_common.h"
 #include "../unit_test/unit_test.h"
+#include "libyuv/video_common.h"
 
 namespace libyuv {
 
 // Tests FourCC codes in video common, which are used for ConvertToI420().
 
-static bool TestValidChar(uint32 onecc) {
-  if ((onecc >= '0' && onecc <= '9') ||
-      (onecc >= 'A' && onecc <= 'Z') ||
-      (onecc >= 'a' && onecc <= 'z') ||
-      (onecc == ' ') || (onecc == 0xff)) {
-    return true;
-  }
-  return false;
+static bool TestValidChar(uint32_t onecc) {
+  return (onecc >= '0' && onecc <= '9') || (onecc >= 'A' && onecc <= 'Z') ||
+         (onecc >= 'a' && onecc <= 'z') || (onecc == ' ') || (onecc == 0xff);
 }
 
-static bool TestValidFourCC(uint32 fourcc, int bpp) {
-  if (!TestValidChar(fourcc & 0xff) ||
-      !TestValidChar((fourcc >> 8) & 0xff) ||
+static bool TestValidFourCC(uint32_t fourcc, int bpp) {
+  if (!TestValidChar(fourcc & 0xff) || !TestValidChar((fourcc >> 8) & 0xff) ||
       !TestValidChar((fourcc >> 16) & 0xff) ||
       !TestValidChar((fourcc >> 24) & 0xff)) {
     return false;
   }
-  if (bpp < 0 || bpp > 32) {
+  if (bpp < 0 || bpp > 64) {
     return false;
   }
   return true;
 }
 
 TEST_F(LibYUVBaseTest, TestCanonicalFourCC) {
-  EXPECT_EQ(FOURCC_I420, CanonicalFourCC(FOURCC_IYUV));
-  EXPECT_EQ(FOURCC_I420, CanonicalFourCC(FOURCC_YU12));
-  EXPECT_EQ(FOURCC_I422, CanonicalFourCC(FOURCC_YU16));
-  EXPECT_EQ(FOURCC_I444, CanonicalFourCC(FOURCC_YU24));
-  EXPECT_EQ(FOURCC_YUY2, CanonicalFourCC(FOURCC_YUYV));
-  EXPECT_EQ(FOURCC_YUY2, CanonicalFourCC(FOURCC_YUVS));
-  EXPECT_EQ(FOURCC_UYVY, CanonicalFourCC(FOURCC_HDYC));
-  EXPECT_EQ(FOURCC_UYVY, CanonicalFourCC(FOURCC_2VUY));
-  EXPECT_EQ(FOURCC_MJPG, CanonicalFourCC(FOURCC_JPEG));
-  EXPECT_EQ(FOURCC_MJPG, CanonicalFourCC(FOURCC_DMB1));
-  EXPECT_EQ(FOURCC_RAW,  CanonicalFourCC(FOURCC_RGB3));
-  EXPECT_EQ(FOURCC_24BG, CanonicalFourCC(FOURCC_BGR3));
-  EXPECT_EQ(FOURCC_BGRA, CanonicalFourCC(FOURCC_CM32));
-  EXPECT_EQ(FOURCC_RAW,  CanonicalFourCC(FOURCC_CM24));
-  EXPECT_EQ(FOURCC_RGBO, CanonicalFourCC(FOURCC_L555));
-  EXPECT_EQ(FOURCC_RGBP, CanonicalFourCC(FOURCC_L565));
-  EXPECT_EQ(FOURCC_RGBO, CanonicalFourCC(FOURCC_5551));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_I420), CanonicalFourCC(FOURCC_IYUV));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_I420), CanonicalFourCC(FOURCC_YU12));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_I422), CanonicalFourCC(FOURCC_YU16));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_I444), CanonicalFourCC(FOURCC_YU24));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_YUY2), CanonicalFourCC(FOURCC_YUYV));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_YUY2), CanonicalFourCC(FOURCC_YUVS));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_UYVY), CanonicalFourCC(FOURCC_HDYC));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_UYVY), CanonicalFourCC(FOURCC_2VUY));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_MJPG), CanonicalFourCC(FOURCC_JPEG));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_MJPG), CanonicalFourCC(FOURCC_DMB1));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_RAW), CanonicalFourCC(FOURCC_RGB3));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_24BG), CanonicalFourCC(FOURCC_BGR3));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_BGRA), CanonicalFourCC(FOURCC_CM32));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_RAW), CanonicalFourCC(FOURCC_CM24));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_RGBO), CanonicalFourCC(FOURCC_L555));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_RGBP), CanonicalFourCC(FOURCC_L565));
+  EXPECT_EQ(static_cast<uint32_t>(FOURCC_RGBO), CanonicalFourCC(FOURCC_5551));
 }
 
 TEST_F(LibYUVBaseTest, TestFourCC) {
@@ -66,23 +60,34 @@ TEST_F(LibYUVBaseTest, TestFourCC) {
   EXPECT_TRUE(TestValidFourCC(FOURCC_I420, FOURCC_BPP_I420));
   EXPECT_TRUE(TestValidFourCC(FOURCC_I422, FOURCC_BPP_I422));
   EXPECT_TRUE(TestValidFourCC(FOURCC_I444, FOURCC_BPP_I444));
-  EXPECT_TRUE(TestValidFourCC(FOURCC_I411, FOURCC_BPP_I411));
   EXPECT_TRUE(TestValidFourCC(FOURCC_I400, FOURCC_BPP_I400));
   EXPECT_TRUE(TestValidFourCC(FOURCC_NV21, FOURCC_BPP_NV21));
   EXPECT_TRUE(TestValidFourCC(FOURCC_NV12, FOURCC_BPP_NV12));
   EXPECT_TRUE(TestValidFourCC(FOURCC_YUY2, FOURCC_BPP_YUY2));
   EXPECT_TRUE(TestValidFourCC(FOURCC_UYVY, FOURCC_BPP_UYVY));
-  EXPECT_TRUE(TestValidFourCC(FOURCC_M420, FOURCC_BPP_M420));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_M420, FOURCC_BPP_M420));  // deprecated.
   EXPECT_TRUE(TestValidFourCC(FOURCC_Q420, FOURCC_BPP_Q420));  // deprecated.
   EXPECT_TRUE(TestValidFourCC(FOURCC_ARGB, FOURCC_BPP_ARGB));
   EXPECT_TRUE(TestValidFourCC(FOURCC_BGRA, FOURCC_BPP_BGRA));
   EXPECT_TRUE(TestValidFourCC(FOURCC_ABGR, FOURCC_BPP_ABGR));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_AR30, FOURCC_BPP_AR30));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_AB30, FOURCC_BPP_AB30));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_AR64, FOURCC_BPP_AR64));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_AB64, FOURCC_BPP_AB64));
   EXPECT_TRUE(TestValidFourCC(FOURCC_24BG, FOURCC_BPP_24BG));
-  EXPECT_TRUE(TestValidFourCC(FOURCC_RAW,  FOURCC_BPP_RAW));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_RAW, FOURCC_BPP_RAW));
   EXPECT_TRUE(TestValidFourCC(FOURCC_RGBA, FOURCC_BPP_RGBA));
   EXPECT_TRUE(TestValidFourCC(FOURCC_RGBP, FOURCC_BPP_RGBP));
   EXPECT_TRUE(TestValidFourCC(FOURCC_RGBO, FOURCC_BPP_RGBO));
   EXPECT_TRUE(TestValidFourCC(FOURCC_R444, FOURCC_BPP_R444));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_H420, FOURCC_BPP_H420));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_H422, FOURCC_BPP_H422));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_H010, FOURCC_BPP_H010));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_H210, FOURCC_BPP_H210));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_I010, FOURCC_BPP_I010));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_I210, FOURCC_BPP_I210));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_P010, FOURCC_BPP_P010));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_P210, FOURCC_BPP_P210));
   EXPECT_TRUE(TestValidFourCC(FOURCC_MJPG, FOURCC_BPP_MJPG));
   EXPECT_TRUE(TestValidFourCC(FOURCC_YV12, FOURCC_BPP_YV12));
   EXPECT_TRUE(TestValidFourCC(FOURCC_YV16, FOURCC_BPP_YV16));
@@ -101,7 +106,7 @@ TEST_F(LibYUVBaseTest, TestFourCC) {
   EXPECT_TRUE(TestValidFourCC(FOURCC_RGB3, FOURCC_BPP_RGB3));
   EXPECT_TRUE(TestValidFourCC(FOURCC_BGR3, FOURCC_BPP_BGR3));
   EXPECT_TRUE(TestValidFourCC(FOURCC_H264, FOURCC_BPP_H264));
-  EXPECT_TRUE(TestValidFourCC(FOURCC_ANY,  FOURCC_BPP_ANY));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_ANY, FOURCC_BPP_ANY));
 }
 
 }  // namespace libyuv
diff --git a/media/libyuv/util/Makefile b/media/libyuv/util/Makefile
index 6044d2adf6..40e74b6509 100644
--- a/media/libyuv/util/Makefile
+++ b/media/libyuv/util/Makefile
@@ -4,3 +4,6 @@ ifeq ($(CXX),icl)
 else
 	$(CXX) -msse2 -O3 -fopenmp -static -o psnr psnr.cc ssim.cc psnr_main.cc -Wl,--strip-all
 endif
+
+# for MacOS
+# /usr/local/bin/g++-7 -msse2 -O3 -fopenmp -Bstatic -o psnr psnr.cc ssim.cc psnr_main.cc
diff --git a/media/libyuv/util/android/test_runner.py b/media/libyuv/util/android/test_runner.py
deleted file mode 100644
index 8b06b7eabe..0000000000
--- a/media/libyuv/util/android/test_runner.py
+++ /dev/null
@@ -1,37 +0,0 @@
-#!/usr/bin/env python
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""
-Runs tests on Android devices.
-
-This script exists to avoid Libyuv being broken by changes in the Chrome Android
-test execution toolchain. It also conveniently sets the CHECKOUT_SOURCE_ROOT
-environment variable.
-"""
-
-import os
-import sys
-
-SCRIPT_DIR = os.path.dirname(__file__)
-ROOT_DIR = os.path.abspath(os.path.join(SCRIPT_DIR, os.pardir, os.pardir))
-CHROMIUM_BUILD_ANDROID_DIR = os.path.join(ROOT_DIR, 'build', 'android')
-sys.path.insert(0, CHROMIUM_BUILD_ANDROID_DIR)
-
-
-import test_runner  # pylint: disable=W0406
-
-def main():
-  # Override environment variable to make it possible for the scripts to find
-  # the root directory (our symlinking of the Chromium build toolchain would
-  # otherwise make them fail to do so).
-  os.environ['CHECKOUT_SOURCE_ROOT'] = ROOT_DIR
-  return test_runner.main()
-
-if __name__ == '__main__':
-  sys.exit(main())
diff --git a/media/libyuv/util/color.cc b/media/libyuv/util/color.cc
new file mode 100644
index 0000000000..8c3bbefd27
--- /dev/null
+++ b/media/libyuv/util/color.cc
@@ -0,0 +1,120 @@
+/*
+ *  Copyright 2021 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+// This utility computes values needed to generate yuvconstants based on
+// white point values.
+// The yuv formulas are tuned for 8 bit YUV channels.
+
+// For those MCs that can be represented as kr and kb:
+// Full range
+// float M[3][3]
+// {{1,0,2*(1-kr)},{1,-((2*kb)/((2-kb)*(1-kb-kr))),-((2*kr)/((2-kr)*(1-kb-kr)))},{1,2*(1-kb),0}};
+// float B[3]
+// {1+(256*(1-kr))/255,1-(256*kb)/(255*(2-kb)*(1-kb-kr))-(256*kr)/(255*(2-kr)*(1-kb-kr)),1+(256*(1-kb))/255};
+// Limited range
+// float M[3][3]
+// {{85/73,0,255/112-(255*kr)/112},{85/73,-((255*kb)/(112*(2-kb)*(1-kb-kr))),-((255*kr)/(112*(2-kr)*(1-kb-kr)))},{85/73,255/112-(255*kb)/112,0}};
+// float B[3]
+// {77662/43435-(1537*kr)/1785,203/219-(1537*kb)/(1785*(2-kb)*(1-kb-kr))-(1537*kr)/(1785*(2-kr)*(1-kb-kr)),77662/43435-(1537*kb)/1785};
+
+// mc bt
+// 1 bt.709      KR = 0.2126; KB = 0.0722
+// 4 fcc         KR = 0.30;   KB = 0.11
+// 6 bt.601      KR = 0.299;  KB = 0.114
+// 7 SMPTE 240M  KR = 0.212;  KB = 0.087
+// 10 bt2020     KR = 0.2627; KB = 0.0593
+
+// BT.709 full range YUV to RGB reference
+//  R = Y               + V * 1.5748
+//  G = Y - U * 0.18732 - V * 0.46812
+//  B = Y + U * 1.8556
+//  KR = 0.2126
+//  KB = 0.0722
+
+// https://mymusing.co/bt601-yuv-to-rgb-conversion-color/
+
+// // Y contribution to R,G,B.  Scale and bias.
+// #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+// #define YB 32    /* 64 / 2 */
+//
+// // U and V contributions to R,G,B.
+// #define UB 113 /* round(1.77200 * 64) */
+// #define UG 22  /* round(0.34414 * 64) */
+// #define VG 46  /* round(0.71414 * 64) */
+// #define VR 90  /* round(1.40200 * 64) */
+//
+// // Bias values to round, and subtract 128 from U and V.
+// #define BB (-UB * 128 + YB)
+// #define BG (UG * 128 + VG * 128 + YB)
+// #define BR (-VR * 128 + YB)
+
+int round(float v) {
+  return (int)(v + 0.5);
+}
+
+int main(int argc, const char* argv[]) {
+  if (argc < 2) {
+    printf("color kr kb\n");
+    return -1;
+  }
+  float kr = atof(argv[1]);
+  float kb = atof(argv[2]);
+  float kg = 1 - kr - kb;
+
+  float vr = 2 * (1 - kr);
+  float ug = 2 * ((1 - kb) * kb / kg);
+  float vg = 2 * ((1 - kr) * kr / kg);
+  float ub = 2 * (1 - kb);
+
+  printf("Full range\n");
+  printf("R = Y                + V * %5f\n", vr);
+  printf("G = Y - U * %6f - V * %6f\n", ug, vg);
+  printf("B = Y + U * %5f\n", ub);
+
+  printf("KR = %4f; ", kr);
+  printf("KB = %4f\n", kb);
+  //  printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3d /* round(%f * 64) */\n", round(ub * 64), ub);
+  printf("UG %-3d /* round(%f * 64) */\n", round(ug * 64), ug);
+  printf("VG %-3d /* round(%f * 64) */\n", round(vg * 64), vg);
+  printf("VR %-3d /* round(%f * 64) */\n", round(vr * 64), vr);
+
+  vr = 255.f / 224.f * 2 * (1 - kr);
+  ug = 255.f / 224.f * 2 * ((1 - kb) * kb / kg);
+  vg = 255.f / 224.f * 2 * ((1 - kr) * kr / kg);
+  ub = 255.f / 224.f * 2 * (1 - kb);
+
+  printf("Limited range\n");
+  printf("R = (Y - 16) * 1.164                + V * %5f\n", vr);
+  printf("G = (Y - 16) * 1.164 - U * %6f - V * %6f\n", ug, vg);
+  printf("B = (Y - 16) * 1.164 + U * %5f\n", ub);
+
+  //  printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3d /* round(%f * 64) */\n", round(ub * 64), ub);
+  printf("UG %-3d /* round(%f * 64) */\n", round(ug * 64), ug);
+  printf("VG %-3d /* round(%f * 64) */\n", round(vg * 64), vg);
+  printf("VR %-3d /* round(%f * 64) */\n", round(vr * 64), vr);
+
+  return 0;
+}
diff --git a/media/libyuv/util/compare.cc b/media/libyuv/util/compare.cc
index c36c0fa5f3..a16613ee2f 100644
--- a/media/libyuv/util/compare.cc
+++ b/media/libyuv/util/compare.cc
@@ -29,20 +29,24 @@ int main(int argc, char** argv) {
   FILE* fin2 = name2 ? fopen(name2, "rb") : NULL;
 
   const int kBlockSize = 32768;
-  uint8 buf1[kBlockSize];
-  uint8 buf2[kBlockSize];
-  uint32 hash1 = 5381;
-  uint32 hash2 = 5381;
-  uint64 sum_square_err = 0;
-  uint64 size_min = 0;
+  uint8_t buf1[kBlockSize];
+  uint8_t buf2[kBlockSize];
+  uint32_t hash1 = 5381;
+  uint32_t hash2 = 5381;
+  uint64_t sum_square_err = 0;
+  uint64_t size_min = 0;
   int amt1 = 0;
   int amt2 = 0;
   do {
     amt1 = static_cast<int>(fread(buf1, 1, kBlockSize, fin1));
-    if (amt1 > 0) hash1 = libyuv::HashDjb2(buf1, amt1, hash1);
+    if (amt1 > 0) {
+      hash1 = libyuv::HashDjb2(buf1, amt1, hash1);
+    }
     if (fin2) {
       amt2 = static_cast<int>(fread(buf2, 1, kBlockSize, fin2));
-      if (amt2 > 0) hash2 = libyuv::HashDjb2(buf2, amt2, hash2);
+      if (amt2 > 0) {
+        hash2 = libyuv::HashDjb2(buf2, amt2, hash2);
+      }
       int amt_min = (amt1 < amt2) ? amt1 : amt2;
       size_min += amt_min;
       sum_square_err += libyuv::ComputeSumSquareError(buf1, buf2, amt_min);
@@ -52,8 +56,8 @@ int main(int argc, char** argv) {
   printf("hash1 %x", hash1);
   if (fin2) {
     printf(", hash2 %x", hash2);
-    double mse = static_cast<double>(sum_square_err) /
-                 static_cast<double>(size_min);
+    double mse =
+        static_cast<double>(sum_square_err) / static_cast<double>(size_min);
     printf(", mse %.2f", mse);
     double psnr = libyuv::SumSquareErrorToPsnr(sum_square_err, size_min);
     printf(", psnr %.2f\n", psnr);
diff --git a/media/libyuv/util/cpuid.c b/media/libyuv/util/cpuid.c
index 94e245b11d..b618bb106e 100644
--- a/media/libyuv/util/cpuid.c
+++ b/media/libyuv/util/cpuid.c
@@ -12,20 +12,25 @@
 #include <stdlib.h>
 #include <string.h>
 
-#define INCLUDE_LIBYUV_COMPARE_H_
-#include "libyuv.h"
-#include "./psnr.h"
-#include "./ssim.h"
+#include "libyuv/cpu_id.h"
+
+#ifdef __cplusplus
+using namespace libyuv;
+#endif
 
 int main(int argc, const char* argv[]) {
   int cpu_flags = TestCpuFlag(-1);
   int has_arm = TestCpuFlag(kCpuHasARM);
   int has_mips = TestCpuFlag(kCpuHasMIPS);
   int has_x86 = TestCpuFlag(kCpuHasX86);
+  int has_loongarch = TestCpuFlag(kCpuHasLOONGARCH);
+  (void)argc;
+  (void)argv;
+
 #if defined(__i386__) || defined(__x86_64__) || \
     defined(_M_IX86) || defined(_M_X64)
   if (has_x86) {
-    uint32 family, model, cpu_info[4];
+    int family, model, cpu_info[4];
     // Vendor ID:
     // AuthenticAMD AMD processor
     // CentaurHauls Centaur processor
@@ -61,13 +66,20 @@ int main(int argc, const char* argv[]) {
   printf("Has ARM %x\n", has_arm);
   printf("Has MIPS %x\n", has_mips);
   printf("Has X86 %x\n", has_x86);
+  printf("Has LOONGARCH %x\n", has_loongarch);
   if (has_arm) {
     int has_neon = TestCpuFlag(kCpuHasNEON);
     printf("Has NEON %x\n", has_neon);
   }
   if (has_mips) {
-    int has_dspr2 = TestCpuFlag(kCpuHasDSPR2);
-    printf("Has DSPR2 %x\n", has_dspr2);
+    int has_msa = TestCpuFlag(kCpuHasMSA);
+    printf("Has MSA %x\n", has_msa);
+  }
+  if (has_loongarch) {
+    int has_lsx  = TestCpuFlag(kCpuHasLSX);
+    printf("Has LSX %x\n", has_lsx);
+    int has_lasx = TestCpuFlag(kCpuHasLASX);
+    printf("Has LASX %x\n", has_lasx);
   }
   if (has_x86) {
     int has_sse2 = TestCpuFlag(kCpuHasSSE2);
@@ -76,18 +88,34 @@ int main(int argc, const char* argv[]) {
     int has_sse42 = TestCpuFlag(kCpuHasSSE42);
     int has_avx = TestCpuFlag(kCpuHasAVX);
     int has_avx2 = TestCpuFlag(kCpuHasAVX2);
-    int has_avx3 = TestCpuFlag(kCpuHasAVX3);
     int has_erms = TestCpuFlag(kCpuHasERMS);
     int has_fma3 = TestCpuFlag(kCpuHasFMA3);
+    int has_f16c = TestCpuFlag(kCpuHasF16C);
+    int has_gfni = TestCpuFlag(kCpuHasGFNI);
+    int has_avx512bw = TestCpuFlag(kCpuHasAVX512BW);
+    int has_avx512vl = TestCpuFlag(kCpuHasAVX512VL);
+    int has_avx512vnni = TestCpuFlag(kCpuHasAVX512VNNI);
+    int has_avx512vbmi = TestCpuFlag(kCpuHasAVX512VBMI);
+    int has_avx512vbmi2 = TestCpuFlag(kCpuHasAVX512VBMI2);
+    int has_avx512vbitalg = TestCpuFlag(kCpuHasAVX512VBITALG);
+    int has_avx512vpopcntdq = TestCpuFlag(kCpuHasAVX512VPOPCNTDQ);
     printf("Has SSE2 %x\n", has_sse2);
     printf("Has SSSE3 %x\n", has_ssse3);
     printf("Has SSE4.1 %x\n", has_sse41);
     printf("Has SSE4.2 %x\n", has_sse42);
     printf("Has AVX %x\n", has_avx);
     printf("Has AVX2 %x\n", has_avx2);
-    printf("Has AVX3 %x\n", has_avx3);
     printf("Has ERMS %x\n", has_erms);
     printf("Has FMA3 %x\n", has_fma3);
+    printf("Has F16C %x\n", has_f16c);
+    printf("Has GFNI %x\n", has_gfni);
+    printf("Has AVX512BW %x\n", has_avx512bw);
+    printf("Has AVX512VL %x\n", has_avx512vl);
+    printf("Has AVX512VNNI %x\n", has_avx512vnni);
+    printf("Has AVX512VBMI %x\n", has_avx512vbmi);
+    printf("Has AVX512VBMI2 %x\n", has_avx512vbmi2);
+    printf("Has AVX512VBITALG %x\n", has_avx512vbitalg);
+    printf("Has AVX512VPOPCNTDQ %x\n", has_avx512vpopcntdq);
   }
   return 0;
 }
diff --git a/media/libyuv/util/i444tonv12_eg.cc b/media/libyuv/util/i444tonv12_eg.cc
new file mode 100644
index 0000000000..0fcb4095a8
--- /dev/null
+++ b/media/libyuv/util/i444tonv12_eg.cc
@@ -0,0 +1,28 @@
+
+#include "libyuv/convert.h"
+
+#include <stdio.h>   // for printf
+#include <string.h>  // for memset
+
+int main(int, char**) {
+  unsigned char src_i444[640 * 400 * 3];
+  unsigned char dst_nv12[640 * 400 * 3 / 2];
+
+  for (size_t i = 0; i < sizeof(src_i444); ++i) {
+    src_i444[i] = i & 255;
+  }
+  memset(dst_nv12, 0, sizeof(dst_nv12));
+  libyuv::I444ToNV12(&src_i444[0], 640,              // source Y
+                     &src_i444[640 * 400], 640,      // source U
+                     &src_i444[640 * 400 * 2], 640,  // source V
+                     &dst_nv12[0], 640,              // dest Y
+                     &dst_nv12[640 * 400], 640,      // dest UV
+                     640, 400);                      // width and height
+
+  int checksum = 0;
+  for (size_t i = 0; i < sizeof(dst_nv12); ++i) {
+    checksum += dst_nv12[i];
+  }
+  printf("checksum %x %s\n", checksum, checksum == 0x2ec0c00 ? "PASS" : "FAIL");
+  return 0;
+}
\ No newline at end of file
diff --git a/media/libyuv/util/psnr.cc b/media/libyuv/util/psnr.cc
index 52b04bd516..c7bee7f97d 100644
--- a/media/libyuv/util/psnr.cc
+++ b/media/libyuv/util/psnr.cc
@@ -21,14 +21,14 @@
 extern "C" {
 #endif
 
-typedef unsigned int uint32;  // NOLINT
+typedef unsigned int uint32_t;  // NOLINT
 #ifdef _MSC_VER
-typedef unsigned __int64 uint64;
+typedef unsigned __int64 uint64_t;
 #else  // COMPILER_MSVC
 #if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
-typedef unsigned long uint64;  // NOLINT
-#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
-typedef unsigned long long uint64;  // NOLINT
+typedef unsigned long uint64_t;  // NOLINT
+#else   // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long long uint64_t;  // NOLINT
 #endif  // __LP64__
 #endif  // _MSC_VER
 
@@ -38,86 +38,82 @@ typedef unsigned long long uint64;  // NOLINT
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
     !defined(__aarch64__)
 #define HAS_SUMSQUAREERROR_NEON
-static uint32 SumSquareError_NEON(const uint8* src_a,
-                                  const uint8* src_b, int count) {
-  volatile uint32 sse;
-  asm volatile (
-    "vmov.u8    q7, #0                         \n"
-    "vmov.u8    q9, #0                         \n"
-    "vmov.u8    q8, #0                         \n"
-    "vmov.u8    q10, #0                        \n"
+static uint32_t SumSquareError_NEON(const uint8_t* src_a,
+                                    const uint8_t* src_b,
+                                    int count) {
+  volatile uint32_t sse;
+  asm volatile(
+      "vmov.u8    q7, #0                         \n"
+      "vmov.u8    q9, #0                         \n"
+      "vmov.u8    q8, #0                         \n"
+      "vmov.u8    q10, #0                        \n"
 
-  "1:                                          \n"
-    "vld1.u8    {q0}, [%0]!                    \n"
-    "vld1.u8    {q1}, [%1]!                    \n"
-    "vsubl.u8   q2, d0, d2                     \n"
-    "vsubl.u8   q3, d1, d3                     \n"
-    "vmlal.s16  q7, d4, d4                     \n"
-    "vmlal.s16  q8, d6, d6                     \n"
-    "vmlal.s16  q8, d5, d5                     \n"
-    "vmlal.s16  q10, d7, d7                    \n"
-    "subs       %2, %2, #16                    \n"
-    "bhi        1b                             \n"
+      "1:                                        \n"
+      "vld1.u8    {q0}, [%0]!                    \n"
+      "vld1.u8    {q1}, [%1]!                    \n"
+      "vsubl.u8   q2, d0, d2                     \n"
+      "vsubl.u8   q3, d1, d3                     \n"
+      "vmlal.s16  q7, d4, d4                     \n"
+      "vmlal.s16  q8, d6, d6                     \n"
+      "vmlal.s16  q8, d5, d5                     \n"
+      "vmlal.s16  q10, d7, d7                    \n"
+      "subs       %2, %2, #16                    \n"
+      "bhi        1b                             \n"
 
-    "vadd.u32   q7, q7, q8                     \n"
-    "vadd.u32   q9, q9, q10                    \n"
-    "vadd.u32   q10, q7, q9                    \n"
-    "vpaddl.u32 q1, q10                        \n"
-    "vadd.u64   d0, d2, d3                     \n"
-    "vmov.32    %3, d0[0]                      \n"
-    : "+r"(src_a),
-      "+r"(src_b),
-      "+r"(count),
-      "=r"(sse)
-    :
-    : "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
+      "vadd.u32   q7, q7, q8                     \n"
+      "vadd.u32   q9, q9, q10                    \n"
+      "vadd.u32   q10, q7, q9                    \n"
+      "vpaddl.u32 q1, q10                        \n"
+      "vadd.u64   d0, d2, d3                     \n"
+      "vmov.32    %3, d0[0]                      \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
   return sse;
 }
 #elif !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 #define HAS_SUMSQUAREERROR_NEON
-static uint32 SumSquareError_NEON(const uint8* src_a,
-                                  const uint8* src_b, int count) {
-  volatile uint32 sse;
-  asm volatile (
-    "eor        v16.16b, v16.16b, v16.16b      \n"
-    "eor        v18.16b, v18.16b, v18.16b      \n"
-    "eor        v17.16b, v17.16b, v17.16b      \n"
-    "eor        v19.16b, v19.16b, v19.16b      \n"
+static uint32_t SumSquareError_NEON(const uint8_t* src_a,
+                                    const uint8_t* src_b,
+                                    int count) {
+  volatile uint32_t sse;
+  asm volatile(
+      "eor        v16.16b, v16.16b, v16.16b      \n"
+      "eor        v18.16b, v18.16b, v18.16b      \n"
+      "eor        v17.16b, v17.16b, v17.16b      \n"
+      "eor        v19.16b, v19.16b, v19.16b      \n"
 
-  "1:                                          \n"
-    "ld1        {v0.16b}, [%0], #16            \n"
-    "ld1        {v1.16b}, [%1], #16            \n"
-    "subs       %w2, %w2, #16                  \n"
-    "usubl      v2.8h, v0.8b, v1.8b            \n"
-    "usubl2     v3.8h, v0.16b, v1.16b          \n"
-    "smlal      v16.4s, v2.4h, v2.4h           \n"
-    "smlal      v17.4s, v3.4h, v3.4h           \n"
-    "smlal2     v18.4s, v2.8h, v2.8h           \n"
-    "smlal2     v19.4s, v3.8h, v3.8h           \n"
-    "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld1        {v0.16b}, [%0], #16            \n"
+      "ld1        {v1.16b}, [%1], #16            \n"
+      "subs       %w2, %w2, #16                  \n"
+      "usubl      v2.8h, v0.8b, v1.8b            \n"
+      "usubl2     v3.8h, v0.16b, v1.16b          \n"
+      "smlal      v16.4s, v2.4h, v2.4h           \n"
+      "smlal      v17.4s, v3.4h, v3.4h           \n"
+      "smlal2     v18.4s, v2.8h, v2.8h           \n"
+      "smlal2     v19.4s, v3.8h, v3.8h           \n"
+      "b.gt       1b                             \n"
 
-    "add        v16.4s, v16.4s, v17.4s         \n"
-    "add        v18.4s, v18.4s, v19.4s         \n"
-    "add        v19.4s, v16.4s, v18.4s         \n"
-    "addv       s0, v19.4s                     \n"
-    "fmov       %w3, s0                        \n"
-    : "+r"(src_a),
-      "+r"(src_b),
-      "+r"(count),
-      "=r"(sse)
-    :
-    : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
+      "add        v16.4s, v16.4s, v17.4s         \n"
+      "add        v18.4s, v18.4s, v19.4s         \n"
+      "add        v19.4s, v16.4s, v18.4s         \n"
+      "addv       s0, v19.4s                     \n"
+      "fmov       %w3, s0                        \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
+      :
+      : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
   return sse;
 }
 #elif !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
 #define HAS_SUMSQUAREERROR_SSE2
-__declspec(naked)
-static uint32 SumSquareError_SSE2(const uint8* /*src_a*/,
-                                  const uint8* /*src_b*/, int /*count*/) {
+__declspec(naked) static uint32_t SumSquareError_SSE2(const uint8_t* /*src_a*/,
+                                                      const uint8_t* /*src_b*/,
+                                                      int /*count*/) {
   __asm {
-    mov        eax, [esp + 4]    // src_a
-    mov        edx, [esp + 8]    // src_b
-    mov        ecx, [esp + 12]   // count
+    mov        eax, [esp + 4]  // src_a
+    mov        edx, [esp + 8]  // src_b
+    mov        ecx, [esp + 12]  // count
     pxor       xmm0, xmm0
     pxor       xmm5, xmm5
     sub        edx, eax
@@ -150,46 +146,48 @@ static uint32 SumSquareError_SSE2(const uint8* /*src_a*/,
 }
 #elif !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 #define HAS_SUMSQUAREERROR_SSE2
-static uint32 SumSquareError_SSE2(const uint8* src_a,
-                                  const uint8* src_b, int count) {
-  uint32 sse;
-  asm volatile (  // NOLINT
-    "pxor      %%xmm0,%%xmm0                   \n"
-    "pxor      %%xmm5,%%xmm5                   \n"
-    "sub       %0,%1                           \n"
+static uint32_t SumSquareError_SSE2(const uint8_t* src_a,
+                                    const uint8_t* src_b,
+                                    int count) {
+  uint32_t sse;
+  asm volatile(  // NOLINT
+      "pxor      %%xmm0,%%xmm0                   \n"
+      "pxor      %%xmm5,%%xmm5                   \n"
+      "sub       %0,%1                           \n"
 
-  "1:                                          \n"
-    "movdqu    (%0),%%xmm1                     \n"
-    "movdqu    (%0,%1,1),%%xmm2                \n"
-    "lea       0x10(%0),%0                     \n"
-    "movdqu    %%xmm1,%%xmm3                   \n"
-    "psubusb   %%xmm2,%%xmm1                   \n"
-    "psubusb   %%xmm3,%%xmm2                   \n"
-    "por       %%xmm2,%%xmm1                   \n"
-    "movdqu    %%xmm1,%%xmm2                   \n"
-    "punpcklbw %%xmm5,%%xmm1                   \n"
-    "punpckhbw %%xmm5,%%xmm2                   \n"
-    "pmaddwd   %%xmm1,%%xmm1                   \n"
-    "pmaddwd   %%xmm2,%%xmm2                   \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "paddd     %%xmm2,%%xmm0                   \n"
-    "sub       $0x10,%2                        \n"
-    "ja        1b                              \n"
+      "1:                                        \n"
+      "movdqu    (%0),%%xmm1                     \n"
+      "movdqu    (%0,%1,1),%%xmm2                \n"
+      "lea       0x10(%0),%0                     \n"
+      "movdqu    %%xmm1,%%xmm3                   \n"
+      "psubusb   %%xmm2,%%xmm1                   \n"
+      "psubusb   %%xmm3,%%xmm2                   \n"
+      "por       %%xmm2,%%xmm1                   \n"
+      "movdqu    %%xmm1,%%xmm2                   \n"
+      "punpcklbw %%xmm5,%%xmm1                   \n"
+      "punpckhbw %%xmm5,%%xmm2                   \n"
+      "pmaddwd   %%xmm1,%%xmm1                   \n"
+      "pmaddwd   %%xmm2,%%xmm2                   \n"
+      "paddd     %%xmm1,%%xmm0                   \n"
+      "paddd     %%xmm2,%%xmm0                   \n"
+      "sub       $0x10,%2                        \n"
+      "ja        1b                              \n"
 
-    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
-    "paddd     %%xmm1,%%xmm0                   \n"
-    "movd      %%xmm0,%3                       \n"
+      "pshufd    $0xee,%%xmm0,%%xmm1             \n"
+      "paddd     %%xmm1,%%xmm0                   \n"
+      "pshufd    $0x1,%%xmm0,%%xmm1              \n"
+      "paddd     %%xmm1,%%xmm0                   \n"
+      "movd      %%xmm0,%3                       \n"
 
-  : "+r"(src_a),      // %0
-    "+r"(src_b),      // %1
-    "+r"(count),      // %2
-    "=g"(sse)         // %3
-  :
-  : "memory", "cc"
+      : "+r"(src_a),  // %0
+        "+r"(src_b),  // %1
+        "+r"(count),  // %2
+        "=g"(sse)     // %3
+      :
+      : "memory", "cc"
 #if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+        ,
+        "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
 #endif
   );  // NOLINT
   return sse;
@@ -199,20 +197,22 @@ static uint32 SumSquareError_SSE2(const uint8* src_a,
 #if defined(HAS_SUMSQUAREERROR_SSE2)
 #if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
 static __inline void __cpuid(int cpu_info[4], int info_type) {
-  asm volatile (  // NOLINT
-    "mov %%ebx, %%edi                          \n"
-    "cpuid                                     \n"
-    "xchg %%edi, %%ebx                         \n"
-    : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type));
+  asm volatile(  // NOLINT
+      "mov %%ebx, %%edi                          \n"
+      "cpuid                                     \n"
+      "xchg %%edi, %%ebx                         \n"
+      : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]),
+        "=d"(cpu_info[3])
+      : "a"(info_type));
 }
 // For gcc/clang but not clangcl.
-#elif (defined(__i386__) || defined(__x86_64__)) && !defined(_MSC_VER)
+#elif !defined(_MSC_VER) && (defined(__i386__) || defined(__x86_64__))
 static __inline void __cpuid(int cpu_info[4], int info_type) {
-  asm volatile (  // NOLINT
-    "cpuid                                     \n"
-    : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type));
+  asm volatile(  // NOLINT
+      "cpuid                                     \n"
+      : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]),
+        "=d"(cpu_info[3])
+      : "a"(info_type));
 }
 #endif
 
@@ -228,20 +228,22 @@ static int CpuHasSSE2() {
 }
 #endif  // HAS_SUMSQUAREERROR_SSE2
 
-static uint32 SumSquareError_C(const uint8* src_a,
-                               const uint8* src_b, int count) {
-  uint32 sse = 0u;
+static uint32_t SumSquareError_C(const uint8_t* src_a,
+                                 const uint8_t* src_b,
+                                 int count) {
+  uint32_t sse = 0u;
   for (int x = 0; x < count; ++x) {
     int diff = src_a[x] - src_b[x];
-    sse += static_cast<uint32>(diff * diff);
+    sse += static_cast<uint32_t>(diff * diff);
   }
   return sse;
 }
 
-double ComputeSumSquareError(const uint8* src_a,
-                             const uint8* src_b, int count) {
-  uint32 (*SumSquareError)(const uint8* src_a,
-                           const uint8* src_b, int count) = SumSquareError_C;
+double ComputeSumSquareError(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count) {
+  uint32_t (*SumSquareError)(const uint8_t* src_a, const uint8_t* src_b,
+                             int count) = SumSquareError_C;
 #if defined(HAS_SUMSQUAREERROR_NEON)
   SumSquareError = SumSquareError_NEON;
 #endif
@@ -251,9 +253,9 @@ double ComputeSumSquareError(const uint8* src_a,
   }
 #endif
   const int kBlockSize = 1 << 15;
-  uint64 sse = 0;
+  uint64_t sse = 0;
 #ifdef _OPENMP
-#pragma omp parallel for reduction(+: sse)
+#pragma omp parallel for reduction(+ : sse)
 #endif
   for (int i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
     sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
@@ -278,8 +280,9 @@ double ComputeSumSquareError(const uint8* src_a,
 // Returns 128.0 (kMaxPSNR) if sse is 0 (perfect match).
 double ComputePSNR(double sse, double size) {
   const double kMINSSE = 255.0 * 255.0 * size / pow(10.0, kMaxPSNR / 10.0);
-  if (sse <= kMINSSE)
+  if (sse <= kMINSSE) {
     sse = kMINSSE;  // Produces max PSNR of 128
+  }
   return 10.0 * log10(255.0 * 255.0 * size / sse);
 }
 
diff --git a/media/libyuv/util/psnr.h b/media/libyuv/util/psnr.h
index 0816b97600..aac128cbca 100644
--- a/media/libyuv/util/psnr.h
+++ b/media/libyuv/util/psnr.h
@@ -20,7 +20,7 @@ extern "C" {
 #endif
 
 #if !defined(INT_TYPES_DEFINED) && !defined(UINT8_TYPE_DEFINED)
-typedef unsigned char uint8;
+typedef unsigned char uint8_t;
 #define UINT8_TYPE_DEFINED
 #endif
 
@@ -31,7 +31,9 @@ static const double kMaxPSNR = 128.0;
 #if !defined(HAVE_JPEG)
 // Computer Sum of Squared Error (SSE).
 // Pass this to ComputePSNR for final result.
-double ComputeSumSquareError(const uint8* org, const uint8* rec, int size);
+double ComputeSumSquareError(const uint8_t* src_a,
+                             const uint8_t* src_b,
+                             int count);
 #endif
 
 // PSNR formula: psnr = 10 * log10 (Peak Signal^2 * size / sse)
diff --git a/media/libyuv/util/psnr_main.cc b/media/libyuv/util/psnr_main.cc
index 0518ab84e0..8b9fd97246 100644
--- a/media/libyuv/util/psnr_main.cc
+++ b/media/libyuv/util/psnr_main.cc
@@ -71,8 +71,8 @@ bool ExtractResolutionFromFilename(const char* name,
   // Isolate the .width_height. section of the filename by searching for a
   // dot or underscore followed by a digit.
   for (int i = 0; name[i]; ++i) {
-    if ((name[i] == '.' || name[i] == '_') &&
-        name[i + 1] >= '0' && name[i + 1] <= '9') {
+    if ((name[i] == '.' || name[i] == '_') && name[i + 1] >= '0' &&
+        name[i + 1] <= '9') {
       int n = sscanf(name + i + 1, "%dx%d", width_ptr, height_ptr);  // NOLINT
       if (2 == n) {
         return true;
@@ -88,11 +88,11 @@ bool ExtractResolutionFromFilename(const char* name,
     return false;
   }
   fseek(file_org, 0, SEEK_END);
-  size_t total_size  = ftell(file_org);
+  size_t total_size = ftell(file_org);
   fseek(file_org, 0, SEEK_SET);
-  uint8* const ch_org = new uint8[total_size];
+  uint8_t* const ch_org = new uint8_t[total_size];
   memset(ch_org, 0, total_size);
-  size_t bytes_org = fread(ch_org, sizeof(uint8), total_size, file_org);
+  size_t bytes_org = fread(ch_org, sizeof(uint8_t), total_size, file_org);
   fclose(file_org);
   if (bytes_org == total_size) {
     if (0 == libyuv::MJPGSize(ch_org, total_size, width_ptr, height_ptr)) {
@@ -107,11 +107,15 @@ bool ExtractResolutionFromFilename(const char* name,
 
 // Scale Y channel from 16..240 to 0..255.
 // This can be useful when comparing codecs that are inconsistant about Y
-uint8 ScaleY(uint8 y) {
+uint8_t ScaleY(uint8_t y) {
   int ny = (y - 16) * 256 / 224;
-  if (ny < 0) ny = 0;
-  if (ny > 255) ny = 255;
-  return static_cast<uint8>(ny);
+  if (ny < 0) {
+    ny = 0;
+  }
+  if (ny > 255) {
+    ny = 255;
+  }
+  return static_cast<uint8_t>(ny);
 }
 
 // MSE = Mean Square Error
@@ -119,16 +123,18 @@ double GetMSE(double sse, double size) {
   return sse / size;
 }
 
-void PrintHelp(const char * program) {
+void PrintHelp(const char* program) {
   printf("%s [-options] org_seq rec_seq [rec_seq2.. etc]\n", program);
 #ifdef HAVE_JPEG
   printf("jpeg or raw YUV 420 supported.\n");
 #endif
   printf("options:\n");
-  printf(" -s <width> <height> .... specify YUV size, mandatory if none of the "
-         "sequences have the\n");
-  printf("                          resolution embedded in their filename (ie. "
-         "name.1920x800_24Hz_P420.yuv)\n");
+  printf(
+      " -s <width> <height> .... specify YUV size, mandatory if none of the "
+      "sequences have the\n");
+  printf(
+      "                          resolution embedded in their filename (ie. "
+      "name.1920x800_24Hz_P420.yuv)\n");
   printf(" -psnr .................. compute PSNR (default)\n");
   printf(" -ssim .................. compute SSIM\n");
   printf(" -mse ................... compute MSE\n");
@@ -146,7 +152,9 @@ void PrintHelp(const char * program) {
 }
 
 void ParseOptions(int argc, const char* argv[]) {
-  if (argc <= 1) PrintHelp(argv[0]);
+  if (argc <= 1) {
+    PrintHelp(argv[0]);
+  }
   for (int c = 1; c < argc; ++c) {
     if (!strcmp(argv[c], "-v")) {
       verbose = true;
@@ -168,16 +176,16 @@ void ParseOptions(int argc, const char* argv[]) {
     } else if (!strcmp(argv[c], "-h") || !strcmp(argv[c], "-help")) {
       PrintHelp(argv[0]);
     } else if (!strcmp(argv[c], "-s") && c + 2 < argc) {
-      image_width = atoi(argv[++c]);    // NOLINT
-      image_height = atoi(argv[++c]);   // NOLINT
+      image_width = atoi(argv[++c]);   // NOLINT
+      image_height = atoi(argv[++c]);  // NOLINT
     } else if (!strcmp(argv[c], "-skip") && c + 2 < argc) {
-      num_skip_org = atoi(argv[++c]);   // NOLINT
-      num_skip_rec = atoi(argv[++c]);   // NOLINT
+      num_skip_org = atoi(argv[++c]);  // NOLINT
+      num_skip_rec = atoi(argv[++c]);  // NOLINT
     } else if (!strcmp(argv[c], "-frames") && c + 1 < argc) {
-      num_frames = atoi(argv[++c]);     // NOLINT
+      num_frames = atoi(argv[++c]);  // NOLINT
 #ifdef _OPENMP
     } else if (!strcmp(argv[c], "-t") && c + 1 < argc) {
-      num_threads = atoi(argv[++c]);    // NOLINT
+      num_threads = atoi(argv[++c]);  // NOLINT
 #endif
     } else if (argv[c][0] == '-') {
       fprintf(stderr, "Unknown option. %s\n", argv[c]);
@@ -206,11 +214,9 @@ void ParseOptions(int argc, const char* argv[]) {
     int org_width, org_height;
     int rec_width, rec_height;
     bool org_res_avail = ExtractResolutionFromFilename(argv[fileindex_org],
-                                                       &org_width,
-                                                       &org_height);
+                                                       &org_width, &org_height);
     bool rec_res_avail = ExtractResolutionFromFilename(argv[fileindex_rec],
-                                                       &rec_width,
-                                                       &rec_height);
+                                                       &rec_width, &rec_height);
     if (org_res_avail) {
       if (rec_res_avail) {
         if ((org_width == rec_width) && (org_height == rec_height)) {
@@ -234,24 +240,28 @@ void ParseOptions(int argc, const char* argv[]) {
   }
 }
 
-bool UpdateMetrics(uint8* ch_org, uint8* ch_rec,
-                   const int y_size, const int uv_size, const size_t total_size,
+bool UpdateMetrics(uint8_t* ch_org,
+                   uint8_t* ch_rec,
+                   const int y_size,
+                   const int uv_size,
+                   const size_t total_size,
                    int number_of_frames,
                    metric* cur_distortion_psnr,
-                   metric* distorted_frame, bool do_psnr) {
+                   metric* distorted_frame,
+                   bool compute_psnr) {
   const int uv_offset = (do_swap_uv ? uv_size : 0);
-  const uint8* const u_org = ch_org + y_size + uv_offset;
-  const uint8* const u_rec = ch_rec + y_size;
-  const uint8* const v_org = ch_org + y_size + (uv_size - uv_offset);
-  const uint8* const v_rec = ch_rec + y_size + uv_size;
-  if (do_psnr) {
+  const uint8_t* const u_org = ch_org + y_size + uv_offset;
+  const uint8_t* const u_rec = ch_rec + y_size;
+  const uint8_t* const v_org = ch_org + y_size + (uv_size - uv_offset);
+  const uint8_t* const v_rec = ch_rec + y_size + uv_size;
+  if (compute_psnr) {
 #ifdef HAVE_JPEG
     double y_err = static_cast<double>(
-      libyuv::ComputeSumSquareError(ch_org, ch_rec, y_size));
+        libyuv::ComputeSumSquareError(ch_org, ch_rec, y_size));
     double u_err = static_cast<double>(
-      libyuv::ComputeSumSquareError(u_org, u_rec, uv_size));
+        libyuv::ComputeSumSquareError(u_org, u_rec, uv_size));
     double v_err = static_cast<double>(
-      libyuv::ComputeSumSquareError(v_org, v_rec, uv_size));
+        libyuv::ComputeSumSquareError(v_org, v_rec, uv_size));
 #else
     double y_err = ComputeSumSquareError(ch_org, ch_rec, y_size);
     double u_err = ComputeSumSquareError(u_org, u_rec, uv_size);
@@ -265,17 +275,17 @@ bool UpdateMetrics(uint8* ch_org, uint8* ch_rec,
     distorted_frame->y = ComputePSNR(y_err, static_cast<double>(y_size));
     distorted_frame->u = ComputePSNR(u_err, static_cast<double>(uv_size));
     distorted_frame->v = ComputePSNR(v_err, static_cast<double>(uv_size));
-    distorted_frame->all = ComputePSNR(total_err,
-                                       static_cast<double>(total_size));
+    distorted_frame->all =
+        ComputePSNR(total_err, static_cast<double>(total_size));
   } else {
     distorted_frame->y = CalcSSIM(ch_org, ch_rec, image_width, image_height);
-    distorted_frame->u = CalcSSIM(u_org, u_rec, (image_width + 1) / 2,
-                                 (image_height + 1) / 2);
-    distorted_frame->v = CalcSSIM(v_org, v_rec, (image_width + 1) / 2,
-                                 (image_height + 1) / 2);
+    distorted_frame->u =
+        CalcSSIM(u_org, u_rec, (image_width + 1) / 2, (image_height + 1) / 2);
+    distorted_frame->v =
+        CalcSSIM(v_org, v_rec, (image_width + 1) / 2, (image_height + 1) / 2);
     distorted_frame->all =
-      (distorted_frame->y + distorted_frame->u + distorted_frame->v)
-        / total_size;
+        (distorted_frame->y + distorted_frame->u + distorted_frame->v) /
+        total_size;
     distorted_frame->y /= y_size;
     distorted_frame->u /= uv_size;
     distorted_frame->v /= uv_size;
@@ -294,12 +304,15 @@ bool UpdateMetrics(uint8* ch_org, uint8* ch_rec,
   cur_distortion_psnr->all += distorted_frame->all;
 
   bool ismin = false;
-  if (distorted_frame->y < cur_distortion_psnr->min_y)
+  if (distorted_frame->y < cur_distortion_psnr->min_y) {
     cur_distortion_psnr->min_y = distorted_frame->y;
-  if (distorted_frame->u < cur_distortion_psnr->min_u)
+  }
+  if (distorted_frame->u < cur_distortion_psnr->min_u) {
     cur_distortion_psnr->min_u = distorted_frame->u;
-  if (distorted_frame->v < cur_distortion_psnr->min_v)
+  }
+  if (distorted_frame->v < cur_distortion_psnr->min_v) {
     cur_distortion_psnr->min_v = distorted_frame->v;
+  }
   if (distorted_frame->all < cur_distortion_psnr->min_all) {
     cur_distortion_psnr->min_all = distorted_frame->all;
     cur_distortion_psnr->min_frame = number_of_frames;
@@ -330,8 +343,8 @@ int main(int argc, const char* argv[]) {
   }
 
   // Open all files to compare to
-  FILE** file_rec = new FILE* [num_rec];
-  memset(file_rec, 0, num_rec * sizeof(FILE*)); // NOLINT
+  FILE** file_rec = new FILE*[num_rec];
+  memset(file_rec, 0, num_rec * sizeof(FILE*));  // NOLINT
   for (int cur_rec = 0; cur_rec < num_rec; ++cur_rec) {
     file_rec[cur_rec] = fopen(argv[fileindex_rec + cur_rec], "rb");
     if (file_rec[cur_rec] == NULL) {
@@ -347,27 +360,28 @@ int main(int argc, const char* argv[]) {
 
   const int y_size = image_width * image_height;
   const int uv_size = ((image_width + 1) / 2) * ((image_height + 1) / 2);
-  const size_t total_size = y_size + 2 * uv_size;    // NOLINT
+  const size_t total_size = y_size + 2 * uv_size;  // NOLINT
 #if defined(_MSC_VER)
-  _fseeki64(file_org,
-            static_cast<__int64>(num_skip_org) *
-            static_cast<__int64>(total_size), SEEK_SET);
+  _fseeki64(
+      file_org,
+      static_cast<__int64>(num_skip_org) * static_cast<__int64>(total_size),
+      SEEK_SET);
 #else
   fseek(file_org, num_skip_org * total_size, SEEK_SET);
 #endif
   for (int cur_rec = 0; cur_rec < num_rec; ++cur_rec) {
 #if defined(_MSC_VER)
-    _fseeki64(file_rec[cur_rec],
-              static_cast<__int64>(num_skip_rec) *
-              static_cast<__int64>(total_size),
-              SEEK_SET);
+    _fseeki64(
+        file_rec[cur_rec],
+        static_cast<__int64>(num_skip_rec) * static_cast<__int64>(total_size),
+        SEEK_SET);
 #else
     fseek(file_rec[cur_rec], num_skip_rec * total_size, SEEK_SET);
 #endif
   }
 
-  uint8* const ch_org = new uint8[total_size];
-  uint8* const ch_rec = new uint8[total_size];
+  uint8_t* const ch_org = new uint8_t[total_size];
+  uint8_t* const ch_rec = new uint8_t[total_size];
   if (ch_org == NULL || ch_rec == NULL) {
     fprintf(stderr, "No memory available\n");
     fclose(file_org);
@@ -420,29 +434,24 @@ int main(int argc, const char* argv[]) {
   }
 
   int number_of_frames;
-  for (number_of_frames = 0; ; ++number_of_frames) {
-    if (num_frames && number_of_frames >= num_frames)
+  for (number_of_frames = 0;; ++number_of_frames) {
+    if (num_frames && number_of_frames >= num_frames) {
       break;
+    }
 
-    size_t bytes_org = fread(ch_org, sizeof(uint8), total_size, file_org);
+    size_t bytes_org = fread(ch_org, sizeof(uint8_t), total_size, file_org);
     if (bytes_org < total_size) {
 #ifdef HAVE_JPEG
       // Try parsing file as a jpeg.
-      uint8* const ch_jpeg = new uint8[bytes_org];
+      uint8_t* const ch_jpeg = new uint8_t[bytes_org];
       memcpy(ch_jpeg, ch_org, bytes_org);
       memset(ch_org, 0, total_size);
 
-      if (0 != libyuv::MJPGToI420(ch_jpeg, bytes_org,
-                                  ch_org,
-                                  image_width,
-                                  ch_org + y_size,
-                                  (image_width + 1) / 2,
+      if (0 != libyuv::MJPGToI420(ch_jpeg, bytes_org, ch_org, image_width,
+                                  ch_org + y_size, (image_width + 1) / 2,
                                   ch_org + y_size + uv_size,
-                                  (image_width + 1) / 2,
-                                  image_width,
-                                  image_height,
-                                  image_width,
-                                  image_height)) {
+                                  (image_width + 1) / 2, image_width,
+                                  image_height, image_width, image_height)) {
         delete[] ch_jpeg;
         break;
       }
@@ -453,26 +462,20 @@ int main(int argc, const char* argv[]) {
     }
 
     for (int cur_rec = 0; cur_rec < num_rec; ++cur_rec) {
-      size_t bytes_rec = fread(ch_rec, sizeof(uint8),
-                               total_size, file_rec[cur_rec]);
+      size_t bytes_rec =
+          fread(ch_rec, sizeof(uint8_t), total_size, file_rec[cur_rec]);
       if (bytes_rec < total_size) {
 #ifdef HAVE_JPEG
         // Try parsing file as a jpeg.
-        uint8* const ch_jpeg = new uint8[bytes_rec];
+        uint8_t* const ch_jpeg = new uint8_t[bytes_rec];
         memcpy(ch_jpeg, ch_rec, bytes_rec);
         memset(ch_rec, 0, total_size);
 
-        if (0 != libyuv::MJPGToI420(ch_jpeg, bytes_rec,
-                                    ch_rec,
-                                    image_width,
-                                    ch_rec + y_size,
-                                    (image_width + 1) / 2,
+        if (0 != libyuv::MJPGToI420(ch_jpeg, bytes_rec, ch_rec, image_width,
+                                    ch_rec + y_size, (image_width + 1) / 2,
                                     ch_rec + y_size + uv_size,
-                                    (image_width + 1) / 2,
-                                    image_width,
-                                    image_height,
-                                    image_width,
-                                    image_height)) {
+                                    (image_width + 1) / 2, image_width,
+                                    image_height, image_width, image_height)) {
           delete[] ch_jpeg;
           break;
         }
@@ -486,12 +489,10 @@ int main(int argc, const char* argv[]) {
         printf("%5d", number_of_frames);
       }
       if (do_psnr) {
-        metric distorted_frame;
+        metric distorted_frame = {};
         metric* cur_distortion_psnr = &distortion_psnr[cur_rec];
-        bool ismin = UpdateMetrics(ch_org, ch_rec,
-                                   y_size, uv_size, total_size,
-                                   number_of_frames,
-                                   cur_distortion_psnr,
+        bool ismin = UpdateMetrics(ch_org, ch_rec, y_size, uv_size, total_size,
+                                   number_of_frames, cur_distortion_psnr,
                                    &distorted_frame, true);
         if (verbose) {
           printf("\t%10.6f", distorted_frame.y);
@@ -502,12 +503,10 @@ int main(int argc, const char* argv[]) {
         }
       }
       if (do_ssim) {
-        metric distorted_frame;
+        metric distorted_frame = {};
         metric* cur_distortion_ssim = &distortion_ssim[cur_rec];
-        bool ismin = UpdateMetrics(ch_org, ch_rec,
-                                   y_size, uv_size, total_size,
-                                   number_of_frames,
-                                   cur_distortion_ssim,
+        bool ismin = UpdateMetrics(ch_org, ch_rec, y_size, uv_size, total_size,
+                                   number_of_frames, cur_distortion_ssim,
                                    &distorted_frame, false);
         if (verbose) {
           printf("\t%10.6f", distorted_frame.y);
@@ -543,24 +542,20 @@ int main(int argc, const char* argv[]) {
     }
 
     if (do_psnr) {
-      const double global_psnr_y = ComputePSNR(
-          cur_distortion_psnr->global_y,
-          static_cast<double>(y_size) * number_of_frames);
-      const double global_psnr_u = ComputePSNR(
-          cur_distortion_psnr->global_u,
-          static_cast<double>(uv_size) * number_of_frames);
-      const double global_psnr_v = ComputePSNR(
-          cur_distortion_psnr->global_v,
-          static_cast<double>(uv_size) * number_of_frames);
-      const double global_psnr_all = ComputePSNR(
-          cur_distortion_psnr->global_all,
-          static_cast<double>(total_size) * number_of_frames);
-      printf("Global:\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d",
-          global_psnr_y,
-          global_psnr_u,
-          global_psnr_v,
-          global_psnr_all,
-          number_of_frames);
+      const double global_psnr_y =
+          ComputePSNR(cur_distortion_psnr->global_y,
+                      static_cast<double>(y_size) * number_of_frames);
+      const double global_psnr_u =
+          ComputePSNR(cur_distortion_psnr->global_u,
+                      static_cast<double>(uv_size) * number_of_frames);
+      const double global_psnr_v =
+          ComputePSNR(cur_distortion_psnr->global_v,
+                      static_cast<double>(uv_size) * number_of_frames);
+      const double global_psnr_all =
+          ComputePSNR(cur_distortion_psnr->global_all,
+                      static_cast<double>(total_size) * number_of_frames);
+      printf("Global:\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d", global_psnr_y,
+             global_psnr_u, global_psnr_v, global_psnr_all, number_of_frames);
       if (show_name) {
         printf("\t%s", argv[fileindex_rec + cur_rec]);
       }
@@ -570,20 +565,14 @@ int main(int argc, const char* argv[]) {
     if (!quiet) {
       printf("Avg:");
       if (do_psnr) {
-        printf("\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d",
-             cur_distortion_psnr->y,
-             cur_distortion_psnr->u,
-             cur_distortion_psnr->v,
-             cur_distortion_psnr->all,
-             number_of_frames);
+        printf("\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d", cur_distortion_psnr->y,
+               cur_distortion_psnr->u, cur_distortion_psnr->v,
+               cur_distortion_psnr->all, number_of_frames);
       }
       if (do_ssim) {
-        printf("\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d",
-             cur_distortion_ssim->y,
-             cur_distortion_ssim->u,
-             cur_distortion_ssim->v,
-             cur_distortion_ssim->all,
-             number_of_frames);
+        printf("\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d", cur_distortion_ssim->y,
+               cur_distortion_ssim->u, cur_distortion_ssim->v,
+               cur_distortion_ssim->all, number_of_frames);
       }
       if (show_name) {
         printf("\t%s", argv[fileindex_rec + cur_rec]);
@@ -594,19 +583,15 @@ int main(int argc, const char* argv[]) {
       printf("Min:");
       if (do_psnr) {
         printf("\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d",
-            cur_distortion_psnr->min_y,
-            cur_distortion_psnr->min_u,
-            cur_distortion_psnr->min_v,
-            cur_distortion_psnr->min_all,
-            cur_distortion_psnr->min_frame);
+               cur_distortion_psnr->min_y, cur_distortion_psnr->min_u,
+               cur_distortion_psnr->min_v, cur_distortion_psnr->min_all,
+               cur_distortion_psnr->min_frame);
       }
       if (do_ssim) {
         printf("\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d",
-            cur_distortion_ssim->min_y,
-            cur_distortion_ssim->min_u,
-            cur_distortion_ssim->min_v,
-            cur_distortion_ssim->min_all,
-            cur_distortion_ssim->min_frame);
+               cur_distortion_ssim->min_y, cur_distortion_ssim->min_u,
+               cur_distortion_ssim->min_v, cur_distortion_ssim->min_all,
+               cur_distortion_ssim->min_frame);
       }
       if (show_name) {
         printf("\t%s", argv[fileindex_rec + cur_rec]);
@@ -615,20 +600,20 @@ int main(int argc, const char* argv[]) {
     }
 
     if (do_mse) {
-      double global_mse_y = GetMSE(cur_distortion_psnr->global_y,
-        static_cast<double>(y_size) * number_of_frames);
-      double global_mse_u = GetMSE(cur_distortion_psnr->global_u,
-        static_cast<double>(uv_size) * number_of_frames);
-      double global_mse_v = GetMSE(cur_distortion_psnr->global_v,
-        static_cast<double>(uv_size) * number_of_frames);
-      double global_mse_all = GetMSE(cur_distortion_psnr->global_all,
-        static_cast<double>(total_size) * number_of_frames);
-      printf("MSE:\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d",
-          global_mse_y,
-          global_mse_u,
-          global_mse_v,
-          global_mse_all,
-          number_of_frames);
+      double global_mse_y =
+          GetMSE(cur_distortion_psnr->global_y,
+                 static_cast<double>(y_size) * number_of_frames);
+      double global_mse_u =
+          GetMSE(cur_distortion_psnr->global_u,
+                 static_cast<double>(uv_size) * number_of_frames);
+      double global_mse_v =
+          GetMSE(cur_distortion_psnr->global_v,
+                 static_cast<double>(uv_size) * number_of_frames);
+      double global_mse_all =
+          GetMSE(cur_distortion_psnr->global_all,
+                 static_cast<double>(total_size) * number_of_frames);
+      printf("MSE:\t%10.6f\t%10.6f\t%10.6f\t%10.6f\t%5d", global_mse_y,
+             global_mse_u, global_mse_v, global_mse_all, number_of_frames);
       if (show_name) {
         printf("\t%s", argv[fileindex_rec + cur_rec]);
       }
diff --git a/media/libyuv/util/ssim.cc b/media/libyuv/util/ssim.cc
index 5a6399b782..096fbcf061 100644
--- a/media/libyuv/util/ssim.cc
+++ b/media/libyuv/util/ssim.cc
@@ -16,11 +16,11 @@
 extern "C" {
 #endif
 
-typedef unsigned int uint32;     // NOLINT
-typedef unsigned short uint16;   // NOLINT
+typedef unsigned int uint32_t;    // NOLINT
+typedef unsigned short uint16_t;  // NOLINT
 
 #if !defined(LIBYUV_DISABLE_X86) && !defined(__SSE2__) && \
-  (defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)))
+    (defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)))
 #define __SSE2__
 #endif
 #if !defined(LIBYUV_DISABLE_X86) && defined(__SSE2__)
@@ -38,56 +38,69 @@ enum { KERNEL = 3, KERNEL_SIZE = 2 * KERNEL + 1 };
 // The maximum value (11 x 11) must be less than 128 to avoid sign
 // problems during the calls to _mm_mullo_epi16().
 static const int K[KERNEL_SIZE] = {
-  1, 3, 7, 11, 7, 3, 1    // ~11 * exp(-0.3 * i * i)
+    1, 3, 7, 11, 7, 3, 1  // ~11 * exp(-0.3 * i * i)
 };
 static const double kiW[KERNEL + 1 + 1] = {
-  1. / 1089.,   // 1 / sum(i:0..6, j..6) K[i]*K[j]
-  1. / 1089.,   // 1 / sum(i:0..6, j..6) K[i]*K[j]
-  1. / 1056.,   // 1 / sum(i:0..5, j..6) K[i]*K[j]
-  1. / 957.,    // 1 / sum(i:0..4, j..6) K[i]*K[j]
-  1. / 726.,    // 1 / sum(i:0..3, j..6) K[i]*K[j]
+    1. / 1089.,  // 1 / sum(i:0..6, j..6) K[i]*K[j]
+    1. / 1089.,  // 1 / sum(i:0..6, j..6) K[i]*K[j]
+    1. / 1056.,  // 1 / sum(i:0..5, j..6) K[i]*K[j]
+    1. / 957.,   // 1 / sum(i:0..4, j..6) K[i]*K[j]
+    1. / 726.,   // 1 / sum(i:0..3, j..6) K[i]*K[j]
 };
 
 #if !defined(LIBYUV_DISABLE_X86) && defined(__SSE2__)
 
-#define PWEIGHT(A, B)  static_cast<uint16>(K[(A)] * K[(B)])   // weight product
-#define MAKE_WEIGHT(L)                                               \
-  { { { PWEIGHT(L, 0), PWEIGHT(L, 1), PWEIGHT(L, 2), PWEIGHT(L, 3),  \
-        PWEIGHT(L, 4), PWEIGHT(L, 5), PWEIGHT(L, 6), 0 } } }
+#define PWEIGHT(A, B) static_cast<uint16_t>(K[(A)] * K[(B)])  // weight product
+#define MAKE_WEIGHT(L)                                                \
+  {                                                                   \
+    {                                                                 \
+      {                                                               \
+        PWEIGHT(L, 0)                                                 \
+        , PWEIGHT(L, 1), PWEIGHT(L, 2), PWEIGHT(L, 3), PWEIGHT(L, 4), \
+            PWEIGHT(L, 5), PWEIGHT(L, 6), 0                           \
+      }                                                               \
+    }                                                                 \
+  }
 
 // We need this union trick to be able to initialize constant static __m128i
 // values. We can't call _mm_set_epi16() for static compile-time initialization.
 static const struct {
   union {
-    uint16 i16_[8];
+    uint16_t i16_[8];
     __m128i m_;
   } values_;
-} W0 = MAKE_WEIGHT(0),
-  W1 = MAKE_WEIGHT(1),
-  W2 = MAKE_WEIGHT(2),
+} W0 = MAKE_WEIGHT(0), W1 = MAKE_WEIGHT(1), W2 = MAKE_WEIGHT(2),
   W3 = MAKE_WEIGHT(3);
-  // ... the rest is symmetric.
+// ... the rest is symmetric.
 #undef MAKE_WEIGHT
 #undef PWEIGHT
 #endif
 
 // Common final expression for SSIM, once the weighted sums are known.
-static double FinalizeSSIM(double iw, double xm, double ym,
-                           double xxm, double xym, double yym) {
+static double FinalizeSSIM(double iw,
+                           double xm,
+                           double ym,
+                           double xxm,
+                           double xym,
+                           double yym) {
   const double iwx = xm * iw;
   const double iwy = ym * iw;
   double sxx = xxm * iw - iwx * iwx;
   double syy = yym * iw - iwy * iwy;
   // small errors are possible, due to rounding. Clamp to zero.
-  if (sxx < 0.) sxx = 0.;
-  if (syy < 0.) syy = 0.;
+  if (sxx < 0.) {
+    sxx = 0.;
+  }
+  if (syy < 0.) {
+    syy = 0.;
+  }
   const double sxsy = sqrt(sxx * syy);
   const double sxy = xym * iw - iwx * iwy;
   static const double C11 = (0.01 * 0.01) * (255 * 255);
   static const double C22 = (0.03 * 0.03) * (255 * 255);
   static const double C33 = (0.015 * 0.015) * (255 * 255);
   const double l = (2. * iwx * iwy + C11) / (iwx * iwx + iwy * iwy + C11);
-  const double c = (2. * sxsy      + C22) / (sxx + syy + C22);
+  const double c = (2. * sxsy + C22) / (sxx + syy + C22);
   const double s = (sxy + C33) / (sxsy + C33);
   return l * c * s;
 }
@@ -98,15 +111,22 @@ static double FinalizeSSIM(double iw, double xm, double ym,
 // Note: worst case of accumulation is a weight of 33 = 11 + 2 * (7 + 3 + 1)
 // with a diff of 255, squared. The maximum error is thus 0x4388241,
 // which fits into 32 bits integers.
-double GetSSIM(const uint8 *org, const uint8 *rec,
-               int xo, int yo, int W, int H, int stride) {
-  uint32 ws = 0, xm = 0, ym = 0, xxm = 0, xym = 0, yym = 0;
+double GetSSIM(const uint8_t* org,
+               const uint8_t* rec,
+               int xo,
+               int yo,
+               int W,
+               int H,
+               int stride) {
+  uint32_t ws = 0, xm = 0, ym = 0, xxm = 0, xym = 0, yym = 0;
   org += (yo - KERNEL) * stride;
   org += (xo - KERNEL);
   rec += (yo - KERNEL) * stride;
   rec += (xo - KERNEL);
   for (int y_ = 0; y_ < KERNEL_SIZE; ++y_, org += stride, rec += stride) {
-    if (((yo - KERNEL + y_) < 0) || ((yo - KERNEL + y_) >= H)) continue;
+    if (((yo - KERNEL + y_) < 0) || ((yo - KERNEL + y_) >= H)) {
+      continue;
+    }
     const int Wy = K[y_];
     for (int x_ = 0; x_ < KERNEL_SIZE; ++x_) {
       const int Wxy = Wy * K[x_];
@@ -114,8 +134,8 @@ double GetSSIM(const uint8 *org, const uint8 *rec,
         const int org_x = org[x_];
         const int rec_x = rec[x_];
         ws += Wxy;
-        xm  += Wxy * org_x;
-        ym  += Wxy * rec_x;
+        xm += Wxy * org_x;
+        ym += Wxy * rec_x;
         xxm += Wxy * org_x * org_x;
         xym += Wxy * org_x * rec_x;
         yym += Wxy * rec_x * rec_x;
@@ -125,10 +145,13 @@ double GetSSIM(const uint8 *org, const uint8 *rec,
   return FinalizeSSIM(1. / ws, xm, ym, xxm, xym, yym);
 }
 
-double GetSSIMFullKernel(const uint8 *org, const uint8 *rec,
-                         int xo, int yo, int stride,
+double GetSSIMFullKernel(const uint8_t* org,
+                         const uint8_t* rec,
+                         int xo,
+                         int yo,
+                         int stride,
                          double area_weight) {
-  uint32 xm = 0, ym = 0, xxm = 0, xym = 0, yym = 0;
+  uint32_t xm = 0, ym = 0, xxm = 0, xym = 0, yym = 0;
 
 #if defined(LIBYUV_DISABLE_X86) || !defined(__SSE2__)
 
@@ -161,8 +184,8 @@ double GetSSIMFullKernel(const uint8 *org, const uint8 *rec,
       const int ll2 = rec[dy2 - x];
       const int lr2 = rec[dy2 + x];
 
-      xm  += Wxy * (ul1 + ur1 + ll1 + lr1);
-      ym  += Wxy * (ul2 + ur2 + ll2 + lr2);
+      xm += Wxy * (ul1 + ur1 + ll1 + lr1);
+      ym += Wxy * (ul2 + ur2 + ll2 + lr2);
       xxm += Wxy * (ul1 * ul1 + ur1 * ur1 + ll1 * ll1 + lr1 * lr1);
       xym += Wxy * (ul1 * ul2 + ur1 * ur2 + ll1 * ll2 + lr1 * lr2);
       yym += Wxy * (ul2 * ul2 + ur2 * ur2 + ll2 * ll2 + lr2 * lr2);
@@ -189,8 +212,8 @@ double GetSSIMFullKernel(const uint8 *org, const uint8 *rec,
     const int l2 = rec[-y];
     const int r2 = rec[y];
 
-    xm  += Wxy * (u1 + d1 + l1 + r1);
-    ym  += Wxy * (u2 + d2 + l2 + r2);
+    xm += Wxy * (u1 + d1 + l1 + r1);
+    ym += Wxy * (u2 + d2 + l2 + r2);
     xxm += Wxy * (u1 * u1 + d1 * d1 + l1 * l1 + r1 * r1);
     xym += Wxy * (u1 * u2 + d1 * d2 + l1 * l2 + r1 * r2);
     yym += Wxy * (u2 * u2 + d2 * d2 + l2 * l2 + r2 * r2);
@@ -201,13 +224,13 @@ double GetSSIMFullKernel(const uint8 *org, const uint8 *rec,
   const int s1 = org[0];
   const int s2 = rec[0];
 
-  xm  += Wxy * s1;
-  ym  += Wxy * s2;
+  xm += Wxy * s1;
+  ym += Wxy * s2;
   xxm += Wxy * s1 * s1;
   xym += Wxy * s1 * s2;
   yym += Wxy * s2 * s2;
 
-#else   // __SSE2__
+#else  // __SSE2__
 
   org += (yo - KERNEL) * stride + (xo - KERNEL);
   rec += (yo - KERNEL) * stride + (xo - KERNEL);
@@ -221,29 +244,31 @@ double GetSSIMFullKernel(const uint8 *org, const uint8 *rec,
 
 // Read 8 pixels at line #L, and convert to 16bit, perform weighting
 // and acccumulate.
-#define LOAD_LINE_PAIR(L, WEIGHT) do {                                       \
-  const __m128i v0 =                                                         \
-      _mm_loadl_epi64(reinterpret_cast<const __m128i*>(org + (L) * stride)); \
-  const __m128i v1 =                                                         \
-      _mm_loadl_epi64(reinterpret_cast<const __m128i*>(rec + (L) * stride)); \
-  const __m128i w0 = _mm_unpacklo_epi8(v0, zero);                            \
-  const __m128i w1 = _mm_unpacklo_epi8(v1, zero);                            \
-  const __m128i ww0 = _mm_mullo_epi16(w0, (WEIGHT).values_.m_);              \
-  const __m128i ww1 = _mm_mullo_epi16(w1, (WEIGHT).values_.m_);              \
-  x = _mm_add_epi32(x, _mm_unpacklo_epi16(ww0, zero));                       \
-  y = _mm_add_epi32(y, _mm_unpacklo_epi16(ww1, zero));                       \
-  x = _mm_add_epi32(x, _mm_unpackhi_epi16(ww0, zero));                       \
-  y = _mm_add_epi32(y, _mm_unpackhi_epi16(ww1, zero));                       \
-  xx = _mm_add_epi32(xx, _mm_madd_epi16(ww0, w0));                           \
-  xy = _mm_add_epi32(xy, _mm_madd_epi16(ww0, w1));                           \
-  yy = _mm_add_epi32(yy, _mm_madd_epi16(ww1, w1));                           \
-} while (0)
-
-#define ADD_AND_STORE_FOUR_EPI32(M, OUT) do {                                \
-  uint32 tmp[4];                                                             \
-  _mm_storeu_si128(reinterpret_cast<__m128i*>(tmp), (M));                    \
-  (OUT) = tmp[3] + tmp[2] + tmp[1] + tmp[0];                                 \
-} while (0)
+#define LOAD_LINE_PAIR(L, WEIGHT)                                            \
+  do {                                                                       \
+    const __m128i v0 =                                                       \
+        _mm_loadl_epi64(reinterpret_cast<const __m128i*>(org + (L)*stride)); \
+    const __m128i v1 =                                                       \
+        _mm_loadl_epi64(reinterpret_cast<const __m128i*>(rec + (L)*stride)); \
+    const __m128i w0 = _mm_unpacklo_epi8(v0, zero);                          \
+    const __m128i w1 = _mm_unpacklo_epi8(v1, zero);                          \
+    const __m128i ww0 = _mm_mullo_epi16(w0, (WEIGHT).values_.m_);            \
+    const __m128i ww1 = _mm_mullo_epi16(w1, (WEIGHT).values_.m_);            \
+    x = _mm_add_epi32(x, _mm_unpacklo_epi16(ww0, zero));                     \
+    y = _mm_add_epi32(y, _mm_unpacklo_epi16(ww1, zero));                     \
+    x = _mm_add_epi32(x, _mm_unpackhi_epi16(ww0, zero));                     \
+    y = _mm_add_epi32(y, _mm_unpackhi_epi16(ww1, zero));                     \
+    xx = _mm_add_epi32(xx, _mm_madd_epi16(ww0, w0));                         \
+    xy = _mm_add_epi32(xy, _mm_madd_epi16(ww0, w1));                         \
+    yy = _mm_add_epi32(yy, _mm_madd_epi16(ww1, w1));                         \
+  } while (0)
+
+#define ADD_AND_STORE_FOUR_EPI32(M, OUT)                    \
+  do {                                                      \
+    uint32_t tmp[4];                                        \
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(tmp), (M)); \
+    (OUT) = tmp[3] + tmp[2] + tmp[1] + tmp[0];              \
+  } while (0)
 
   LOAD_LINE_PAIR(0, W0);
   LOAD_LINE_PAIR(1, W1);
@@ -266,10 +291,14 @@ double GetSSIMFullKernel(const uint8 *org, const uint8 *rec,
   return FinalizeSSIM(area_weight, xm, ym, xxm, xym, yym);
 }
 
-static int start_max(int x, int y) { return (x > y) ? x : y; }
+static int start_max(int x, int y) {
+  return (x > y) ? x : y;
+}
 
-double CalcSSIM(const uint8 *org, const uint8 *rec,
-                const int image_width, const int image_height) {
+double CalcSSIM(const uint8_t* org,
+                const uint8_t* rec,
+                const int image_width,
+                const int image_height) {
   double SSIM = 0.;
   const int KERNEL_Y = (image_height < KERNEL) ? image_height : KERNEL;
   const int KERNEL_X = (image_width < KERNEL) ? image_width : KERNEL;
@@ -284,7 +313,7 @@ double CalcSSIM(const uint8 *org, const uint8 *rec,
   }
 
 #ifdef _OPENMP
-  #pragma omp parallel for reduction(+: SSIM)
+#pragma omp parallel for reduction(+ : SSIM)
 #endif
   for (int j = KERNEL_Y; j < image_height - KERNEL_Y; ++j) {
     for (int i = 0; i < KERNEL_X; ++i) {
@@ -302,8 +331,8 @@ double CalcSSIM(const uint8 *org, const uint8 *rec,
       // NOTE: we could use similar method for the left-most pixels too.
       const int kScratchWidth = 8;
       const int kScratchStride = kScratchWidth + KERNEL + 1;
-      uint8 scratch_org[KERNEL_SIZE * kScratchStride] = { 0 };
-      uint8 scratch_rec[KERNEL_SIZE * kScratchStride] = { 0 };
+      uint8_t scratch_org[KERNEL_SIZE * kScratchStride] = {0};
+      uint8_t scratch_rec[KERNEL_SIZE * kScratchStride] = {0};
 
       for (int k = 0; k < KERNEL_SIZE; ++k) {
         const int offset =
@@ -311,9 +340,9 @@ double CalcSSIM(const uint8 *org, const uint8 *rec,
         memcpy(scratch_org + k * kScratchStride, org + offset, kScratchWidth);
         memcpy(scratch_rec + k * kScratchStride, rec + offset, kScratchWidth);
       }
-      for (int k = 0;  k <= KERNEL_X + 1; ++k) {
-        SSIM += GetSSIMFullKernel(scratch_org, scratch_rec,
-                                  KERNEL + k, KERNEL, kScratchStride, kiW[k]);
+      for (int k = 0; k <= KERNEL_X + 1; ++k) {
+        SSIM += GetSSIMFullKernel(scratch_org, scratch_rec, KERNEL + k, KERNEL,
+                                  kScratchStride, kiW[k]);
       }
     }
   }
@@ -333,4 +362,3 @@ double CalcLSSIM(double ssim) {
 #ifdef __cplusplus
 }  // extern "C"
 #endif
-
diff --git a/media/libyuv/util/ssim.h b/media/libyuv/util/ssim.h
index 430eb71c3d..a855f1d123 100644
--- a/media/libyuv/util/ssim.h
+++ b/media/libyuv/util/ssim.h
@@ -10,7 +10,7 @@
 
 // Get SSIM for video sequence. Assuming RAW 4:2:0 Y:Cb:Cr format
 
-#ifndef UTIL_SSIM_H_  // NOLINT
+#ifndef UTIL_SSIM_H_
 #define UTIL_SSIM_H_
 
 #include <math.h>  // For log10()
@@ -20,12 +20,14 @@ extern "C" {
 #endif
 
 #if !defined(INT_TYPES_DEFINED) && !defined(UINT8_TYPE_DEFINED)
-typedef unsigned char uint8;
+typedef unsigned char uint8_t;
 #define UINT8_TYPE_DEFINED
 #endif
 
-double CalcSSIM(const uint8* org, const uint8* rec,
-                const int image_width, const int image_height);
+double CalcSSIM(const uint8_t* org,
+                const uint8_t* rec,
+                const int image_width,
+                const int image_height);
 
 double CalcLSSIM(double ssim);
 
@@ -33,4 +35,4 @@ double CalcLSSIM(double ssim);
 }  // extern "C"
 #endif
 
-#endif  // UTIL_SSIM_H_  // NOLINT
+#endif  // UTIL_SSIM_H_
diff --git a/media/libyuv/util/yuvconstants.c b/media/libyuv/util/yuvconstants.c
new file mode 100644
index 0000000000..bc0c0de28e
--- /dev/null
+++ b/media/libyuv/util/yuvconstants.c
@@ -0,0 +1,106 @@
+/*
+ *  Copyright 2021 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+// This utility computes values needed to generate yuvconstants based on
+// white point values.
+// The yuv formulas are tuned for 8 bit YUV channels.
+
+// See Also
+// https://mymusing.co/bt601-yuv-to-rgb-conversion-color/
+
+// BT.709 full range YUV to RGB reference
+//  R = Y               + V * 1.5748
+//  G = Y - U * 0.18732 - V * 0.46812
+//  B = Y + U * 1.8556
+//  KR = 0.2126
+//  KB = 0.0722
+
+// // Y contribution to R,G,B.  Scale and bias.
+// #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+// #define YB 32    /* 64 / 2 */
+//
+// // U and V contributions to R,G,B.
+// #define UB 113 /* round(1.77200 * 64) */
+// #define UG 22  /* round(0.34414 * 64) */
+// #define VG 46  /* round(0.71414 * 64) */
+// #define VR 90  /* round(1.40200 * 64) */
+//
+// // Bias values to round, and subtract 128 from U and V.
+// #define BB (-UB * 128 + YB)
+// #define BG (UG * 128 + VG * 128 + YB)
+// #define BR (-VR * 128 + YB)
+
+int main(int argc, const char* argv[]) {
+  if (argc < 3) {
+    printf("yuvconstants [KR] [KB]\n");
+    printf("  e.g. yuvconstants 0.2126 0.0722\n");
+    printf("  MC BT          KR           KB\n");
+    printf("  1  BT.709      KR = 0.2126; KB = 0.0722\n");
+    printf("  4  FCC         KR = 0.30;   KB = 0.11\n");
+    printf("  6  BT.601      KR = 0.299;  KB = 0.114\n");
+    printf("  7  SMPTE 240M  KR = 0.212;  KB = 0.087\n");
+    printf("  9  BT.2020     KR = 0.2627; KB = 0.0593\n");
+    return -1;
+  }
+  float kr = atof(argv[1]);
+  float kb = atof(argv[2]);
+  float kg = 1 - kr - kb;
+
+  float vr = 2 * (1 - kr);
+  float ug = 2 * ((1 - kb) * kb / kg);
+  float vg = 2 * ((1 - kr) * kr / kg);
+  float ub = 2 * (1 - kb);
+
+  printf("Full range\n");
+  printf("R = Y                + V * %5f\n", vr);
+  printf("G = Y - U * %6f - V * %6f\n", ug, vg);
+  printf("B = Y + U * %5f\n", ub);
+
+  printf("KR = %4f; ", kr);
+  printf("KB = %4f\n", kb);
+  // printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ub * 64), ub, ub * 64);
+  printf("UG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ug * 64), ug, ug * 64);
+  printf("VG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vg * 64), vg, vg * 64);
+  printf("VR %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vr * 64), vr, vr * 64);
+
+  vr = 255.f / 224.f * 2 * (1 - kr);
+  ug = 255.f / 224.f * 2 * ((1 - kb) * kb / kg);
+  vg = 255.f / 224.f * 2 * ((1 - kr) * kr / kg);
+  ub = 255.f / 224.f * 2 * (1 - kb);
+
+  printf("\nLimited range\n");
+  printf("R = (Y - 16) * 1.164                + V * %5f\n", vr);
+  printf("G = (Y - 16) * 1.164 - U * %6f - V * %6f\n", ug, vg);
+  printf("B = (Y - 16) * 1.164 + U * %5f\n", ub);
+
+  // printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ub * 64), ub, ub * 64);
+  printf("UG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ug * 64), ug, ug * 64);
+  printf("VG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vg * 64), vg, vg * 64);
+  printf("VR %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vr * 64), vr, vr * 64);
+
+  return 0;
+}
diff --git a/media/libyuv/util/convert.cc b/media/libyuv/util/yuvconvert.cc
index 5f071416da..93b5266853 100644
--- a/media/libyuv/util/convert.cc
+++ b/media/libyuv/util/yuvconvert.cc
@@ -9,7 +9,7 @@
  */
 
 // Convert an ARGB image to YUV.
-// Usage: convert src_argb.raw dst_yuv.raw
+// Usage: yuvconvert src_argb.raw dst_yuv.raw
 
 #ifndef _CRT_SECURE_NO_WARNINGS
 #define _CRT_SECURE_NO_WARNINGS
@@ -29,27 +29,27 @@ bool verbose = false;
 bool attenuate = false;
 bool unattenuate = false;
 int image_width = 0, image_height = 0;  // original width and height
-int dst_width = 0, dst_height = 0;  // new width and height
+int dst_width = 0, dst_height = 0;      // new width and height
 int fileindex_org = 0;  // argv argument contains the original file name.
 int fileindex_rec = 0;  // argv argument contains the reconstructed file name.
-int num_rec = 0;  // Number of reconstructed images.
-int num_skip_org = 0;  // Number of frames to skip in original.
-int num_frames = 0;  // Number of frames to convert.
-int filter = 1;  // Bilinear filter for scaling.
+int num_rec = 0;        // Number of reconstructed images.
+int num_skip_org = 0;   // Number of frames to skip in original.
+int num_frames = 0;     // Number of frames to convert.
+int filter = 1;         // Bilinear filter for scaling.
 
-static __inline uint32 Abs(int32 v) {
+static __inline uint32_t Abs(int32_t v) {
   return v >= 0 ? v : -v;
 }
 
 // Parse PYUV format. ie name.1920x800_24Hz_P420.yuv
-bool ExtractResolutionFromFilename(const char* name,
-                                   int* width_ptr,
-                                   int* height_ptr) {
+static bool ExtractResolutionFromFilename(const char* name,
+                                          int* width_ptr,
+                                          int* height_ptr) {
   // Isolate the .width_height. section of the filename by searching for a
   // dot or underscore followed by a digit.
   for (int i = 0; name[i]; ++i) {
-    if ((name[i] == '.' || name[i] == '_') &&
-        name[i + 1] >= '0' && name[i + 1] <= '9') {
+    if ((name[i] == '.' || name[i] == '_') && name[i + 1] >= '0' &&
+        name[i + 1] <= '9') {
       int n = sscanf(name + i + 1, "%dx%d", width_ptr, height_ptr);  // NOLINT
       if (2 == n) {
         return true;
@@ -59,13 +59,14 @@ bool ExtractResolutionFromFilename(const char* name,
   return false;
 }
 
-void PrintHelp(const char * program) {
+static void PrintHelp(const char* program) {
   printf("%s [-options] src_argb.raw dst_yuv.raw\n", program);
-  printf(" -s <width> <height> .... specify source resolution.  "
-         "Optional if name contains\n"
-         "                          resolution (ie. "
-         "name.1920x800_24Hz_P420.yuv)\n"
-         "                          Negative value mirrors.\n");
+  printf(
+      " -s <width> <height> .... specify source resolution.  "
+      "Optional if name contains\n"
+      "                          resolution (ie. "
+      "name.1920x800_24Hz_P420.yuv)\n"
+      "                          Negative value mirrors.\n");
   printf(" -d <width> <height> .... specify destination resolution.\n");
   printf(" -f <filter> ............ 0 = point, 1 = bilinear (default).\n");
   printf(" -skip <src_argb> ....... Number of frame to skip of src_argb\n");
@@ -77,8 +78,10 @@ void PrintHelp(const char * program) {
   exit(0);
 }
 
-void ParseOptions(int argc, const char* argv[]) {
-  if (argc <= 1) PrintHelp(argv[0]);
+static void ParseOptions(int argc, const char* argv[]) {
+  if (argc <= 1) {
+    PrintHelp(argv[0]);
+  }
   for (int c = 1; c < argc; ++c) {
     if (!strcmp(argv[c], "-v")) {
       verbose = true;
@@ -89,17 +92,17 @@ void ParseOptions(int argc, const char* argv[]) {
     } else if (!strcmp(argv[c], "-h") || !strcmp(argv[c], "-help")) {
       PrintHelp(argv[0]);
     } else if (!strcmp(argv[c], "-s") && c + 2 < argc) {
-      image_width = atoi(argv[++c]);    // NOLINT
-      image_height = atoi(argv[++c]);   // NOLINT
+      image_width = atoi(argv[++c]);   // NOLINT
+      image_height = atoi(argv[++c]);  // NOLINT
     } else if (!strcmp(argv[c], "-d") && c + 2 < argc) {
-      dst_width = atoi(argv[++c]);    // NOLINT
-      dst_height = atoi(argv[++c]);   // NOLINT
+      dst_width = atoi(argv[++c]);   // NOLINT
+      dst_height = atoi(argv[++c]);  // NOLINT
     } else if (!strcmp(argv[c], "-skip") && c + 1 < argc) {
-      num_skip_org = atoi(argv[++c]);   // NOLINT
+      num_skip_org = atoi(argv[++c]);  // NOLINT
     } else if (!strcmp(argv[c], "-frames") && c + 1 < argc) {
-      num_frames = atoi(argv[++c]);     // NOLINT
+      num_frames = atoi(argv[++c]);  // NOLINT
     } else if (!strcmp(argv[c], "-f") && c + 1 < argc) {
-      filter = atoi(argv[++c]);     // NOLINT
+      filter = atoi(argv[++c]);  // NOLINT
     } else if (argv[c][0] == '-') {
       fprintf(stderr, "Unknown option. %s\n", argv[c]);
     } else if (fileindex_org == 0) {
@@ -127,11 +130,9 @@ void ParseOptions(int argc, const char* argv[]) {
   int org_width, org_height;
   int rec_width, rec_height;
   bool org_res_avail = ExtractResolutionFromFilename(argv[fileindex_org],
-                                                     &org_width,
-                                                     &org_height);
+                                                     &org_width, &org_height);
   bool rec_res_avail = ExtractResolutionFromFilename(argv[fileindex_rec],
-                                                     &rec_width,
-                                                     &rec_height);
+                                                     &rec_width, &rec_height);
   if (image_width == 0 || image_height == 0) {
     if (org_res_avail) {
       image_width = org_width;
@@ -158,26 +159,29 @@ void ParseOptions(int argc, const char* argv[]) {
 static const int kTileX = 32;
 static const int kTileY = 32;
 
-static int TileARGBScale(const uint8* src_argb, int src_stride_argb,
-                         int src_width, int src_height,
-                         uint8* dst_argb, int dst_stride_argb,
-                         int dst_width, int dst_height,
+static int TileARGBScale(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         int src_width,
+                         int src_height,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int destination_width,
+                         int destination_height,
                          libyuv::FilterMode filtering) {
-  for (int y = 0; y < dst_height; y += kTileY) {
-    for (int x = 0; x < dst_width; x += kTileX) {
+  for (int y = 0; y < destination_height; y += kTileY) {
+    for (int x = 0; x < destination_width; x += kTileX) {
       int clip_width = kTileX;
-      if (x + clip_width > dst_width) {
-        clip_width = dst_width - x;
+      if (x + clip_width > destination_width) {
+        clip_width = destination_width - x;
       }
       int clip_height = kTileY;
-      if (y + clip_height > dst_height) {
-        clip_height = dst_height - y;
+      if (y + clip_height > destination_height) {
+        clip_height = destination_height - y;
       }
-      int r = libyuv::ARGBScaleClip(src_argb, src_stride_argb,
-                                    src_width, src_height,
-                                    dst_argb, dst_stride_argb,
-                                    dst_width, dst_height,
-                                    x, y, clip_width, clip_height, filtering);
+      int r = libyuv::ARGBScaleClip(src_argb, src_stride_argb, src_width,
+                                    src_height, dst_argb, dst_stride_argb,
+                                    destination_width, destination_height, x, y,
+                                    clip_width, clip_height, filtering);
       if (r) {
         return r;
       }
@@ -197,8 +201,8 @@ int main(int argc, const char* argv[]) {
   }
 
   // Open all files to convert to
-  FILE** file_rec = new FILE* [num_rec];
-  memset(file_rec, 0, num_rec * sizeof(FILE*)); // NOLINT
+  FILE** file_rec = new FILE*[num_rec];
+  memset(file_rec, 0, num_rec * sizeof(FILE*));  // NOLINT
   for (int cur_rec = 0; cur_rec < num_rec; ++cur_rec) {
     file_rec[cur_rec] = fopen(argv[fileindex_rec + cur_rec], "wb");
     if (file_rec[cur_rec] == NULL) {
@@ -222,8 +226,8 @@ int main(int argc, const char* argv[]) {
   // Input is YUV
   if (org_is_yuv) {
     const int y_size = Abs(image_width) * Abs(image_height);
-    const int uv_size = ((Abs(image_width) + 1) / 2) *
-        ((Abs(image_height) + 1) / 2);
+    const int uv_size =
+        ((Abs(image_width) + 1) / 2) * ((Abs(image_height) + 1) / 2);
     org_size = y_size + 2 * uv_size;  // YUV original.
   }
 
@@ -233,15 +237,15 @@ int main(int argc, const char* argv[]) {
   const size_t total_size = y_size + 2 * uv_size;
 #if defined(_MSC_VER)
   _fseeki64(file_org,
-            static_cast<__int64>(num_skip_org) *
-            static_cast<__int64>(org_size), SEEK_SET);
+            static_cast<__int64>(num_skip_org) * static_cast<__int64>(org_size),
+            SEEK_SET);
 #else
   fseek(file_org, num_skip_org * total_size, SEEK_SET);
 #endif
 
-  uint8* const ch_org = new uint8[org_size];
-  uint8* const ch_dst = new uint8[dst_size];
-  uint8* const ch_rec = new uint8[total_size];
+  uint8_t* const ch_org = new uint8_t[org_size];
+  uint8_t* const ch_dst = new uint8_t[dst_size];
+  uint8_t* const ch_rec = new uint8_t[total_size];
   if (ch_org == NULL || ch_rec == NULL) {
     fprintf(stderr, "No memory available\n");
     fclose(file_org);
@@ -256,20 +260,22 @@ int main(int argc, const char* argv[]) {
   }
 
   if (verbose) {
-    printf("Size: %dx%d to %dx%d\n", image_width, image_height,
-           dst_width, dst_height);
+    printf("Size: %dx%d to %dx%d\n", image_width, image_height, dst_width,
+           dst_height);
   }
 
   int number_of_frames;
-  for (number_of_frames = 0; ; ++number_of_frames) {
-    if (num_frames && number_of_frames >= num_frames)
+  for (number_of_frames = 0;; ++number_of_frames) {
+    if (num_frames && number_of_frames >= num_frames) {
       break;
+    }
 
     // Load original YUV or ARGB frame.
-    size_t bytes_org = fread(ch_org, sizeof(uint8),
-                             static_cast<size_t>(org_size), file_org);
-    if (bytes_org < static_cast<size_t>(org_size))
+    size_t bytes_org =
+        fread(ch_org, sizeof(uint8_t), static_cast<size_t>(org_size), file_org);
+    if (bytes_org < static_cast<size_t>(org_size)) {
       break;
+    }
 
     // TODO(fbarchard): Attenuate doesnt need to know dimensions.
     // ARGB attenuate frame
@@ -290,22 +296,17 @@ int main(int argc, const char* argv[]) {
         int half_src_height = (src_height + 1) / 2;
         int half_dst_width = (dst_width + 1) / 2;
         int half_dst_height = (dst_height + 1) / 2;
-        I420Scale(ch_org, src_width,
-                  ch_org + src_width * src_height, half_src_width,
-                  ch_org + src_width * src_height +
-                      half_src_width * half_src_height,  half_src_width,
-                  image_width, image_height,
-                  ch_rec, dst_width,
-                  ch_rec + dst_width * dst_height, half_dst_width,
-                  ch_rec + dst_width * dst_height +
-                      half_dst_width * half_dst_height,  half_dst_width,
-                  dst_width, dst_height,
-                      static_cast<libyuv::FilterMode>(filter));
+        I420Scale(
+            ch_org, src_width, ch_org + src_width * src_height, half_src_width,
+            ch_org + src_width * src_height + half_src_width * half_src_height,
+            half_src_width, image_width, image_height, ch_rec, dst_width,
+            ch_rec + dst_width * dst_height, half_dst_width,
+            ch_rec + dst_width * dst_height + half_dst_width * half_dst_height,
+            half_dst_width, dst_width, dst_height,
+            static_cast<libyuv::FilterMode>(filter));
       } else {
-        TileARGBScale(ch_org, Abs(image_width) * 4,
-                      image_width, image_height,
-                      ch_dst, dst_width * 4,
-                      dst_width, dst_height,
+        TileARGBScale(ch_org, Abs(image_width) * 4, image_width, image_height,
+                      ch_dst, dst_width * 4, dst_width, dst_height,
                       static_cast<libyuv::FilterMode>(filter));
       }
       bool rec_is_yuv = strstr(argv[fileindex_rec + cur_rec], "_P420.") != NULL;
@@ -321,27 +322,28 @@ int main(int argc, const char* argv[]) {
       if (!org_is_yuv && rec_is_yuv) {
         int half_width = (dst_width + 1) / 2;
         int half_height = (dst_height + 1) / 2;
-        libyuv::ARGBToI420(ch_dst, dst_width * 4,
-                           ch_rec, dst_width,
-                           ch_rec + dst_width * dst_height, half_width,
-                           ch_rec + dst_width * dst_height +
-                               half_width * half_height,  half_width,
-                           dst_width, dst_height);
+        libyuv::ARGBToI420(
+            ch_dst, dst_width * 4, ch_rec, dst_width,
+            ch_rec + dst_width * dst_height, half_width,
+            ch_rec + dst_width * dst_height + half_width * half_height,
+            half_width, dst_width, dst_height);
       }
 
       // Output YUV or ARGB frame.
       if (rec_is_yuv) {
-        size_t bytes_rec = fwrite(ch_rec, sizeof(uint8),
-                                  static_cast<size_t>(total_size),
-                                  file_rec[cur_rec]);
-        if (bytes_rec < static_cast<size_t>(total_size))
+        size_t bytes_rec =
+            fwrite(ch_rec, sizeof(uint8_t), static_cast<size_t>(total_size),
+                   file_rec[cur_rec]);
+        if (bytes_rec < static_cast<size_t>(total_size)) {
           break;
+        }
       } else {
-        size_t bytes_rec = fwrite(ch_dst, sizeof(uint8),
-                                  static_cast<size_t>(dst_size),
-                                  file_rec[cur_rec]);
-        if (bytes_rec < static_cast<size_t>(dst_size))
+        size_t bytes_rec =
+            fwrite(ch_dst, sizeof(uint8_t), static_cast<size_t>(dst_size),
+                   file_rec[cur_rec]);
+        if (bytes_rec < static_cast<size_t>(dst_size)) {
           break;
+        }
       }
       if (verbose) {
         printf("%5d", number_of_frames);
diff --git a/media/libyuv/winarm.mk b/media/libyuv/winarm.mk
index c4307a431f..b0a344ae06 100644
--- a/media/libyuv/winarm.mk
+++ b/media/libyuv/winarm.mk
@@ -31,6 +31,7 @@ LOCAL_OBJ_FILES = \
 	source/scale_any.o\
 	source/scale_argb.o\
 	source/scale_common.o\
+	source/scale_uv.o\
 	source/video_common.o
 
 .cc.o: